CoChat
Sign Up
Cited by 372Industrial Vision Systems and Defect DetectionAdvanced Neural Network ApplicationsInfrastructure Maintenance and Monitoring

A review on modern defect detection models using DCNNs – Deep convolutional neural networks

Open access
Tulbure, Andrei-Alexandru, Tulbure, Adrian-Alexandru, Dulf, Eva-Henrietta·2022·Journal of Advanced Research
DOI 10.1016/j.jare.2021.03.015PMC PMC8721352
Open this paper in CoChat — free accountOpen PDF

Abstract

Background: Over the last years Deep Learning has shown to yield remarkable results when compared to traditional computer vision algorithms, in a large variety of computer vision applications. The deeplearning models outperformed in both accuracy and processing time. Thus, once a deeplearning models won the Image Net Large Scale Visual Recognition Contest, it proved that this area of research is of great potential. Furthermore, these increases in recognition performance resulted in more applied research and thus, more applications where deeplearning is useful: one of which is defect detection (or visual defect detection). In the last few years, deeplearning models achieved higher and higher accuracy on the complex testing datasets used for benchmarking. This surge in accuracy and usage is also supported (besides swarms of researchers pouring into the race), by incremental breakthroughs in computing hardware: such as more powerful GPUs(Graphical processing units), CPUs(central processing units) and better computing procedures (libraries and frameworks). Aim of the review: To offer a structured and analytical overview(stating both advantages and disadvantages) of the existing popular object detection models that can be re-purposed for defect detection: such as Region based CNNs(Convolutional neural networks), YOLO(You only look once), SSD(single shot detectors) and cascaded architectures. A further brief summary on model compression and acceleration techniques that enabled the portability of deeplearning detection models is included. Key Scientific Concepts of Review: It is of great use for future developments in the manufacturing industry that many of the popular, above mentioned models are easy to re-purpose for defect detection and, thus could really contribute to the overall increase in productivity of this sector. Moreover, in the experiment performed the YOLOv4 model was trained and re-purposed for industrial cable detection in several hours. The computing needs could be fulfilled by a general purpose computer or by a high-performance desktop setup, depending on the specificity of the application. Hence, the barrier of computing shall be somewhat easier to climb for all types of businesses.

AI snapshot

Written from the paper’s full text.

TL;DR

This review surveys deep convolutional neural network models for industrial defect detection, covering traditional computer vision, modern object detectors, model compression, and benchmarking datasets. It argues that dataset quality and model choice are central to performance, and it reports a small YOLOv4-based cable-defect experiment.

Key findings

  • Modern detectors are compared on standard benchmarks: YOLOv4-608 reaches 43.5% box AP and 47.0% AP75 on COCO, PP-YOLO reaches 45.2% mAP, and EfficientDet-D7x reaches 55.1% mAP; YOLO-family detectors are noted as fastest, with frame rates up to 120 FPS.
  • Defect-detection applications report high task-specific results: PCB defect detection with YOLOv2 reached 98.79% accuracy; steel strip surface defect detection reached up to 99% detection rate, 95.86% recall, and 83 FPS; wood defect detection with an SSD/DenseNet backbone reached 96.1% mean average precision at 56 ms per image.
  • A compact CNN for surface defects achieved 99.29% accuracy with only 60 k weights and 76 million FLOPs, performing detection and classification in 75 ms, compared with ResNetV2-50 at 99.60% accuracy and 893 ms.
  • In the authors’ cable experiment, labeling technique affected mAP: labeling long cables multiple times or not labeling every instance decreased mAP; augmentation improved results but not significantly, mainly because the sample datasets were small.

Methods

  • Design: narrative review of traditional/shallow computer vision methods, deep-learning object detection frameworks (RCNN, SPPNet, Fast/Faster R-CNN, YOLO variants, SSD, SqueezeDet, Mask R-CNN, EfficientDet, CSPNet, MnasFPN, SpineNet, PP-YOLO), model compression/acceleration, and benchmarking datasets.
  • Data: the authors’ own experiment used two self-labeled cable datasets of 164 images (400 after augmentation) and 342 images (806 after augmentation), with YOLOv4 as the detector; training used a low-cost Nvidia GF750 and a remote station with 4 Nvidia P106-100 GPUs.
  • Analysis: benchmark comparison using reported metrics such as Top1 accuracy, box AP, AP75, mAP, FPS, parameter counts, and FLOPs; the cable experiment used horizontal/vertical flip, rotations between −20 and +20 degrees, blur up to 6.5 px, and random noise up to 10% of pixels.

