CoChat
Sign Up
Cited by 85Industrial Vision Systems and Defect DetectionAdvanced Neural Network ApplicationsInfrastructure Maintenance and Monitoring

A Compact Convolutional Neural Network for Surface Defect Inspection

Open access
Huang, Yibin, Qiu, Congying, Wang, Xiaonan, Wang, Shijun, Yuan, Kui·2020·Sensors
DOI 10.3390/s20071974PMC PMC7180441
Open this paper in CoChat — free accountOpen PDF

Abstract

The advent of convolutional neural networks (CNNs) has accelerated the progress of computer vision from many aspects. However, the majority of the existing CNNs heavily rely on expensive GPUs (graphics processing units). to support large computations. Therefore, CNNs have not been widely used to inspect surface defects in the manufacturing field yet. In this paper, we develop a compact CNN-based model that not only achieves high performance on tiny defect inspection but can be run on low-frequency CPUs (central processing units). Our model consists of a light-weight (LW) bottleneck and a decoder. By a pyramid of lightweight kernels, the LW bottleneck provides rich features with less computational cost. The decoder is also built in a lightweight way, which consists of an atrous spatial pyramid pooling (ASPP) and depthwise separable convolution layers. These lightweight designs reduce the redundant weights and computation greatly. We train our models on groups of surface datasets. The model can successfully classify/segment surface defects with an Intel i3-4010U CPU within 30 ms. Our model obtains similar accuracy with MobileNetV2 while only has less than its 1/3 FLOPs (floating-point operations per second) and 1/8 weights. Our experiments indicate CNNs can be compact and hardware-friendly for future applications in the automated surface inspection (ASI).

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a compact CNN with a lightweight bottleneck and lightweight decoder for surface defect classification and segmentation, targeting low-power CPUs in automated surface inspection. It reports competitive accuracy to MobileNetV2 with far fewer weights and FLOPs, and strong results across texture, steel, DAGM, and wood defect datasets.

Key findings

  • The model can run on an Intel i3-4010U CPU: the abstract claims classification/segmentation within 30 ms, while the conclusions report ∼29 ms for 200×200 NEU images, ∼140 ms for 331×331 Textures images (@0.75, 99.33% Acc), and 217 ms for 512×512 DAGM images (@0.35, 99.43% mAcc); GPU inference is <30 ms.
  • On Textures classification, ours@0.75 is ∼4.9× faster than MobileNetV2@1.4 with only ∼1/10 weights, ∼1/4 FLOPs, and a slight ∼0.41% accuracy decrease; the abstract states less than 1/3 FLOPs and 1/8 weights of MobileNetV2.
  • On NEU classification, both MobileNetV2@0.35 and ours@0.35 reach 100% test accuracy, but the paper states ours has the least parameter, computational cost, and running time.
  • On DAGM segmentation, ours@0.35 achieves 99.43% mAcc; ImageNet-pretrained MobileNetV2@0.35 with the LW decoder reaches 99.90% mAcc. Gradual training plus ImageNet pretraining improves tiny-defect TPR2 by 4.77% and mAcc by 0.61%.

Methods

  • Architecture: a lightweight bottleneck uses parallel depthwise convolutions with kernels 5×1, 1×5, 3×1, 1×3, 2×1, 1×2, 2×2, and 3×3, concatenated and followed by pointwise convolution; classification adds global average pooling and a fully connected layer, while segmentation uses an ASPP decoder with depthwise separable convolutions and atrous rates 6, 12, and 18.
  • Data: Textures dataset (8674 images, 64 classes from KTH, Kyberge, and UIUC, resized to 331×331); NEU steel defect dataset (6 classes, 1800 grayscale 200×200 images, split 150 train/150 test per class); DAGM 2007 (6 textured-surface classes, 575 training images per class, about 80 with defects); Wood defect dataset (18 defect types, 839 images of size 488×512).
  • Training and analysis: TensorFlow with softmax cross-entropy, Adam, learning rate 0.0001 decaying 10% every 10,000 steps, batch normalization, L2 regularization, strong imgaug-based augmentation, fine-tuning from ImageNet/Flower102/Textures/DAGM, and gradual training where defective images are zoomed with ratio linearly reduced from 3.6 to 1 over the first 20K steps and non-defect images are added after 80K steps. Evaluation compares accuracy, mAcc, TPR/TNR, weights, FLOPs, and runtime on CPU/GPU.

