Stain Defect Classification by Gabor Filter and Dual-Stream Convolutional Neural Network

Round 1

Reviewer 1 Report

The article proposes a classification network that integrates Gabor filters. The article is well-organized and written, but there are a few minor issues:

1) Are there any other image pre-processing methods besides Gabor filters that could be used to compare the model's performance?

2) Some relevant methods, such as "Combining Prior Knowledge With CNN for Weak Scratch Inspection of Optical Components" should be mentioned in the introduction. (Reference: IEEE Transactions on Instrumentation and Measurement, 2020, 70: 1-11.)

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

1. According to the evaluation indicators in Table 3 and Table 6, the dual-stream network is even worse than the single-stream network in most cases. What is the reason for this? Whether it is consistent with the main innovations described by the author.

2. Model and training parameter selection are crucial to network performance, but the author does not mention it in the paper, and it is suggested to add.

3. For the data set (small sample data set) used in this paper, the author chooses Resnet50 as the backbone network to design the single-stream and dual-stream networks, but whether Resnet50 is the best choice should be clarified by the author.

4. In Section 3.1, the author describes "the smaller the IDM value, the image has a relatively clear boundary of the stain defect",but in Section 3.2," N images with the highest IDM value are selected ", Why?

 

5. The resolution of FIG. 8 and FIG. 9 does not seem to meet the requirements of the journal.

6. Repeat the spelling of "Google GoogLeNet" in line 185.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

In this paper, the authors proposes a CNN for stain defect classification using inspection images and Gabor filter images to improve the performance of defect classification. IDM value was used to select the Gabor filter images and features were extracted by a Dual-stream CNN structure. 

(1) In Figure 2 and Figure 3, is there a best value for  ?

(2) Please note the training platform you use in the paper.

(3) Table 3 and Table 6 show that it seems Single-stream network with inputting Gabor images has a high accuracy, so what are the advantages of Dual-stream network in specific deployment?

(4) The size and training time of Single-stream and Dual-stream models should be comprehensively considered in the table3 and table 6.

(5) Figures 8 and 9 are not very clear, please upload clearer Figures.

(6) The Authors should discuss/resume the limitations/disadvantages of the proposed solution.

(7) How to choose configurations for the proposed model?

Author Response

Please see the attachment

Author Response File: Author Response.docx