CoChat
Sign Up
Cited by 151Industrial Vision Systems and Defect DetectionNon-Destructive Testing TechniquesInfrastructure Maintenance and Monitoring

Automated Visual Defect Classification for Flat Steel Surface: A Survey

Luo, Qiwu, Fang, Xiaoxin, Su, Jiaojiao, Zhou, Jian, Zhou, Bingxing, Yang, Chunhua, Liu, Li, Gui, Weihua +1 more·2020·IEEE Transactions on Instrumentation and Measurement
DOI 10.1109/tim.2020.3030167
Open this paper in CoChat — free account

Abstract

For a typical surface automated visual inspection (AVI) instrument of planar materials, defect classification is an indispensable part after defect detection, which acts as a crucial precondition for achieving the online quality inspection of end products. In the industrial environment of manufacturing flat steels, this task is awfully difficult due to diverse defect appearances, ambiguous intraclass, and interclass distances. This article attempts to present a focused but systematic review of the traditional and emerging automated computer-vision-based defect classification methods by investigating approximately 140 studies on three specific flat steel products of con-casting slabs, hot-rolled steel strips, and cold-rolled steel strips. According to the natural image processing procedure of defect recognition, the diverse approaches are grouped into five successive parts: image acquisition, image preprocessing, feature extraction, feature selection, and defect classifier. Recent literature has been reviewed from an industrial goal-oriented perspective to provide some guidelines for future studies and recommend suitable methods for boosting the surface quality inspection level of AVI instruments.

AI snapshot

Written from the paper’s full text.

TL;DR

This survey reviews approximately 140 studies on automated visual defect classification for flat steel surfaces, covering continuous casting slabs, hot-rolled steel strips, and cold-rolled steel strips. It synthesizes traditional and deep-learning methods across five processing stages and identifies future directions such as multimode fusion, small-sample learning, and better time-cost evaluation.

Key findings

  • The survey investigates approximately 140 studies and states that literature over the last ten years accounts for approximately 75% of the reviewed advancements.
  • Methods are grouped into five successive components: image acquisition, image preprocessing, feature extraction, feature selection, and defect classifier.
  • In feature selection, PCA is reported as the most frequently used dimension-reduction method, and GA is preferred for feature optimization; deep-learning methods such as CNN, CAE, and GAN are widely applied but often depend on labeled data.
  • Table IV reports typical classification accuracies, including 99.50% for a deep CNN with majority voting on hot-rolled strips, 99.11% for generalized completed LBP + NNC on hot-rolled strips (266.74 ms per image), 97.5% for a SqueezeNet-based model (8.0 ms per image), and 98.12% for multiple support vector hypersphere with feature and sample weights (1.17 ms per image).

Methods

  • Design: a focused systematic review of traditional and emerging computer-vision-based defect classification methods for flat steel, investigating approximately 140 studies on continuous casting slabs, hot-rolled steel strips, and cold-rolled steel strips.
  • Organization: the review follows the natural image-processing procedure and groups approaches into image acquisition, image preprocessing, feature extraction, feature selection, and defect classifier.
  • Analysis: it compares strengths and weaknesses of method families and includes a statistical list of typical methods with application, difficulty, performance, and classification time where available; no new experimental dataset is introduced.

Limitations

  • Stated: most current classification methods emphasize accuracy and noise robustness while rarely involving time-cost evaluation, which matters for in-situ AVI on high-speed production lines.
  • Stated: feature selection algorithms are generally identical and somewhat unitary in steel surface defect recognition, and the feature selection process is discounted; unsupervised learning results still show some disparity from the ultimate goal.
  • Evident: as a survey, the paper reports no primary experiments, and many entries in Table IV have “T: NA,” limiting cross-study comparison of real-time suitability; its scope is also limited to three flat steel product types rather than all planar materials.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tim.2020.3030167

Cites

137 of 137 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 221 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning
    § B. Dimension Reduction · p. 11

    PCA is the most frequently used method to reduce dimensions in the problem of strip steel surface defect classification because of its reliable performance.

    “Throughout the literature we consulted, we found that the PCA is the most frequently used method to reduce the dimensions in the problem of strip steel surface defect classification because of its reliable performance.”
    Quote in the paperNot a substantial claim
  • Findingcitation
    § B. Dimension Reduction · p. 11

    The most popular algorithms applied to flat steel surface defect classification are PCA and LPP.

    “Of these methods, the most popular algorithms applied to flat steel surface defect classification are PCA and LPP [112, 113].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [112] [112] A. Cord, F. Bach, and D. Jeulin, 'Texture classification by statistical learning from morphological image processing: application to metallic surfaces, ' J Microsc, vol. 239, no. 2, pp. 159-166, Aug. 2010. source

      Not checked in the free run.

    • cites [113] [113] M. Chu, A. Wang, R. Gong, and M. Sha, 'Multi-class classification methods of enhanced LS-TWSVM for strip steel surface defects,' J. Iron Steel Res. Int., vol. 21, no. 2, pp. 174-180, Feb. 2014. source

      Not checked in the free run.

  • Findingreasoning
    § D. Brief Summary · p. 11

    The algorithms used in feature selection are generally identical and a little bit unitary in the topic of steel surface defect recognition.

    “In general, the algorithms used in the feature selection are generally identical and a little bit unitary in the topic of steel surface defect recognition.”
    Quote in the paperSubstantial claim
  • Findingown data
    § VIII. SUMMARY AND DISCUSSION · p. 17

    Most current classification methods place emphasis on classification accuracy and noise robustness while rarely involving time cost evaluation.

    “The statistical table reflects that most of the current classification methods place emphasis on classification accuracy and noise robustness while rarely involving the time cost evaluation of the designed classifiers.”
    Quote in the paperSubstantial claim
  • Premisecitation
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    Feature extraction and feature selection are dedicated to learning temporal, spatial and spectral features in images and even intrinsic priors in the production line to narrow intraclass variation and expand interclass distance.

    “Essentially, feature extraction and feature selection are dedicated to learning the temporal, spatial [13-15] and spectral features [16, 17] in images and even the intrinsic priors in the production line [18] to narrow the intraclass variation and expand the interclass distance.”
    Quote in the paperNot a substantial claim0 of 6 sources checked
    • cites [13] [13] L. Fang, S. Li, X. Kang, and J. A. Benediktsson, 'Spectral-spatial hyperspectral image classification via multiscale adaptive sparse representation,' IEEE Trans. Geosci. Remote Sensing, vol. 52, no. 12, pp. 7738-7749, Dec. 2014. - source

      Not checked in the free run.

    • cites [14] [14] L. Fang, C. Wang, S. Li, and J. A. Benediktsson, 'Hyperspectral image classification via multiple-feature-based adaptive sparse representation,' IEEE Trans. Instrum. Meas., vol. 66, no. 7, pp. 1646-1657, Jul. 2017. - source

      Not checked in the free run.

    • cites [15] [15] X. Kang, S. Li, and J. A. Benediktsson, 'Spectral-spatial hyperspectral image classification wth edge-preserving filtering,' IEEE Trans. Geosci. Remote Sensing, vol. 52, no. 5, pp. 2666-2677, May. 2014. - source

      Not checked in the free run.

    • cites [16] [16] L. He, J. Li, C. Liu, and S. Li, 'Recent advances on spectral-spatial hyperspectral image classification: An overview and new guidelines,' IEEE Trans. Geosci. Remote Sensing, vol. 56, no. 3, pp. 1579-1597, Mar. 2018. - source

      Not checked in the free run.

    • cites [17] [17] X. Kang, C. Li, S. Li, and H. Lin, 'Classification of hyperspectral images by gabor filtering based deep network,' IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., vol. 11, no. 4, pp. 1166-1178, Apr. 2018. - source

      Not checked in the free run.

    • cites [18] [18] J. Wang, Q. Li, J. Gan, H. Yu, and X. Yang, 'Surface defect detection via entity sparsity pursuit with intrinsic priors,' IEEE Trans. Ind. Inform., vol. 16, no. 1, pp. 141-150, Jan. 2020. - source

      Not checked in the free run.

  • Classified as probably not a substantial claim. 146
  • May combine two claims. 109
  • The quote only partly states this claim. 1

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    The literature over the last ten years accounts for approximately 75% of the advancements reviewed in this survey.

    “In particular, the literature over the last ten years accounts for approximately 75% of the abovementioned advancements.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Findingreasoning
    § C. Feature Optimization · p. 11

    Studies have a preference for GA when considering steel feature optimization, mainly because there are few superior methods for query.

    “According to the review, we can find that studies have a preference for GA when considering steel feature optimization, which is mainly because there are few superior methods for query.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given
    § A. Feature Evaluation · p. 10

    Feature evaluation approaches existing in the literature can be categorized into three groups: filter method, wrapper method and embedding method.

    “The adequacy of evaluation approaches existing in the literature can be categorized into three groups, namely, filter method , wrapper method and embedding method .”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given
    § 4) Brief Summary · p. 10

    Deep learning can make full use of big data to automatically learn multilayer depiction of features, involving thousands of features, and the expression ability of learned features is stronger than that of manually designed features.

    “Deep learning can make full use of the advantages of big data to automatically learn the multilayer depiction of features, which can involve thousands of features, and the expression ability of these learned features is stronger than that of manually designed features.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given
    § D. Brief Summary · p. 11

    After evaluation, PCA is mostly used to reduce dimensions, and almost all studies choose GA for feature optimization; both are traditional and reliable methods.

    “After evaluation, the PCA is mostly used to reduce dimensions, and almost all studies choose the GA for feature optimization; both are traditional and reliable methods.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given
    § A. Supervised Learning · p. 11

    The most commonly used classifiers in supervised learning are neural networks, support vector machines, distance functions, dictionary learning, sparse representation and multi-classifier fusion.

    “The most commonly used classifiers in supervised learning are neural networks (NN), support vector machines (SVM), distance functions, dictionary learning, sparse representation and multi-classifier fusion.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § I. INTRODUCTION · p. 1

    Any surface defects not treated in time will threaten steel product quality and might cause substantial economic and reputation cost to both steel manufacturers and end customers.

    “Any surface defects not treated in time will threaten the steel product quality, which might cause substantial economic and reputation cost to both the steel manufacturers and end customers [1-5].”
    Quote in the paperSubstantial claim0 of 5 sources checked
    • cites [1] [1] K. Song, and Y. Yan, 'A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,' Appl. Surf. Sci., vol. 285, pp. 858-864, Sep. 2013. -

      Not checked in the free run.

    • cites [2] [2] K. Song, and Y. Yan, 'Micro surface defect detection method for silicon steel strip based on saliency convex active Contour model,' Math. Probl. Eng., vol. 2013, pp. 1-13, Nov. 2013. -

      Not checked in the free run.

    • cites [3] [3] K. Song, S. Hu, Y. Yan, and J. Li, 'Surface defect detection method using saliency linear scanning morphology for silicon steel strip under oil pollution interference,' ISIJ Int., vol. 54, no. 11, pp. 2598-2607, Jul. 2014. -

      Not checked in the free run.

    • cites [4] [4] S. Ghorai, A. Mukherjee, M. Gangadaran, and P. K. Dutta, 'Automatic defect detection on hot-rolled flat steel products,' IEEE Trans. Instrum. Meas., vol. 62, no. 3, pp. 612-621, Mar. 2013. -

      Not checked in the free run.

    • cites [5] [5] Q. Luo, and Y. He, 'A cost-effective and automatic surface defect inspection system for hot-rolled flat steel,' Robot. Comput.-Integr. Manuf., vol. 38, pp. 16-30, Apr. 2016. -

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    In the industrial environment of manufacturing flat steels, defect classification is awfully difficult due to diverse defect appearances and ambiguous intraclass and interclass distances.

    “In the industrial environment of manufacturing flat steels, this task is awfully difficult due to diverse defect appearances, ambiguous intraclass and interclass distances.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    For a typical surface automated visual inspection instrument of planar materials, defect classification is an indispensable part after defect detection and a crucial precondition for achieving on-line quality inspection of end products.

    “For a typical surface automated visual inspection (AVI) instrument of planar materials, defect classification is an indispensable part after defect detection, which acts as a crucial precondition for achieving the on-line quality inspection of end products.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [5] [5] Q. Luo, and Y. He, 'A cost-effective and automatic surface defect inspection system for hot-rolled flat steel,' Robot. Comput.-Integr. Manuf., vol. 38, pp. 16-30, Apr. 2016. -

      Not checked in the free run.

    • cites [10] [10] Q. Luo, X. Fang, Y. Sun, L. Liu, J. Ai, C. Yang, and O. Simpson, 'Surface defect classification for hot-rolled steel strips by selectively dominant local binary patterns,' IEEE Access, vol. 7, pp. 23488-23499, Feb. 2019. -

      Not checked in the free run.

    • cites [11] [11] M. Chu, R. Gong, S. Gao, and J. Zhao, 'Steel surface defects recognition based on multi-type statistical features and enhanced twin support vector machine,' Chemometrics Intell. Lab. Syst., vol. 171, pp. 140-150, Dec. 2017. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    A typical AVI instrument mainly realizes the two fundamental functions of defect detection and classification, whose results are used to adjust the relevant configurations of the production line to guarantee the quality of final steel product.

    “A typical AVI instrument mainly realizes the two fundamental functions of defect detection and classification, whose results are used to adjust the relevant configurations of the production line to guarantee the quality of final steel product [5, 10, 11].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [5] [5] Q. Luo, and Y. He, 'A cost-effective and automatic surface defect inspection system for hot-rolled flat steel,' Robot. Comput.-Integr. Manuf., vol. 38, pp. 16-30, Apr. 2016. -

      Not checked in the free run.

    • cites [10] [10] Q. Luo, X. Fang, Y. Sun, L. Liu, J. Ai, C. Yang, and O. Simpson, 'Surface defect classification for hot-rolled steel strips by selectively dominant local binary patterns,' IEEE Access, vol. 7, pp. 23488-23499, Feb. 2019. -

      Not checked in the free run.

