A Fabric Defect Segmentation Model Based on Improved Swin-Unet with Gabor Filter
, Chengming Liu 1,†,‡, Shuya Duan 1,†,‡, Liangpin Ren 1,*,†,‡, Guozhen Cheng 2 and Bing Hao 3Round 1
Reviewer 1 Report
1. More citations can be added to the paragraph 3 of the introduction to back what was mentioned.
2. In the literature review section, some paragraphs are 3 lines, please re structure it.
3. The sensitivity of the models as observed was a bit low compared to the accuracy, can you explain that.
4. The future work section can be extended and be more precise.
5. The abstract should have a numerical conclusion.
To be improvred
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
-Reference --Include more relevant and recent sources --Organize the references section alphabetically --Update some references
-Minor editing of English language required
Some minor edits could be made to improve the clarity of the text, such as adding a comma after "algorithm" and rephrasing "traditional operators" to be more specific.
Author Response
We sincerely thank you to the reviewer for your careful reading and recognition of our work. Here is our answer to your question:
There are many inaccuracies and missing annotations in the original dataset. Failure to filter will result in the model being unable to learn correctly. In addition, we filtered the dataset before dividing it into a training set and a validation set. This is not a specific selection of the dataset based on the model.
In addition, in the original AITEX dataset, the image resolution is 4096*256. During the process of cutting it to 256*256, we found that there are a large number of samples without defects. Adding fabric images without defects to the training set will make the already imbalanced positive and negative samples even more imbalanced. Among them, there are 105 pictures of cloth with defects before image cutting. After cutting, there should be 16*105=1680 pictures. After selecting those cloth pictures with defects, there are 184 pictures. The selection and cropping work we do on the data set is just to make the training process better and prevent annotation errors or flawless pictures in the dataset from affecting the learning of the model.
Author Response File: Author Response.docx
Reviewer 3 Report
The paper reads well and appears to outperform the state of the art. The contruibutions that are claimed are :
1. A Transformer-based fabric defect segmentation network is proposed to solve the local optimal fitting problem of the previous transformer model.
2. Embed traditional spectral methods into the model, and use fixed Gabor filter to eliminate the influence of complex textured backgrounds on defect segmentation.
3. They have designed a U-shaped architecture that is not completely symmetrical to make model training easier. Meanwhile, multi-stage result fusion is proposed for precise locating defects.
4. A new loss function is proposed to solve the imbalance problem between defect samples and non-defect samples. The loss function proposed in this paper uses a combination of weighted Focal loss and Dice loss to improve the convergence speed and detection accuracy of the model
The only question or issue I have is with the datasets. It appears that a selection is made of images from those datasets and also cropping etc. How does this influence the results, since due to the selection the algorithms work better than other algorithms. So this should be explained what the influence is.
No special some proofreading perhaps
Author Response
We sincerely thank you to the reviewer for your careful reading and recognition of our work. Here is our answer to your question:
There are many inaccuracies and missing annotations in the original dataset. Failure to filter will result in the model being unable to learn correctly. In addition, we filtered the dataset before dividing it into a training set and a validation set. This is not a specific selection of the dataset based on the model.
In addition, in the original AITEX dataset, the image resolution is 4096*256. During the process of cutting it to 256*256, we found that there are a large number of samples without defects. Adding fabric images without defects to the training set will make the already imbalanced positive and negative samples even more imbalanced. Among them, there are 105 pictures of cloth with defects before image cutting. After cutting, there should be 16*105=1680 pictures. After selecting those cloth pictures with defects, there are 184 pictures. The selection and cropping work we do on the data set is just to make the training process better and prevent annotation errors or flawless pictures in the dataset from affecting the learning of the model.
Author Response File: Author Response.docx