Limitations

  • The paper states that surface defect datasets are small and difficult to collect, with severe image and pixel imbalance; tiny defects can be only 3–4 pixels wide in 512×512 DAGM images, and gradual-training setups are empirically designed.
  • The paper states that ImageNet pretraining is costly, taking days to weeks and over 130 GB disk storage, though Flower102 or Textures pretraining is reported as a competitive compromise.
  • Evident from the text: there is no dedicated limitations section, evaluation is limited to four surface datasets and selected comparison models, and CPU inference times vary by input size, with 331×331 and 512×512 inputs exceeding the abstract’s 30 ms claim.

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

Integrity signals

Registry

Crossref · 10.3390/s20071974

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 35 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The paper's compact CNN model achieves high performance on tiny defect inspection.

    “we develop a compact CNN-based model that not only achieves high performance on tiny defect inspection but can be run on low-frequency CPUs (central processing units).”
    Quote in the paperSubstantial claim
  • Findingown data

    Compared with MobileNetV2, our backbone is more compact in terms of the number of layers and blocks (MobileNetV2 has 53 convolution layers and 19 blocks, ours has 21 and 12).

    “Compared with MobileNetV2, our backbone is more compact in terms of the number of layers and blocks. For example, the MobileNetV2 has 53 convolution layers and 19 blocks, while our model has 21 and 12.”
    Quote in the paperSubstantial claim
  • Findingown data

    The paper's compact CNN model can be run on low-frequency CPUs.

    “we develop a compact CNN-based model that not only achieves high performance on tiny defect inspection but can be run on low-frequency CPUs (central processing units).”
    Quote in the paperSubstantial claim
  • Findingown data

    The model has 1/8 weights of MobileNetV2.

    “Our model obtains similar accuracy with MobileNetV2 while only has less than its 1/3 FLOPs (floating-point operations per second) and 1/8 weights.”
    Quote in the paperSubstantial claim
  • Findingnone given

    The model has satisfactory results for different materials with small training sets.

    “Following proper training strategies, our model has satisfactory results for different materials with small training sets.”
    Quote in the paperSubstantial claim
  • May combine two claims. 35
  • Classified as probably not a substantial claim. 27
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingreasoning

    The model significantly reduces the number of weights, computation and time cost.

    “Our model significantly reduces the number of weights, computation and time cost.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The model can successfully classify/segment surface defects with an Intel i3-4010U CPU within 30 ms.

    “The model can successfully classify/segment surface defects with an Intel i3-4010U CPU within 30 ms.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The gradual training strategy prevents the network from overfitting and improves model accuracy.

    “This strategy has proved to be a valid way to prevent the network from overfitting. It also improves model accuracy.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The defect pixels only take a small proportion of the whole image, so the backpropagation gradient of defects is overwhelmed by the background gradient, causing low accuracy.
    • Rests on an assumption: The images with defects are usually far less than non-defect, so networks do not have sufficient data to learn defect features, resulting in accuracy significantly decreases.

  • May combine two claims.

    Findingnone given

    MobileNetV2@1.4 and ResNetV2-50 contain a large number of parameters (and operations) which require GPUs.

    “Regardless of the good accuracy performance of the MobileNetV2@1.4 and ResNetV2-50, these models contain a large number of parameters (and operations) which require GPUs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Channel design leads to ∼1 M of weights reduction without hurting model performance.

    “Another improvement is on channel design, which leads to ∼1 M of weights reduction without hurting model performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    MobileNetV2@1.4 performs better than MobileNetV2@0.35; ours@0.75 performs better than ours@0.35.

    “A model with more weights is more likely to obtain higher accuracy. For example, MobileNetV2@1.4 performs better than MobileNetV2@0.35; ours@0.75 performs better than ours@0.35.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The model obtains state-of-the-art in both classification and segmentation.

    “Our model obtains the state-of-the-art in both classification and segmentation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The compact design remarkably reduces the number of weights and computation while still achieving high performance.

    “this work introduces our novel lightweight CNN, the compact design remarkably reduces the number of weights and computation while still achieving a high performance.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    The DAGM full input image size is 512×512, but the width of the smallest defects only consists of 3-4 pixels.

    “Noticeably, the full input image size is 512×512, but the width of the smallest defects only consists of 3–4 pixels.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Compared with random initialized ResNetV2-50 and MobileNetV2@1.4, ours@0.75 (ImageNet pretrained) has much fewer weights and FLOPs, but higher accuracy.