    • cites [11] [11] M. Chu, R. Gong, S. Gao, and J. Zhao, 'Steel surface defects recognition based on multi-type statistical features and enhanced twin support vector machine,' Chemometrics Intell. Lab. Syst., vol. 171, pp. 140-150, Dec. 2017. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § D. Quick Glance of Defect Types · p. 3

    Hot-rolled steel strip surfaces are usually covered with many scales, which increase the possibility of misclassifying other types of defects.

    “Fig. 2 (b) shows eighteen types of image samples of hot-rolled steel strips, because the surfaces usually covered with many scales, which increase the possibility of misclassifying other types of defects.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § D. Quick Glance of Defect Types · p. 3

    The most critical goal of defect classification for continuous casting slabs is to distinguish cracks from the other three similar pseudo-defects.

    “The most critical goal of defect classification for continuous casting slabs is to distinguish cracks from the other three similar pseudo-defects.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    Compared with defect detection, which mainly affects the time-efficiency and missed detection rate of the AVI instrument, defect classification directly determines its final user experience, as the cognitive performance of defect patterns represents the troubleshooting ability of the AVI instrument.

    “Compared with defect detection, which mainly affects the time-efficiency and missed detection rate of the AVI instrument, defect classification directly determines its final user experience, as the cognitive performance of defect patterns represents the troubleshooting ability of the AVI instrument.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § A. Image Denoising · p. 3

    Median filters have the best performance on the suppression of salt and pepper noise, and Gaussian filter and Wiener filter have better effects on Gaussian noise.

    “Among them, it is proved that the median filters have the best performance on the suppression of salt and pepper noise, and the Gaussian filter and Wiener filter have better effects on Gaussian noise.”
    Quote in the paperNot a substantial claim0 of 10 sources checked
    • cites [46] [46] A. Landstrom, and M. J. Thurley, 'Morphology-based crack detection for steel slabs,' IEEE J. Sel. Top. Signal Process., vol. 6, no. 7, pp. 866-875, Nov. 2012. - source

      Not checked in the free run.

    • cites [47] [47] S. G. Ryu, D. c. Choi, Y. J. Jeon, S. J. Lee, J. P. Yun, and S. W. Kim, 'Detection of scarfing faults on the edges of slabs,' ISIJ Int., vol. 54, no. 1, pp. 112-118, 2014. - source

      Not checked in the free run.

    • cites [48] [48] F. G. Bulnes, D. F. Garcia, F. Javier De la Calle, R. Usamentiaga, and J. Molleda, 'A Non-invasive technique for online defect detection on steel strip surfaces,' J. Nondestruct. Eval., vol. 35, no. 4, pp. 1-18, Dec. 2016. - source

      Not checked in the free run.

    • cites [49] [49] Y. Liu, Y. Hsu, Y. Sun, S. Tsai, C. Ho, and C. Chen, 'A computer vision system for automatic steel surface inspection,' in 2010 5th IEEE Conf. Ind. Electron. Appl., Taichung, Taiwan, 2010, pp. 1667-1670. - source

      Not checked in the free run.

    • cites [50] [50] L. A. O. Martins, F. L. C. Padua, and P. E. M. Almeida, 'Automatic detection of surface defects on rolled steel using computer vision and artificial neural networks,' in 36th Annual Conf. IEEE Ind. Electron., Glendale, AZ, USA, 2010, pp. 1081-1086. - source

      Not checked in the free run.

    • cites [51] [51] Y. Wang, H. Xia, X. Yuan, L. Ling, and S. Bei, 'Distributed defect recognition on steel surfaces using an improved random forest algorithm with optimal multi-feature-set fusion,' Multimed. Tools Appl., vol. 77, no. 13, pp. 16741-16770, Jul. 2017. - source

      Not checked in the free run.

    • cites [52] [52] T. Bo, J. Kong, X. Wang, and C. Li, 'Surface inspection system of steel strip based on machine vision,' in 1st Int. Workshop. Database Technol. Appl., Wuhan, Hubei, China, Apr. 2009, pp. 359-362. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [53] [53] Y. Chen, L. Chen, X. Liu, S. Ding, and H. Zhang, 'Real-time steel inspection system based on support vector machine and multiple kernel learning,' in 6th Int. Conf. Intell. Syst. Knowl. Eng., Shanghai, PEOPLES R CHINA, 2011, pp. 185-190. - source

      Not checked in the free run.

    • cites [54] [54] R. Gong, C. Wu, and M. Chu, 'Steel surface defect classification using multiple hyper-spheres support vector machine with additional information,' Chemometrics Intell. Lab. Syst., vol. 172, pp. 109-117, Jan. 2018. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [55] [55] M. Ricci, A. Ficola, M. L. Fravolini, L. Battaglini, A. Palazzi, P. Burrascano, P. Valigi, L. Appolloni, S. Cervo, and C. Rocchi, 'Magnetic imaging and machine vision NDT for the on-line inspection of stainless steel strips,' Meas. Sci. Technol., vol. 24, no. 2, pp. 1-17, Feb, 2013. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § B. Image Enhancement · p. 4

    Because the global gray value of steel surface images is ordinarily low but the dynamic range is large, the effect of histogram equalization is not as good as that of homomorphic filtering.

    “Because the global gray value of steel surface images is ordinarily low but the dynamic range is large, the effect of histogram equalization is not as good as that of homomorphic filtering.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given
    § IV. IMAGE PREPROCESSING · p. 3

    Weak correlation information and strong interference information influence the reliability of feature extraction and defect classification.

    “Weak correlation information and strong interference information influence the reliability of feature extraction and defect classification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § C. Brief Summary of Image Acquisition · p. 3

    The challenges that image acquisition faces are the greatest in all parts of flat steel surface defect classification.

    “Generally, the challenges that image acquisition faces are the greatest in all parts of flat steel surface defect classification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning
    § D. Quick Glance of Defect Types · p. 3

    Cracks are challenging to detect because of the interference of scales, slag marks, and uneven illumination, which are similar to the appearance of cracks.

    “However, cracks are challenging to detect because of the interference of scales, slag marks, and uneven illumination, which are similar to the appearance of cracks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § C. Brief Summary of Image Acquisition · p. 3

    Intensity imaging technology is largely dependent on the illumination condition, which is the most vulnerable to interference in production.

    “Intensity imaging technology is largely dependent on the illumination condition, which is the most vulnerable to interference in production.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § IMAGE ACQUISITION · p. 2

    Intensity imaging is most commonly used in real-world flat steel production lines because range imaging is limited to surface defects with three-dimensional properties and is not competitive with intensity imaging in spatial resolution and acquisition speed.

    “Because the former is limited to surface defects with three-dimensional properties, e.g. cavities, scratches, and nicks, and it is not competitive with the latter in terms of spatial resolution and acquisition speed, intensity imaging is most commonly used in real-world flat steel production lines.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] [32] F. Pernkopf, and P. O'Leary, 'Image acquisition techniques for automatic visual inspection of metallic surfaces,' NDT E Int., vol. 36, no. 8, pp. 609-617, Dec. 2003. -

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given
    § II. PRIOR LITERATURE REVIEW · p. 2

    Existing AVI surveys are insufficiently updated, and the methods involved do not reflect the latest level of algorithms in contemporary development.

    “However, these surveys are insufficiently updated, and the methods involved do not reflect the latest level of algorithms in contemporary development.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [19] [19] R. T. Chin, and C. A. Harlow, 'Automated visual inspection: a survey,' IEEE Trans. Pattern Anal. Mach. Intell., vol. 4, no. 6, pp. 557-573, Dec. 1982. - source

      Not checked in the free run.

    • cites [20] [20] T. S. Newman, and A. K. Jain, 'A survey of automated visual inspection,' Comput. Vis. Image Underst., vol. 61, no. 2, pp. 231-262, Mar. 1995. - source

      Not checked in the free run.

    • cites [21] [21] X. Xie, 'A review of recent advances in surface defect detection using texture analysis techniques,' Electron. Lett. Comput. Vis. Image Anal., vol. 7, no. 3, pp. 1-22, 2008. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § B. Image Enhancement · p. 5

    The DHC algorithm explores the intrinsic prior that image intensity is varying actively but horizontal intensity distribution is extremely stable, which helped the DHC algorithm win a competition against other image enhancement methods.

    “The resource-saving algorithm of dynamic homogenizing compensation (DHC) [5] explores the intrinsic prior that the image intensity is varying actively but horizontal intensity distribution is extremely stable, which helped the DHC algorithm win a competition against other image enhancement methods.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5] [5] Q. Luo, and Y. He, 'A cost-effective and automatic surface defect inspection system for hot-rolled flat steel,' Robot. Comput.-Integr. Manuf., vol. 38, pp. 16-30, Apr. 2016. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Automated Visual Defect Classification for Flat Steel Surface: A Survey · p. 1

    Advanced imaging techniques and emerging machine learning methods jointly resist the challenges of unsatisfactory imaging environments and quasi real-time requirements, forcefully driving the progress of defect recognition, especially in learning classification.

    “Fortunately, advanced imaging techniques and emerging machine learning methods jointly resist the challenges of unsatisfactory imaging environments and quasi real-time requirements [5], forcefully driving the progress of defect recognition, especially in learning classification [8].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [5] [5] Q. Luo, and Y. He, 'A cost-effective and automatic surface defect inspection system for hot-rolled flat steel,' Robot. Comput.-Integr. Manuf., vol. 38, pp. 16-30, Apr. 2016. - source

      Not checked in the free run.

    • cites [8] [8] L. Liu, W. Ouyang, X. Wang, P. Fieguth, J. Chen, X. Liu, and M. Pietikainen, 'Deep learning for generic object detection: A survey,' Int. J. Comput. Vis., vol. 128, no. 2, pp. 261-318, Feb. 2020. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § II. PRIOR LITERATURE REVIEW · p. 2

    Prior surveys paid more attention to detection methods and simply summarized the research progress of classification methods through supervised and unsupervised classifiers.

    “However, these surveys paid more attention to the detection methods, and they simply summarized the research progress of the classification methods through supervised and unsupervised classifiers.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [29] [29] N. Neogi, D. K. Mohanta, and P. K. Dutta, 'Review of vision-based steel surface inspection systems,' EURASIP J. Image Video Process., vol. 2014, no. 1, pp. 1-19, Nov. 2014. - source

      Not checked in the free run.

    • cites [30] [30] X. Sun, J. Gu, S. Tang, and J. Li, 'Research progress of visual inspection technology of steel products-A review,' Appl. Sci.-Basel, vol. 8, no. 11, pp. 1-25, Nov, 2018. - source

      Not checked in the free run.

    • cites [31] [31] T. Czimmermann, G. Ciuti, M. Milazzo, M. Chiurazzi, S. Roccella, C. M. Oddo, and P. Dario, 'Visual-based defect detection and classification approaches for industrial applications-a survey,' Sensors, vol. 20, no. 5, pp. 1-25, Mar, 2020. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § IMAGE ACQUISITION · p. 2

    Image acquisition occurs during the first step of defect classification, and the quality of the acquired images directly affects the performance of the subsequent processing.

    “Image acquisition occurs during the first step of defect classification, and the quality of the acquired images directly affects the performance of the subsequent processing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § D. Quick Glance of Defect Types · p. 3

    The surface quality of cold-rolled strip is usually better than the other two flat steel, so the number, size and degree of defects are also the most demanding.

    “The surface quality of the cold-rolled strip is usually better than the other two flat steel. Therefore, the number, size and degree of defects are also the most demanding.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § A. Image Denoising · p. 3

    Wavelet smoothing can sharpen an object's edges and restrain noise; moreover, it can enhance the positioning precision of the edges and the depiction of the images.

    “Wavelet smoothing can sharpen an object's edges and restrain noise; moreover, it can enhance the positioning precision of the edges and the depiction of the images.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [58] [58] Y. Xu, J. B. Weaver, D. M. Healy, and J. Lu, 'Wavelet transform domain filters: a spatially selective noise filtration technique,' IEEE Trans. Image Process., vol. 3, no. 6, pp. 747-758, Nov. 1994. - source

      Not checked in the free run.

    • cites [59] [59] X. Zhang, Y. Ding, Y. Lv, A. Shi, and R. Liang, 'A vision inspection system for the surface defects of strongly reflected metal based on multi-class SVM,' Expert Syst. Appl., vol. 38, no. 5, pp. 5930-5939, May. 2011. - source

      Not checked in the free run.

    • cites [60] [60] C. S. Lee, C. H. Choi, J. Y. Choi, Y. K. Kim, and S. H. Choi, 'Feature extraction algorithm based on adaptive wavelet packet for surface defect classification,' in Int. Conf. Image Processing LAUSANNE, SWITZERLAND 1996, pp. 673-676. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [61] [61] X. Wu, K. Xu, and J. Xu, 'Application of undecimated wavelet transform to surface defect detection of hot rolled steel plates,' in Int. Congress Image Signal Process., Sanya, China, May. 2008, pp. 528-532. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given
    § II. PRIOR LITERATURE REVIEW · p. 2

    In addition to the performance of classifiers, the characteristic of the extracted features is another important factor influencing the precision of classification methods.

    “In addition to the performance of the classifiers, the characteristic of the extracted features is another important factor influencing the precision of classification methods.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § A. Image Denoising · p. 3

    A single filter may not only smooth the image but also miss the details of the image.

    “However, a single filter may not only smooth the image but also miss the details of the image.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § B. Light Source and Lighting Method · p. 3

    LED is the most widely used luminaire in machine vision applications due to its advantages of longevity and low heat production.

    “Among them, LED is the most widely used luminaire in machine vision applications due to its advantages of longevity and low heat production [35, 37, 38].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [35] [35] K. Liu, H. Wang, H. Chen, E. Qu, Y. Tian, and H. Sun, 'Steel surface defect detection using a new Haar-Weibull-variance model in unsupervised manner,' IEEE Trans. Instrum. Meas., vol. 66, no. 10, pp. 2585-2596, Jul. 2017. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [37] [37] H. Zheng, L. X. Kong, and S. Nahavandi, 'Automatic inspection of metallic surface defects using genetic algorithms,' in 9th Internet. Conf. Metal Forming, UNIV BRIMINGHAM, BIRMINGHAM, ENGLAND, Sep. 2002, pp. 427-433. - source

      Not checked in the free run.

