PCB Defect Images Super-Resolution Reconstruction Based on Improved SRGAN
Round 1
Reviewer 1 Report
Authors presented an interesting proposal. I believe it is an interesting contribution, but I have several points to be improved to its final version.
-
It is necessary to correct some typos, and to make extensive language and logic consistency checking.
-
In my opinion SRVIT can be understood as an improvement of SRGAN. Thus, I believe it is necessary to present an empirical ablation in order to highlight the individual performance contributions of the improved discriminator network and the improved generator network in comparison to the ones employed in SRGAN.
-
In a not so exhausting search I’ve found the following references which address the use of transformers in super resolution:
https://arxiv.org/pdf/2108.07597.pdf
https://www.hindawi.com/journals/bmri/2022/4431536/
https://arxiv.org/pdf/2108.11084.pdf
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10007518/
https://www.mdpi.com/1424-8220/22/21/8126
-
So, it is necessary expanding your bibliographic review and perhaps including such proposals in the experiments
The paper contains some typos and have some sections that should be carefully rewritten.
Author Response
Dear reviewer,
Re: Manuscript ID: applsci-2391307 and Title: PCB defect images super resolution reconstruction based on improved SRGAN
Thanks for your comments concerning our manuscript. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the paper. The responds to the reviewer’s comments are as following:
- Comment: It is necessary to correct some typos, and to make extensive language and logic consistency checking.
Response: Typos and some grammatical errors have been corrected.
- Comment: In my opinion SRVIT can be understood as an improvement of SRGAN. Thus, I believe it is necessary to present an empirical ablation in order to highlight the individual performance contributions of the improved discriminator network and the improved generator network in comparison to the ones employed in SRGAN.
Response: Dear teacher, I would like to explain the reason for not conducting the generator and discriminator ablation experiments. If we consider the GAN network as a whole, we may not be able to achieve good results by improving the generator or discriminator individually. The number of parameters in the generator is not reduced as effectively as in the discriminator. So it is more convincing to combine the generator and the discriminator for comparison. In addition, the effect of combining the VIT and HFB and IRB networks with each other in the net clusters looks more obvious, and the picture reconstruction effect or lightness cannot be well demonstrated by using only one of the networks.
- Comment: In a not so exhausting search I’ve found the following references which address the use of transformers in super resolution.
Response: Thank you very much for the references, they have been added to the paper.
- Comment: So, it is necessary expanding your bibliographic review and perhaps including such proposals in the experiments.
Response: References and new comparative experimental analyses are added to the paper.
Reviewer 2 Report
The paper presents an original algorithm in the field of super-resolution image reconstruction. The proposed SRVIT algorithm can be used for detecting PCB defects from high resolution pictures. The authors obtained remarkable results (PSNR, SSIM, spatial complexity) by integrating VIT into SRGAN in one improved super resolution reconstruction algorithm.
The scientific level of the paper is good, and it may be accepted after the minor corrections suggested to the authors. In this sense here there are few questions and remarks for authors:
1. Please include, in the text, the meaning of MOS abbreviation
2. Line 152: Please put: Where “a” represents the value mapped…. instead of Where a represents the value mapped…
3. When using different models, are the PSNR and SSIM metrics enough to appreciate and compare the quality of PCB reconstructed images?
Author Response
Dear reviewer,
Re: Manuscript ID: applsci-2391307 and Title: PCB defect images super resolution reconstruction based on improved SRGAN
Thanks for your comments concerning our manuscript. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the paper. The responds to the reviewer’s comments are as following:
- Comment: Please include, in the text, the meaning of MOS abbreviation
Response: Thank you for your comments. MOS means Mean Opinion Score and is based on the human eye of the observer scoring the quality of the image, which to some extent reflects the true visual effect of the image. In this paper, 40 image evaluators were selected and each evaluator scored all the reconstructed images of the model with a score of 5 for HR images and between 1 and 5 for reconstructed images, and the final MOS value was calculated by averaging the scores of all the evaluators.
- Comment: Line 152: Please put: Where “a”represents the value mapped…. instead of Where a represents the value mapped…
Response: Already revised in the paper.
- Comment: When using different models, are the PSNR and SSIM metrics enough to appreciate and compare the quality of PCB reconstructed images?
Response: PSNR and SSIM are the main evaluation criteria in the field of image overscoring, PSNR is expressed as a discriminatory criterion based on the pixel level in measuring image quality and reflects the degree of image degradation. There are strong correlations between pixels in natural images, and these correlations often contain important information. detecting transformations of structural information can perceive the degree of distortion and thus measure the phase speed of the original and reconstructed images. the SSIM evaluation criterion consists of three components: brightness, contrast and structure. In addition subjective evaluation metrics are included in the text to measure the different models.
Reviewer 3 Report
Please use formal standard flow chart shape. (As the True/False in Fig. 1 should not be a box shape )
Please explain more on the QKV space and the correlation between Q and K.
Is this approach applicable to an image with uneven lighting condition ? or can color/texture recognition help?
For Eq. (4), where is 10^(-3) from ?
Need some major English revise. (Like Line 122 to Line 130 as well as many other paragraph)
If a sentence is too long, readers have trouble to understand. As in Line 122 to Line 130, that sentence has 9 lines and may capital words. Hard to read. This is happened many places in your paper. It is a good paper.
Author Response
Dear reviewer,
Re: Manuscript ID: applsci-2391307 and Title: PCB defect images super resolution reconstruction based on improved SRGAN
Thanks for your comments concerning our manuscript. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in red in the paper. The responds to the reviewer’s comments are as following:
- Comment: Please use formal standard flow chart shape. (As the True/False in Fig. 1 should not be a box shape )
Response: Thank you for your correction, it has been corrected in the paper.
- Comment: Please explain more on the QKV space and the correlation between Q and K.
Response: The self-attentive mechanism for capturing global information relies on a modelling approach consisting of three elements: query, key and value. q represents the query, k represents the key and v represents the value. QKT represents the inner product of two feature matrices. softmax represents the normalization operation such that dK is the dimension of the key and the elements of the normalization function are divided by with a variance of 1. This allows the normalized distribution to fluctuate to the extent that d is decoupled so that the gradient values remain stable during training. the normalized distribution to be decoupled from d so that the gradient values remain stable during training.
- Comment: Is this approach applicable to an image with uneven lighting condition ? or can color/texture recognition help?
Response: Dear teacher, as most of the PCB next time images are produced by the plant output, they are generally taken away from light and the bare board is green. Will explore related work in future research work.
- Comment: For Eq. (4), where is 10^(-3) from
Response: The loss function is divided into content loss and adversarial loss, and 10-3 is the weighting factor for adversarial loss
- Comment: Need some major English revise. (Like Line 122 to Line 130 as well as many other paragraph)
Response: Adjustments have been made in the paper.
Round 2
Reviewer 1 Report
I recommend that this paper be accepted in its current form.