    “Compared with the random initialized ResNetV2-50 and MobileNetV2@1.4, ours@0.75 model (ImageNet pretrained) has much fewer weights and FLOPs, but higher accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Ours@0.35 is 3 times faster than MobileNetV2@0.35 on a low-cost IPC, with only 0.47% mAcc drop.

    “Ours@0.35 is 3 times faster than the MobileNetV2@0.35 on a low-cost IPC, which make ∼5 fps surface defect inspection is possible, with only 0.47% mAcc drop.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Ours@0.35 model costs the least computation with the least weights, while still achieves a mAcc of 99.43% on DAGM dataset.

    “Ours@0.35 model costs the least computation with the least weights, while still achieves a mAcc of 99.43%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Gradual training improves TPR2 of the tiny defect but does not improve mAcc.

    “As shown in Table 6, gradual training improves TPR2 of the tiny defect (the True Positive Rate of the second type texture surface), but cannot improve the mAcc.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Ours@0.35 has the same mAcc as ShuffleNet@0.25 but is 1.7 times faster on the IPC.

    “Ours@0.35 has same mAcc with the ShuffleNet@0.25 but 1.7 times faster on the IPC.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Both MobileNetV2@0.35 and ours@0.35 achieve 100% test accuracy on the NEU dataset.

    “Both MobileNetV2@0.35 and the ours@0.35 achieve exceptionally high performance with the best record of 100% test accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Our model makes a fine prediction of elliptical defect masks, and sometimes the outputs are more accurate than the ground truths.

    “As shown in Figure 9, our model makes a fine prediction of elliptical defect masks, and sometimes the outputs are more accurate than the ground truths.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    Our model has much higher accuracy than the Decaf and similar Error escape rate on the Wood defect dataset.

    “Our model has much higher accuracy than the Decaf and similar Error escape rate.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Compared with modern CNNs, the LW model has less than 1 M weights, which is no longer needed to run on expensive GPUs.

    “Compared with modern CNNs, the LW model has less than 1 M weights, which is no longer needed to run on expensive GPUs.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    In the Wood defect dataset, the evenly distributed bounding boxes sometimes do not match the defects well.

    “These evenly distributed bounding boxes sometimes do not match the defects well.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    The inference time will be <30 ms if using a GPU.

    “The inference time will be <30 ms if using a GPU.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Our model tests an image in 397 ms with an Intel i3-4010U(1.7 GHz) CPU or 27 ms with a GTX1080 GPU.

    “our model tests an image in 397 ms with an Intel i3-4010U(1.7 GHz) CPU or 27 ms with a GTX1080 GPU.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    The output channels of the last layer are shrunk into 320, which is sufficient enough for surface defect inspection.

    “In our design, the output channels of the last layer are shrunk into 320, which is sufficient enough for surface defect inspection.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Compact models are also used in surface defect inspection; Reference [29] introduces a compact CNN to perform defect segmentation and classification simultaneously.

    “Compact models are also used in surface defect inspection, for instance, Reference [29] introduces a compact CNN to perform defect segmentation and classification simultaneously.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [29] Racki, D., Tomazevic, D., and Skocaj, D. (2018, January 12–15). A compact convolutional neural network for textured surface anomaly detection. Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, USA. source

      Not checked in the free run.

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

    Premisecitation

    Many compact DNN models can still achieve excellent performance on image classification.

    “Interestingly, many compact deep neural network (DNN) models can still achieve an excellent performance on image classification, such as SquezeNet [21], Compact DNN [22], MobileNetV1 [23], ShuffleNet [24], and MobileNetV2 [25].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [21] Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., and Keutzer, K. (2016). Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5 mb model size. arXiv.

      Not checked in the free run.

    • cites [22] Wu, C., Wen, W., Afzal, T., Zhang, Y., Chen, Y., and Li, H. (2017, January 21–26). A Compact DNN: Approaching GoogLeNet-Level Accuracy of Classification and Domain Adaptation. Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA. source

      Not checked in the free run.

    • cites [23] Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., and Adam, H. (2017). Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv.

      Not checked in the free run.

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

    Premisecitation

    Fine-tuned models significantly outperform models trained with random initialization or fixed feature extractors.

    “The study shows that fine-tuned models significantly outperform others.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30] Kornblith, S., Shlens, J., and Le, Q.V. (2018). Do Better ImageNet Models Transfer Better?. arXiv. source

      Not checked in the free run.

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

    Premisecitation

    Models that achieve higher accuracy on ImageNet also exceed others on different datasets.