Limitations

  • Paper-stated: the cable datasets are small and take hours to acquire and label correctly; the largest dataset could not be uploaded; augmentation did not significantly affect mAP mainly because the sample datasets were small.
  • Paper-stated: there is no rule of thumb for choosing which object detection model will generalize best to a particular dataset.
  • Evident from the text: the review relies on results reported by cited papers rather than re-benchmarking all methods, and the authors’ own experiments are limited to a cable-defect case with small datasets, so the experimental findings may not generalize to other defect types or industrial settings.

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

Integrity signals

Registry

Crossref · 10.1016/j.jare.2021.03.015

Cites

95 of 95 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 57 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    In Table 2, box average precision increases with every single iteration of an improvement or a new model framework.

    “Thus, one can easily see the box average precision(correct class predicted and bounding box that encapsulates the class is correct) increases with every single iteration of an improvement or a new model framework, which is party enabled by the increases in performance of image classification algorithms.”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    The mean average precision on the COCO real-time object detection database increases over time.

    “Thus, one can easily see the mean average precision(correct class predicted and bounding box that encapsulates the class is correct) increases over time”
    Quote in the paperNot a substantial claim
  • Findingcitation

    There can be a smaller, more efficient network that displays better accuracy than a larger, more clogged one, as with ResNet-50 and VGG-19.

    “There can be a smaller, more efficient network that displays better accuracy than a larger, more clogged one (see the comparison of ResNet-50 [49] and VGG-19 [36] CNN models).”
    Quote in the paperNot a substantial claim
  • Findingnone given

    A small dataset performs significantly worse than a more comprehensive dataset that includes lots of edge cases.

    “Finally, a small dataset performs significantly worse, than a more comprehensive dataset, which includes lots of edge cases in which we can find the suggested defect.”
    Quote in the paperSubstantial claim
  • Findingown data

    In the authors' cable experiment, the labeling technique impacts detector accuracy.

    “The accuracy of the detector is impacted by the labeling technique: if you label a long cable twice or more and label multiple segments of a cable(especially is you label one part good and one part faulty), the mAP will decrease, both because of localization errors and because of classification errors.”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 50
  • May combine two claims. 39
  • The quote only partly states this claim. 3

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

    Findingreasoning

    The FPS rate is usually mirrored against the mean AP.

    “Thus, the FPS rate is usually mirrored against the mean AP.”
    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

    The performance of real time object detectors does not seem to have significantly slowed down and a great chunk of progress is still possible.

    “Encouraging is the fact that, the performance of real time object detectors does not seem to have significantly slowed down and a great chunk of progress is still possible.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    YOLO framework detectors are currently the fastest detectors with frame rates up to 120 FPS.

    “The YOLO framework detectors are currently the fastest detectors with frame rates up to 120 FPS.”
    Quote in the paperNot a substantial claim

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

    Findingreasoning

    Generally more parameters mean a better accuracy performance of the network, but this is not an inherent rule of thumb for convolutional neural networks.

    “Also, one can easily see that generally more parameters mean a better accuracy performance of the network, but this is not an inherent rule of thumb for convolutional neural networks!”
    Quote in the paperNot a substantial claim

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

    Findingown data

    In Table 1, Top1-accuracy increases with every single iteration of an improvement or a new model framework.

    “Thus, one can easily see the Top1-accuracy(correct class predicted) increases with every single iteration of an improvement or a new model framework.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The rate of improvement of Top1 accuracy is slowing down.

    “Thus, one can easily see that the rate of improvement of Top1 accuracy is slowing down.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    Balancing the dataset needs to be considered because too many or too few samples of an object may hurt overall test time accuracy.

    “Moreover, the balancing of the dataset needs to be considered, as too many samples of an object or too few, may further hurt overall test time accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Labeling a long cable twice or more and labeling multiple segments of a cable decreases mAP.

    “if you label a long cable twice or more and label multiple segments of a cable(especially is you label one part good and one part faulty), the mAP will decrease, both because of localization errors and because of classification errors.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    If every single instance of the object in the sample images is not labeled, the mAP will decrease.

