SCDNet: Self-Calibrating Depth Network with Soft-Edge Reconstruction for Low-Light Image Enhancement

Round 1

Reviewer 1 Report

The paper is well written and presents interesting result.

Only one remark: in both results and discussion overexposure problem should be better discussed.

Best Regards,
The Reviewer

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Title:

------

SCDNet: Self-calibrating depth network with soft-edge reconstruction for low-light image enhancement

Overall Contribution:

----------------------

This paper proposes a low light-image enhancement approach that combines an edge reconstruction module with a self-calibration module and a final deep curve estimation network. The paper is well written. The experimental results look promising. 

 

Limitations and Improvements:

---------------------------------

Comment 1: The authors should clearly explain the difference between the proposed DCE-Net and ZeroDCE model. Also they should reference the loss functions used since they are the same as those used in ZeroDCE paper.

 

Comment 2: The authors mention that their model has low time consumption yet they do not show it in the results. A runtime analysis of the proposed model compared with the state of the art solutions needs to be added.

 

Comment 3: In the ablation study, it is better to include sample enhanced images for each of the ablated components so that readers can better visualize the importance of these components.

 

Comment 4: The authors should consider the effect of their enhancement approach on high level computer vision tasks like objects detection and classification. The authors can refer to the below papers for their reference:

- Comparing Deep Learning Models for Low-light Image Enhancement and their Impact on Object Detection and Classification: Overview, Empirical Evaluation, and Challenges, Signal Processing: Image Communications, 109:116848 2022 

- Dynamic Low-Light Image Enhancement for Object Detection via End-to-End Training. ICPR 2020: 5611-5618

- Low-Light Homomorphic Filtering Network for integrating image enhancement and classification. Signal Processing: Image Communications, 100: 116527 (2022)

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, the authors focus on improving low light image enhancement. Soft edge reconstruction module is proposed to reconstruct the soft edges. Self-calibration module is introduced to improve the computation. And iterative light enhancement curve is proposed to get high quality images. 

1. For the figure 1, it is not really clear about which part is the soft-edge reconstruction network and which part is self-calibration module. Please add more information to explain this.

2. Is the self-calibration module proposed from this paper or cited from another paper? I saw it is from the reference [25]. Can authors explain what is the changes they made in this paper? 

3. Is the soft-edge reconstruction module proposed from this paper or reference [24]? Can authors explain what is the changes they made in this paper? 

4. Is this claim "Meanwhile, our SCDNet is significantly 18 better than the state-of-the-art methods in objective matrices" true? I found the SOTA results on LOL dataset is much higher. If it is not true, please change this sentence. 

5. The table 3 shows that SER and SC independently cannot improve the results much. However, the combination of them improves the results a lot. Is there any other parameters change in the experiment? Why this is the case? Can authors further explain this? 

6. The paper also mentions that there are several losses are applied to the model. However, there is not results to prove the importance of the loss. 

7. There are also some writing and grammar errors in the paper that need to be improved. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf