Super-Resolution Model Quantized in Multi-Precision
Round 1
Reviewer 1 Report
In this work, the authors investigate the problem of super-resolution for images, focusing in particular on the single-image problem, i.e., reconstructing high-resolution images from individual low-resolution ones. Deep learning methods have recently shown remarkable performance for this task. In particular, this work studies size and energy consumption limitations of these models.
The topic of this paper is relevant, as super-resolution methods are becoming central in many applications and need to be efficiently integrated into low-memory low-energy devices.
However, I think that there are several major issues in the manuscript in its current form:
- First of all, the quality of the presentation is low. This severely impacts comprehensibility in many points. Several parts are difficult to understand, and, in general, the goals, the method and the results are hard to understand.
- The novelty of this work appears to be very limited. Most of the manuscript is focused on describing related work or implementation details.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The english is insufficient. There are minor typos (space before comas, missing spaces, etc.) but there are also sentences without verbs, truncated, or that seem to finish some other sentence, or ends in an unexpected way. For e.g., lines 97-98: the first sentence speaks of user needs but the next speaks of "processing requirements".
The results are insufficiently explained. Some tables are quite mysterious. What does a checkmark mean?
Table 1 and table 3 are mostly redundant (1 line difference)
The experiments seem to be correctly conducted and the results promising. I would really recommend finding a copy-editor for the english, rewrite the parts that need to, and then resubmit. I think this is a good paper, just very hard to read in its current form.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This is good topic chosen by authors. Few things need to be clarified before acceptance.
- Line 161- Figure 2 I think authors missed the min to represent. or it is represented with negative values of -|Max| I am bit sceptical if it is discussed . discuss to avoid confusion to authors.
- Figure 4 is too blurred make it clear.
- Figure 5 represent that a small figure is enlarged and has more clarity , does your methods enlarge the images , or this is just for representation. If So please adjust the scale accordingly.
- Discussion written part is very small. i need authors to discuss the results with previously published results and how it is improved or better than those mentioned methods.
- Conclusion should be more intense as compared to the provided one. It need to provide the robust and future work also associated with the present work/ methodology.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
In the paper, a quantization aware training of two deep learning models, SRGAN and ESRGAN, is shown. Comments:
1. The references should be extended with more works published in the last two years. At least 10 works should be added and suitably described.
2. The large part of Section 2.1 does not contain any reference.
3. The paper lacks a discussion useful for readers who would like to quantize their models of interest. What indicated a model suitable for quantization in the way shown in this paper?
4. Figures:
a. Fig. 1 can be applied to many concepts that use deep learning. It should be corrected considering super-resolution-related ideas.
b. Fig. 4 is illegible. Also, the used blocks are not common or their functions are wrongly applied.
c. Fig. 5 is illegible. Please change colors. Are the colors meaningful somehow here? Do they correspond to the colored blocks in other images?
5. Numbers should be rounded to significant digits.
6. The paper does not contain any examples of obtained images.
7. The paper requires proofreading fixing linguistic and grammar errors (e.g., the first person in line 33; two close sentences starting with ‘but’ in lines 33-34 and many, many more), or typos (mostly problems with additional space, e.g., line 33, commas or lack of spaces before a bracket, e.g., line 26).
8. The presented results cannot be replicated by a reader. Note that a link to a webpage with source code would greatly promote the findings and facilitate the replication of the results.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have improved the manuscript with respect to the previous version. I suggest further improving the quality of presentation before this manuscript is accepted.
Reviewer 3 Report
May be accepted
Reviewer 4 Report
My comments have been addressed.