    “Also, if you don‘t label every single instance of the object in the sample images, the mAP will decrease.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In the authors' cable experiment, the augmentation techniques used did not have a significant effect on mAP, mainly because the datasets were small, but mAP did improve.

    “In terms of mAP it did not had a significant effect on the results, mainly because we had small sample datasets, but it did improve.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The authors' results show that one can quickly get a proof of concept and can easily see in which direction to look in order to improve performances.

    “Our results show that, one can quickly get the proof of concept and can easily see in which direction to look in order to improve the performances.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    The authors achieved this without a large expense in the name of a workstation equipped with lots of GPUs.

    “All this, without a large expense in the name of workstation equiped with lots of GPUs.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    SPPNet achieved 60.9% mAP on the PASCAL VOC 2007 dataset (best one).

    “The accuracy obtained on the PASCAL VOC 2007 dataset was 60.9% mAP (best one)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [79] He K. Zhang X. Ren S. Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition IEEE Trans Pattern Anal Mach Intell 37 2015 1904 1916 26353135 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The accuracy for image classification tasks is higher than the one for detection, so detection models can be significantly improved or the technology did not achieve enough.

    “The accuracy for Image classification tasks is higher than the one for detection, thus, one can conclude that detection models can be significantly improved or that the technology did not achieve ”enough”.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Benchmark performance on ImageNet or COCO is indicative of suitability for industrial defect detection.

  • May combine two claims.

    Interpretationreasoning

    The most accurate detection models are also the largest ones, so embedded devices need more progress in efficiency.

    “The most accurate ones are also the largest ones, thus in order to enable embedded devices like cars, to benefit from the technology, more and more progress needs to be done in the sphere of efficiency.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Benchmark performance on ImageNet or COCO is indicative of suitability for industrial defect detection.

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

    Premisecitation

    A deep learning model for detecting defects in printed circuit boards got up to 98.79% accuracy by training on a TITAN V GPU.

    “In [39] the authors developed a deeplearning model for detecting defects in printed circuit boards(PCBs), that got up to an accuracy of 98.79% by training it on a TITAN V GPU.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [39] Adibhatla V.A. Chih H.-C. Hsu C.-C. Cheng J. Abbod M. Shieh J. Defect detection in printed circuit boards using you-only-look-once convolutional neural networks Electronics 9 2020 1547 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The compact CNN performed detection and classification in 75 ms, while ResNetV2-50 did it in 893 ms and MobileNetV2 did it in 259 ms while obtaining 99.63% accuracy with 267 million FLOPs and 480k trainable weights.

    “Also, when taking into account processing time, their model performed the detection and classification in 75 ms, while ResNetV2-50 did it in 893 ms. For comparison, MobileNetV2 [48] did it in 259 ms, while obtaining 99.63% accuracy, with 267 million FLOPs and 480 k trainable weights.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [47] Huang Y, Qiu C, Wang X, Wang S, Yuan K. A compact convolutional neural network for surface defect inspection. Sensors (Basel, Switzerland) 20. PMC7180441 32244764 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The compact CNN obtained 99.29% accuracy with 60k weights and 76 million FLOPs, while ResNetV2-50 with 16082 million FLOPs and 23.65 million weights obtained only 99.60%.

    “Together with a architecture design based on depth-wise convolution and a light wight backbone they obtained 99.29 % accuracy with only 60 k weights and only 76 million FLOPs, while ResNetV2-50 [49], with 16082 million FLOPs and 23.65 million weights obtained only 99.60%.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [47] Huang Y, Qiu C, Wang X, Wang S, Yuan K. A compact convolutional neural network for surface defect inspection. Sensors (Basel, Switzerland) 20. PMC7180441 32244764 source

      Not checked in the free run.

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

    Interpretationcitation

    A three stage cascaded detection architecture for fasteners on catenary support device yielded encouraging results and can reduce human labour needed for maintenance and enable processing lots of data accurately.

    “The results are encouraging and thus, present the opportunity to reduce the human labour needed for maintenance and enable the system to process lots of data accurately.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [45] Chen J. Liu Z. Wang H. Núñez A. Han Z. Automatic defect detection of fasteners on the catenary support device using deep convolutional neural network IEEE Trans Instrum Meas 67 2018 257 269 source

      Not checked in the free run.

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

    Interpretationnone given

    Having a strong backbone is one of the key design choices for a high accuracy defect detection model.

