Blind Image Super Resolution Using Deep Unsupervised Learning
Round 1
Reviewer 1 Report
The manuscript “Blind Image Super-Resolution using Deep Unsupervised Learning” is interesting and well written. However, I am writing some concerns and suggestions for improvement of the manuscript before final acceptance.
1. Section 2: Please summarize the major shortcomings of existing studies with a short paragraph.
2. Section 3: Please discuss the hyperparameter selection for the proposed technique.
3. Section 4: Please provide the details about the simulation platform for the proposed algorithm.
4. Section 4: Shortly describe the main contents of each dataset (including features and classes)
5. Section 4: Details about preprocessing of datasets and feature selection is missing.
6. Section 4: In comparison with state of the art, authors only considered 3 studies only please add some more latest studies to prove the effectiveness of the proposed scheme.
7. Add an additional section before the conclusion to discuss the potential key challenges and future directions.
Author Response
We appreciate the expert reviewer’s helpful comments and constructive suggestions. The manuscript has been significantly improved based on reviewers’ comments and suggestions. We will supply our detailed point-by-point responses to the reviewer. The manuscript has also been revised as suggested by the reviewer (marked in red for the required revisions).
- Please summarize the major shortcomings of existing studies with a short paragraph (The first comment of Review #1)
Response: Thanks very much for pointing this issue. In the revised manuscript, we summarize the limitations of the existing methods in lines 211-230 of Page 5.
- Section 3: Please discuss the hyperparameter selection for the proposed technique. (The second comment of Review #1)
Response: Thanks for the expert reviewer’s comment. In section 3, there is only one hyper-parameter: the perturbation degree $\beta$. We added some explanation about the hyper-parameter value to be usually set from 0.01 to 0.08 in lines 302 to 306 of page 8 in the revised manuscript. We also give some compared experimental results with different values of $\beta$ on the Set5 dataset in Table 3, which manifests there are no large effect on the SR performance with different values of $\beta$.
- Section 4: Please provide the details about the simulation platform for the proposed algorithm. (The third comment of Review #1)
Response: Thanks for the expert reviewer’s comment. In the original manuscript, we described our proposed algorithm was implemented using Pytorch (Line 353). In the revised manuscript, we further provided the information of the used computer on lines 358 to 360 of page 10.
- Section 4: Shortly describe the main contents of each dataset (including features and classes) (The fourth comment of Review #1)
Response: Thanks for carefully checking our manuscript. We provided the detailed description of all used dataset on lines 335 to 342 of page 9 in the revised manuscript.
- Section 4: Details about preprocessing of datasets and feature selection is missing. (The fifth comment of Review #1)
Response: Thanks for the expert reviewer’s comment. We added the pre-processing for the images of all used dataset on lines 342 to 346 of page 9 in the revised manuscript. However, there is no feature section procedure in our proposed BSR-DUL method.
- Section 4: In comparison with state of the art, authors only considered 3 studies only please add some more latest studies to prove the effectiveness of the proposed scheme. (The sixth comment of Review #1)
Response: Thanks very much for pointing this issue. Since our proposed method is a complete blind and unsupervised method for being flexibly adapted to the LR observation captured under uncontrolled conditions, which is more challenge than the conditions used in conventional SR task. It would be unfair to naively compared our proposed method with the dominated deep learning models in a fully-supervised way. Thus, we did not give a lot of comparisons with the methods belonging to the similar research paradigm, and only conducted comparisons with the representative methods in different research directions. For example, we provided the compared results with the representative deep supervised non-blind methods: EDSR and LapSRN, the deep unsupervised non-blind methods: DIP and ZSSR (The new added ono in Table 2 of the revised paper) and the prior-based unsupervised non-blind method: TV_Prior. - Add an additional section before the conclusion to discuss the potential key challenges and future directions. (The seventh comment of Review #1)
Response: Thanks for the expert reviewer’s comment. We added ‘Discussion’ subsection (before the ‘Conclusion’ section) on lines 433 to 450 of page 15 in the revised manuscript.
