Lightweight Implicit Blur Kernel Estimation Network for Blind Image Super-Resolution

Round 1

Reviewer 1 Report

The paper presents a method for Blind single image super resolution. the paper is well written and the methodology is solid. The section of the related work is well detailed and exhaustive. 

The methodology is rigorous and the provided experimental validation confirms the viability of the proposed approach. 

As a minor remark, in the methodology section, I have the impression that the presentation of the loss functions should be done BEFORE the final loss equation, as the reader does not know what the single terms mean, otherwise. 

 

At line 203 the figure is cited without the terms "Figure 3 (c)", but only as 3(c). 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper has described an image super-resolution architecture to be used under unknown degradation (blind) scenarios. The paper targets an interesting research topic. However, there are certain shortcomings in the manuscript which need to be addressed.

1-     Two different contributions related to blur kernel estimation have been listed on lines 75 and 77. Please elaborate the difference between these two and various previous works which have also estimated blur kernels.

2-     It is advisable to keep figure captions short. E.g. figure 2 caption is too long and describes the whole operation. It’s better to keep the description in the main text.

3-     In the introduction it is stated that “the Estimator predicts a blur kernel k from low-resolution image LR input, which the Super Resolver uses to restore the SR image”. however, in both figures 1c and figure 2, SR and k are independently estimated i.e. k is not used to restore SR. Please elaborate what is the use of estimating kernel k separately when it is not actually required to estimate SR?

4-     It has been claimed that proposed network has fewer number of parameters than the state-of-the-art works. However, a quantitative measure has not been provided. E.g. the execution times on PC and embedded systems. In the abstract, it has been mentioned that the proposed technique is suitable for embedded systems. However, the text has not substantiated this claim with any experimental results.

5-     Using fewer parameters for SR estimation has been originally claimed by ref 31. The current woks has simply fine-tuned this network. Moreover, the generator and discriminator structures shown in fig. 3 have been verbatim copied from ref 31. Please cite these properly and provide a copyright waiver if applicable.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this manuscript, the authors propose a lightweight blur kernel estimation approach for blind single-image super-resolution (SISR) based on a deep convolutional neural network (CNN) and a deep super-resolution residual convolutional generative adversarial network. The results are proved compelling through comparative experiments with some state-of-the-art methods. But the writing of the work needs to be standardized, the experiment is not enough, and there is still a lot of room for improvement. My recommendation for this paper is “Major revision”. Some comments are as follows.

(1)   The modules in Figures 2 and 3 are too small to be clear. Please readjust the layout to make it clearer.

(2)   The data sets selected in Seting1 and Seting2 are different. Explain why. Moreover, we have noticed that the range of scale factors 4 and scale factors 2 are different in Seting1. Please explain how they are defined.

(3)   It is suggested to add more fish SR methods especially blind image SR methods in Table 1 and Table 2 to demonstrate the superiority of the proposed methods.

(4)   It is suggested that the author introduce Multi-Adds or FLOPS as indicators to measure the computational performance of the model.

(5)   The effects of the unknown blur kernels can be explored further in ablation experiments.

(6)   The paper has some typos and misnomers. The authors are suggested to do the proofreading carefully.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have adequately addressed the concerns raised in the previous review cycle. 

Author Response

Thanks

Reviewer 3 Report

Please provide the file of response to the reviewer.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report

My recommendation for this paper is accept.