Image Enhancement of Steel Plate Defects Based on Generative Adversarial Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a particular but interesting problem which is the detection and classification of surface defects of steel sheets under conditions of small sample sizes. In reality, it is about the limited number of data samples that poses a problem.

To deal with this difficulty, the authors propose a bidirectional attention

mechanism (DAPT), specifically designed to improve the capacity of a classification model by identifying weak defects. The model in question is based on generative adversarial networks (GAN).

The experiment, which combines the proposed DAPT mechanism with a modified U-Net discriminator, demonstrates an improvement in accuracy from 2.7% to 9% (Figure 7) and consequently a better classification of surface defects on steel plates.

However, on the technical and editorial presentation, we raised certain questions:

1- Figures 1 and 2 must be very explicit and well explained to allow the reader to understand the proposed method, for example:

- How can we explain the calculation of y3 of formulas 1) and 2) and that of y2 then?

- we are talking about the LeakyReLU activation function while, in the diagram, there are only Sigmoids.

- there is a lack of a general schema (or architecture) integrating the different components (DAPT, U-Net, and the classification part).

2- Problem with the numbering of figures: no Figure 6 but Figure 4 is duplicated)

3- Sometimes some sentences are too long(e.g the last sentence of the introduction)

4- Some parameters (labels) (tables 1 and 2, figures 3 and 4) such as SC and PS are not described in the text. It is necessary to emphasize the importance of each parameter in the defect of metal surfaces.

5- Figure 7 presents results on the single GC10-DET dataset (high resolution). Why not on the other NEU dataset? Does the resolution have a determining role on mAP?

6- In the introduction, the authors mainly cite traditional deep-learning classification work in the field (Steel Plate Defects) but not those based on GANs for example:

-“Steel Surface Defect Detection Using GAN and One-Class Classifier”, Kun Liu; Aimei Li

-”Strip Steel Defect Classification Using the Improved GAN and EfficientNet”, Shengqi Guan

7. The background must discuss those works and the difference must be noted in the proposal, as well as the result comparisons.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This study proposes a two-way attention mechanism tailored to enhance the model's capacity in detecting subtle defects and refining the network discriminator's architecture. Experimental results confirm its superiority over several related approaches. However, there remain minor issues to be addressed. Firstly, this work needs to introduce comparative analyses delineating the relationships and distinctions with works such as "Learning Semantically Enhanced Feature for Fine-grained Image Classification," "Adaptive Noise Dictionary Construction via IRRPCA for Face Recognition," and "Three-dimensional softmax mechanism guided bidirectional GRU networks for hyperspectral remote sensing image classification." Secondly, this work needs to emphasize the necessity for ablation analyses to further elucidate the method's contributions.

Comments on the Quality of English Language

Minor editing of English language required

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I reviewed the corrected version. Overall, the authors responded and modified the majority of questions.

Globally, the current version can be published.

But, If possible:

Despite the authors' argument, I think it is interesting to cite in 2 or 3 sentences (Discussion) the results (at least the accuracy) of certain works in the field to appreciate the contribution better.