    • cites [38] [38] Q. Li, and S. Ren, 'A Real-time visual inspection system for discrete surface defects of rail heads,' IEEE Trans. Instrum. Meas., vol. 61, no. 8, pp. 2189-2199, Aug. 2012. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § A. Image Denoising · p. 3

    Optimized filters are effective not only in filtering mixed noise, but also in maintaining edge details.

    “The optimized filters have been very effective not only in filtering mixed noise, but also in maintaining edge details.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [39] [39] R. Gong, C. Wu, M. Chu, and X. Liu, 'The strip steel surface defect recognition based on multiple support vector hyper-sphere with feature and sample weights,' Steel Res. Int., vol. 87, no. 12, pp. 1678-1685, Dec. 2016. - source

      Not checked in the free run.

    • cites [57] [57] M. Chu, A. Wang, and R. Gong, 'An image filtering algorithm for removing mixed noise,' in 2nd Int. Conf. Mechatronics Control Eng., Dalian, PEOPLES R CHINA 2013, pp. 383-388. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Fourier Transform (FT) · p. 7

    The Fourier amplitude spectrum is translational invariant and is often used to assess the directional information of carbide distribution images.

    “the Fourier amplitude spectrum is translational invariant and is often used to assess the directional information of carbide distribution images [68].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [68] [68] K. Wiltschi, A. Pinz, and T. Lindeberg, 'An automatic assessment scheme for steel quality inspection,' Mach. Vis. Appl., vol. 12, no. 3, pp. 113-128, Oct. 2000. -

      Not checked in the free run.

  • May combine two claims.

    Premisenone given
    § 2) Gray-level Co-occurrence Matrix (GLCM) · p. 7

    GLCM is sensitive to rotation, which could bring about information redundancy.

    “However, it is sensitive to rotation, which could bring about information redundancy.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    Edge-based operators can locate edges accurately but cannot guarantee the continuity and closure of edges.

    “Edge-based operators can locate edges accurately but cannot guarantee the continuity and closure of edges.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 3) Histogram of Oriented Gradients (HOG) · p. 7

    PHOG has an excellent feature description ability and was applied to describe steel texture features.

    “the PHOG has an excellent feature description ability and was applied to describe steel texture features.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [94] [94] J. Masci, U. Meier, D. Ciresan, J. Schmidhuber, and G. Fricout, 'Steel defect classification with max-pooling convolutional neural networks,' in 2012 Int. Joint Conf. Neural Network, Brisbane, AUSTRALIA, 2012, pp. 1-6. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § B. Shape-based Methods · p. 7

    Luo et al.'s extraction of forgotten useful information hidden in nonuniform patterns improves accuracy and calculation time simultaneously and has better robustness to noise.

    “Luo et al. innovatively extracted the forgotten useful information hidden in nonuniform patterns in [10, 90], which not only improves the accuracy and calculation time simultaneously but also has better robustness to noise.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [10] [10] Q. Luo, X. Fang, Y. Sun, L. Liu, J. Ai, C. Yang, and O. Simpson, 'Surface defect classification for hot-rolled steel strips by selectively dominant local binary patterns,' IEEE Access, vol. 7, pp. 23488-23499, Feb. 2019. -

      Not checked in the free run.

    • cites [90] [90] Q. Luo, Y. Sun, P. Li, O. Simpson, L. Tian, and Y. He, 'Generalized completed local binary patterns for time-efficient steel surface defect classification,' IEEE Trans. Instrum. Meas., vol. 68, no. 3, pp. 667-679, Mar. 2019. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    The watershed algorithm is able to detect weak edges but is still sensitive to noise.

    “The watershed algorithm is able to detect weak edges but is still sensitive to noise.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    The Laplacian operator is of poor robustness to noise and has a large amount of calculation.

    “The Laplacian operator is of poor robust to noise and has a large amount of calculation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § 2) Gabor Filter (GF) · p. 7

    FT only depicts the spatial-frequency distribution without regard to spatial domain information.

    “The FT only depicts the spatial-frequency distribution without regard to the spatial domain information.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 2) Gabor Filter (GF) · p. 8

    Local binary pattern is sensitive to scale variation and noise interference.

    “Sensitive to scale variation and noise interference.”
    Quote in the paperNot a substantial claim0 of 10 sources checked
    • cites [1] [1] K. Song, and Y. Yan, 'A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,' Appl. Surf. Sci., vol. 285, pp. 858-864, Sep. 2013. - source

      Not checked in the free run.

    • cites [10] [10] Q. Luo, X. Fang, Y. Sun, L. Liu, J. Ai, C. Yang, and O. Simpson, 'Surface defect classification for hot-rolled steel strips by selectively dominant local binary patterns,' IEEE Access, vol. 7, pp. 23488-23499, Feb. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [11] [11] M. Chu, R. Gong, S. Gao, and J. Zhao, 'Steel surface defects recognition based on multi-type statistical features and enhanced twin support vector machine,' Chemometrics Intell. Lab. Syst., vol. 171, pp. 140-150, Dec. 2017. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [18] [18] J. Wang, Q. Li, J. Gan, H. Yu, and X. Yang, 'Surface defect detection via entity sparsity pursuit with intrinsic priors,' IEEE Trans. Ind. Inform., vol. 16, no. 1, pp. 141-150, Jan. 2020. - source

      Not checked in the free run.

    • cites [81] [81] M. Xiao, M. Jiang, G. Li, L. Xie, and L. Yi, 'An evolutionary classifier for steel surface defects with small sample set,' EURASIP J. Image Video Process., vol. 2017, no. 1, pp. 1-13, Jul. 2017. - source

      Not checked in the free run.

    • cites [82] [82] S. R. Aghdam, E. Amid, and M. F. Imani, 'A fast method of steel surface defect detection using decision trees applied to LBP based features, ' in 7th IEEE Conf. Ind. Electron. Appl., Singapore, Singapore, 2012, pp. 1447-1452. - source

      Not checked in the free run.

    • cites [83] [83] S. Tian, and K. Xu, 'An algorithm for surface defect identification of steel plates based on genetic algorithm and extreme learning machine,' Metals, vol. 7, no. 8, pp. 1-11, Aug. 2017. - source

      Not checked in the free run.

    • cites [84] [84] R. Ren, T. Hung, and K. C. Tan, 'A generic deep-learning-based approach for automated surface inspection,' IEEE Trans. Cybern., vol. 48, no. 3, pp. 929-940, Mar. 2018. - source

      Not checked in the free run.

    • cites [85] [85] X. Tao, D. Zhang, W. Ma, X. Liu, and D. Xu, 'Automatic metallic surface defect detection and recognition with convolutional neural networks,' Appl. Sci.-Basel, vol. 8, no. 9, pp. 1-15, Sep. 2018. - source

      Not checked in the free run.

    • cites [86] [86] Y. Liu, K. Xu, and D. Wang, 'Online surface defect identification of cold rolled strips based on local binary pattern and extreme learning machine,' Metals, vol. 8, no. 3, pp. 1-18, Mar. 2018. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § 4) Others · p. 7

    LBP, GLCM and HOG are the three reliable methods most widely used in steel due to their adaptability and robustness.

    “In contrast, LBP, GLCM and HOG are the three reliable methods most widely used in steel due to their adaptability and robustness.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § C. Image Segmentation · p. 5

    Otsu is a classical adaptive threshold method for separating defects from the background in flat steel images, obtaining a threshold value based on the characteristic of the large variance between the background and foreground.

    “Otsu is a classical adaptive threshold method for separating defects from the background in flat steel images [59, 69, 70], which obtains a threshold value based on the characteristic of the large variance between the background and foreground.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [59] [59] X. Zhang, Y. Ding, Y. Lv, A. Shi, and R. Liang, 'A vision inspection system for the surface defects of strongly reflected metal based on multi-class SVM,' Expert Syst. Appl., vol. 38, no. 5, pp. 5930-5939, May. 2011. - source

      Not checked in the free run.

    • cites [69] [69] M. R. Yazdchi, A. G. Mahyari, and A. Nazeri, 'Detection and classification of surface defects of cold rolling mill steel using morphology and neural network,' in Int. Conf. Comput. Intell. for Modelling, Control and Automation, Vienna, AUSTRIA, 2008, pp. 1071-1076. - source

      Not checked in the free run.

    • cites [70] [70] M. Win, A. R. Bushroa, M. A. Hassan, N. M. Hilman, and A. Ide-Ektessabi, 'A contrast adjustment thresholding method for surface defect detection based on mesoscopy,' IEEE Trans. Ind. Informat., vol. 11, no. 3, pp. 642-649, Jun. 2015. - source

      Not checked in the free run.

  • May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    The grayscale of steel strip images is ordinarily nonuniform, the gray value variation across the background and the defect is sometimes gradual, and the size of the defect area is very small, not easy to be recognized by the computer.

    “The grayscale of steel strip images is ordinarily nonuniform, the gray value variation cross the background and the defect is sometimes gradual, and the size of defect area is very small, not easy to be recognized by the computer.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § C. Image Segmentation · p. 5

    The watershed algorithm can well detect weak edges, and even though noise in an image may cause over-segmentation, it can still ensure a closed continuous edge.

    “The watershed algorithm can well detect weak edges, even though the noise in an image may cause over-segmentation, it can still ensure a closed continuous edge.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [45] [45] R. Gong, M. Chu, Y. Yang, and Y. Feng, 'A multi-class classifier based on support vector hyper-spheres for steel plate surface defects,' Chemometrics Intell. Lab. Syst., vol. 188, pp. 70-78, May. 2019. - source

      Not checked in the free run.

    • cites [54] [54] R. Gong, C. Wu, and M. Chu, 'Steel surface defect classification using multiple hyper-spheres support vector machine with additional information,' Chemometrics Intell. Lab. Syst., vol. 172, pp. 109-117, Jan. 2018. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    Kirsch operator has a good effect in maintaining details and anti-noise, but it depends on the edge direction and cannot guarantee the continuity and closure of the edge, making it difficult to form a large region.

    “Kirsch operator has a good effect in maintaining details and anti-noise, but it depends on the edge direction and cannot guarantee the continuity and closure of the edge, which make it difficult to form a large region.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    In many cases, the contrast between the defect and the background is not the same everywhere in the image, so it is challenging to separate the abnormal area and the background with a unified threshold.

    “Although this method is simple and efficient, in many cases, the contrast between the defect and the background is not the same everywhere in the image, so it is challenging to separate the abnormal area and the background with a unified threshold.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2) Gabor Filter (GF) · p. 7

    Choi et al. combined two Gabor filters to detect seam cracks on steel plates, offering high detection performance and effective reduction of noise components.

    “Choi et al. [63] combined two Gabor filters to detect seam cracks on steel plates, which offers high detection performance and the effective reduction of noise components.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [63] [63] D. C. Choi, Y. J. Jeon, S. J. Lee, J. P. Yun, and S. W. Kim, 'Algorithm for detecting seam cracks in steel plates using a Gabor filter combination method,' Appl. Optics, vol. 53, no. 22, pp. 4865-4872, Aug. 2014. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § D. Transform-based methods · p. 7

    Features in the spatial-frequency domain show stronger robustness to noise and intensity variation than in the spatial domain, and transform features are useful in improving the representation of spectral data and increasing classification accuracy.

    “Many works suggest that the features in the spatial-frequency domain show stronger robust to noise and intensity variation than in the spatial domain, and transform features have been reported to be useful in improving the representation of spectral data and increasing the classification accuracy [16].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [16] [16] L. He, J. Li, C. Liu, and S. Li, 'Recent advances on spectral-spatial hyperspectral image classification: An overview and new guidelines,' IEEE Trans. Geosci. Remote Sensing, vol. 56, no. 3, pp. 1579-1597, Mar. 2018. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 1) Local Binary Pattern (LBP) · p. 6

    AECLBPs and SLBP are both robust to noise to a large extent.

    “These two methods are both robust to noise to a large extent.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [1] [1] K. Song, and Y. Yan, 'A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,' Appl. Surf. Sci., vol. 285, pp. 858-864, Sep. 2013. - source

      Not checked in the free run.

    • cites [88] [88] M. Chu, and R. Gong, 'Invariant feature extraction method based on smoothed local binary pattern for strip steel surface defect,' ISIJ Int., vol. 55, no. 9, pp. 1956-1962, May 2015. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation
    § 1) Fourier Transform (FT) · p. 7

    FT can reflect local outstanding feature information and is widely used in the extraction of features of steel surface images.

    “FT can reflect the local outstanding feature information, which is widely used in the extraction of features of steel surface images [96, 97].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [96] [96] C. Dongyan, X. Kewen, N. Aslam, and H. Jingzhong, 'Defect classification recognition on strip steel surface using second-order cone programming-relevance vector machine algorithm,' J. Comput. Theor. Nanosci., vol. 13, no. 9, pp. 6141-6148, Step. 2016. - source

      Not checked in the free run.

    • cites [97] [97] Y. h. Ai, and K. Xu, 'Surface detection of continuous casting slabs based on curvelet transform and kernel locality preserving projections,' J. Iron Steel Res. Int., vol. 20, no. 5, pp. 80-86, May. 2013. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Fourier Transform (FT) · p. 7

    Wu et al. introduced SVP and RRCR to reflect global and local statistic feature information, excavating more deep information from spectrum images.

    “Wu et al. [98] first obtained an original feature set by the FFT, and then , two extended features (Sum of Valid Pixels (SVP) and Repletion Ration of Center Region (RRCR)) were introduced to reflect the global and local statistic feature information, respectively, which excavated more deep information from the spectrum images.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [98] [98] G. Wu, H. Zhang, X. Sun, J. Xu, and K. Xu, 'A Bran-new feature extraction method and its application to surface defect recognition of hot rolled strips,' in IEEE Int. Conf. Autom. Logis., Jinan, China, Aug. 2007, pp. 2069-2074. -

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning
    § 2) Gray-level Co-occurrence Matrix (GLCM) · p. 7

    GLCM and its variants merely describe spatial features and cannot reflect the high- or low-frequency components of the multidirectional defects on a surface, while steel strip surface defects vary widely.

