An Improved VGG19 Transfer Learning Strip Steel Surface Defect Recognition Deep Neural Network Based on Few Samples and Imbalanced Datasets

May combine two claims.

The improved gray-scale projection algorithm can filter out defect-free surfaces, edge false defects, and illumination interference, and extract defects on the surface effectively.

“The algorithm can filter out defect-free surfaces, edge false defects, and illumination interference, and extract defects on the surface effectively.”

• Also stated in: § 3.4. Strip Steel Surface Rapid Quality Screening and Defect Feature Extraction Experiment · plain

  “It can be seen that our improved gray projection algorithm can quickly screen defect and defect-free surfaces, and extract defect features.”

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

For defect images with high brightness background or high reflective surface, the algorithm will have some degradation in detection accuracy.

“For the defect images with high brightness background or high reflective surface, the algorithm will have some degradation in detection accuracy.”

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

The mixed image dataset is a strip steel defect dataset with full annotation and class balance.

“This is a strip steel defect dataset with full annotation and class balance.”

May combine two claims.

The baseline VGG19 network's low learning rate resulted in slow convergence, falling to local optimal, large loss curve shocks on the testing dataset, and poor generalization.

“This resulted in slow convergence and falling to the local optimal value of the network, while the loss curve shocked greatly on the testing dataset, and achieved poor generalization.”

• AssumptionRests on an assumption: A more complex model is more able to fit training data in traditional machine learning algorithms.
• AssumptionRests on an assumption: If the training dataset is too small, the model will become over-fitted.

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

In the medium term, there were occasional increases in loss.

“In the medium term, there as occasional increases in loss.”

Classified as probably not a substantial claim.

The mixed image dataset was pre-processed to create a new class balanced dataset.

“The mixed image dataset was pre-processed to create a new class balanced dataset”

May combine two claims.

The improved VGG19 network performs slightly better than baseline VGG19 in six types of defects, but significantly better in surface seams defects.

“The improved VGG19 network performs slightly better than the baseline VGG19 in these six types of defects, but the improved VGG19 performs significantly better in the surface seams defects.”

• Also stated in: § 7. Conclusions · plain

  “The improved VGG19 network performs slightly better than the baseline VGG19 in six types of defects, but the improved VGG19 performs significantly better in the surface seams defects.”

May combine two claims.

The accuracy of baseline VGG19 and improved VGG19 on surface seams dataset are 70% and 90%, respectively.

“Therefore, the accuracy of the baseline VGG19 and the improved VGG19 model on surface seams dataset are 70% and 90%, respectively.”

May combine two claims.

Traditional detection algorithms can achieve high detection accuracy with fewer images, but their real-time performance is poor.

“As can be seen from Table 3, traditional detection algorithms can achieve high detection accuracy with fewer images, but their real-time performance is poor.”

May combine two claims.

Our network's real-time performance was weaker than M-pooling CNN and improved YOLO, but its detection speed of 54.6 FPS meets the basic requirements of online real-time detection.

“Its real-time performance was weaker than those two algorithms, but the detection speed of our network was 54.6 FPS, which means it meets the basic requirements of online real-time detection.”

• AssumptionRests on an assumption: The speed of a steel rolling production line is 10-30 m/s and the view field of a single camera is 50-100 cm.

May combine two claims.

The improved VGG19 network's convergence and accuracy were greatly improved compared to the basic network.

“Figure 15 shows that the convergence and accuracy of the network was greatly improved on the basic network.”

May combine two claims.

Our network's average detection time for a single image was 0.0183 s, and it can detect about 54 images per second.

“Our network's average detection time for a single image was 0.0183 s, and it can detect about 54 images per second.”

May combine two claims.

Our algorithm achieved 97.75% accuracy, 0.0183 s average detection per image (54.6 FPS), on 1850 samples.

“| Ours | DNN Classification | 97.75% | 0.0183 s 54.6 FPS | 1850 |”

May combine two claims.

The surface seams dataset has only 40 pictures and only 10 pictures for the testing dataset.

“However, the surface seams dataset has only 40 pieces of pictures and only 10 pieces of pictures for the testing dataset.”

May combine two claims.

The improved VGG19 model occasionally had a sudden increase in loss during training, but the verification loss decreased and tended to converge as training progressed.

“The model occasionally had a sudden increase in loss during the training period, but the verification function's loss decreased and tended to converge as the training progressed.”

May combine two claims.

For surface scratch, baseline VGG19 misidentified four pictures and improved VGG19 misidentified three pictures.