    “At the same time, those models that achieve higher accuracy on ImageNet also exceed others on different datasets.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30] Kornblith, S., Shlens, J., and Le, Q.V. (2018). Do Better ImageNet Models Transfer Better?. arXiv. source

      Not checked in the free run.

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

    Premisecitation

    In surface inspection, a fine-tuned ImageNet pretrained VGG-16 is used as a patch classifier; fine-tuning all layers is the best valid way for models to achieve high accuracy.

    “In the practice of surface inspection [15], a fine-tuned ImageNet pretrained VGG-16 is used as a patch classifier on surface defect inspection. The experiment shows that fine-tuning all layers is the best valid way for models to achieve high accuracy.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] Kim, S., Kim, W., Noh, Y.K., and Park, F.C. (2017, January 14–19). Transfer learning for automated optical inspection. Proceedings of the International Joint Conference on Neural Networks, Anchorage, AK, USA. source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    In our network design, we hold the condition Cout≤2×Cin, thus WLW<WMB.

    “In our network design, we hold the condition Cout≤2×Cin, thus WLW<WMB.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Reference [2] model gets 99.05% accuracy on the NEU dataset.

    “their model gets a better accuracy at 99.05%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [2] Yi, 2017, An End-to-End Steel Strip Surface Defects Recognition System Based on Convolutional Neural Networks, Steel Res. Int. source

      Not checked in the free run.

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

    Study design and conductnone given

    The Textures dataset was split randomly into two halves, one for training and one for validation.

    “We split the Textures dataset into two halves randomly, one for the training set, and another for the validation.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The model was pretrained on Oxford Flower102 and ImageNet.

    “As shown in Table 3, we pretrain our model on Oxford Flower102 [53] and ImageNet.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [53] Nilsback, M.E., and Zisserman, A. (2008, January 16–19). Automated Flower Classification over a Large Number of Classes. Proceedings of the 2018 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, Bhubaneswar, India. source

      Not checked in the free run.

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

    Premisecitation

    Context information is critical to increasing the mAcc of the sliding window classifier.

    “Context information is critical to increasing the mAcc this sliding window classifier.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [13] Weimer, D., Benggolo, A.Y., and Freitag, M. (December, January 30). Context-aware Deep Convolutional Neural Networks for Industrial Inspection. Proceedings of the Australasian Conference on Artificial Intelligence, Canberra, Australia.

      Not checked in the free run.

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

    Premisecitation

    Data augmentation on the NEU dataset is crucial for performance improvement.

    “Their work proves that data augmentation on this dataset is crucial for performance improvement”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [2] Yi, 2017, An End-to-End Steel Strip Surface Defects Recognition System Based on Convolutional Neural Networks, Steel Res. Int. source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Sliding window based models are unstable and hard to train.

    “So these sliding window based models are unstable and hard to train.”
    Quote in the paperSubstantial claim

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

    Study design and conductreasoning

    L2 regularization was used in each convolution layer to avoid overfitting.

    “To avoid overfitting, we use L2 regularization in each convolution layer.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    VGG-16 parameters are 134.31 M, causing 9024 M FLOPs in one sub-image, and 9 times FLOPs are needed in an entire input image.

    “However, the parameters of a VGG-16 are 134.31 M, which causes very large computational cost—9024 M FLOPs in one sub-image, 9 times FLOPs are needed in an entire input image.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [15] Kim, S., Kim, W., Noh, Y.K., and Park, F.C. (2017, January 14–19). Transfer learning for automated optical inspection. Proceedings of the International Joint Conference on Neural Networks, Anchorage, AK, USA. source

      Not checked in the free run.

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

    Interpretationreasoning

    To extract positive samples, more accurate manual re-label ground truths are necessary.

    “To extract positive samples, more accurate manual re-label ground truths are necessary.”
    Quote in the paperNot a substantial claim

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

    Study design and conductreasoning

    Batch normalization layers were attached to each convolution layer to speed up training.

    “To speed up training, batch normalization layers are attached to each convolution layer, with decay = 0.9997 and epsilon = 0.00001.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    C-CNN training strategy enables the model to fit a good classifier without affecting the segmentation accuracy.

    “This training strategy enables the model to fit a good classifier without affecting the segmentation accuracy.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [29] Racki, D., Tomazevic, D., and Skocaj, D. (2018, January 12–15). A compact convolutional neural network for textured surface anomaly detection. Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, USA. source

      Not checked in the free run.