    “However, since GLCM and its variants are merely describe spatial features, they cannot reflect the highor low-frequency components of the multidirectional defects on a surface, while steel strip surface defects vary widely (various types, shapes and orientations).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § C. Image Segmentation · p. 5

    Threshold-based methods are simple and efficient, but the selection of the optimal threshold value is usually labor-intensive.

    “In contrast, threshold-based methods are simple and efficient, but the selection of the optimal threshold value is usually labor-intensive.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 2) Gabor Filter (GF) · p. 7

    The 2D Gabor function contains real and imaginary parts; the real part detects blob and the imaginary part detects edge.

    “The 2D Gabor function contains real and imaginary parts. The former can be utilized to detect blob, and the latter can be used to detect edge.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 2) Gabor Filter (GF) · p. 7

    The Gabor filter is obtained by superimposing a trigonometric function and a Gaussian function and is an effective texture detection tool for extracting features of a specific scale and orientation.

    “The GF can be obtained by superimposing a trigonometric function and a Gaussian function; which is an effective texture detection tool to extract the features of a specific of specific scale and orientation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2) Gabor Filter (GF) · p. 8

    Local binary pattern is rotation and gray invariant and has fast calculation speed.

    “Be of rotation and gray invariance, fast calculation speed.”
    Quote in the paperNot a substantial claim0 of 10 sources checked
    • cites [1] [1] K. Song, and Y. Yan, 'A noise robust method based on completed local binary patterns for hot-rolled steel strip surface defects,' Appl. Surf. Sci., vol. 285, pp. 858-864, Sep. 2013. - source

      Not checked in the free run.

    • cites [10] [10] Q. Luo, X. Fang, Y. Sun, L. Liu, J. Ai, C. Yang, and O. Simpson, 'Surface defect classification for hot-rolled steel strips by selectively dominant local binary patterns,' IEEE Access, vol. 7, pp. 23488-23499, Feb. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [11] [11] M. Chu, R. Gong, S. Gao, and J. Zhao, 'Steel surface defects recognition based on multi-type statistical features and enhanced twin support vector machine,' Chemometrics Intell. Lab. Syst., vol. 171, pp. 140-150, Dec. 2017. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [18] [18] J. Wang, Q. Li, J. Gan, H. Yu, and X. Yang, 'Surface defect detection via entity sparsity pursuit with intrinsic priors,' IEEE Trans. Ind. Inform., vol. 16, no. 1, pp. 141-150, Jan. 2020. - source

      Not checked in the free run.

    • cites [81] [81] M. Xiao, M. Jiang, G. Li, L. Xie, and L. Yi, 'An evolutionary classifier for steel surface defects with small sample set,' EURASIP J. Image Video Process., vol. 2017, no. 1, pp. 1-13, Jul. 2017. - source

      Not checked in the free run.

    • cites [82] [82] S. R. Aghdam, E. Amid, and M. F. Imani, 'A fast method of steel surface defect detection using decision trees applied to LBP based features, ' in 7th IEEE Conf. Ind. Electron. Appl., Singapore, Singapore, 2012, pp. 1447-1452. - source

      Not checked in the free run.

    • cites [83] [83] S. Tian, and K. Xu, 'An algorithm for surface defect identification of steel plates based on genetic algorithm and extreme learning machine,' Metals, vol. 7, no. 8, pp. 1-11, Aug. 2017. - source

      Not checked in the free run.

    • cites [84] [84] R. Ren, T. Hung, and K. C. Tan, 'A generic deep-learning-based approach for automated surface inspection,' IEEE Trans. Cybern., vol. 48, no. 3, pp. 929-940, Mar. 2018. - source

      Not checked in the free run.

    • cites [85] [85] X. Tao, D. Zhang, W. Ma, X. Liu, and D. Xu, 'Automatic metallic surface defect detection and recognition with convolutional neural networks,' Appl. Sci.-Basel, vol. 8, no. 9, pp. 1-15, Sep. 2018. - source

      Not checked in the free run.

    • cites [86] [86] Y. Liu, K. Xu, and D. Wang, 'Online surface defect identification of cold rolled strips based on local binary pattern and extreme learning machine,' Metals, vol. 8, no. 3, pp. 1-18, Mar. 2018. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 1) Fourier Transform (FT) · p. 7

    The Fast Fourier Transform represents images' gray and geometrical features and realizes fast convolution and object recognition simultaneously.

    “The Fast Fourier Transform (FFT) not only represents the images' gray features and geometrical features but also realizes fast convolution and object recognition simultaneously.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    The biorthogonal wavelet transform can be constructed using the lifting-scheme and has compact support, time-frequency localization, high vanishing moments, anisotropy, strict sampling and adaptability.

    “The BWT can be constructed using the lifting-scheme, not only has the properties of compactly support, time-frequency localization, high vanishing moments and anisotropy but also provides the specific characteristics of strict sampling and adaptability.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Premisenone given
    § 3) Wavelet Transform (WT) · p. 8

    A wavelet transform may be an ideal candidate because it presents powerful judgment on the spatial-frequency characteristics both horizontally and vertically.

    “A WT may be an ideal candidate because it presents powerful judgment on the spatial-frequency characteristics both horizontally and vertically.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 3) Wavelet Transform (WT) · p. 8

    The undecimated wavelet transform was put forward to overcome the discrete wavelet transform's different wavelet coefficients issue and can produce more accurate frequency localization and is robust to scales, water marks and uneven illumination.

    “The discrete wavelet transform (DWT) results in different wavelet coefficients when transforming the original signal, the undecimated wavelet transform (UWT) was put forward to overcome this issue in [61]. Which can produce more accurate information for frequency localization and is robust to scales, water marks and uneven illumination.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [61] [61] X. Wu, K. Xu, and J. Xu, 'Application of undecimated wavelet transform to surface defect detection of hot rolled steel plates,' in Int. Congress Image Signal Process., Sanya, China, May. 2008, pp. 528-532. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § 7) Discrete Cosine Transform (DCT) · p. 8

    The FFT can increase operation speed but is not very competitive in image coding, especially in real-time processing.

    “Despite the FFT can faster the operation speed, it is not very competitive in image coding, especially in real-time processing.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 3) Wavelet Transform (WT) · p. 8

    The multi-scale and multidirection localization method was suggested to represent high-dimensional data of hot-rolled steels and has been proved to be less redundant.

    “The multi-scale and multidirection localization method (MGA) was suggested to represent the high-dimensional data of hot-rolled steels in [100], which has been proved to be less redundancy.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [100] [100] K. Xu, Y.-h. Ai, and X.-y. Wu, 'Application of multi-scale feature extraction to surface defect classification of hot-rolled steels,' Int. J. Minerals, Metallurgy, Mater., vol. 20, no. 1, pp. 37-41, Jan. 2013. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 2) Gabor Filter (GF) · p. 8

    Histogram of Oriented Gradients is invariant to geometric and optical deformations.

    “Invariant to geometric and optical deformations.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [93] [93] N. Dalal, and B. Triggs, 'Histograms of oriented gradients for human detection,' in Conf. Comput. Vis. Pattern Recognition, San Diego, CA, 2005, pp. 886-893. - source

      Not checked in the free run.

    • cites [94] [94] J. Masci, U. Meier, D. Ciresan, J. Schmidhuber, and G. Fricout, 'Steel defect classification with max-pooling convolutional neural networks,' in 2012 Int. Joint Conf. Neural Network, Brisbane, AUSTRALIA, 2012, pp. 1-6. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    The biorthogonal wavelet transform behaves remarkably for hot-rolled steel.

    “Based on the above advantages, it behaves remarkably for hot-rolled steel [92].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [92] [92] M. W. Ashour, F. Khalid, A. Abdul Halin, L. N. Abdullah, and S. H. Darwish, 'Surface defects classification of hot-rolled steel strips using multi-directional Shearlet features,' Arab. J. Sci. Eng., vol. 44, no. 4, pp. 2925-2932, Apr. 2018. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 5) Shearlet Transform (ST) · p. 8

    The shearlet transform sets up disparate direction amounts at diverse decomposition scales and is preeminent when approximating 2D smooth functions with discontinuities along C2-curves; it is fit to analyze images with complicated backgrounds and has been successfully applied to steel defects.

    “Compared to other methods, it sets up disparate direction amounts at diverse decomposition scales and is preeminent when approximating 2D smooth functions with discontinuities along the C2-curves; the ST is fit to analyze images with complicated backgrounds and has been successfully applied to defecting steel defects in [101, 102].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [101] [101] K. Xu, S. Liu, and Y. Ai, 'Application of Shearlet transform to classification of surface defects for metals,' Image Vis. Comput., vol. 35, pp. 23-30, Mar 2015. - source

      Not checked in the free run.

    • cites [102] [102] X. Liu, K. Xu, P. Zhou, D. Zhou, and Y. Zhou, 'Surface defect identification of aluminium strips with non-subsampled shearlet transform,' Opt. Lasers Eng., vol. 127, pp. 1-9, Apr. 2020. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 5) Shearlet Transform (ST) · p. 8

    Gabor wavelets cannot effectively depict directional properties because of their isotropic support and limited directivity.

    “Gabor wavelets cannot effectively depict the directional properties because of their isotropic support and limited directivity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 2) Gabor Filter (GF) · p. 8

    Gray-level Co-occurrence Matrix has a wide selection range of parameters and large calculation.

    “The selection range of parameters is wide and the calculation is large.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [34] [34] M. Chu, A. Wang, R. Gong, and M. Sha, 'Strip steel surface defect recognition based on novel feature extraction and enhanced least squares twin support vector machine,' ISIJ Int., vol. 54, no. 7, pp. 1638-1645, 2014. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [53] [53] Y. Chen, L. Chen, X. Liu, S. Ding, and H. Zhang, 'Real-time steel inspection system based on support vector machine and multiple kernel learning,' in 6th Int. Conf. Intell. Syst. Knowl. Eng., Shanghai, PEOPLES R CHINA, 2011, pp. 185-190. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [91] [91] X. Zhang, W. Li, J. Xi, Z. Zhang, and X. Fan, 'Surface defect target identification on copper strip based on adaptive genetic algorithm and feature saliency,' Math. Probl. Eng., vol. 2013, pp. 1-10, 2013. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [92] [92] M. W. Ashour, F. Khalid, A. Abdul Halin, L. N. Abdullah, and S. H. Darwish, 'Surface defects classification of hot-rolled steel strips using multi-directional Shearlet features,' Arab. J. Sci. Eng., vol. 44, no. 4, pp. 2925-2932, Apr. 2018. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 8

    The discrete cosine transform is constructed as a real domain transform based on the discrete Fourier transform, and most discriminative information about the steel surface is concentrated in a few DCT coefficients.

    “Based on the discrete Fourier transform, the DCT is constructed as a real domain transform, which has the characteristic that most of the discriminative information about the steel surface is concentrated in a few coefficients of the DCT [99].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] [99] F. Pernkopf, 'Detection of surface defects on raw steel blocks using Bayesian network classifiers,' Pattern Anal. Appl., vol. 7, no. 3, pp. 333-342, Dec. 2004. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 5) Shearlet Transform (ST) · p. 8

    The shearlet transform provides efficient multiscale directional representation and is a relatively new multiscale geometric analysis method.

    “The ST provides efficient multiscale directional representation, which is a relatively new MGA method.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    The curvelet transform is a higher-dimensional extension of the wavelet transform and was created to describe images of different scales and angles.

    “The CT is a higher-dimensional extension of the WT and was created to describe images of different scales and angles.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    The curvelet transform is a nonadaptive image representation method with two essential features: anisotropy scaling law and directionality.

    “CT is a nonadaptive image representation method with two essential features: anisotropy scaling law and directionality.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning
    § 3) Wavelet Transform (WT) · p. 8

    Because the wavelet basis function can only represent 1D directions, it is challenging for wavelet-based methods to sparsely express high-dimensional data due to line singularity and surface singularity.

    “Since the wavelet basis function can only represent 1D directions, it is challenging for wavelet based methods to sparsely express high-dimensional data due to their line singularity and surface singularity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2) Gabor Filter (GF) · p. 8

    Histogram of Oriented Gradients has high feature dimension and a large amount of calculation.

    “High feature dimension and large amount of calculation.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [93] [93] N. Dalal, and B. Triggs, 'Histograms of oriented gradients for human detection,' in Conf. Comput. Vis. Pattern Recognition, San Diego, CA, 2005, pp. 886-893. - source

      Not checked in the free run.

    • cites [94] [94] J. Masci, U. Meier, D. Ciresan, J. Schmidhuber, and G. Fricout, 'Steel defect classification with max-pooling convolutional neural networks,' in 2012 Int. Joint Conf. Neural Network, Brisbane, AUSTRALIA, 2012, pp. 1-6. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    The curvelet transform is applicable to characterize and analyze edge features with curved or linear shapes and performs well in continuous casting slabs.

    “It is applicable to characterize and analyze edge features with curved or linear shapes that perform well in continuous casting slabs [97].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [97] [97] Y. h. Ai, and K. Xu, 'Surface detection of continuous casting slabs based on curvelet transform and kernel locality preserving projections,' J. Iron Steel Res. Int., vol. 20, no. 5, pp. 80-86, May. 2013. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § 3) Wavelet Transform (WT) · p. 8

    Classical wavelet transforms can realize time-frequency localization and 1D data sparse representation.

    “Classical WTs can realize time-frequency localization and 1D data sparse representation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 2) Gabor Filter (GF) · p. 8

    Gray-level Co-occurrence Matrix can reflect the comprehensive distribution characteristic of the image gray level, such as direction, adjacent interval, and change range.