“For the defect categories of surface scratch, the baseline VGG19 misidentified four pieces of pictures and the improved VGG19 model misidentified three pieces of pictures.”

May combine two claims.

Our algorithm achieved 97.62% accuracy and 0.0192 s/52.1 FPS on the NEU database.

“| Ours | 97.62% | 0.0192 s/52.1 FPS |”

May combine two claims.

For inclusions, baseline VGG19 misidentified four pictures and improved VGG19 misidentified three pictures.

“For the defect categories of inclusions, the baseline VGG19 misidentified four pieces of pictures and the improved VGG19 model misidentified three pieces of pictures.”

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

For deep learning detection algorithms, detection accuracy increases with data volume and real-time performance is better than traditional algorithms.

“For deep learning detection algorithms (such as M-Pooling CNN [17], Improved YOLO [19]), their detection accuracy increases with data volume and their real-time performance is better than that of the traditional algorithms.”

• cites [17] Masci, J., Meier, U., Ciresan, D., Schmidhuber, J., and Fricout, G. (2012, January 10–15). Steel defect classification with Max-Pooling Convolutional Neural Networks. Proceedings of the 2012 International Joint Conference on Neural Networks (IJCNN), Brisbane, QLD, Australia. sourceNot checked yet

  Not checked in the free run.

• cites [19] Li, 2018, Real-time Detection of Steel Strip Surface Defects Based on Improved YOLO Detection Network, IFAC-Pap.Not checked yet

  Not checked in the free run.

May combine two claims.

For products surface defect detection carried on a conveyor belt, the human eye can detect only 60% of the surface defects, and the width of the product cannot exceed 2 m and the moving speed of the products cannot exceed 30 m/s, in the best case.

“For products surface defect detection carried on a conveyor belt, the human eye can detect only 60% of the surface defects, and the width of the product cannot exceed 2 m and, the moving speed of the products cannot exceed 30 m/s, in the best case [3].”

• cites [3] Srinivasan, 1992, FDAS: A Knowledge-based Framework for Analysis of Defects in Woven Textile Structures, J. Text. Inst. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The proposed ROI region random crop algorithm can extract details of sample characteristics and enlarge edge and fine trivial details in the ROI region.

“The industrial image data augmentation algorithm based on ROI region random crop in this paper can extract details of samples' characteristics and enlarge images' edge and other fine trivial details in the ROI region.”

May combine two claims.

This kind of model generalization on the testing dataset is poor, and it can lead the model to prefer a dataset with more samples or samples with a low identification degree.

“This kind of model generalization on the testing dataset is poor, it can lead the model to prefer a dataset with more samples or samples with a low identification degree [41].”

• cites [41] Springer Nature Editorial (2016). More accountability for big-data algorithms. Nature, 537, 449. sourceNot checked yet

  Not checked in the free run.

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

On the NEU database, ResNet50 with the MFN module achieves a 99.7% detection rate with a speed of 6.1 FPS.

“On the NEU database, ResNet50 with the MFN module achieves a 99.7% detection rate with a speed of 6.1 FPS.”

• cites [22] He, 2020, An End-to-End Steel Surface Defect Detection Approach via Fusing Multiple Hierarchical Features, IEEE Trans. Instrum. Meas. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The category imbalance problem is mainly reflected in two aspects: positive and negative samples imbalance, with the positive and negative samples ratio reaching 1:100, and uneven difficulty samples.

“The category imbalance problem is mainly reflected in two aspects: positive and negative samples imbalance (the positive and negative samples ratio reached 1:100) and uneven difficulty samples (simple samples leading to loss function).”

May combine two claims.

A CCD camera captures the edge region and background image simultaneously during image acquisition, which introduces false edge defects.

“Therefore, in order to ensure full coverage of the strip steel surface, a CCD camera captures the edge region and background image simultaneously during the image acquisition process.However, it will introduce false edge defects”

May combine two claims.

Compared with VGG19, VGG11, VGG13 and VGG16 are shallow networks.

“Compared with VGG19 and, VGG11, VGG13 and VGG16 are shallow networks.”

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

Due to the large number of parameters in the full connection layer, three full connection layers will greatly slow down the efficiency of network training and recognition.

“However, due to the large number of parameters in the full connection layer, three full connection layers will greatly slow down the efficiency of network training and recognition.”

May combine two claims.

GAN-based image data augmentation algorithms are mainly used for relatively few samples datasets with high-resolution, but they are not applied to few samples datasets images with low-resolution on an industrial scale.

