Automated Visual Defect Detection for Flat Steel Surface: A Survey

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

The combinational approach based on Gabor filter and volumetric fractal dimension possesses promising ability of obtaining rich texture features.

“Alvaro et al. [67] confirmed that the combinational approach based on Gabor filter and volumetric fractal dimension possesses promising ability of obtaining rich texture features.”

• cites [67] 67. G. Alvaro, B. Joao, and M. Odemir, "Gabor wavelets combined with volumetric fractal dimension applied to texture analysis," Pattern. Recogn. Lett., vol. 36, pp. 135-143, Oct. 2014. sourceSupported

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

Zhou et al. presented a double low-rank and sparse decomposition (DLRSD) model to obtain the defective region of steel sheet surface.

“Zhou et al. [99] presented a double low-rank and sparse decomposition (DLRSD) model to obtain the defective region of steel sheet surface.”

• cites [99] 99. S. Zhou, S. Wu, H. Liu, Y. Lu, and N. Hu, "Double low-rank and sparse decomposition for surface defect segmentation of steel sheet," Appl. Sci., vol. 8, no. 9, pp. 1628-1643, Sep. 2018. sourceSupported

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

These low-rank based approaches are reported to perform well.

“These approaches are reported to perform well.”

• cites [97] 97. H. Yan, K. Paynabar, and J. Shi, "Anomaly Detection in Images With Smooth Background via Smooth-Sparse Decomposition," Technometrics, vol. 59, no. 1, pp. 102-114, 2017. sourceNot checked — no full text of the source
• cites [98] 98. Q. Huangpeng, H. Zhang, X. Zeng, and W. Huang, "Automatic visual defect detection using texture prior and low-rank representation," IEEE Access, vol. 6, pp. 37965-37976, Jul. 2018. sourceNot checked — no full text of the source
• cites [99] 99. S. Zhou, S. Wu, H. Liu, Y. Lu, and N. Hu, "Double low-rank and sparse decomposition for surface defect segmentation of steel sheet," Appl. Sci., vol. 8, no. 9, pp. 1628-1643, Sep. 2018. sourceSupported

May combine two claims.

DCGAN builds constraints on traditional GANs to overcome unstable output and often works in unsupervised learning manner for defect detection.

“The deep convolutional generative adversarial network (DCGAN) [111] is a kind of CNN, which build certain constraints on traditional generative adversarial networks (GANs) to overcome its drawback of unstable output, and it often works in unsupervised learning manner for defect detection [112,113].”

• cites [111] 111. A. Radford, and L. Metz, "Unsupervised representation learning with deep convolutional generative adversarial networks," in Comput. Sci., 2015. >Paper ID: IM-19-22905R Title: Automated Visual Defect Detection for Flat Steel Surface: A Survey< 18 sourceSupported
• cites [112] 112. T. Schlegl, P. Seeböck, S. Waldstein, U. Schmidt-Erfurth, and G. Langs, "Unsupervised anomaly detection with generative adversarial networks to guide marker discovery," in IPMI 2017: Inf. Process. Med. Imag., 2017, pp. 146-157. sourceNot checked — no full text of the source
• cites [113] 113. Z. Zhao, B. Li, R. Dong, and P. Zhao, "A surface defect detection method based on positive samples," PRICAI 2018: Trends. AI, Lecture Notes in Computer Science, pp. 473-481, 2018. sourceNot checked — no full text of the source

May combine two claims.

The SCACM yielded good performance on both spot-defect and steel-pit-defect because the fused visual saliency map highlights potential defects and suppresses clutter background.

“The SCACM yielded good performance on both spot-defect and steel-pit-defect as the fused visual saliency map highlights the potential defects and suppresses the clutter background as well.”

• cites [94] 94. 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 matched

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

May combine two claims.

The SCN method has been successfully applied on surface defect detection for hot-rolled strips and obtained an average correct recognition accuracy of 97.22%.

“This method has been successfully applied on surface defect detection for hot-rolled strips and obtained an average correct recognition accuracy of 97.22%.”

• cites [75] 75. 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 matched

  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).

The machine learning trend is gradually sinking to the defect detection topic.

“it is not surprising that the machine learning trend is gradually sinking to the discussed defect detection topic.”

• AssumptionRests on an assumption: The defect detection task can be treated as a bi-classification problem.