Reviewer 2 Report
This manuscript presents a technique to produce high-resolution images from low-resolution images using a deep unsupervised learning network. Unlike previous techniques, for which high-resolution images are converted to lower resolutions using bicubic downsampling, this paper simultaneously estimates a high-resolution image and the image degradation process without prior knowledge of the true degradation.
The manuscript is well organized and presents the findings of the research. Nevertheless, some points still need to be revised before the paper can be considered for publication.
1. Please revise the manuscript for readability. Many sentences are very long and make it difficult for the reader to understand the content.
2. Please reference the studies that explicitly used bicubic down-sampling to generate the training images (line 33).
3. Algorithm 1 should be inserted after the whole description of the algorithm itself (line 305).
4. Please explain why ADAM was used as optimization method.
5. The measurement results could be presented more clearly.
6. It is not clear from the text whether the model was tested with all images of each dataset. The presented values for PSNR and SSIM do not say whether they are averaged values of the whole set or only the results for the presented images.
7. Please check the references list.
Author Response
We appreciate the expert reviewer’s helpful comments and constructive suggestions. The manuscript has been significantly improved based on reviewers’ comments and suggestions. We will supply our detailed point-by-point responses to the reviewer. The manuscript has also been revised as suggested by the reviewer (marked in red for the required revisions).
- Revise the manuscript for readability. (The first comment of Review #2)
Response: Thanks very much for pointing this issue. We revised some sentences in the revised mapper to improve the readability.
- Reference the studies that explicitly used bicubic down-sampling to generate the training images (line 33). (The second comment of Review #2)
Response: Thanks for the expert reviewer’s comment. The Bicubic down-sampling for synthesizing the LR images from the HR images usually is a default setting in the dominated SR algorithms, and many publications may not explicitly explain the used down-sampling operation with ‘Bicubic’. However, it is a common view in the image SR research community using ‘Bicubic’ down-sampling. We also added some references (Refs 21 and 22) in the revised paper, which explicitly claimed that most of deep learning-based methods adopted the ‘Bicubic’ down-sampling operation to synthesize the LR images.
- Algorithm 1 should be inserted after the whole description of the algorithm itself. (The third comment of Review #2)
Response: Thanks for the expert reviewer’s comment. We modified the layout and position of algorithm 1 in the revised manuscript. Thanks.
- Please explain why ADAM was used as optimization method. (The fourth comment of Review #2)
Response: Thanks very much for pointing this issue. We added some explanation about why we select the ADAM as the optimizer of our method on lines 325 to 329 of page 9 in the revised paper. And we also provided the experimental results using different optimizers on the Set5 dataset with the upscale factor 8 in Table 3(b).
- The measurement results could be presented more clearly. (The fifth comment of Review #2)
Response: Thanks the comments. We adjusted the layout of the visualization results especially for Fig. 4 for clear presentation. Thanks.
- It is not clear from the text whether the model was tested with all images of each dataset. (The sixth comment of Review #2)
Response: Thanks for the expert reviewer’s comment. We clearly explained the evaluation results are the average values of all images in each dataset on lines 349 to 352 in the revised manuscript.
- Please check the references list. (The seventh comment of Review #2)
Response: Thanks the comments. We carefully checked the reverences, and corrected some misses in the revised manuscript. Thanks.
Round 2
Reviewer 2 Report
The authors took my comments into account and improved their manuscript. The text presents the research, the procedure and its results in good detail, so that the reader gets a good overview. In some places, minor adjustments should be made to the text (e.g., consistent capitalization of abbreviations such as "ADAM", etc.).
Author Response
- In some places, minor adjustments should be made to the text. (e.g., consistent capitalization of abbreviations such as "ADAM", etc.).
Response: Thanks for the expert reviewer’s comment. 'Adam' has been unified with 'ADAM' in the revised manuscript. We also modified all words ‘'bicubic' to 'Bicubic' in the revised paper. Thanks.