    “Can reflect the comprehensive distribution characteristic of the image gray level, such as direction, adjacent interval, change range, etc.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [34] [34] M. Chu, A. Wang, R. Gong, and M. Sha, 'Strip steel surface defect recognition based on novel feature extraction and enhanced least squares twin support vector machine,' ISIJ Int., vol. 54, no. 7, pp. 1638-1645, 2014. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [53] [53] Y. Chen, L. Chen, X. Liu, S. Ding, and H. Zhang, 'Real-time steel inspection system based on support vector machine and multiple kernel learning,' in 6th Int. Conf. Intell. Syst. Knowl. Eng., Shanghai, PEOPLES R CHINA, 2011, pp. 185-190. - source

      Not checked in the free run.

    • cites [91] [91] X. Zhang, W. Li, J. Xi, Z. Zhang, and X. Fan, 'Surface defect target identification on copper strip based on adaptive genetic algorithm and feature saliency,' Math. Probl. Eng., vol. 2013, pp. 1-10, 2013. - source

      Not checked in the free run.

    • cites [92] [92] M. W. Ashour, F. Khalid, A. Abdul Halin, L. N. Abdullah, and S. H. Darwish, 'Surface defects classification of hot-rolled steel strips using multi-directional Shearlet features,' Arab. J. Sci. Eng., vol. 44, no. 4, pp. 2925-2932, Apr. 2018. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    The biorthogonal wavelet transform is symmetric, preventing image content from shifting between subbands while allowing extensions at the boundaries.

    “BWT is symmetric, preventing image content from shifting between subbands while allowing for extensions at the boundaries.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 3) Wavelet Transform (WT) · p. 8

    Gabor wavelets have many useful characteristics and have performed well for defect classification in textured materials, but deep multilayer architecture can extract more influential features.

    “Gabor wavelets have many useful characteristics and have been performed well for defect classification in textured materials, which is naturally a single layer architecture, yet deep multilayer architecture is in a position to extract more influential features [96].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [96] [96] C. Dongyan, X. Kewen, N. Aslam, and H. Jingzhong, 'Defect classification recognition on strip steel surface using second-order cone programming-relevance vector machine algorithm,' J. Comput. Theor. Nanosci., vol. 13, no. 9, pp. 6141-6148, Step. 2016. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 4) Curvelet Transform (CT) and Biorthogonal Wavelet transform (BWT) · p. 8

    Curvelets need only a few coefficients to approximate curved singularities, making curvelet coefficients for pixels pertaining to a particular object particular.

    “Curvelets have very interesting properties, especially, only a few coefficients are needed to approximate the curved singularities, which makes the curvelet coefficients for pixels pertaining to a particular object particular.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 6) Scattering Transform (SCT) · p. 8

    The scattering transform improves tolerance to local deformations and builds nonlinear invariant representations for defects on hot-rolled steel strips by cascading wavelet transforms and modulus pooling operators.

    “By contrast, the SCT improves the tolerance ability of local deformations for current feature extraction and builds nonlinear invariant representations for the defects on hot-rolled steel strips by cascading wavelet transforms and modulus pooling operators [6], providing a new idea for defect classification.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] [6] K. Song, S. Hu, and Y. Yan, 'Automatic recognition of surface defects on hot-rolled steel strip using scattering convolution network,' J. Comput. Inf. Syst., vol. 10, no. 7, pp. 3049-3055, Apr. 2014. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    The Fourier transform is invariant to translation, expansion and rotation.

    “Invariance to translation, expansion and rotation.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [68] [68] K. Wiltschi, A. Pinz, and T. Lindeberg, 'An automatic assessment scheme for steel quality inspection,' Mach. Vis. Appl., vol. 12, no. 3, pp. 113-128, Oct. 2000. - source

      Not checked in the free run.

    • cites [96] [96] C. Dongyan, X. Kewen, N. Aslam, and H. Jingzhong, 'Defect classification recognition on strip steel surface using second-order cone programming-relevance vector machine algorithm,' J. Comput. Theor. Nanosci., vol. 13, no. 9, pp. 6141-6148, Step. 2016. - source

      Not checked in the free run.

    • cites [97] [97] Y. h. Ai, and K. Xu, 'Surface detection of continuous casting slabs based on curvelet transform and kernel locality preserving projections,' J. Iron Steel Res. Int., vol. 20, no. 5, pp. 80-86, May. 2013. - source

      Not checked in the free run.

    • cites [98] [98] G. Wu, H. Zhang, X. Sun, J. Xu, and K. Xu, 'A Bran-new feature extraction method and its application to surface defect recognition of hot rolled strips,' in IEEE Int. Conf. Autom. Logis., Jinan, China, Aug. 2007, pp. 2069-2074. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 7) Discrete Cosine Transform (DCT) · p. 8

    The discrete cosine transform requires a complicated number operation.

    “The DCT requires a complicated number operation.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisenone given
    § 2) Convolutional Autoencoder (CAE) · p. 9

    Autoencoder as an unsupervised learning method can automatically learn from a large amount of unlabelled data to obtain effective features.

    “As an unsupervised learning method, Autoencoder (AE) can automatically learn from a large amount of unlabelled data to obtain the effective features.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisereasoning
    § 2) Convolutional Autoencoder (CAE) · p. 9

    Transfer learning is often considered for classification tasks because most steel samples are unlabelled.

    “A novel method such as transfer learning is often been considered for classification tasks due to the most samples of steel are unlabelled.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Discrete cosine transform has great description and anti-noise ability.

    “Great description and anti-noise ability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] [99] F. Pernkopf, 'Detection of surface defects on raw steel blocks using Bayesian network classifiers,' Pattern Anal. Appl., vol. 7, no. 3, pp. 333-342, Dec. 2004. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 1) Convolutional Neural Network (CNN) · p. 9

    Transfer learning makes full use of previously labelled data and guarantees the precision of the model on new tasks with limited training samples, broadening the application prospect of the CNN.

    “At present, transfer learning makes full use of the previously labelled data and guarantees the precision of the model on new tasks with limited training samples [107], which broadens the application prospect of the CNN.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [107] [107] V. Natarajan, T.-Y. Hung, S. Vaikundam, L.-T. Chia, and Ieee, 'Convolutional networks for voting-based anomaly classification in metal surface inspection,' in IEEE Int. Conf. Ind. Technol., Toronto, CANADA, 2017, pp. 986-991. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 2) Convolutional Autoencoder (CAE) · p. 9

    A group of AutoEncoders were trained to reduce the dimension of extracted multiscale features, improving performance under inadequate training samples.

    “A group of AutoEncoders were also trained to reduce the dimension of the extracted multiscale features in [109], which improved the performance under inadequate training samples.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [109] [109] D. He, K. Xu, and D. Wang, 'Design of multi-scale receptive field convolutional neural network for surface inspection of hot rolled steels,' Image Vis. Comput., vol. 89, pp. 12-20, Sep. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Curvelet transform and biorthogonal wavelet transform have anisotropic scaling law and directionality.

    “Anisotropic scaling law and directionality.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [92] [92] M. W. Ashour, F. Khalid, A. Abdul Halin, L. N. Abdullah, and S. H. Darwish, 'Surface defects classification of hot-rolled steel strips using multi-directional Shearlet features,' Arab. J. Sci. Eng., vol. 44, no. 4, pp. 2925-2932, Apr. 2018. - source

      Not checked in the free run.

    • cites [97] [97] Y. h. Ai, and K. Xu, 'Surface detection of continuous casting slabs based on curvelet transform and kernel locality preserving projections,' J. Iron Steel Res. Int., vol. 20, no. 5, pp. 80-86, May. 2013. - source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Discrete cosine transform lacks good direction selectivity and multi-resolution analytical ability.

    “Lack of good direction selectivity and multi-resolution analytical ability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] [99] F. Pernkopf, 'Detection of surface defects on raw steel blocks using Bayesian network classifiers,' Pattern Anal. Appl., vol. 7, no. 3, pp. 333-342, Dec. 2004. -

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning
    § E. Deep-learning-based methods · p. 9

    Traditional features rely on manual design, tend to be simpler, mainly depend on prior information, and are difficult to use the advantage of big data.

    “Due to the reliance on manual design, these features tend to be simpler and mainly depend on some prior information, and it is difficult to use the advantage of big data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 1) Convolutional Neural Network (CNN) · p. 9

    CNNs can learn deep level features that cannot be extracted by traditional manual feature extraction under supervised learning, showing stronger discrimination.

    “They are the most basic network framework for deep learning, they can learn deep level features that cannot be extracted by traditional manual feature extraction methods under supervised learning manner, which show stronger discrimination[104].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [104] [104] S. Zhou, Y. Chen, D. Zhang, J. Xie, and Y. Zhou, 'Classification of surface defects on steel sheet using convolutional neural networks,' Mater. Tehnol., vol. 51, no. 1, pp. 123-131, 2017. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Scattering transform has high computational complexity.

    “High computational complexity”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] [6] K. Song, S. Hu, and Y. Yan, 'Automatic recognition of surface defects on hot-rolled steel strip using scattering convolution network,' J. Comput. Inf. Syst., vol. 10, no. 7, pp. 3049-3055, Apr. 2014. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Gabor filters have difficulty determining the optimal filter parameters and are not robust to rotation invariance.

    “Difficulty in determining the optimal filter parameters and not robust to rotation invariance.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [62] [62] R. Medina, F. Gayubo, L. M. Gonzá lez-Rodrigo, D. Olmedo, J. Gó mez-Garcí a-Bermejo, E. Zalama, and J. R. Perán, 'Automated visual classification of frequent defects in flat steel coils,' Int. J. Adv. Manuf. Technol., vol. 57, no. 9-12, pp. 1087-1097, Dec. 2011. - source

      Not checked in the free run.

    • cites [63] [63] D. C. Choi, Y. J. Jeon, S. J. Lee, J. P. Yun, and S. W. Kim, 'Algorithm for detecting seam cracks in steel plates using a Gabor filter combination method,' Appl. Optics, vol. 53, no. 22, pp. 4865-4872, Aug. 2014. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § E. Deep-learning-based methods · p. 9

    Deep-learning-based feature extraction can learn automatically from the massive data characteristics of multilayers.

    “By contrast, the deep-learning-based feature extraction method can learn automatically from the massive data characteristics of multilayers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2) Convolutional Autoencoder (CAE) · p. 9

    Xu et al. trained CAE to depict fine-grained features and fed them into a softmax classification layer to form a classification network.

    “Xu et al. [108] trained CAE to depict fine-grained features and fed them into a softmax classification layer to form a classification network.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [108] [108] D. He, K. Xu, P. Zhou, and D. Zhou, 'Surface defect classification of steels with a new semi-supervised learning method,' Opt. Lasers Eng., vol. 117, pp. 40-48, Feb. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Shearlet transform has good sparsity.

    “Good sparsity.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [101] [101] K. Xu, S. Liu, and Y. Ai, 'Application of Shearlet transform to classification of surface defects for metals,' Image Vis. Comput., vol. 35, pp. 23-30, Mar 2015. - source

      Not checked in the free run.

    • cites [102] [102] X. Liu, K. Xu, P. Zhou, D. Zhou, and Y. Zhou, 'Surface defect identification of aluminium strips with non-subsampled shearlet transform,' Opt. Lasers Eng., vol. 127, pp. 1-9, Apr. 2020. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 2) Convolutional Autoencoder (CAE) · p. 9

    Convolutional autoencoder combines the convolution and pooling operations of the convolutional neural network to extract more robust and discriminative representations.

    “CAE combines the convolution and pooling operations of the convolutional neural network to extract more robust and discriminative representations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Wavelet transform is capable of multi-resolution analysis and representing local signal features.

    “Capable of multi-resolution analysis and represent local signal features.”
    Quote in the paperNot a substantial claim0 of 5 sources checked
    • cites [4] [4] S. Ghorai, A. Mukherjee, M. Gangadaran, and P. K. Dutta, 'Automatic defect detection on hot-rolled flat steel products,' IEEE Trans. Instrum. Meas., vol. 62, no. 3, pp. 612-621, Mar. 2013. -

      Not checked in the free run.

    • cites [61] [61] X. Wu, K. Xu, and J. Xu, 'Application of undecimated wavelet transform to surface defect detection of hot rolled steel plates,' in Int. Congress Image Signal Process., Sanya, China, May. 2008, pp. 528-532. -

      Not checked in the free run.

    • cites [96] [96] C. Dongyan, X. Kewen, N. Aslam, and H. Jingzhong, 'Defect classification recognition on strip steel surface using second-order cone programming-relevance vector machine algorithm,' J. Comput. Theor. Nanosci., vol. 13, no. 9, pp. 6141-6148, Step. 2016. -

      Not checked in the free run.

    • cites [99] [99] F. Pernkopf, 'Detection of surface defects on raw steel blocks using Bayesian network classifiers,' Pattern Anal. Appl., vol. 7, no. 3, pp. 333-342, Dec. 2004. -

      Not checked in the free run.

    • cites [100] [100] K. Xu, Y.-h. Ai, and X.-y. Wu, 'Application of multi-scale feature extraction to surface defect classification of hot-rolled steels,' Int. J. Minerals, Metallurgy, Mater., vol. 20, no. 1, pp. 37-41, Jan. 2013. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § 1) Convolutional Neural Network (CNN) · p. 9

    The performance of CNN-based methods mainly depends on plentiful training samples, which stunts the utilization of CNNs in industrial scenes with small datasets.

    “However, the performance of CNN-based methods mainly depend on plentiful training samples, which stunts the utilization of CNNs in industrial scenes with small datasets.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    The process of curvelet transform and biorthogonal wavelet transform is complicated and not easy to realize.

    “The process is complicated and not easy realize.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [92] [92] M. W. Ashour, F. Khalid, A. Abdul Halin, L. N. Abdullah, and S. H. Darwish, 'Surface defects classification of hot-rolled steel strips using multi-directional Shearlet features,' Arab. J. Sci. Eng., vol. 44, no. 4, pp. 2925-2932, Apr. 2018. - source

      Not checked in the free run.