    “Having a strong backbone is one of the key design choices for a high accuracy defect detection model.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Benchmark performance on ImageNet or COCO is indicative of suitability for industrial defect detection.

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

    Premisecitation

    A modified YOLO for steel strip surface defects obtained up to 99% detection rate, 95.86% recall, at 83 FPS.

    “Furthermore, they obtained great results, on a particular solution they obtained up to 99% detection rate, with 95.86% recall rate, all while their detector is running at 83 FPS.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [41] Li J. Su Z. Geng J. Yin Y. Real-time detection of steel strip surface defects based on improved yolo detection network IFAC-PapersOnLine 51 2018 76 81 source

      Not checked in the free run.

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

    Interpretationnone given

    When a network has more parameters, its accuracy tends to increase, but this is not mandatory as inefficiencies can easily occur.

    “Generally, as previously specified, when a network has more parameters, it‘s accuracy tends to increase, but it is not mandatory as inefficiencies can easily occur.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    RCNN achieved 31.4% on the ILSVRC2013 detection dataset.

    “and 31.4% on the ILSVRC2013 detection dataset.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] Girshick R.B. Donahue J. Darrell T. Malik J. Rich feature hierarchies for accurate object detection and semantic segmentation 2014 IEEE conference on computer vision and pattern recognition 2014 580 587 source

      Not checked in the free run.

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

    Premisecitation

    SPPNet achieved 35.11% mAP on ISLVRC 2014, coming in second.

    “while on ISLVRC 2014 it obtained 35.11% mAP, coming in second.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [79] He K. Zhang X. Ren S. Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition IEEE Trans Pattern Anal Mach Intell 37 2015 1904 1916 26353135 source

      Not checked in the free run.

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

    Interpretationreasoning

    YOLOv4 is the fastest model and its accuracy is more than decent and given the easy training procedure it is a top candidate for defect detection models.

    “YOLOv4 is the fastest model and it‘s accuracy is more than decent and given the easy training procedure it is a top candidate for defect detection models(as were their previous versions v2,v3 etc.).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Benchmark performance on ImageNet or COCO is indicative of suitability for industrial defect detection.

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

    Interpretationreasoning

    AP75 is needed for defect detection because the process needs high accuracy and reliability to replace traditional industrial inspection methods.

    “We need this kind of accuracy for defect detection because the process needs high accuracy and reliability in order for it to replace traditional industrial inspection methods.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Defects in industrial images can be treated as objects for object detection.

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

    Premisecitation

    An automated detection and classification system for infrastructure surface defects using deep residual network and active training obtained 87.5% detection accuracy.

    “The results are encouraging, they obtained 87.5% detection accuracy.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46] Feng C. Liu M.-Y. Kao C.-C. Lee T.-Y. Deep active learning for civil infrastructure defect detection and classification 2017

      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

    A lightweight YOLOv3-Lite combined with depthwise separable convolution and feature pyramids obtained 50% better speed than basic YOLOv3 on aircraft structure crack detection.

    “They evaluated the method on fuselage, wing, aircraft tail or engine blades and they obtained 50% better speed than the basic YOLOv3 [31].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] Li Y. Han Z. Xu H. Liu L. Li X. Zhang K. Yolov3-lite: A lightweight crack detection network for aircraft structure based on depthwise separable convolutions Appl Sci 9 2019 3781 source

      Not checked in the free run.

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

    Premisecitation

    Fast R-CNN achieved state-of-the-art 68.4% mAP on the PASCAL VOC 2012 dataset.

    “while on the PASCAL VOC 2012 dataset, it obtained state-of-the-art 68.4% mAP.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [80] Girshick RB. Fast r-cnn. In: 2015 IEEE International conference on computer vision (ICCV); 2015. p. 1440–8. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    A compact CNN for surface defect detection achieved comparable results to MobileNetV2 while having 33% of the FLOPs and 1/8 of the weights.

    “They did achieve comparable results to MobileNetV2 [48], while only having 33% of the FLOPs(Floating point operations per second) and 1/8 of the weights.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [47] Huang Y, Qiu C, Wang X, Wang S, Yuan K. A compact convolutional neural network for surface defect inspection. Sensors (Basel, Switzerland) 20. PMC7180441 32244764 source

      Not checked in the free run.