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

    Premisecitation

    In Reference [14], 324,000 training patches are cropped to ensure model accuracy.

    “in Reference [14] 324,000 training patches are cropped to ensure model accuracy.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [14] Weimer, 2016, Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection, CIRP Ann. source

      Not checked in the free run.

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

    Premisecitation

    Ruoxu, Ren et al. [54] transfer learning method reaches 99.27% accuracy on the NEU dataset.

    “This transfer learning method helps them reach a higher accuracy at 99.27%.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

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

    Premisecitation

    C-CNN mAcc is 99.43% with small computational cost and very few parameters.

    “The mAcc of this model is 99.43% with small computational cost and very few parameters.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [29] Racki, D., Tomazevic, D., and Skocaj, D. (2018, January 12–15). A compact convolutional neural network for textured surface anomaly detection. Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, USA. source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The conservative gradual training strategy enables the network to learn tiny-defect features and avoid foreground underfitting that would result in predicting all images as background.

    “This conservative training strategy enables the network to learn the feature of tiny defects, to avoid foreground underfitting and result in predict all images as the background.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    Strong data augmentation was implemented to balance the image number of each class and avoid overfitting.

    “We implement strong data augmentation to balance the image number of each class and avoid the effect of overfitting.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    Some strategies, e.g., transfer learning and zoom-in training, are critical to building the networks of tiny defect inspection.

    “we observe that some strategies, e.g., transfer learning and zoom-in training, are critical to building the networks of tiny defect inspection.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Reference [54] used an ImageNet pretrained Decaf as the fixed feature extractor and trained a multinomial logistic regression (MLR) classifier with patches.

    “In Reference [54], an ImageNet pretrained Decaf is used as the fixed feature extractor, and a multinomial logistic regression (MLR) classifier is trained with patches.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

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

    Premisecitation

    Reference [54] found that patch size and patch stride greatly affect the result: small patches lack enough information, small stride leads to high computational cost, while large patch size or stride leads to imprecise segmentation.

    “They found that patch size and patch stride will be a greatly affect the result, small patches will not include enough information, and small stride will lead to high computational cost, while large patch size or stride will lead to imprecise segmentation result.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

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

    Premisecitation

    Reference [54] chose a patch size roughly the same as the size of the defect (50×50) and a stride of 10 on this dataset.

    “they choose a patch size roughly the same as the size of the defect (50×50) and a stride of 10.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Reference [54] uses a 24 core workstation and 2 min to generate heatmaps per image.

    “they use a 24 core workstation and 2 min to generate heatmaps per image”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    We introduce a novel compact CNN model with low computation and memory cost.

    “we introduce a novel compact CNN model with low computation and memory cost”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    In contrast to traditional approaches, our model gets rid of hand-crafted features and patch size selection.

    “In contrast to the traditional approaches, our model gets rid of hand-crafted features and patch size selection.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Reference [54] got the best-reported result in all the defect types, with the highest A and the lowest E.

    “They got the best-reported result in all the defect types, which has the highest A, and the lowest E.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

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

    Interpretationnone given

    Our model has no patch size selection process and no post-process, differing from the segmentation method in Reference [54].

    “Our model has no patch size selection process and no post-process, differ from the segmentation method in Reference [54].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [54] Ren, 2017, A Generic Deep-Learning-Based Approach for Automated Surface Inspection, IEEE Trans. Cybern.

      Not checked in the free run.

53 of 53

  • The paper relies on this without support.

    High-performance cross-product ASI algorithms are urgently needed in manufacturing.

    “High-performance cross-product ASI algorithms are urgently needed in manufacturing.”

  • The paper relies on this without support.

    Most surface defects are tiny and very similar to their background, making inspection difficult for traditional methods.

    “Besides, most surface defects are tiny and very similar to their background, even the inspection for a single type of defect can be very difficult for these traditional methods.”

  • The paper relies on this without support.

    Traditional algorithms can barely achieve the standard of industrial applications.

    “Although traditional algorithms are designed for specific surfaces, effective measures such as recall, precision, and accuracy can barely achieve the standard of industrial applications.”

  • The paper relies on this without support.

    Expensive GPUs with high computing capabilities are necessary to support large CNN workloads.

    “In many cases, expensive GPUs with high computing capabilities are necessary to support large workloads.”

  • The paper relies on this without support.

    Most core computing resources of ASI platforms are low-power CPUs of industrial personal computers or FPGAs.