    • cites [97] [97] Y. h. Ai, and K. Xu, 'Surface detection of continuous casting slabs based on curvelet transform and kernel locality preserving projections,' J. Iron Steel Res. Int., vol. 20, no. 5, pp. 80-86, May. 2013. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § 2) Convolutional Autoencoder (CAE) · p. 9

    Steel surface image information is distinctive from most pretrained models, breaching transfer learning utilization conditions and making transfer learning to steel defect detection not as good as in other fields.

    “However, the image information of the steel surface is actually distinctive from most pretrained models, which breach the utilization conditions of transfer learning and makes the applications of transfer learning to steel defect detection not as good as the applications in other fields.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Convolutional Neural Network (CNN) · p. 9

    Multigroup CNN was created to build more effective and explainable feature map groups and can be used for feature extraction of hot-rolled steel images.

    “Taking the errors of localization result and background into consideration, multigroup CNN (MG-CNN) was created, which can build more effective and explainable feature map groups and can be used for feature extraction of hot-rolled steel images [106].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [106] [106] D. He, K. Xu, and P. Zhou, 'Defect detection of hot rolled steels with a new object detection framework called classification priority network,' Comput. Ind. Eng., vol. 128, pp. 290-297, Feb. 2019. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Convolutional Neural Network (CNN) · p. 9

    Standard CNNs lack multiple resolution pooling and are restricted to a constant size of input images; a multiscale pyramidal pooling network was presented to solve these problems.

    “Standard CNNs lack multiple resolution pooling and are restricted to a constant size of input images, A multiscale pyramidal pooling network (MSPPN) was presented to solve the above problems [105].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [105] [105] J. Masci, U. Meier, G. Fricout, and J. Schmidhuber, 'Multi-scale pyramidal pooling network for generic steel defect classification,' in Int. Joint Conf. Neural Networks, Dallas, TX, USA 2013, pp. 1-8. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 7) Discrete Cosine Transform (DCT) · p. 9

    Scattering transform has local translation invariance and elastic deformation stability.

    “Local translation invariance and elastic deformation stability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] [6] K. Song, S. Hu, and Y. Yan, 'Automatic recognition of surface defects on hot-rolled steel strip using scattering convolution network,' J. Comput. Inf. Syst., vol. 10, no. 7, pp. 3049-3055, Apr. 2014. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 1) Convolutional Neural Network (CNN) · p. 9

    Variants of CNNs have been proven to have record-breaking performance.

    “Variants of CNNs have been proven to have record-breaking performance.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § C. Feature Optimization · p. 11

    Recursive feature elimination ranks all features in descending order and selects an appropriate number of top features.

    “recursive feature elimination (RFE), which ranks all features in descending order, and an appropriate number of top features are selected [114].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [114] [114] S. Zhou, Y. Chen, D. Zhang, J. Xie, and Y. Zhou, 'Learning a class-specific and shared dictionary for classifying surface defects of steel sheet,' ISIJ Int., vol. 57, no. 1, pp. 123-130, Nov. 2017.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Premisenone given
    § 4) Brief Summary · p. 10

    Deep-learning-based methods can quickly learn modern and powerful feature representations from training data for strange applications.

    “Deep-learning-based methods can quickly learn modern and powerful feature representations from training data for strange applications.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 3) Manifold Learning (ML) · p. 9

    Zhao et al. proposed the discriminant manifold regularized local descriptor based on a new viewpoint learning mechanism, applying manifold structure to regularize the local descriptor for describing image features.

    “In contrast, Zhao et al. [87] proposed the discriminant manifold regularized local descriptor (DMRLD) based on the new viewpoint learning mechanism, which applies the manifold structure to regularize the local descriptor for describing the features of the image.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] [87] J. Zhao, Y. Peng, and Y. Yan, 'Steel surface defect classification based on discriminant manifold regularized local descriptor,' IEEE Access, vol. 6, pp. 71719-71731, Nov. 2018. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 3) Manifold Learning (ML) · p. 9

    DMRLD is designed for depicting useful and completed local features for steel surface defect classification without a high defect image quality requirement.

    “DMRLD is purposefully designed for depicting the useful and completed local feature for steel surface defect classification without a high defect image quality requirement.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] [87] J. Zhao, Y. Peng, and Y. Yan, 'Steel surface defect classification based on discriminant manifold regularized local descriptor,' IEEE Access, vol. 6, pp. 71719-71731, Nov. 2018. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § C. Feature Optimization · p. 11

    The adaptive genetic algorithm and the hybrid chromosome genetic algorithm were proposed to enhance visual feature selection.

    “To enhance the visual feature selection, the adaptive genetic algorithm (AGA) [91] and the hybrid chromosome genetic algorithm (HCGA) [67] were proposed.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [91] [91] X. Zhang, W. Li, J. Xi, Z. Zhang, and X. Fan, 'Surface defect target identification on copper strip based on adaptive genetic algorithm and feature saliency,' Math. Probl. Eng., vol. 2013, pp. 1-10, 2013. - source

      Not checked in the free run.

    • cites [67] [67] H. Hu, Y. Liu, M. Liu, and L. Nie, 'Surface defect classification in large-scale strip steel image collection via hybrid chromosome genetic algorithm,' Neurocomputing, vol. 181, pp. 86-95, Mar. 2016. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4) Brief Summary · p. 10

    Compared with a fully connected network, CNN reduces many parameters and simplifies calculation by adopting local connection, weight sharing and down sampling operations.

    “Compare with fully connected network, it reduces many parameters and simplifies the calculation by adopting local connection, weight sharing and down sampling operations.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [104] [104] S. Zhou, Y. Chen, D. Zhang, J. Xie, and Y. Zhou, 'Classification of surface defects on steel sheet using convolutional neural networks,' Mater. Tehnol., vol. 51, no. 1, pp. 123-131, 2017. - source

      Not checked in the free run.

    • cites [105] [105] J. Masci, U. Meier, G. Fricout, and J. Schmidhuber, 'Multi-scale pyramidal pooling network for generic steel defect classification,' in Int. Joint Conf. Neural Networks, Dallas, TX, USA 2013, pp. 1-8. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [106] [106] D. He, K. Xu, and P. Zhou, 'Defect detection of hot rolled steels with a new object detection framework called classification priority network,' Comput. Ind. Eng., vol. 128, pp. 290-297, Feb. 2019. - source

      Not checked in the free run.

    • cites [107] [107] V. Natarajan, T.-Y. Hung, S. Vaikundam, L.-T. Chia, and Ieee, 'Convolutional networks for voting-based anomaly classification in metal surface inspection,' in IEEE Int. Conf. Ind. Technol., Toronto, CANADA, 2017, pp. 986-991. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Premisecitation
    § 4) Brief Summary · p. 10

    Because all local parts share weights, convolutional autoencoder does not take into account the difference of contribution to the whole between each local part.

    “Because all local parts share weights, it does not take into account the difference of contribution to the whole between each local part.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [108] [108] D. He, K. Xu, P. Zhou, and D. Zhou, 'Surface defect classification of steels with a new semi-supervised learning method,' Opt. Lasers Eng., vol. 117, pp. 40-48, Feb. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [109] [109] D. He, K. Xu, and D. Wang, 'Design of multi-scale receptive field convolutional neural network for surface inspection of hot rolled steels,' Image Vis. Comput., vol. 89, pp. 12-20, Sep. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § B. Dimension Reduction · p. 11

    Kernel LPP was designed by implementing LPP in a kernel space and was reported to solve nonlinear problems well when applied on surface inspection for con-casting slabs.

    “Xu et al. designed a kernel LPP (KLPP) by implementing the LPP in a kernel space, which was reported to solve nonlinear problems well when applied on the surface inspection for con-casting slabs [97].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [97] [97] Y. h. Ai, and K. Xu, 'Surface detection of continuous casting slabs based on curvelet transform and kernel locality preserving projections,' J. Iron Steel Res. Int., vol. 20, no. 5, pp. 80-86, May. 2013. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § C. Feature Optimization · p. 11

    The genetic algorithm is a typical random search algorithm.

    “The genetic algorithm (GA) is a typical random search algorithm [98].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [98] [98] G. Wu, H. Zhang, X. Sun, J. Xu, and K. Xu, 'A Bran-new feature extraction method and its application to surface defect recognition of hot rolled strips,' in IEEE Int. Conf. Autom. Logis., Jinan, China, Aug. 2007, pp. 2069-2074. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § B. Dimension Reduction · p. 11

    The incorporation of PCA with bootstrap aggregating reduces not only the dimensions but also the variance in the decision trees.

    “A variety of improved methods based on concrete scenes have emerged, such as the incorporation of the PCA with bootstrap aggregating, which reduces not only the dimensions but also the variance in the decision trees [82].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [82] [82] S. R. Aghdam, E. Amid, and M. F. Imani, 'A fast method of steel surface defect detection using decision trees applied to LBP based features, ' in 7th IEEE Conf. Ind. Electron. Appl., Singapore, Singapore, 2012, pp. 1447-1452. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § C. Feature Optimization · p. 11

    Relief is a representative heuristic algorithm mainly based on sample learning; it is straightforward and efficient but restricted to binary classification.

    “A representative heuristic algorithm, Relief, is mainly based on sample learning. This method is straightforward and efficient but is restricted to binary classification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § VI. FEATURE SELECTION · p. 10

    A large number of features may increase the complexity of the classifier, slow down the training process and seriously affect the final classification accuracy and efficiency.

    “A large number of features may increase the complexity of the classifier, slow down the training process and seriously affect the final classification accuracy and efficiency.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given
    § 4) Brief Summary · p. 10

    Research methods used to design networks automatically are helpful to design space exploration of deep learning networks and play an essential role in accelerating network design and promoting application of deep learning in engineering.

    “The research methods used to design the network automatically is helpful to the design space exploration of deep learning networks and plays an essential role in accelerating the process of network design and promoting the application of deep learning in engineering.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § 3) Manifold Learning (ML) · p. 9

    Manifold regularization adds items related to manifolds to regularization items to play the role of semisupervision by using geometric structures in the data.

    “Correlatively, manifold regularization adds items related to the manifolds to the regularization items to play the role of semisupervision by using the geometric structures in the data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 4) Brief Summary · p. 10

    Manifold learning has high computational burden and poor classification ability.

    “High computational burden and poor classification ability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] [87] J. Zhao, Y. Peng, and Y. Yan, 'Steel surface defect classification based on discriminant manifold regularized local descriptor,' IEEE Access, vol. 6, pp. 71719-71731, Nov. 2018. -

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning
    § B. Dimension Reduction · p. 11

    After feature evaluation, the dimension of feature subsets may be too large and redundant, reducing the robustness and generalization speed of the classifier, so dimension reduction is necessary.

    “After feature evaluation, the dimension of the feature subsets may be too large, and there exists much redundancy, which will reduce the robustness and generalization speed of the classifier, so the process of dimension reduction is necessary.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisereasoning
    § C. Feature Optimization · p. 11

    After feature evaluation and dimension reduction, the acquired subsets may not be optimal for the classifiers; thus, feature optimization is essential in improving overall performance.

    “After the processing of the two steps mentioned above, the subsets acquired may not be optimal for the classifiers; thus, feature optimization is essential in improving the overall performance.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § C. Feature Optimization · p. 11

    There are two traditional optimization algorithms based on searching strategies: random search algorithms and heuristic algorithms.

    “There are two traditional optimization algorithms based on searching strategies: random search algorithms and heuristic algorithms .”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § B. Dimension Reduction · p. 11

    The calculation of weight matrices and the selection of constraint matrices are the main differences in graph embedding dimensionality reduction methods.

    “The calculation of weight matrices and the selection of constraint matrices are the main differences in these methods [100].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [100] [100] K. Xu, Y.-h. Ai, and X.-y. Wu, 'Application of multi-scale feature extraction to surface defect classification of hot-rolled steels,' Int. J. Minerals, Metallurgy, Mater., vol. 20, no. 1, pp. 37-41, Jan. 2013. -

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning
    § VI. FEATURE SELECTION · p. 10

    Feature selection is needed to avoid the curse of dimensionality and to enhance the generalization ability of the designed classifier.

    “One reason feature selection is needed is to avoid the so-called curse of dimensionality, and another is to enhance the generalization ability of the designed classifier.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 4) Brief Summary · p. 10

    Compared with the traditional autoencoder, convolutional autoencoder can well retain the spatial information of 2D signals and can be used in unsupervised learning.

    “Compared with the traditional autoencoder, it can well retain the spatial information of 2D signals and can be used in unsupervised learning.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [108] [108] D. He, K. Xu, P. Zhou, and D. Zhou, 'Surface defect classification of steels with a new semi-supervised learning method,' Opt. Lasers Eng., vol. 117, pp. 40-48, Feb. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [109] [109] D. He, K. Xu, and D. Wang, 'Design of multi-scale receptive field convolutional neural network for surface inspection of hot rolled steels,' Image Vis. Comput., vol. 89, pp. 12-20, Sep. 2019. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § A. Feature Evaluation · p. 10

    In [110], the filter approach assigned weights before induction and the wrapper approach ran an induction algorithm on the training set and used the accuracy of the resulting description to evaluate the mill scale defect feature set.

    “Two main approaches were used in [110], where the filter approach assigned the weights before induction, the wrapper approach ran an induction algorithm on the training set and used the accuracy of the resulting description to evaluate the mill scale defect feature set.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [110] [110] S. Lechwar, Ł. Rauch, and M. Pietrzyk, 'Use of Artificial Intelligence in Classification of Mill Scale Defects,' Steel Res. Int., vol. 86, no. 3, pp. 266-277, Mar. 2015. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § 4) Brief Summary · p. 10

    Academia and industry have carried out much research on structure optimization to speed up the promotion and application of deep learning features.

    “To speed up the promotion and application of deep learning features, academia and industry have carried out much research on structure optimization.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 4) Brief Summary · p. 10

    Manifold learning can find the essence from observed phenomena and find the internal law of data generation.

    “Be able to find the essence of from the observed phenomena, and find the internal law of data generation.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] [87] J. Zhao, Y. Peng, and Y. Yan, 'Steel surface defect classification based on discriminant manifold regularized local descriptor,' IEEE Access, vol. 6, pp. 71719-71731, Nov. 2018. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 3) Manifold Learning (ML) · p. 9

    Earlier local descriptors such as LBP and HOG are subject to hand-crafted definition and application limitations.