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

The research trend in automated visual defect detection of flat steel surface has gradually shifted from previous theoretical study to on-site application.

“The research trend has gradually shifted from previous theoretical study to on-site application.”

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

Equipping a high pressure air-gun to remove surface water droplets is far more effective than developing advanced water removal algorithms to eliminate false alarms triggered by pseudo-defects.

“For example, equipping high pressure air-gun removing surface water droplets is far more effective than to develop advanced water removal algorithms to eliminate false alarm triggered by pseudo-defect.”

May combine two claims.

Song et al. realized precise detection of weak scratches on metal surface by combining deep CNN and skeleton extraction, with strong robustness to background noises.

“Moreover, Song et al. [108] realized precise detection of weak scratches on metal surface by confusing deep CNN and skeleton extraction, the experimental results indicate its strong robustness to background noises.”

• cites [108] 108. L. Song, W. Lin, Y. Yang, X. Zhu, Q. Guo, and J. Xi, "Weak micro-scratch detection based on deep convolutional neural network," IEEE Access., vol. 7, pp. 27547-27554, Feb. 2019. sourceNot checked — no full text of the source

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

The largest volume of published reports in the literature on automated visual defect detection of flat steel surface in industrial manufacturing belongs to the last five years.

“where the largest volume of published reports in this literature belongs to the last five years.”

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

Ghorai et al. found that the performance of VVRKFA with one-level Haar features ranks first among all feature-classifier combinations they tested.

“The authors fused the classifiers (i.e., SVM and VVRKFA) with different feature sets (i.e., Haar, DB2, DN4) to divide the test images into defective and normal ones, finding that the performance of VVRKFA with one-level Haar features ranks first among all the feature-classifier combinations.”

• cites [1] 1. 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. sourceNot checked — no full text of the source

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

Different references have different standards of detection accuracy, such as TPR, FNR, FPR, EER, AUC, MAE, G-mean, F-measure and so on.

“While, different references have different standard of detection accuracy, such as true positive rate (TPR), false negative rate (FNR), false positive rate (FPR), equal error rate (EER), area under curve (AUC), mean absolute error (MAE), G-mean, F-measure and so on.”

May combine two claims.

Sobel is sensitive to uneven illumination on flat steel surface, which easily leads to false edge detection.

“Sobel is specialized in weighing the influence of pixel position to reduce the ambiguity of edge, but it is sensitive to uneven illumination on flat steel surface, which easily leads to false edge detection.”

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

Clustering method is more suitable for defect classification than defect detection.

“For the nature of clustering, it is more suitable for defect classification than defect detection.”

May combine two claims.

The approach combining HOG and GLCM is sensitive to background noises and ununiform gray level changes, and its computation is relatively complex.

“But this approach is sensitive to background noises and ununiform gray level changes, moreover, the computation is relatively complex.”

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

Fourier transform realizes image features of translation invariance, expansion invariance and rotation invariance.

“With the appearance of the Fourier transform (FT), image features of translation invariance, expansion invariance and rotation invariance are realized.”

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

Statistical approaches are frequently used to detect defects of flat steel surface by evaluating the regular and periodic distribution of pixel intensities.

“Statistical approaches are frequently used to detect defects of flat steel surface by evaluating the regular and periodic distribution of pixel intensities.”

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

Mathematical morphology is widely used in nearly all aspects in image processing, including image segmentation, feature extraction, edge detection, image filtering, and image enhancement.

“With the firm and complete theory basis, mathematical morphology is widely used in nearly all aspects in image processing, including image segmentation, feature extraction, edge detection, image filtering, image enhancement, and so on.”

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

The undesirable image quality requires preeminent detection algorithms to resist large intra-class variation and minor inter-class distance.

“The undesirable image quality requires preeminent detection algorithms to resist large intra-class variation and minor inter-class distance [1][2][3][4].”

• cites [1][2][3][4]Not matched

  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), S4 (based).

Earlier defect detection allows less economic losses.

“earlier defect detection allows less economic losses.”

Classified as probably not a substantial claim.

AVI techniques are more suitable to inspect surface defects on sheet materials than on wire rod/bar with minor diameter or even special-shaped structures.

“It is widely recognized that AVI techniques are more suitable to inspect surface defects on sheet materials than on wire rod/bar with minor diameter or even special-shaped structures [15].”