    “However, in considering manufacturing cost, most core computing resources of ASI platforms are mainly the low-power CPUs of industrial personal computers (IPC), or even FPGAs (field programmable gate array).”

  • The paper relies on this without support.

    There exist few online datasets for surface defect images.

    “Meanwhile, a CNN with high performance needs large datasets to train, while there exist few online datasets for surface defect images.”

  • The paper relies on this without support.

    The defect pixels only take a small proportion of the whole image, so the backpropagation gradient of defects is overwhelmed by the background gradient, causing low accuracy.

    “It is worth pointing out that the defect pixels only take a small proportion of the whole image, therefore, the backpropagation gradient of defects will be overwhelmed by the background gradient, and subsequently cause low accuracy.”

  • The paper relies on this without support.

    The images with defects are usually far less than non-defect, so networks do not have sufficient data to learn defect features, resulting in accuracy significantly decreases.

    “Moreover, the images with defects are usually far less than the non-defect, the networks do not have sufficient data to learn the defect features, resulting in accuracy significantly decreases.”

  • The paper relies on this without support.

    Many ASI platforms are not equipped with expensive hardware devices.

    “However, an ideal ASI model should not only perform well inaccuracy but also takes less computational resources, because many ASI platforms usually are not equipped with such expensive hardware devices.”

  • The paper relies on this without support.

    Deep neural networks are computationally expensive and memory expensive, while hardware resources are very limited in reality.

    “Deep neural networks are computationally expensive and memory expensive, while hardware resources are very limited in reality.”

  • The paper relies on this without support.

    Pretraining a designed model on ImageNet requires days to weeks and high-performance computers with strong GPUs and over 130 GB disk storage.

    “However, it takes days to weeks to pretrain a designed model on ImageNet with a high-performance computer equipped with exceptionally strong GPUs and over 130 GB disk storage.”
11 of 11

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?

  • The compact CNN can classify and segment correctly surface defects in 30 ms on an Intel i3-4010U CPU.

    “This means it can classify and segment correctly surface defects in 30 ms on an common use CPU like Intel i3-4010U.”

    From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [47]

  • 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.”

    From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [47]

  • 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.”

    From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [47]

  • 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%.”

    From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [47]

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.

  • CPU inference is not within 30 ms for all reported tasks

    Major

    The abstract says the model can classify or segment defects on the specified CPU within 30 ms. The conclusion reports 140 ms for Textures classification and 217 ms for DAGM segmentation on that CPU.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 5

  1. Weimer, D., Benggolo, A.Y., and Freitag, M. (December, January 30). Context-aware Deep Convolutional Neural Networks for Industrial Inspection. Proceedings of the Australasian Conference on Artificial Intelligence, Canberra, Australia.
  2. Iandola, F.N., Han, S., Moskewicz, M.W., Ashraf, K., Dally, W.J., and Keutzer, K. (2016). Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5 mb model size. arXiv.
  3. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., and Adam, H. (2017). Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv.
  4. Molchanov, P., Tyree, S., Karras, T., Aila, T., and Kautz, J. (2016). Pruning convolutional neural networks for resource efficient inference. arXiv.
  5. Krizhevsky, 2010, Convolutional deep belief networks on cifar-10, Unpubl. Manuscr., 40, 1

Cited by

Sorted by citations

Related

Authors
Huang, Y., Qiu, C., Wang, X., Wang, S., & Yuan, K.
Affiliations
  • Chinese Academy of Sciences (CN)
  • Columbia University (US)
Corresponding author
Wang, Xiaonan
ORCID
Venue
Sensors
Publisher
MDPI AG
Published
2020-04-01
Volume
20
Issue
7
Pages
1974
ISSN
1424-8220 (ISSN-L)
DOI
10.3390/s20071974
PMC
PMC7180441
OpenAlex
W3015187397
Semantic Scholar
4675b6b4075ddf91df71cc0677c3b36df0ac46bc
Cited by
85
References
52
Citations by year
2026: 6 · 2025: 8 · 2024: 18 · 2023: 15 · 2022: 17 · 2021: 19 · 2020: 2
FWCI
8.45
Type
journal article
Language
en
Open access
Licence
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Natural Science Founda- ion of China (61421004)
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
surface defect detection, lightweight convolutional neural network, CPU-only inference, automated surface inspection, atrous spatial pyramid pooling, depthwise separable convolution, defect segmentation, defect classification, manufacturing quality inspection

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