    “The earlier local descriptors, such as the LBP and HOG, subject to the hand-crafted definition and the limitations of the applications.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning
    § 6) Other Classifiers and brief summary · p. 13

    Supervised learning has strong adaptability to data, especially with sufficient labelled data, to obtain better results.

    “This kind of mind-set has strong adaptability to the data, especially in the case of sufficient labelled data to obtain better results.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Classifiers based on NN · p. 12

    Neuhauser et al. altered the GoogLeNet architecture and enabled transform learning to expedite the training process and improve performance of the network by compensating for the limited training set.

    “Neuhauser et al. [117] altered the GoogLeNet architecture and enabled transform learning to expedite the training process and to improve performance of the network by compensating for the limited training set.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [117] [117] F. M. Neuhauser, G. Bachmann, and P. Hora, 'Surface defect classification and detection on extruded aluminum profiles using convolutional neural networks,' Int. J. Mater. Form., vol. 13, no. 4, pp. 591-603, Jul. 2020.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2) Classifiers based on SVM · p. 12

    Chu's team came up with a series of SVM variants: enhanced least squares twin SVM, multiple support vector hypersphere with feature and sample weights, machine learning with quantile hyperspheres, and multiple hyperspheres vector machine with additional information.

    “Chu's team kept improving the algorithm and came up with a series of SVM variants, such as the enhanced least squares twin SVM [34], multiple support vector hypersphere with feature and sample weights [39], machine learning with quantile hyperspheres [121] and multiple hyperspheres vector machine with additional information [54].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [34] [34] M. Chu, A. Wang, R. Gong, and M. Sha, 'Strip steel surface defect recognition based on novel feature extraction and enhanced least squares twin support vector machine,' ISIJ Int., vol. 54, no. 7, pp. 1638-1645, 2014. -

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [39] [39] R. Gong, C. Wu, M. Chu, and X. Liu, 'The strip steel surface defect recognition based on multiple support vector hyper-sphere with feature and sample weights,' Steel Res. Int., vol. 87, no. 12, pp. 1678-1685, Dec. 2016. - source

      Not checked in the free run.

    • cites [121] [121] M. Chu, J. Zhao, X. Liu, and R. Gong, 'Multi-class classification for steel surface defects based on machine learning with quantile hyper-spheres,' Chemometrics Intell. Lab. Syst., vol. 168, pp. 15-27, Sep. 2017.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Classifiers based on NN · p. 12

    He et al. designed a hierarchical learning scheme combining CNN and AutoEncoder, which greatly improved the performance of the model with insufficient training samples.

    “Moreover, He et al. [109] designed a hierarchical learning scheme by combining the CNN and the AutoEncoder, which has greatly improved the performance of the model with insufficient training samples.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [109] [109] D. He, K. Xu, and D. Wang, 'Design of multi-scale receptive field convolutional neural network for surface inspection of hot rolled steels,' Image Vis. Comput., vol. 89, pp. 12-20, Sep. 2019. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisereasoning
    § 1) Classifiers based on NN · p. 12

    Compressing and accelerating the model under the premise of ensuring the accuracy of the network has become a hot spot of discussion in the field of network structure optimization.

    “Therefore, compressing and accelerating the model under the premise of ensuring the accuracy of the network has become a hot spot of discussion in the field of network structure optimization.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning
    § B. Unsupervised Learning · p. 13

    Studying unsupervised classification methods greatly benefits surface defect recognition for flat steel.

    “Thus, studying unsupervised classification methods greatly benefits surface defect recognition for flat steel.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: In real-world flat steel industrial production lines, collecting and manually labeling a large number of flat steel surface defect data is extremely labour intensive and time consuming, compounded by harsh manufacturing environments.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 1) Classifiers based on NN · p. 12

    Networks with multiple convolutional layers can do better in regression and classification tasks than shallow networks.

    “In general, networks with multiple convolutional layers can do better in regression and classification tasks than shallow networks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 3) Classifiers based on the Distance Function · p. 12

    The distance function is one of the simplest tools for defect classification.

    “The distance function is one of the simplest tools for defect classification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 3) Classifiers based on the Distance Function · p. 12

    Luo et al. selected the nearest neighbour classifier as the dissimilarity metric between two multi-region histograms to identify LBP-like histograms.

    “Luo et al. [10, 90] selected the NNC as the dissimilarity metric between two multi-region histograms to identify LBP-like histograms.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [10] [10] Q. Luo, X. Fang, Y. Sun, L. Liu, J. Ai, C. Yang, and O. Simpson, 'Surface defect classification for hot-rolled steel strips by selectively dominant local binary patterns,' IEEE Access, vol. 7, pp. 23488-23499, Feb. 2019. -

      Not checked in the free run.

    • cites [90] [90] Q. Luo, Y. Sun, P. Li, O. Simpson, L. Tian, and Y. He, 'Generalized completed local binary patterns for time-efficient steel surface defect classification,' IEEE Trans. Instrum. Meas., vol. 68, no. 3, pp. 667-679, Mar. 2019. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 3) Classifiers based on the Distance Function · p. 12

    Distance is an important measure to describe the relationship between pixels.

    “Distance is an important measure to describe the relationship between pixels.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 1) Classifiers based on NN · p. 12

    Fu et al. presented a compact yet powerful CNN model by training low-level features and incorporating multiple receptive fields; this method was reported to realize high classification accuracy of steel surface defects based on limited defect-specific training samples.

    “To achieve high classification accuracy with a small training sample set, Fu et al. [116] presented a compact yet powerful CNN model by training the low-level features and incorporating multiple receptive fields. This method was reported to realize high classification accuracy of steel surface defects based on limited defect-specific training samples.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [116] [116] G. Fu, P. Sun, W. Zhu, J. Yang, Y. Cao, M. Y. Yang, and Y. Cao, 'A deep-learning-based approach for fast and robust steel surface defects classification,' Opt. Lasers Eng., vol. 121, pp. 397-405, Oct. 2019.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given
    § 5) Classifiers based on Multi-Classifier Fusion · p. 13

    The iteration scale of multi-classifier fusion expands with the increase in the number of weak classifiers, and more training time is spent.

    “The iteration scale expands with the increase of the number in weak classifiers, and correspondingly, more training time will be spent.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 3) Classifiers based on the Distance Function · p. 12

    Zhao et al. simultaneously considered the variation and exemplar distances of the nearest neighbour classifier to measure the similarity between local models.

    “Moreover, Zhao et al. [87] simultaneously considered the variation and exemplar distances of the NNC to measure the similarity between the local models.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] [87] J. Zhao, Y. Peng, and Y. Yan, 'Steel surface defect classification based on discriminant manifold regularized local descriptor,' IEEE Access, vol. 6, pp. 71719-71731, Nov. 2018. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § D. Brief Summary · p. 11

    Feature selection plays a critical role in enhancing the performance of the classifier and indirectly influences the process of steel defect recognition.

    “Feature selection plays a critical role in enhancing the performance of the classifier and indirectly influences the process of steel defect recognition.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § VII. DEFECT CLASSIFIER · p. 11

    Current classifiers can be classified into supervised, unsupervised and semi-supervised learning-based classifiers according to whether the training sample has a label or not.

    “According to whether the training sample has a label or not, the current classifiers can be classified into supervised, unsupervised and semi-supervised learning-based classifiers.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § 2) Classifiers based on SVM · p. 12

    Agarwal et al. proposed a process-knowledge-based multiclass SVM combining feature extraction with process knowledge; the PK-MSVM performed better than the traditional multiclass SVM.

    “Inspired by [59], Agarwal et al. [119] proposed a classification scheme of the process-knowledge-based multiclass SVM (called PK-MSVM) by combining the feature extraction task of the automated defect inspection with the process knowledge; the PK-MSVM performed better than the traditional multiclass SVM.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [119] [119] K. Agarwal, R. Shivpuri, Y. Zhu, T.-S. Chang, and H. Huang, 'Process knowledge based multi-class support vector classification (PK-MSVM) approach for surface defects in hot rolling,' Expert Syst. Appl., vol. 38, no. 6, pp. 7251-7262, Jun. 2011. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisereasoning
    § 3) Classifiers based on the Distance Function · p. 12

    Further increasing the classification accuracy requires exploring high-dimensional features, in return sacrificing the classification speed.

    “Further increasing the classification accuracy requires exploring high-dimensional features, in return, sacrificing the classification speed.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given
    § 3) Classifiers based on the Distance Function · p. 12

    The nearest neighbour classifier is the commonly used distance classifier for surface defect classification, where chi-square distance is a simple but effective criterion for measuring similarity between training and test images.

    “The nearest neighbour classifier (NNC) is the commonly used distance classifier for surface defect classification, where the chi-square distance is a simple but effective criterion for measuring the similarity between the training images and the test images.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 1) Classifiers based on NN · p. 12

    The expensive training step limits the generalization of nascent neural networks.

    “However, the expensive training step limits the generalization of nascent neural networks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 6) Other Classifiers and brief summary · p. 13

    Supervised learning classification methods can execute sufficient training and learning on images to obtain the most effective representation.

    “The classification methods based on supervised learning can excute sufficient training and learning on the images to obtain the most effective representation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § 3) Classifiers based on the Distance Function · p. 12

    The accuracy of this kind of conventional classifier primarily relies on the precision of the extracted features.

    “The accuracy of this kind of conventional classifier primarily relies on the precision of the extracted features.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisereasoning
    § 4) Classifiers based on Sparse Representation · p. 13

    Classification methods based on dictionary learning and sparse representation do not have an explicit stage of feature extraction, which might be more effective in surface defect classification for flat steel.

    “Different from the conventional classification process that treats feature extraction and classifier training as two separate steps, classification methods based on dictionary learning and sparse representation do not have an explicit stage of feature extraction, which might be more effective in surface defect classification for flat steel.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Premisereasoning
    § 1) Classifiers based on NN · p. 12

    A large number of application scenarios cannot provide the necessary resources.

    “However, a large number of application scenarios cannot provide the necessary resources.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 5) Classifiers based on Multi-Classifier Fusion · p. 13

    The performance of multi-classifier fusion mainly depends on the selection of weak classifiers.

    “However, the performance of multi-classifier fusion mainly depends on the selection of weak classifiers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § 1) Classifiers based on NN · p. 12

    There exist network architectures containing more than 100 convolutional layers.

    “At present, there exists network architectures containing more than 100 convolutional layers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § A. Supervised Learning · p. 11

    The performance of supervised classification methods mainly relies on two factors: the quality of the features and the capability of the classifier.

    “The performance of supervised classification methods mainly relies on two factors: the quality of the aforementioned features and the capability of the classifier.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 4) Classifiers based on Sparse Representation · p. 13

    Zhou et al. utilized a high-quality dictionary to extract compact, reconstructive and discriminative features of the test images.

    “Furthermore, Zhou et al. [114] utilized the high-quality dictionary to extract the compact, reconstructive and discriminative features of the test images.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [114] [114] S. Zhou, Y. Chen, D. Zhang, J. Xie, and Y. Zhou, 'Learning a class-specific and shared dictionary for classifying surface defects of steel sheet,' ISIJ Int., vol. 57, no. 1, pp. 123-130, Nov. 2017.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § 3) Classifiers based on the Distance Function · p. 12

    In [74], the authors recommended K nearest neighbour as a classifier to realize defect classification for cold-rolled steel with a co-occurrence matrix.

    “In [74], the authors recommended the K nearest neighbour (KNN) as a classifier to realize defect classification for cold-rolled steel with a co-occurrence matrix.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [74] [74] F. Dupont, C. Odet, and M. Cartont, 'Optimization of the recognition of defects in flat steel products with the cost matrices theory,' Ndt E Int., vol. 30, no. 1, pp. 3-10, Feb. 1997. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § 2) Classifiers based on SVM · p. 12

    Xiao et al. constructed a multiple classifier system for classifying steel surface defects, containing multiple SVM classifiers and a Bayes kernel classifier; the Bayes kernel fuses results and adjusts hybrid parameters with only a small sample set.

    “Xiao et al. [81] constructed a multiple classifier system for classifying steel surface defects, which contains multiple SVM classifiers and a Bayes kernel classifier. The Bayes kernel fuses the results from the multiple classifiers and adjusts the hybrid parameters with only a small sample set.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [81] [81] M. Xiao, M. Jiang, G. Li, L. Xie, and L. Yi, 'An evolutionary classifier for steel surface defects with small sample set,' EURASIP J. Image Video Process., vol. 2017, no. 1, pp. 1-13, Jul. 2017. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 2) Classifiers based on SVM · p. 12

    SVM is a generalized linear classifier for binary classification and has been widely used for defect classification of flat steel surfaces.

    “SVM is a generalized linear classifier for binary classification, which has been widely used for the defect classification of flat steel surfaces [67, 100-102, 118].”
    Quote in the paperNot a substantial claim0 of 5 sources checked
    • cites [67] [67] H. Hu, Y. Liu, M. Liu, and L. Nie, 'Surface defect classification in large-scale strip steel image collection via hybrid chromosome genetic algorithm,' Neurocomputing, vol. 181, pp. 86-95, Mar. 2016. - source

      Not checked in the free run.

    • cites [100] [100] K. Xu, Y.-h. Ai, and X.-y. Wu, 'Application of multi-scale feature extraction to surface defect classification of hot-rolled steels,' Int. J. Minerals, Metallurgy, Mater., vol. 20, no. 1, pp. 37-41, Jan. 2013. - source

      Not checked in the free run.

    • cites [101] [101] K. Xu, S. Liu, and Y. Ai, 'Application of Shearlet transform to classification of surface defects for metals,' Image Vis. Comput., vol. 35, pp. 23-30, Mar 2015. - source

      Not checked in the free run.