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

    Interpretationreasoning

    The increase in mean average precision over time is mostly enabled by increases in performance of image classification algorithms and object detection algorithms.

    “which is mostly enabled by the increases in performance of image classification algorithms and object detection algorithms.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    PP-YOLO yields 45.2% mAP on the COCO test-dev dataset at a faster speed than YOLO-v4.

    “The PP(Paddle-Paddle) YOLO model yields 45.2 % mAP on the COCO test-dev dataset, at a faster speed than YOLO-v4 [63].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [65] Long X, Deng K, Wang G, Zhang Y, Dang Q, Gao Y, et al. Pp-yolo: An effective and efficient implementation of object detector, ArXiv abs/2007.12099.

      Not checked in the free run.

    • cites [63] Bochkovskiy A, Wang CY, Liao H. Yolov4: Optimal speed and accuracy of object detection, ArXiv abs/2004.10934.

      Not checked in the free run.

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

    Interpretationreasoning

    Pre-training is not aware of the task of interest and hence fails to adapt.

    “The intuition lies in the fact that pre-training is not aware of the task of interest and hence fails to adapt.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    MnasFPN with a mobile compatible backbone yields state-of-the-art for mobile devices mAP of 26.1% on the COCO test-dev dataset and uses 0.92B MAdds and 2.5 M parameters.

    “this method yields a state-of-the-art for mobile devices mAP of 26.1 % on the COCO test-dev dataset and uses 0.92B MAdds and 2.5 M parameters.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [86] Chen B. Ghiasi G. Liu H. Lin T.-Y. Kalenichenko D. Adam H. Le Q.V. Mnasfpn: Learning latency-aware pyramid architecture for object detection on mobile devices 2020 IEEE/CVF conference on computer vision and pattern Recognition (CVPR) 2020 13604 13613

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    YOLOv4 obtains 38% mAP @ 120 FPS or 43% mAP @ 80 FPS on PASCAL VOC 2012.

    “getting 38% mAP @ 120 FPS or 43% mAP @ 80 FPS on PASCAL VOC 2012”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [63] Bochkovskiy A, Wang CY, Liao H. Yolov4: Optimal speed and accuracy of object detection, ArXiv abs/2004.10934.

      Not checked in the free run.

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

    Premisecitation

    YOLOv2 can run at 76.8 mAP @ 67 FPS or 78.6% @ 40 FPS on the PASCAL VOC 2007 dataset.

    “the YOLOv2 model can run at different trade-offs: 76.8 mAP @ 67 FPS or 78.6% @ 40 FPS on the PASCAL VOC 2007 dataset.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [40] Redmon J, Farhadi A. Yolo9000: Better, faster, stronger. In: 2017 IEEE conference on computer vision and pattern recognition (CVPR); 2017. p. 6517–25. source

      Not checked in the free run.

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

    Premisecitation

    YOLO achieves real-time object detection at 45 frames per second.

    “the detector running at 45 frames per second”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7] Redmon J. Divvala S. Girshick R.B. Farhadi A. You only look once: Unified, real-time object detection 2016 IEEE conference on computer vision and pattern recognition (CVPR) 2016 779 788

      Not checked in the free run.

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

    Premisecitation

    YOLO has a smaller detection network running at 155 frames per second.

    “They also have a smaller detection network running at 155 frames per second.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7] Redmon J. Divvala S. Girshick R.B. Farhadi A. You only look once: Unified, real-time object detection 2016 IEEE conference on computer vision and pattern recognition (CVPR) 2016 779 788

      Not checked in the free run.

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

    Premisecitation

    Mask R-CNN can run at 5 FPS.

    “It can run at 5 FPS”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [73] He K. Gkioxari G. Dollár P. Girshick R.B. Mask r-cnn IEEE Trans Pattern Anal Mach Intell 42 2020 386 397 29994331 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    PP-YOLO outperforms both YOLO-v4 and YOLO-v3 in terms of accuracy and speed.

    “Comparing it with the novel YOLO-v4 or the older YOLO-v3 [31] we can see that it outperform both in terms of accuracy and speed: the best performing YOLOv3 [31] runs at 60 FPS @38% mAP, while the best performing YOLO-v4 [63] runs at 123 FPS @38% mAP.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [31] Redmon J, Farhadi A. Yolov3: An incremental improvement, ArXiv abs/1804.02767.

      Not checked in the free run.