“These kinds of algorithms are mainly used for relatively few samples datasets with high-resolution, but they are not applied to few samples datasets images with low-resolution on an industrial scale [44].”

• cites [44] Zhao, Z., Li, B., and Dong, R. (2018, January 28–31). A Surface Defect Detection Method Based on Positive Samples. Proceedings of the 5th Pacific Rim International Conference on Artificial Intelligence (PRICAI), Nanjing, China. sourceNot checked yet

  Not checked in the free run.

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

A crack detection method based on a deep fully convolutional network with VGG16 backbones achieved about 90% average precision on a subset of 500 annotated 227 × 227-pixel crack-labeled images.

“The network achieved about 90% average precision on a subset of 500 annotated 227 × 227-pixel crack-labeled images.”

• cites [23] Dung, 2018, Autonomous concrete crack detection using deep fully convolutional neural network, Autom. Constr. sourceNot checked yet

  Not checked in the free run.

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

The cost matrix of algorithm level approach methods needs to be provided by experts in the field based on prior knowledge of the task, which in many real-world problems is not achievable.

“The cost matrix of these algorithm needs to be provided by experts in the field based on prior knowledge of the task, which in many real-world problems is not achievable.”

May combine two claims.

Although high-level information of different images is different, underlying features are similar, which is the basis of image data transfer learning.

“Although high-level information of different images is different, underlying features are similar [20].This is the basis of image data transfer learning.”

• cites [20] Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., and Brendel, W. (2019, January 6–9). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. Proceedings of the International Conference on Learning Representations (ICLR 2019), New Orleans, LA, USA.Not checked yet

  Not checked in the free run.

May combine two claims.

The proposed ROI region random crop algorithm is a data-level approach that eases the uneven difficulty samples problem to a certain extent and allows difficult samples to be identified by local details.

“This is a data level approach method to ease the uneven difficulty samples problem to a certain extent, and allow difficult samples to be identified by local details.”

May combine two claims.

Rapid screening of defect-free surfaces can greatly decrease the amount of calculation required and improve detection efficiency.

“Therefore, rapid screening of defect-free surfaces can not only greatly decrease the amount of calculation required, but also improved detection efficiency.”

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

In the distribution of strip steel defects, the edge portion is a defect-prone region of the strip steel surface.

“In the distribution of strip steel defects, the edge portion is a defect-prone region of the strip steel surface.”

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

Data level approach methods rely on a well-defined distance measure, but industrial datasets often contain multiple non-linear features or missing values.

“These methods rely on a well-defined distance measure, but industrial datasets often contain multiple non-linear features or missing values.”

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

The proposed ROI region random cropping data augmentation algorithm can generate high-quality industrial images from original image datasets on the actual images demanded.

“This algorithm can generate high-quality industrial images from original image datasets on the actual images demanded.”

May combine two claims.

The proposed ROI region random cropping data augmentation algorithm can solve the category imbalance problem using a data-level approach and can be combined with traditional image data augmentation techniques to ease few-sample dataset problems to a certain extent.

“It can not only solve the category imbalance problem using the data level approach method but also can be combined with traditional image data augmentation techniques to ease the problems of few samples datasets to a certain extent.”

May combine two claims.

RAdam has the advantages of both Adam and SGD, ensuring fast convergence and avoiding local optimal, as its convergence result is insensitive to the network's initial value learning rate.

“RAdam has the advantages of both Adam and SGD, which can not only ensure fast convergence speed but also avoid dropping into the local optimal, as the convergence result is insensitive to the network's initial value learning rate.”

• cites [46] Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., and Han, J. (2020, January 26–30). On the Variance of the Adaptive Learning Rate and Beyond. Proceedings of the International Conference on Learning Representations (ICLR 2020), Addis Ababa, Ethiopia.Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Transfer learning can be successfully applied using image data from an entirely different domain.

“Author [21] showed that transfer learning can be successfully applied using image data from an entirely different domain.”

• cites [21] 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. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The improved gray projection algorithm can save time on manual image labeling and provide high-quality defect data for the subsequent ROI image augmentation algorithm.

“The algorithm can save time on manual image labeling, and provide high-quality defect data for the subsequent ROI image augmentation algorithm.”

Classified as probably not a substantial claim.

For few samples datasets with low-quality images, it is not enough to support the training of the GAN network.

“However, for few samples datasets (each type of image only has a dozen images or less, each image is a grayscale image of about 30~50 KB) with low-quality images, it is not enough to support the training of the GAN network.”

May combine two claims.

