Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

To adequately address the prevalent hybrid noise in real-world scenarios,the author proposed a convolutional neural network (CNN)-based approach with a refined composite loss function, specifically designed for hybrid noise removal in raw infrared images. Rigorous experiments demonstrate the superior performance of the proposed approach in hybrid noise suppression and detail preservation. But before publication, there are some issues that must be resolved.

 

①The core images (Eg. Figure 1,2) provided by the authors suffer from low resolution and blurred details, which hinder readers' understanding of the methods described in the paper.

②Although the authors propose a weighted strategy using noise estimation term, TV regularization, and edge preservation loss to guide network learning, Figure 2 still fails to clearly illustrate how these components interact with pseudo labels or how they influence the network.

③The main text contains no explanatory notes for Figure 3, leaving its purpose unclear.

④In Section 4.1, the authors mention using PSNR and SSIM as evaluation metrics for quantitative experiments, but they fail to provide their mathematical definitions and specific meanings. This hinders readers from quickly understanding the experimental conclusions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1.    The title says „self supervised”, the keywords say „unsupervised”. Make this consistent (either „self supervised” or „weakly or semi supervised” if you use pseudo targets)
2.    The noise model is too simple. The equation I=C+N does not capture FPN. Justify this choice or extend the model
3.    Number of loss terms. The text announces four but defines three. Fix the description or add the missing term. Replace the odd comma between w1 and w2
4.    Figure 4 vs no ground truth. The phrase „clean reference images” conflicts with the claim „no ground truth”. Change to „pseudo clean not used as ground truth”
5.    Table 2 typo. It should be LNE+LTV+LEP” without a double plus
6.    Lack of real data. Tests use only synthetic noise. Add sequences with real FPN
7.    LLVIP details. State that you used a subset (LLVIP has tens of thousands of pairs). Provide the exact split protocol
8.    Leakage control. Train and test on LLVIP require a sound splitting strategy.
9.    Metrics and significance. In addition to PSNR and SSIM add LPIPS, NIQE, and BRISQUE, and report mean and standard deviation and significance tests
10.    Hyperparameters are questionable. LR=1 for Adam looks unrealistic. Please provide learning and stability curves or correct it.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. The abstract reads clearly but is too descriptive; it lacks emphasis on quantitative improvements and specific novelty claims.
2. Explain the intuition behind your adaptive TV regularization: how it differs from standard anisotropic diffusion or Sobel-based weighting.
3. Include a block diagram or pseudo-code for the adaptive gradient-perception module.
4. Results are strong but presented descriptively; not all claims are statistically validated.
5. To the table 1 Include #Params (M) and Inference Time (ms) columns.
6. Figure captions are too generic.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The author's revised manuscript has improved the quality of the article and can be published.

Author Response

We sincerely thank you for your hard work and many constructive suggestions during the whole review process. These valuable suggestions have greatly improved the quality of this article.

Reviewer 3 Report

Comments and Suggestions for Authors

No more comments need

Author Response

We sincerely thank you for your hard work and many constructive suggestions during the whole review process. These valuable suggestions have greatly improved the quality of this article.