    • cites [102] [102] X. Liu, K. Xu, P. Zhou, D. Zhou, and Y. Zhou, 'Surface defect identification of aluminium strips with non-subsampled shearlet transform,' Opt. Lasers Eng., vol. 127, pp. 1-9, Apr. 2020. - source

      Not checked in the free run.

    • cites [118] [118] X. Liu, K. Xu, D. Zhou, and P. Zhou, 'Improved contourlet transform construction and its application to surface defect recognition of metals,' Multidimens. Syst. Signal Process., vol. 31, no. 3, pp. 951-964, Jul. 2020.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning
    § D. Brief Summary · p. 14

    Unsupervised learning methods separate samples of the same type through different characteristics of the data and do not need label information.

    “Unsupervised learning methods mainly separate samples of the same type through the different characteristics of the data, and therefore, they do not need the label information from the data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § D. Brief Summary · p. 14

    Results reached with a limited sample set can match results obtained with a massive sample set.

    “The results reached with limited sample set can also match the results obtained with massive sample set.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given
    § D. Brief Summary · p. 14

    Supervised learning methods need labelled samples for model training, making full use of prior information in data categories.

    “Generally, supervised learning methods need labelled samples for model training, which makes full use of the prior information in the data categories.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § VIII. SUMMARY AND DISCUSSION · p. 16

    The latest image processing or machine learning methods provide novel ways to enhance the image quality of flat steel.

    “The latest image processing or machine learning methods provide novel way to enhance the image quality of flat steel.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning
    § VIII. SUMMARY AND DISCUSSION · p. 17

    Internal prior utilization, multimode fusion and interdisciplinary crossing can give full play to their respective advantages.

    “The accuracy of a single-mode algorithm for classification is always limited, and internal prior utilization, multimode fusion and interdisciplinary crossing can give full play to their respective advantages, which is the trend of future development.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The accuracy of a single-mode algorithm for classification is always limited.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given
    § C. Semisupervised Learning · p. 14

    When the quantity of original samples is limited and the distribution is uneven, semisupervised learning methods may become unreliable.

    “In specific industrial practice, the quantity of the original samples might be limited (lack of representativeness and with contingencies), and the distribution can be uneven. If these situations happen, the semisupervised learning methods may become unreliable, creating difficulties for its development.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 17

    Achieving high-precision classification with only a small number of labelled samples is still a hot issue.

    “Meanwhile, achieving high-precision classification in the case of only a small number of labelled samples is still a hot issue”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 16

    Excessive deletion of redundant features leads to the loss of a large amount of information.

    “In addition, excessive deletion of redundant features leads to the loss of a large amount of information.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § B. Unsupervised Learning · p. 14

    Current unsupervised learning results show a disparity from the ultimate goal.

    “The current unsupervised learning results show that there is still some disparity from our ultimate goal.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning
    § VIII. SUMMARY AND DISCUSSION · p. 16

    Designing a feature selection method for high-dimensional small sample data without decreasing the prediction accuracy of the feature subset is challenging.

    “Designing a feature selection method for high-dimensional small sample data without decreasing the prediction accuracy of the feature subset is challenging.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § VIII. SUMMARY AND DISCUSSION · p. 16

    GAN has realized extreme achievement in image enhancement, providing a better solution to insufficient samples, feature extraction difficulties, and poor image quality.

    “For instance, the GAN has realized extreme achievement in image enhancement, providing a better solution to the problems of insufficient samples, feature extraction difficulties and poor image quality.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [131] [131] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, 'Generative adversarial nets,' ArXiv , pp. 1-9, 06/01, 2014.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § VIII. SUMMARY AND DISCUSSION · p. 16

    Obtaining a real-time, reliable and stable fusion feature extraction scheme will always be a research hotspot.

    “At the same time, obtaining a real-time, reliable and stable fusion feature extraction scheme will always be a research hotspot.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 17

    Some techniques that have shown prominent results on image enhancement of texture images with rich datasets in multimedia areas are not necessarily suitable for enhancement of images from industrial production lines where labelled image sets are extremely limited.

    “For instance, some techniques have shown prominent results on the image enhancement (i.e., noise removal) of many texture images with rich datasets in multimedia areas, although they are not necessarily suitable for the enhancement of images gathered from industrial production lines where labelled image sets are extremely limited.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 16

    Due to the strong randomness of features of flat steel surface images, it is very difficult to use a single mathematical model to extract features.

    “Due to the strong randomness of the features of the flat steel surface image, it is very difficult to use a single mathematical model to extract features.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 16

    Making full use of prior knowledge of human visual perception characteristics and combining it with feature extraction methods could be helpful to describe the image more effectively.

    “In addition, making full use of the prior knowledge of human visual perception characteristics and combining it with feature extraction methods could be helpful to describe the image more effectively.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 16

    After feature selection, the prediction accuracy of the high-dimensional small sample dataset may decline.

    “After feature selection, the prediction accuracy of the high-dimensional small sample dataset may decline.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 16

    Image preprocessing usually makes the contrast, visual effect and entropy of the image reach a balance while improving image contrast.

    “Image preprocessing usually makes the contrast, visual effect and entropy of the image reach a balance on the premise of improving the image contrast.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning
    § D. Brief Summary · p. 14

    It is challenging for unsupervised learning methods to realize the high accuracy of supervised classification.

    “It is challenging for unsupervised learning methods to realize the high accuracy of supervised classification.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given
    § D. Brief Summary · p. 14

    Semisupervised learning methods utilize labelled samples and massive unlabelled samples to train classifiers.

    “Semisupervised learning methods joint the previous two manners, which utilizes the labelled samples and massive unlabelled samples to train classifiers.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given
    § VIII. SUMMARY AND DISCUSSION · p. 16

    The feature selection algorithm is subject to the data distribution; if the training set changes, the feature subset is also very different.

    “The feature selection algorithm is subject to the data distribution. If the training set changes, the feature subset is also very different.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given
    § D. Brief Summary · p. 14

    Defect classification methods focus on extracting features of different kinds of defects and classifying defects with similar features into one category, compared with machine learning-based methods for defect detection.

    “Compared with machine learning-based methods for defect detection, defect classification methods focus on extracting the features of different kinds of defects and classifying the defects with similar features into one category.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § D. Brief Summary · p. 14

    The number of labels needed to get comparable results to fully supervised learning is decreasing.

    “We noticed that, the number of labels to get comparable results to fully supervised learning is decreasing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationnone given
    § D. Brief Summary · p. 14

    Some common ideas are not often combined, and the combination of broad range and unusual methods is beneficial.

    “We identified that some common ideas are not often combined and that the combination of broad range and unusual methods is beneficial.”
    Quote in the paperNot a substantial claim
183 of 183

  • The paper relies on this without support.

    The natural image processing procedure of defect recognition consists of five successive parts: image acquisition, image preprocessing, feature extraction, feature selection, and defect classifier.

    “According to the natural image processing procedure of defect recognition, the diverse approaches are grouped into five successive parts: image acquisition, image preprocessing, feature extraction, feature selection and defect classifier.”

  • The paper relies on this without support.

    The traditional spatial domain feature set is very complex, very laborious to extract, and difficult to guarantee real-time recognition of flat steel surface defects.

    “The traditional spatial domain feature set is very complex, very laborious to extract and difficult to guarantee the real-time recognition of flat steel surface defects.”

  • The paper relies on this without support.

    To support near-real-time operation, feature extraction needs to be competitive in feature dimensions and algorithm properties.

    “To support near-real-time operation, the feature extraction need to be competitive in feature dimensions and algorithm properties.”

  • The authors state this as an assumption.

    The data are a low-dimensional manifold uniformly sampled in a high-dimensional Euclidean space.

    “Supposing the data are a low-dimensional manifold uniformly sampled in a high-dimensional Euclidean space”

  • The authors state this as an assumption.

    Graph embedding dimensionality reduction methods assume the data set is on a low-dimensional manifold embedded in the raw high-dimensional feature space.

    “which are all based on the assumption that the data set is on a low-dimensional manifold embedded in the raw high-dimensional feature space.”

  • The paper relies on this without support.

    Most existing steel samples are not marked because of the randomness and limitations of the production line, so only a small amount of data with safety labels can be used for learning, resulting in instability of the final classification results.

    “However, most of the existing steel samples are not marked because of the randomness and limitations of the production line, so only a small amount of data with safety labels can be used for learning, resulting in the instability of the final classification results.”

  • The paper relies on this without support.

    In real-world flat steel industrial production lines, collecting and manually labeling a large number of flat steel surface defect data is extremely labour intensive and time consuming, compounded by harsh manufacturing environments.

    “In real-world flat steel industrial production lines, it is extremely labour intensive and time consuming to collect a large number of flat steel surface defect data, and label them manually; this challenge is compounded by the harsh manufacturing environment (e.g., high temperature, poor lighting conditions, mechanical vibrations, and dense dust).”

  • The paper relies on this without support.

    The imbalance of defect samples in actual flat steel production lines is a real problem that prompts researchers to use limited samples to achieve similar results.

    “However, the imbalance of defect samples in actual flat steel production lines has prompted researchers to use limited samples to achieve similar results.”

  • The paper relies on this without support.

    The quality of the input image and the original parameters of the classification model are the keys to determining the classification performance of the unsupervised learning model.

    “However, the quality of the input image and the original parameters of the classification model are the keys to determining the classification performance of the unsupervised learning model.”

  • The paper relies on this without support.

    The accuracy of a single-mode algorithm for classification is always limited.

    “The accuracy of a single-mode algorithm for classification is always limited”
10 of 10

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • In actual steel production, high-quality datasets for training machine learning algorithms are very difficult to obtain because defects are small probability events and obtaining many samples with various defect types is difficult.

    “In the actual production environment, it is very difficult to obtain high quality datasets for training machine learning related algorithms because defects in steel production are originally small probability events, and it is very difficult to obtain many samples with various types of defects”

    From Steel Surface Defect Recognition: A Survey (2022) · cites this paper as [4]

    The discussion of supervised classifiers (p. 13) says that defect samples are imbalanced on actual flat-steel production lines and that researchers work with limited samples. The Summary and Discussion (pp. 16–17) also describes massive labeling as unrealistic and small labeled datasets as an ongoing challenge. These passages support the difficulty of obtaining training data, but do not establish the claim’s specific explanation that steel defects are small-probability events. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Labeling steel surface defect data is costly and labor intensive.

    “the labeling of the data is also costly and labor intensive [4][5][6][7].”

    From Steel Surface Defect Recognition: A Survey (2022) · cites this paper as [4]

    Table V in the Summary and Discussion (p. 16) says it is unrealistic to label massive numbers of flat-steel surface defects, supporting the claim that labeling is burdensome. The discussion of small numbers of labelled samples (p. 17) reinforces that limitation, but the paper does not explicitly say labeling is costly or labor intensive. (LLM, conf=0.81)

    Experimental — accuracy not yet measured

  • The steel surface defect detection method based on deep learning is widely used in steel production.

    “Therefore, the steel surface defect detection method based on deep learning is widely used in steel production, and more scholars begin to improve and perfect it [13, 14].”

    From A New Steel Defect Detection Algorithm Based on Deep Learning (2021) · cites this paper as [14]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Reference [122] has a mismatched lead author

    Minor

    The text attributes [122] to Yan et al., but its reference-list entry begins with Y. Zhang.

  • Reference [135] has a mismatched lead author

    Minor

    The text attributes [135] to Song et al., but its reference-list entry begins with Y. He.

  • Five stages called "three key parts"

    Minor

    The summary calls these three key parts but names five stages.

  • Cold-rolled defect samples assigned to the wrong figure

    Minor

    The text places the cold-rolled defect samples in Fig. 3(c), while the caption places them in Fig. 2(c). Fig. 3 is captioned as a feature-extraction architecture.

  • Reference [108] attributed to different lead authors

    Minor

    Two passages attribute [108] to Xu et al. and Di et al., respectively, although the reference list identifies its lead author as D. He.

  • Reference [94] is assigned incompatible methods

    Minor

    The HOG section attributes the PHOG descriptor to [94], but Table IV identifies [94] as a max-pooling convolutional neural network study.

6 of 6

Cites

Flagged first, then by citations

3 of the 140 references couldn’t be matched to a known paper.

Research works printed but not linked: 2

  1. song, 2014, Automatic recognition of surface defects on hot-rolled steel strip using scattering convolution network, J Comput Inf Syst, 10, 3049
  2. xu, 2008, Application of fractal dimensions based on the optimized scale to classification of surface defects on hot-rolled steel strips, Metallic Equipment, 168, 1

Cited by

Sorted by citations

Related

Authors
Luo, Q., Fang, X., Su, J., Zhou, J., Zhou, B., Yang, C., Liu, L., Gui, W., & Tian, L.
Affiliations
  • Central South University (CN)
  • Hefei University of Technology (CN)
  • National University of Defense Technology (CN)
ORCID
Venue
IEEE Transactions on Instrumentation and Measurement
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2020-12
Volume
69
Issue
12
Pages
9329-9349
ISSN
  • 0018-9456 (ISSN-L)
  • 1557-9662
DOI
10.1109/tim.2020.3030167
OpenAlex
W3102571413
Semantic Scholar
c1898f9b88e92c05fe42b48c960bdb95898963fd
Cited by
148
References
140
Citations by year
2026: 21 · 2025: 34 · 2024: 28 · 2023: 28 · 2022: 24 · 2021: 13
FWCI
10.36
Type
journal article
Language
en
Open access
closed
Licence
Locations
IEEE Transactions on Instrumentation and Measurement (publishedVersion)
Funders
  • National Natural Science Foundation of China (61988101, 61973323, 51704089)
  • Hunan Provincial National Science Foundation of China (2020JJ4747)
  • Innovation and Development Project of Ministry of Industry and Information Technology of the People’s Republic of China (TC19084DY)
Topics
  • Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
  • Physical Sciences › Engineering › Mechanical Engineering › Non-Destructive Testing Techniques
  • Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
Keywords
automated visual inspection, surface defect classification, flat steel, steel slabs, hot-rolled strip, cold-rolled steel strip, image preprocessing, feature extraction, feature selection, defect detection, online quality inspection, computer vision

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review