    • cites [63] Bochkovskiy A, Wang CY, Liao H. Yolov4: Optimal speed and accuracy of object detection, ArXiv abs/2004.10934.

      Not checked in the free run.

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

    Interpretationnone given

    In certain circumstances, pre-training, self-training and joint-training can have additive effects on the performances of the trained model.

    “A case is also made for joint-training, because in certain circumstances pre-training, self-training and joint-training can have additive effects on the performances of the trained model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    On a Tesla V100 GPU, PP-YOLO runs at 72.9 FPS at 45.2% mAP or 132 FPS at 39% mAP.

    “On a Tesla V100 GPU the frames per second for PP-YOLO are 72.9 FPS @45.2% mAP or 132 FPS @39 % mAP.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [65] Long X, Deng K, Wang G, Zhang Y, Dang Q, Gao Y, et al. Pp-yolo: An effective and efficient implementation of object detector, ArXiv abs/2007.12099.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    SqueezeDet achieved 80.4% mAP on the KITTI dataset with a model size of 26.8 MB and a speed of 32.1 FPS.

    “The detection performance attained on the KITTI dataset is of 80.4% mAP, with a model size of only 26.8 MB and a speed of 32.1 FPS.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [81] Wu B. Iandola F.N. Jin P. Keutzer K. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving 2017 IEEE conference on computer vision and pattern recognition workshops (CVPRW) 2017 446 454

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    CSPNet achieves 38.4% AP with a ResNeXt50 backbone while running at 35 FPS on a 608x608 input image size on MS COCO.

    “reaching 38.4 % AP with a ResNeXt50 backbone, while running at 35FPS on a 608x608 input image size.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [84] Wang CY, Liao H, Yeh IH, Wu YH, Chen PY, Hsieh JW. Cspnet: A new backbone that can enhance learning capability of cnn. In: 2020 IEEE/CVF conference on computer vision and pattern recognition workshops (CVPRW); 2020. p. 1571–80. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    On COCO test-dev2015, SSD obtained better results than state-of-the-art Faster R-CNN on IoU precision [0.5:0.95], with mAP 28.8% and 48.5% @0.5, and 30.3% on IoU 0.75.

    “On the COCO test-dev2015 dataset, the model obtained better results than the state-of-the-art Faster R-CNN on the IoU precision [0.5:0.95], getting a mAP of 28.8% and 48.5% @0.5, while on IoU 0.75, 30.3%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu CY, et al. Ssd: Single shot multibox detector. In: ECCV; 2016. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Increasing image size to 512x512 and using image expansion for augmentation increases SSD accuracy on VOC 2012 up to 82.2%.

    “When increasing the image size to 512x512 and using image expansion for augmentation purposes, one can see a slight bump in the accuracy, on VOC 2012 the accuracy increases up to 82.2%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu CY, et al. Ssd: Single shot multibox detector. In: ECCV; 2016. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    SqueezeDet is 19x faster than Faster-RCNN (with VGG16 backbone) and 61x smaller.

    “this one is 19x faster than Faster-RCNN(with a VGG 16 backbone) and 61x smaller.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [81] Wu B. Iandola F.N. Jin P. Keutzer K. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving 2017 IEEE conference on computer vision and pattern recognition workshops (CVPRW) 2017 446 454

      Not checked in the free run.

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

    Interpretationreasoning

    Training hardware does not need to be expensive if the application does not mandate it.

    “The training hardware doesn‘t need to be expensive, if the application does not mandate it.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    For factory applications, power consumption is not a hard constraint, but memory is a problem, so compression is justified.

    “While the power consumption requirement is not ”a hard constraint”, memory is a problem, thus compression is justified also for factory applications.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Defect detection models need high accuracy, high speed, a few MBs or tens of MBs of memory for operation, and as little power consumption as possible per GFLOPS of computation.

    “As a summary of the needs for defect detection: the models needs to be high accuracy, run at high speed, while needing a few MBs of memory or a few tens of MBs for operation and also consuming as little power as possible per every GFLOPS of computation.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Edge devices such as the Intel Neural Compute Stick or the Nvidia Jetson family are custom built and optimized for deep learning, so despite lacking processing power and memory size, they can comfortably run these network models.

    “These are custom built and optimized for deep learning applications, thus even though they lack in processing power and memory size, they are able to comfortably run these network models.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    There are four main categories of compression and acceleration techniques.

