Using Haze Level Estimation in Data Cleaning for Supervised Deep Image Dehazing Models
Round 1
Reviewer 1 Report
This paper proposes a haze level estimation scheme based on dark channel prior for supervised deep image dehazing (SDID) models. A general framework for SDID models is also developed. The effectiveness of the proposed method is verified by experiments. I suggest the authors address the following points.
1. The Abstract should concisely and clearly summarize the difficulties and practical significance of the research.
2. The number of algorithms used in the comparison experiment is too small. More newly published advanced methods should be added for experimental comparison. The experimental results are not convincing due to insufficient comparison experiments.
3. Ablation experiments are recommended to be added. The effectiveness of the proposed method should be verified.
4. Practical experiments are suggested to be conducted to demonstrate the practical application performance of the proposed method. Only the performance of the proposed method in the generated dataset is shown in the paper.
5. The format of the table needs to be optimized, such as in Table 4 and Table 5, and it is recommended not to divide the same table into two pages. The meaning of the numbers in Tables 5 and 8 should be explained.
6. The experimental visualization results do not reflect the superiority of the proposed method. For example, the results of I1, I2 and I6 in Table 5 look the same as the proposed method.
7. The Conclusion section could be strengthened. It is recommended to quantify the improvements resulting from the proposed methodology.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
1. This manuscript proposes a haze level estimation scheme based on the dark channel prior to discarding image pairs with hazy ground truth images in the data set. Then, the cleaned data has been taken as the input in the two existing models of GCAN and REFN to verify the performance improvement. However, the innovation just refers to presenting a hazy level estimation schema with different parameter settings, which has been proposed by reference [1]. However, the related images have been simply discarded as a process of data cleaning, which cannot be considered a great improvement.
2. In the introduction, the contributions can be considered as one item, items 2 and 3 are similar to item 1.
3. In Section 2, compared to the dark channel prior in reference [1], the authors in this manuscript cannot show great improvement, they just adjust some threshold parameters to confirm discard or not the images in the data set.
4. In Section 3, Figure 3 should be deleted, which is a general data prediction process.
5. In Section 4, the clean data has been input into the GCAN and REFN models for verification. The hazy ground truth images have been discarded based on the hazy level estimation schema, and it will absolutely lead to the performance improvement of image identification.
Written English should be polished.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The manuscript proposes a data cleaning scheme to detect and discard image pairs with hazy ground truth images from the training datasets before training supervised deep dehazing (SDID) models. The proposed data cleaning scheme utilizes haze level estimation (HLE) to distinguish between clear and hazy ground-truth images. The HLE is used to clean the data set for SDID models. The authors hypothesize that the dehazing performance of SDID models can be improved by using a cleaner dataset without hazy GT images. The results demonstrated the effectiveness of the proposed framework. While the results and its experimental results may seem promising, I am not entirely convinced about the novelty and effectiveness of this approach regarding robustness. Discarding over 50% of the data could reduce the model’s robustness, especially when starting with a limited dataset. The hypothesis that performance can be improved using a cleaner but smaller dataset might not always hold true. The Pros and cons of deleting data in the dataset should be more thoroughly discussed.
Specific recommendations for revising the manuscript are provided below.
1) Equation 1: Please explain all labels in the equation. “I” is not defined.
2) Line 47 – 66: the delivery of the paragraph needs improvements. Please refine sentences, such as “Some of them…” in a more professional manner.
3) Line 58: “For more references, see [22-23].” References should not be referred in the text. Phrases such as "…in [1]… " or "Reference [2] states…" are not acceptable. Please rewrite the sentence.
4) Line 93-95: what about the robustness of the model? Please elaborate on the pros and cons of using a cleaner but much smaller dataset for training the model in the discussion section.
5) Line 146: Explain ∆?̃dc and ?̃dc (?1). Is ∆?̃dc the rate of the change of ?̃dc ?
6) Line 182: It states that the testing dataset was used to validate the trained SDID model. This is incorrect. There should be a separate validation dataset for this purpose.
7) Line 185: Please elaborate on why “the performance bias is inevitable when hazy ground truth images are used in evaluation.”
8) Line 202: Explain the rational on setting the general rule of thumb. Does the rule of thumb apply to all scenarios regardless of the available data size?
9) Line 294 – 295: The part claims that “it is not necessary to use all training image pairs”. What are the rules in selecting the training data size?
10) Line 313 – 314: Explain why image I1 is better in REFN0 than in REFN20k. This example raises a question on the effectiveness and consistency in the trained model’s performance of using a cleaner but smaller dataset.
Comments were included in the section above.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
- The definition of haze level is ambiguous and needs to be accurate.
- The reasons for setting parameters (e.g., n, etc.) used in HLE are unclear. Bases for setting parameter values should be provided theoretically rather than experimentally.
- In an image, the fog level varies across regions due to the distance between the camera and objects and the presence of air turbulence. Also, the fog impact on image scenes varies depending on the contrast and color of the objects. For example, even under the same fog conditions, regions with colors more similar to air-light colors and less local contrast are more affected by fog. It is good for the authors to mention how to solve these irreguarilty.
-
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I recommend the paper for publication.
Reviewer 2 Report
The revised manuscript has been improved based on the comments.
The details of the manuscript should be checked carefully.
Reviewer 3 Report
My review comments were all addressed.