• cites [15] 15. W. B. Li, C. H. Lu, and J. C. Zhang, "A local annular contrast based real-time inspection algorithm for steel bar surface defects," Appl. Surf. Sci., vol. 258, no. 16, pp. 6080-6086, Mar. 2012. sourceNot checked — no full text of the source

May combine two claims.

Canny suffers with low adaptive ability and sometimes is easy to blur the noise-free region.

“However, it suffers with low adaptive ability and sometimes is easy to blur the noise-free region.”

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

LBP has significant advantages of rotation and gray invariance.

“As a classical operator, local binary pattern (LBP) is widely used to characterize local texture features of images, which has significant advantages of rotation and gray invariance.”

The quote only partly states this claim.

LBP variants are widely applied on defect detection of flat steel surface.

“Later, LBP is frequently used to detect defects on flat steel surface [43][44][45].”

• cites [43][44][45]Not matched

  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).

A general AVI instrument provides two main functions: defect detection and classification.

“A general AVI instrument provides two main functions of defect detection and classification [1][2][3][4].”

• cites [1][2][3][4]Not matched

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

May combine two claims.

Parameters of a Gabor filter are mainly decided by defect size and direction, so it is hard to obtain desirable results for miscellaneous defects with various sizes by a single Gabor filter.

“The parameters of Gabor filter are mainly decided by the defect size and direction, it is thus hard to obtain desirable results for miscellaneous defects with various sizes by a single Gabor filter.”

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

Huangpeng et al. proposed a weighted low-rank reconstruction model for automatic visual defect detection.

“Huangpeng et al. [98] proposed a novel weighted low-rank reconstruction model for automatic visual defect detection”

• cites [98] 98. Q. Huangpeng, H. Zhang, X. Zeng, and W. Huang, "Automatic visual defect detection using texture prior and low-rank representation," IEEE Access, vol. 6, pp. 37965-37976, Jul. 2018. sourceNot checked — no full text of the source

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

Fofi et al. proposed a non-parametric and efficient Weibull-based defect detection method by computing two parameters of a Weibull fit for the distribution of image gradients in local regions.

“Fofi et al. [16] proposed a novel, non-parametric and efficient Weibull-based defect detection method by computing two parameters of a Weibull fit for the distribution of image gradients in local regions.”

• cites [16] 16. D. Fofi, F. Timm, P. R. Bingham, and E. Barth, "Non-parametric texture defect detection using Weibull features," in Proc. SPIE., San Francisco, CA, Feb. 2011. >Paper ID: IM-19-22905R Title: Automated Visual Defect Detection for Flat Steel Surface: A Survey< 16Not checked — no full text of the source

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

Wavelet characteristics are more in line with human visual mechanism.

“Meanwhile, the characteristics of wavelet are more in line with human visual mechanism.”

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

FIR filter performs better both on optimization scale and computational expense than IIR and Gabor filters because FIR filter has more free available turning parameters.

“Kumar pointed out in his Ph.D. dissertation [69] that FIR filter performs better both on optimization scale and computational expense than infinite impulse response (IIR) and Gabor filters as FIR filter has more free available turning parameters.”

• cites [69] 69. A. Kumar, "Automated defect defection in textured materials," Ph.D. dissertation, Dept. Elect. Electron. Eng., Univ.Hong Kong, Hong Kong, China, May 2001.Not matched

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

May combine two claims.

Gayubo et al. utilized MRF to restore flat steel defects (cracks) and eliminate spurious features (pseudo).

“Gayubo et al. [88] utilized MRF to restore flat steel defects (i.e., cracks) and eliminate the spurious features (i.e., pseudo).”

• cites [88] 88. F. Gayubo, J. L. Gonzá lez, E. De La Fuente, F. Miguel, and J. R. Perá n, "On-line machine vision system for detect split defects in sheet-metal forming processes," in Int. Conf. Pattern Recognit, Hong Kong, China, Aug. 2006, pp. 723-726. sourceNot checked — no full text of the source

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

Wang et al. constructed a compact model as guidance information by mining intrinsic image priors, offering good generalization ability for different detection tasks and robustness to noise.

“Wang et al. [26] constructed a compact model to be regarded as a kind of guidance information by mining intrinsic image priors, it offers a good generalization ability for different detection tasks and is sufficiently robust to noise.”