    “There are four main categories of compression and acceleration techniques [87]:”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] Cheng Y, Wang D, Zhou P, Zhang T. A survey of model compression and acceleration for deep neural networks, ArXiv abs/1710.09282.

      Not checked in the free run.

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

    Interpretationreasoning

    There is a real need for model compression and acceleration developments to fit in embedded devices found on autonomous cars or in factories.

    “Moreover, these is a real need for model compression and acceleration developments, in order to be able to fit in the embedded devices that are found on an autonomous car or in a factory.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The weaker the performance of the hardware setup (a small number of TFOPS or GPU RAM), the better the offset performance gain relative to the basis performance will be.

    “Moreover, the weaker the performance of the hardware setup (a small number of TFOPS or GPU RAM), the better the offset performance gain relative to the basis performance will be.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    For factory equipment, the main constraint for deployment of deep learning defect detection algorithms is that industrial computers' computing power is not great and the memory available is questionable.

    “For factory equipment, the main constraint for the deployment of deep learning defect detection algorithms is the fact that for industrial computers the computing power is not that great and the memory available is also questionable.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    The YOLO detection model allows penalizing more either accuracy or localization, giving some freedom in the training process.

    “All in all, a key aspect that the YOLO detection model provides is the fact that you can penalize more, either accuracy or localization, thus there is some freedom with the training process.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [63] Bochkovskiy A, Wang CY, Liao H. Yolov4: Optimal speed and accuracy of object detection, ArXiv abs/2004.10934.

      Not checked in the free run.

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

    Interpretationnone given

    Deep learning methods achieve state-of-the-art performance in defect detection while also having great generalization properties.

    “as demonstrated deep learning methods achieve state-of-the-art performance in defect detection, while also having great generalization properties.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    YOLOv4 and PPYOLO are network models with great generalization properties and can be helpful for getting a proof of concept quickly.

    “These are network models with great generalization properties and could be really helpful when trying to get a proof of concept as quickly as possible, hence we also tried out some experiments with the YOLOv4 framework.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [63] Bochkovskiy A, Wang CY, Liao H. Yolov4: Optimal speed and accuracy of object detection, ArXiv abs/2004.10934.

      Not checked in the free run.

    • cites [65] Long X, Deng K, Wang G, Zhang Y, Dang Q, Gao Y, et al. Pp-yolo: An effective and efficient implementation of object detector, ArXiv abs/2007.12099.

      Not checked in the free run.

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

    Interpretationnone given

    The dataset and the chosen model lead to great performances.

    “The general accepted idea is that the dataset, as well as the chosen model lead to great performances, thus both need to receive attention from the developer.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Smaller, more specific applications can yield great results using general purpose laptops and generic detection models tweaked for the defects looked into.

    “Smaller, more specific applications can yield great results and thus, improve the workings of a small lab or business just by using general purpose laptops and generic detection models which will be tweaked for the defects we look into.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    A key trend in the sector is to obtain higher and higher box accuracy for detection.

    “A key trend in the sector is to obtain higher and higher box accuracy for detection.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    There is no rule of thumb when choosing which object detection model shall generalize best on a particular dataset of interest.

    “There is no rule of thumb when choosing which object detection model shall generalize the best on the particular dataset of interest.”
    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

    Training models will still be performed on large distributed systems or via the cloud, but deployment and operation will be performed cloud free.

    “Training models will be still performed on large distributed systems or via the cloud, but deployment and operation will be performed ”cloud free”.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    To be deployed in applications like autonomous cars, models need to run at a high frame rate, greater than 30 FPS, to be considered real time and to make decisions faster than humans.

    “Also, in order to be deployed in multiple applications like autonomous cars, these models need to run at a high frame rate, greater than 30 FPS, in order to be considered real time and in order to make the decisions faster than humans.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    An upcoming trend in detection is to limit data transfers between the processor responsible for data acquisition and the cloud system.

    “Furthermore, an upcoming trend in detection is to limit the data transfers between the processor responsible for data acquisition(field device) and the cloud system.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    Hardware setup and availability plays a role in the performance of the models.

    “Thus, hardware setup and availability(that‘s why recommended is using virtual machines on a desktop machine learning station for general solutions) plays a role in the performance of the models.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    General applications that aim to detect lots of defects need very large and balanced datasets, a hardware setup with lots of computational power, and a specific detection model built from the ground up.

