TSDNet: A New Multiscale Texture Surface Defect Detection Model^†

Round 1

Reviewer 1 Report

This paper proposes a texture surface defect detection method based on convolutional neural network and wavelet analysis. This is an interesting research topic. However, there are still many places in the paper that need further modification, and the specific suggestions are as follows.

1. The abstract of the paper needs to be further simplified and refined to highlight the significance of the research.

2. This paper gives an overview of the current research, but does not cite and analyze the latest research papers. It is suggested that the author cite some of the latest literature on fault detection, such as Hybrid robust convolutional autoencoder for unsupervised anomaly detection of machine tools under noises, Novel motor fault detection scheme based on one-class tensor hyperdisk and so on.

3. Experiments need more comparison with current advanced methods. At the same time, the conclusion should be clarified more clearly.

Author Response

Point 1: The abstract of the paper needs to be further simplified and refined to highlight the significance of the research.

Response 1: Thanks for your feedback. We optimized the grammar of the abstract and removed detailed data indicators to make it more concise and emphasize the importance of our study. We hope this meets your expectations.

Point 2: This paper gives an overview of the current research, but does not cite and analyze the latest research papers.

Response 2: Thanks for your advice. Because of the early completion of this article, citations and analyzes of more recent research were not included. In this revision we added recent research results and analyzed and summarized them. Such as Hybrid robust convolutional autoencoder for unsupervised anomaly detection of machine tools under noises (2023), Novel motor fault detection scheme based on one-class tensor hyperdisk (2023) and Rethinking unsupervised texture defect detection using PCA (2023), etc.

Point 3: Experiments need more comparison with current advanced methods. At the same time, the conclusion should be clarified more clearly.

Response 3: Thank you for your suggestion. We looked for the latest research results on the DAGM2007 dataset (Rethinking unsupervised texture defect detection using PCA (2023) and A Four-Stage Product Appearance Defect Detection Method With Small Samples (2022), and compared with our method.

Author Response File: Author Response.docx

Reviewer 2 Report

I think that the presented work is interesting e and may be useful to those who are studying the possibility of identifying surface defects, to develop useful systems for industry.

There are just some minor misspellings.   1. What is the main question addressed by the research? This paper proposes a texture surface defect detection method based on convolutional neural network and wavelet analysis.   2. Do you consider the topic original or relevant in the field? Does it address a specific gap in the field? I think that the research topic is interesting, that could be useful for industrial applications, and that the presented results are encouraging.   3. What does it add to the subject area compared with other published material? The work combines known techniques and improves from the point of view of processing time, existing methods. Although the method is well explained, it should be better emphasized in the experimental part what is the advantage of using wavelets.     4. What specific improvements should the authors consider regarding the methodology? What further controls should be considered? The authors compare their method with existing methods, making use of shared image databases.   5. Are the conclusions consistent with the evidence and arguments presented and do they address the main question posed? Conclusions are consistent.   6. Are the references appropriate? Maybe, an update of the bibliography might be opportune: the last paper cited for algorithms is from 2021 anyway.     7. Please include any additional comments on the tables and figures. Tables and figures are appropriate.  

Author Response

Point 1: Although the method is well explained, it should be better emphasized in the experimental part what is the advantage of using wavelets.

Response 1: Thanks for your feedback. We use wavelet transform mainly to compress images，The low-frequency part obtained after image wavelet transform is the compression of the original image, which is only a quarter of the size of the original image, but the defects in the original image are still visible. Using small images can speed up model training and save resources. We will re-emphasize the advantages of using wavelet transforms in the experimental section.

Point 2: Maybe, an update of the bibliography might be opportune: the last paper cited for algorithms is from 2021 anyway.

Response 2: Thanks for your advice. Because of the early completion of this article, citations and analyzes of more recent research were not included. In this revision we added recent research results and analyzed and summarized them. Such as Hybrid robust convolutional autoencoder for unsupervised anomaly detection of machine tools under noises (2023), Novel motor fault detection scheme based on one-class tensor hyperdisk (2023) and Rethinking unsupervised texture defect detection using PCA (2023), etc.

Author Response File: Author Response.docx

Reviewer 3 Report

Authors propose a defect detection model using wavelets and cnns. Literature research is weak. Some important reviews papers are not examined. The proposed methodology is weak and simple. The mathematical contribution of the method is not understood. There are superficial and simple explanations. The texture analysis capability of the wavelet transform is not well reflected. The deep learning model part is quite simple and superficial.

Author Response

Point 1: Literature research is weak. Some important reviews papers are not examined.

Response 1: Thanks for your advice. Because of the early completion of this article, citations and analyzes of more recent research were not included. In this revision we added recent research results and analyzed and summarized them. Such as Hybrid robust convolutional autoencoder for unsupervised anomaly detection of machine tools under noises (2023), Novel motor fault detection scheme based on one-class tensor hyperdisk (2023) and Rethinking unsupervised texture defect detection using PCA (2023), etc.

Point 2: The proposed methodology is weak and simple. The deep learning model part is quite simple and superficial.

Response 2: Our contribution is mainly to use the wavelet transform to preprocess images, use the random window to quickly extract defect images, and use the sliding window to stably judge defects. We did not design carefully for the model, but adopted the most basic convolutional neural network , to highlight the role of other methods. In future research, we will design the network structure for patches’ feature to further improve the accuracy.

Point 3: The texture analysis capability of the wavelet transform is not well reflected.

Response 3: Thanks for your feedback. We use wavelet transform mainly to compress images，The low-frequency part obtained after image wavelet transform is the compression of the original image, which is only a quarter of the size of the original image, but the defects in the original image are still visible. Using small images can speed up model training and save resources. We will re-emphasize the function of using wavelet transforms in the experimental section.

Author Response File: Author Response.docx