• cites [26] 26. 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., pp. 1-1, May 2019. sourceNot checked — no full text of the source

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

FIR filters are very suitable to be embedded in FPGAs, which is compliant with the lightweight trend of the instrumentation and measurement society.

“In addition, FIR filters are very suitable to be embedded in FPGAs, which is compliant with the lightweight trend of the instrumentation and measurement society.”

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

Recent literature exhibits huge application potentiality of MRF on industrial surface defect detection.

“The recent literatures exhibit the huge application potentiality of MRF on industrial surface defect detection.”

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

HT can only track the direction of edges; the length information of line segment is lost.

“However, HT can only track the direction of edges, the length information of line segment is lost.”

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

Ai et al. applied Curvelet enhanced by kernel locality preserving projections to track longitudinal cracks on con-casting slabs.

“Ai et al. [56] applied Curvelet enhanced by kernel locality preserving projections to track longitudinal cracks on con-casting slabs.”

• cites [56] 56. 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. sourceNot checked — no full text of the source

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

Hough transform is considered a powerful tool in well-defined line-feature detection.

“Hough transform (HT) [83] is considered as a powerful tool in well-defined line-feature detection.”

• cites [83] 83. R. C. Gonzalez, and R. E. Woods, Digital Image Processing (3rd Edition), 2007.Not matched

  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), S4 (based).

Spectral methods are dedicated to finding a special transform domain where defect objects can be more easily and completely separated from local and global backgrounds.

“In general, spectral methods are dedicated to find a special transform domain where the defect objects can be more easily and completely separated from the both the local and global backgrounds.”

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

Yan et al. proposed a wavelet-based image filtering algorithm based on anisotropic diffusion.

“Yan et al. [73] proposed a novel wavelet-based image filtering algorithm based on anisotropic diffusion.”

• cites [73] 73. W. Liu, and Y. Yan, "Automated surface defect detection for cold-rolled steel strip based on wavelet anisotropic diffusion method," Int. J. Ind. Syst. Eng., vol. 17, no. 2, pp. 224-239, Feb. 2014. sourceNot checked — no full text of the source

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

Yang et al. developed an ACM-based defect detection method without edges that can effectively segment defect features with inhomogeneous boundaries from complicated surface textures.

“Yang et al. [95] developed a ACM-based defect detection method without edges through incorporating a variable penalty term and a convolution kernel, authors reported that it can effectively segment defect features with inhomogeneous boundaries from complicated surface textures.”

• cites [95] 95. J. Yang, X. Li, J. Xu, Y. Cao, Y. Zhang, L. Wang, and S. Jiang, "Development of an optical defect inspection algorithm based on an active contour model for large steel roller surfaces," Appl. Optics, vol. 57, no. 10, pp. 2490-2498, Apr. 2018. sourceNot checked — no full text of the source

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

Wavelet transform can effectively extract information from signals and perform multi-scale analysis of functions or signals through scaling and shifting operations.

“Consequently, wavelet transform can effectively extract information from signals and perform multi-scale analysis of functions or signals through scaling and shifting operations.”

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

Susan et al. proposed Gaussian mixture entropy model for defect detection, specialized in identifying miscellaneous defects such as holes and stains.

“Susan et al. [96] proposed Gaussian mixture entropy model for defect detection, which is specialized in identifying miscellaneous defects such as holes and stains.”

• cites [96] 96. S. Susan, and M. Sharma, "Automatic texture defect detection using Gaussian mixture entropy modeling," Neurocomputing, vol. 239, pp. 232-237, May. 2017. sourceNot checked — no full text of the source

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

Wavelet transform can move the time-frequency window and automatically adjust the window with the change of frequency in the center of the window.

“Compared with Gabor filters, wavelet transform can not only move the time-frequency window, but also automatically adjust the window with the change of the frequency in the center of the window.”

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

Hough line detection has strong anti-interference ability and is insensitive to noises, incomplete edges, and other coexisting non-linear structures.

“Hough line detection has the advantage of strong anti-interference ability and is also insensitive to noises, incomplete part of edges, and other coexisting non-linear structures.”

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

Zhang et al. proposed an image fusion method based on Bandelet-PCNN model to solve pseudo-Gibbs phenomena around singularities.

“Zhang et al. [78] have proposed an image fusion method based on Bandelet-PCNN (Pulse coupled neural networks) model to solve the problem of the pseudo-Gibbs phenomena around singularities.”