    “While general applications that have a target as to detect lots of defects, need very large and balanced datasets, a hardware setup with lots of computational power and a specific detection model that is not just tweaked for defect detection, but built from the ground up for the specific action that we want it to perform.”
    Quote in the paperNot a substantial claim
67 of 67

  • The paper relies on this without support.

    Benchmark performance on ImageNet or COCO is indicative of suitability for industrial defect detection.

    “Having a strong backbone is one of the key design choices for a high accuracy defect detection model.”
    “Most of these models can be skewed in order to yield great defect detectors.”

  • The paper relies on this without support.

    Defects in industrial images can be treated as objects for object detection.

    “The main topic of interest for defect detection is object detection, because defects are treated as objects and they need to be localized and subsequently classified.”

  • The paper relies on this without support.

    Industrial ruggedness is not a characteristic of deep-learning algorithms.

    “While the industrial ruggedness is not a characteristic of deep-learning algorithms”

  • The authors state this as an assumption.

    Real-time defect detection (e.g., airplane inspections) requires high accuracy and reliability to replace traditional industrial inspection methods.

    “For defect detection in real time (e.g.: for Airplane inspections) we need this kind of accuracy for defect detection because the process needs high accuracy and reliability in order for it to replace traditional industrial inspection methods”

  • The authors state this as an assumption.

    Real-time defect detection requires high speed of processing because maintenance operations cannot take more than a couple of hours when flights are scheduled daily.

    “and also, we need high speed of processing because maintenance operations cannot take more than a couple of hours when flights are scheduled daily.”
5 of 5

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?

  • Deep learning-based object detection methods have been rapidly improved and advanced.

    “With the development of deep learning technology, deep learning-based object detection methods have been rapidly improved and advanced [7, 8].”

    From A deep learning model for steel surface defect detection (2024) · cites this paper as [7]

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.

  • Compact CNN's FLOP share is misstated

    Minor

    The stated models use 76 million and 267 million FLOPs, respectively. The compact CNN therefore uses about 28.5% of MobileNetV2's FLOPs, not 33%.

    Reported 33 · recomputed 28.464419

  • Stated YOLO speed ceiling conflicts with the YOLO section

    Minor

    The paper says YOLO detectors reach up to 120 FPS, but later describes a YOLO detection network running at 155 FPS.

  • Table 1 lists a more accurate model than FixEfficientNet-L2

    Minor

    The caption calls FixEfficient the most accurate model, but Table 1 lists ViT-H/14 at 88.55%, above FixEfficientNet-L2's 88.5%.

3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. Touvron H, Vedaldi A, Douze M, Jégou H. Fixing the train-test resolution discrepancy: Fixefficientnet, ArXiv abs/2003.08237. · arXiv
  2. Iandola FN, Moskewicz MW, Ashraf K, Han S, Dally W, Keutzer K. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <1mb model size, ArXiv abs/1602.07360. · arXiv

Cited by

Sorted by citations

Related

Authors
Tulbure, A.-A., Tulbure, A.-A., & Dulf, E.-H.
Affiliations
Technical University of Cluj-Napoca (RO)
Corresponding author
Dulf, Eva-Henrietta
ORCID
Venue
Journal of Advanced Research
Publisher
Elsevier BV
Published
2022-01
Volume
35
Pages
33-48
ISSN
  • 2090-1224 (ISSN-L)
  • 2090-1232
DOI
10.1016/j.jare.2021.03.015
PMC
PMC8721352
OpenAlex
W3157243312
Semantic Scholar
cc4c8b3ef52594f996807e85c01831ceae7ba8ed
Cited by
372
References
95
Citations by year
2026: 58 · 2025: 98 · 2024: 89 · 2023: 84 · 2022: 37 · 2021: 6
FWCI
5.03
Type
journal article
Language
en
Open access
Licence
Locations
  • Journal of Advanced Research (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
UEFISCDI
Topics
  • Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
Keywords
defect detection, visual defect detection, object detection models, YOLO, YOLOv4, Single Shot MultiBox Detector, R-CNN, model compression, model acceleration, manufacturing quality inspection, GPU computing
MeSH terms
Algorithms, Image Processing, Computer-Assisted, Software, Neural Networks, Computer, Data Compression

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