A More Effective Zero-DCE Variant: Zero-DCE Tiny

Round 1

Reviewer 1 Report

Overall, it is an interesting task. The task can be further improved by

aa-      Result and discussion needed to be more explain and clear.

 

b-      In introduction section, cite some relevant article. 

1)      Iqbal, Muhammad Shahid, et al. "Deep learning recognition of diseased and normal cell representation." Transactions on Emerging Telecommunications Technologies (2020): e4017.

 

c-      Add Background section, it will be easy to understand the article.  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes a variant of previous Zero-DCE and Zero-DCE++ methods called Zero-DCE+, an image illumination enhancement method which achieves results on par with previous works but with lower cost.

My questions are all about algorithm and experiments:

- First, I think the name of this new proposed method "Zero-DCT+" is not very good, since Zero-DCE++ is already a lighter version of Zero-DCE, while the proposed method is a further lighter version of Zero-DCE++. Therefore, it's very weird that it's called "Zero-DCT+", which is naturally regarded as some version whose lightweight level is between "Zero-DCE" and "Zero-DCE++". The new name could be "Zero-DCE+++" (might be weird as well), or some others like lightweight-Zero-DCE++ or something similar.

- It looks like this paper only compares the proposed method with Zero-DCE and Zero-DCE++, and this is clearly not enough. In lines 288 "Since the Zero-DCE++ [4] has been qualitatively and quantitatively compared with other benchmark methods, here we only compare the results of the Zero-DCE series of networks." I don't agree with this at all. Even though Zero-DCE++ paper already compares with many SOTA methods, there are still many things which might be different, such as different dataset, platform, etc. More comparison with other SOTAs are still required and very necessary.

-  It seems only one dataset [22] is used in comparison? Another unsound place in experiments. More dataset are definitely required to prove the efficiency. The authors can try to use all the dataset used in Zero-DCE++ paper.

- In Table 3, the proposed Zero-DCE+ got the best SSIM result. Any specific reason why it is even better than Zero-DCE, the largest model among the three methods and usually should be the best under these quantitative comparison?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Low Illumination Image Enhancement (LLIE) is to improve the perception 8 or interpretability of images taken in low illumination environments. The Introduction is good enough with organization of the manuscript at the end. Few suggestions will be here to improve this paper at some extent:

(1)    How your approach is better over Zero-DCE approaches? Justify.

(2)    How the results vary when your approach (i.e. Zero-DCE Plus) is used in place of Zero-DCE approach? Explain and incorporate at suitable place in the text.

(3)    Why you preferred ReLU activation function? Is there any other activation function may be suitable here?

(4)    Incorporate one Section before Conclusion Section about the significance of your contribution in few lines clearly.

 

(5)    It will be helpful for the researcher in same area, if you will add few line about future directions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The author have done revision according to comments.