• cites [78] 78. X. Zhang, Y. H. Yan, W. H. Chen, and J. J. Chen, "Image fusion method for strip steel surface detect based on Bandelet-PCNN," Adv. Mater. Res., pp. 806-810, Jul. 2012. sourceNot checked — no full text of the source

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

Sharifzadeh et al. applied HT to detect defects of holes, scratches, coil breaks and rusts on cold-rolled steel strips.

“Interestingly, Sharifzadeh et al. [20] applied HT to detect defects of holes, scratches, coil breaks and rusts on cold-rolled steel strips.”

• cites [20] 20. M. Sharifzadeh, S. Alirezaee, R. Amirfattahi, and S. Sadri, "Detection of steel defect using the image processing algorithms," in IEEE Int. Mult. Conf., Bahria Univ, Karachi, PAKISTAN, Feb. 2009, pp. 125-127.Not matched

  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).

Yan et al. utilized smooth-sparse decomposition (SSD) model for anomaly detection in images.

“Based on low-rank representation, Yan et al. [97] utilized smooth-sparse decomposition (SSD) model for anomaly detection in images”

• cites [97] 97. H. Yan, K. Paynabar, and J. Shi, "Anomaly Detection in Images With Smooth Background via Smooth-Sparse Decomposition," Technometrics, vol. 59, no. 1, pp. 102-114, 2017. sourceNot checked — no full text of the source

May combine two claims.

Wavelet transform can optimally characterize point singularity but can hardly characterize lines and surface singularities due to finiteness of separable wavelet directions.

“Wavelet transform can optimally characterize the point singularity but can hardly characterize lines and surface singularities due to the finiteness of separable wavelet directions.”

May combine two claims.

Xu et al. proposed a Shearlet-based feature extraction method (DST-KLPP) and an adaptive MGA method (RNAMlet), both emphasizing detection rates and computation expenses.

“For quality assurance of con-casting slabs and hot strips, Xu et al. successively proposed a Shearlet-based feature extraction method (DST-KLPP) [79] and an adaptive MGA method (RNAMlet) [80], both of them emphasized much on detection rates and computation expenses.”

• cites [79] 79. 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. sourceNot checked — no full text of the source
• cites [80] 80. K. Xu, Y. Xu, P. Zhou, and L. Wang, "Application of RNAMlet to surface defect identification of steels," Opt. Lasers Eng., vol. 105, pp. 110-117, Jun. 2018. sourceNot checked — no full text of the source

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

Five different types of wavelets (Haar, DB2, DB4, Bior, and multiwavelet) have been evaluated by Ghorai et al. to extract features of small-size image blocks.

“Five different types of wavelets, namely, Haar, Daubechies 2 (DB2), Daubechies 4 (DB4), biorthogonal spline (Bior), and multiwavelet, have been evaluated by Ghorai et al. [1] to extract features of small-size image blocks.”

• cites [1] 1. 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. sourceNot checked — no full text of the source

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

The defect detection task is essentially handled as a binary (defective or defect-free) classification problem in machine learning methods.

“As shown in Fig. 8, the defect detection task is essentially handled as a binary (defective or defect-free) classification problem in machine learning methods (or we can call them advanced classifiers).”

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

Fusing features extracted by multiple descriptors to support the final detection decision can yield better results than those produced by a single descriptor in most cases.

“As for algorithm itself, fusing features extracted by multiple descriptors to support final detection decision can yield better results than those produced by a single descriptor in the most cases.”

May combine two claims.

Zhao et al. combined GAN, autoencoder (AE) and LBP to detect defects on texture surface, needing only positive samples without any defect sample nor manual label.

“Notably, Zhao et al. [113] combined GAN and autoencoder (AE) and LBP to detect defects on texture surface, which needs only positive samples without any defect sample nor manual label.”

• cites [113] 113. Z. Zhao, B. Li, R. Dong, and P. Zhao, "A surface defect detection method based on positive samples," PRICAI 2018: Trends. AI, Lecture Notes in Computer Science, pp. 473-481, 2018. sourceNot checked — no full text of the source

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

Reinforcement learning tries to obtain intelligent self-optimization through continuously interacting with its environment, achieving defect detection by making full use of limited labeled and unlabeled samples with low cost.

“In contrast, reinforcement learning tries to obtain intelligent self-optimization through continuously interacting with its environment, so as to achieve defect detection by making full use of limited labeled and unlabeled samples with low cost.”