The unstable loss curve of the baseline VGG19 indicates unreasonable network design and network structure, needing further improvement.

“the unstable loss curve obviously has some unreasonable network design and network structure, needing further improvement.”

• AssumptionRests on an assumption: A more complex model is more able to fit training data in traditional machine learning algorithms.
• AssumptionRests on an assumption: If the training dataset is too small, the model will become over-fitted.

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

The background subtraction method in actual production suffers from the light environment, camera hardware performance and image blurring, and a large number of false defect images are detected with this method.

“However, the background subtraction method in actual production suffers from the light environment, camera hardware performance and image blurring; a large number of false defect images are detected with this method.”

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

The data level approach method and algorithm level approach method must be combined to deal with difficult samples and abnormal samples using defect detection and classification processes.

“The data level approach method and algorithm level approach method must be combined to deal with difficult samples and abnormal samples using defect detection and classification processes.”

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

Ensemble learning does not always lead to a better performance.

“However, ensemble learning does not always lead to a better performance.”

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

To define a reasonable distance measure on these datasets is very difficult, and requires tremendous extra computing resources to meet real-time requirements.

“To define a reasonable distance measure on these datasets is very difficult, and requires tremendous extra computing resources to meet real-time requirements.”

May combine two claims.

The improved VGG19 model's generalization and robustness were greatly improved.

“the generalization and robustness had been greatly improved”

May combine two claims.

Cracks, scabs, and rolled in scales defects have obvious features and are easy for the model to recognize.

“These three types of defects have obvious features and are easy for the model to recognize.”

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

When there is a similarity between a new defect form and a known defect form, the model will make a misjudgment, which will greatly reduce the accuracy of defect recognition.

“When there is a similarity between a new defect form and a known defect form, the model will make a misjudgment, which will greatly reduce the accuracy of defect recognition.”

• AssumptionRests on an assumption: With continued production line developments, various kinds of new defect types will be created.

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

A model with a certain complexity can achieve high accuracy and recognition speed and prevent non-convergence and over-fitting.

“It is clear that a model with a certain complexity can not only achieve high accuracy and recognition speed, but also prevent the problem of non-convergence and over-fitting.”

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

The proposed method is a non-contact and nondestructive machine vision detection method with high detection rate, strong real-time performance, and perfect anti-interference.

“It is a non-contact and nondestructive machine vision detection method that has the advantages of a high detection rate, strong real-time performance, and perfect anti-interference.”

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

The model in Author's work [48] achieved an accuracy of 98.1% and a recognition speed of 476.2 FPS (2.1 ms), using a NVIDIA GeForce GTX 1080Ti.

“Although the model in Author's work [48] achieved an accuracy of 98.1% and a recognition speed of 476.2 FPS (2.1 ms), a NVIDIA GeForce GTX 1080Ti was used to train and test the neural network.”

• cites [48] Kostenetskiy, 2019, Real-time system for automatic cold strip surface defect detection, FME Trans. sourceNot checked yet

  Not checked in the free run.

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

The model in Author's work [48] is simple in structure and easy to build, but its original data were pre-processed by the normalization Z-Score function.

“Although the model is simple in structure and easy to build, the original data were pre-processed by the normalization Z-Score function.”

• cites [48] Kostenetskiy, 2019, Real-time system for automatic cold strip surface defect detection, FME Trans. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The experimental results show that the improved VGG19 network has high practicability and reliability and has practical value for strip steel defects detection.

“The experimental results show that the improved VGG19 network proposed in this paper has high practicability and reliability; it has a certain practical value for strip steel defects detection.”

May combine two claims.

The model in Author's work [48] is not an end-to-end model, and the time to pre-process the original data was much greater than the time for the network to recognize defects.

“It is not an end-to-end model, and the time to pre-process the original data was much greater than the time for the network to recognize defects.”

• cites [48] Kostenetskiy, 2019, Real-time system for automatic cold strip surface defect detection, FME Trans. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A very deep network is not really required for online real-time defect classification; detection accuracy, speed and complexity need to be balanced.

“As can be seen from Table 4, a very deep network is not really required for the online real-time defect classification task; the detection accuracy, speed and complexity of the network need to be balanced.”

May combine two claims.

The proposed improved VGG19 transfer learning network has strong generalization and good convergence.

“In summary, this paper proposed An Improved VGG19 Transfer Learning Strip Steel Surface Defect Recognition Deep Neural Network Based on Few Samples and Imbalanced Datasets, which has a strong generalization and good convergence.”