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

Running time is another vital evaluation criterion because the rapid casting or rolling rhythm of flat steel in real-world industrial sites has high-level time cost requirement on defect detection.

“On the other hand, running time is another vital evaluation criterion, as the rapid casting or rolling rhythm of flat steel in real-world industrial sites has high-level time cost requirement on defect detection.”

• AssumptionRests on an assumption: The rapid casting or rolling rhythm of flat steel in real-world industrial sites has high-level time cost requirement on defect detection.

May combine two claims.

Mei et al. utilized convolutional denoising AE network to reconstruct image patches, and the scheme can reliably learn final detection results without manual intervention.

“Mei et al. [114] utilized convolutional denoising AE network to reconstruct image patches, combined with the reconstruction residual maps, this scheme can reliably learning final detection results, where no manual intervention is needed throughout all the detection process.”

• cites [114] 114. S. Mei, H. Yang, and Z. Yin, "An unsupervised-learning-based approach for automated defect inspection on textured surfaces," IEEE Trans. Instrum. Meas., vol. 67, no. 6, pp. 1266-1277, Jun. 2018. sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

The most effective denoising method for an AVI system is to make the images as clean as possible by feasible engineering measures.

“the most effective denoising method for AVI system is to make the images as clean as possible by some feasible engineering measures.”

May combine two claims.

AE-based algorithms demonstrate strong competitiveness in steel surface defect detection and are reported to be fairly noise-robust.

“Moreover, AE-based algorithms also demonstrate strong competitiveness in steel surface defect detection, which are reported to be fairly noise-robust.”

• cites [114] 114. S. Mei, H. Yang, and Z. Yin, "An unsupervised-learning-based approach for automated defect inspection on textured surfaces," IEEE Trans. Instrum. Meas., vol. 67, no. 6, pp. 1266-1277, Jun. 2018. sourceNot checked — no full text of the source
• cites [18] 18. S. Youkachen, M. Ruchanurucks, T. Phatrapomnant, and H. Kaneko, "Defect segmentation of hot-rolled steel strip surface by using convolutional auto-encoder and conventional image processing," in 10th IC-ICTES Bangkok, Thailand, Mar. 2019. sourceNot checked — no full text of the source

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

Unsupervised learning is dedicated to discover groups of similar examples within input data.

“Here, the unsupervised learning is dedicated to discover groups of similar examples within the input data.”

May combine two claims.

CNN can be used not only for supervised learning, but also for unsupervised learning.

“CNN can be used not only for supervised learning, but also for unsupervised learning.”

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

Machine learning or deep networks are a preeminent alternative for complex multi-class classification problems with rich datasets, such as defect classification.

“While machine learning or deep network is preeminent alternative for complex multi-class classification problem with rich datasets (i.e., defect classification).”

May combine two claims.

Both supervised learning and unsupervised learning methods have obtained rapid progress on surface defect detection of industrial flat steel.

“Both supervised learning and unsupervised learning methods have obtained a rapid progress on surface defect detection of industrial flat steel.”

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

Machine learning tends to accomplish defect detection tasks more intelligently, and such an emerging technique is promising in the application of flat steel surface defect detection.

“As is stated above, machine learning tends to accomplish the defect detection tasks more intelligently, such an emerging technique is promising in the application of flat steel surface defect detection.”

May combine two claims. The quote only partly states this claim.

Detection accuracy evaluated based on Raw images is more reliable and more credible to flat steel manufacturers than detection accuracy evaluated based on Database, when the corresponding detection methods are applied on an AVI system in an actual steel manufacturing line.

“It is worth mentioning that the results of detection accuracy evaluated based on Raw images are more reliable than those evaluated based on Database, when the corresponding detection methods are really applied on AVI system in actual steel manufacturing line.”

• AssumptionRests on an assumption: Actual defects must be found by professional defect inspectors one by one from historical flat steel products, which is extremely labor-intensive and time-consuming.

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

Convolution and subsampling in CNN effectively reduce model size by tailoring model parameters.

“Convolution and subsampling in convolution neural network (CNN) effectively reduce the model size by tailoring the model parameters.”

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

A great quantity of parameters of neural networks leads to huge computational complexity.

“But a great quantity of parameters of neural networks lead to huge computational complexity.”