Joint Deblurring and Destriping for Infrared Remote Sensing Images with Edge Preservation and Ringing Suppression

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Manuscript title: Joint Deblurring and Destriping for Infrared Remote Sensing Images with Edge Preservation and Ringing Suppression

Authors: Ningfeng Wang, Liang Huang, Mingxuan Li, Bin Zhou and Ting Nie*

Manuscript Number: 4027590

The manuscript introduces a novel three-stage variational framework for simultaneously addressing stripe noise and blurring in infrared remote sensing images. This framework incorporates adaptive edge preservation using structural tensors, a stripe direction fidelity term, and a new WCOB hybrid filtering technique. The study aims to achieve concurrent destriping and deblurring while maintaining image details and suppressing ringing artifacts. Comprehensive experiments using both simulated and real datasets demonstrate the method's superior denoising and restoration performance, showing robust performance.

However, the manuscript still has some issues that require further clarification and improvement before it can be considered for publication.

The specific issues are listed as follows:

1. The presentation of the mathematical formulations in the Methods section requires improvement. The objective function, given in Equations (2) and (13), consists of at least eight different regularization and fidelity terms. This high complexity makes the equations challenging to follow and interpret. It is suggested that the authors reorganize the objective function, perhaps by decomposing the mathematical model into logically distinct components, to enhance clarity and significantly improve readability.
2. Figure 1 appears overly simplified and omits several key methodological details. Specifically, the exact operations represented by the "Structure Tensor-based Method" block are unclear, and the processing stage at which the "Directional Fidelity Term" is incorporated remains ambiguous. Moreover, the fusion strategy between the Wiener and Cosine filters within the "WCOB Filtering" module is not adequately illustrated. It is recommended that the authors provide a more detailed algorithm diagram, clearly linking each functional block to the corresponding mathematical formulation or procedural step described in Section 2 of the manuscript.
3. Based on the convolution theorem, the frequency-domain filter iHH(u,v) (defined in Equations (29)-(32)) should be multiplied directly with the Fourier transform of the image. However, Equation (33) appears to incorrectly apply the Fourier transform of iHH itself, suggesting a possible implementation error. The authors are strongly advised to verify and correct this issue. In addition, the variable "I" in this context should be clearly defined—whether it refers to the de-streaked image I or the original observed image B.
4. Table 6 specifies that two deep learning methods, DestripeGAN+PGDN (De_GANv2) and SLDR+NeRf, were executed on CPUs. However, the computing platform remains unspecified for all other methods, including the authors' proposed approach. The computing environment for the proposed method must be clarified, specifically, whether a GPU or any parallel computing architecture was utilized. Without this clarification, the runtime comparison with CPU-based deep learning methods lacks validity. If providing these details is not feasible, we recommend re-evaluating all methods under consistent hardware conditions to ensure a fair performance comparison.
5. The experimental results section currently lacks statistical significance testing. It is important to note that the provided tables report only average performance metrics, such as the mean PSNR. However, comparing mean values alone is insufficient to establish the superiority of one method over another scientifically. To robustly demonstrate that the observed performance improvements are not due to random chance, it is essential to incorporate statistical significance tests. This would significantly strengthen the validity of the conclusions drawn from the experimental data.
6. The use of bold formatting in Tables 2-7 is inconsistent and should be improved. We recommend establishing a uniform rule for emphasizing the best-performing result in each metric column, and clearly stating this rule either in the main text or in each table's caption. Specifically, for metrics where a lower value indicates better performance (e.g., NIQE, MRD), the minimum value should be bolded; and for metrics where a higher value reflects superior performance (e.g., SSIM, PSNR, ICV), the maximum value should be bolded.
7. The visual clarity of figures requires improvement (g., Figures 4, Figures 6, Figures 9, Figures 11-13). The small font size of labels and fine line weights make the textual and graphical elements difficult to distinguish, even when viewed at high resolution. As a result, the figure's current presentation compromises its readability and effectiveness in conveying comparative results. It is recommended that the figure be reformatted with larger fonts and more distinct line styles to ensure all components remain clearly legible at its intended publication size.
8. The red annotation boxes in Figures 5, 10, 15, and 16 are not prominent enough. They should be made thicker or changed to a high-contrast color. This will ensure they clearly highlight the relevant areas.
9. The manuscript requires thorough proofreading to correct grammatical errors and spelling mistakes. Additionally, the meaning of the symbol "?" above Figure 9 (line 353) should be clearly explained. Finally, the citation format throughout the references section needs to be made consistent.

Comments for author File: Comments.pdf

Author Response

Comments 1: The presentation of the mathematical formulations in the Methods section requires improvement. The objective function, given in Equations (2) and (13), consists of at least eight different regularization and fidelity terms. This high complexity makes the equations challenging to follow and interpret. It is suggested that the authors reorganize the objective function, perhaps by decomposing the mathematical model into logically distinct components, to enhance clarity and significantly improve readability.

Responses 1: Thank you for pointing this out. We agree with this comment. Therefore, We have revised the manuscript accordingly to improve the clarity and readability of the mathematical formulations in the methodology section. Specifically, we acknowledge that Equations (2) and (13) contain multiple regularization and fidelity terms. This is mainly because the proposed framework simultaneously addresses stripe noise removal and blur degradation restoration. In order to fully describe the constraint mechanisms involved in both processes, the objective function is inevitably relatively complex in structure. To make the design logic of each regularization term clearer, we have unified the noise-related objective terms in Equation (2) and denoted them as Rs(S). A detailed explanation of this term has been added immediately below Equation (2), and the same notation has been consistently adopted in the revised Equation (13). In addition, the original Equation (13) (renumbered as Equation (15) in the revised manuscript) corresponds to the augmented Lagrangian formulation derived from the original objective function. Due to the nature of this transformation, the number of regularization terms is unavoidably increased. To address potential confusion, we have added explicit transitional explanations and derivation descriptions between these equations, which significantly improve the coherence and readability of the formulation and help readers better understand the model structure and optimization process. Furthermore, we have added explicit explanations of the roles, physical meanings, and parameter-setting rationale in Section 3.1.1. The different components of the objective function are also clearly labeled and explained in the main text, emphasizing their individual functions and interrelationships to further enhance overall readability.

The modification can be found in the revised manuscript on Page 5, Paragraph 2,  Line 161; Page 8, Paragraph 2.2 ,  Line 257; Page 14, Paragraph3.1.1,  Line387-395.

 

Comments 2: Figure 1 appears overly simplified and omits several key methodological details. Specifically, the exact operations represented by the "Structure Tensor-based Method" block are unclear, and the processing stage at which the "Directional Fidelity Term" is incorporated remains ambiguous. Moreover, the fusion strategy between the Wiener and Cosine filters within the "WCOB Filtering" module is not adequately illustrated. It is recommended that the authors provide a more detailed algorithm diagram, clearly linking each functional block to the corresponding mathematical formulation or procedural step described in Section 2 of the manuscript.

Responses 2: Thank you for pointing this out. We agree with this comment. Therefore, we have revised the overall framework diagram to provide a more detailed and structured workflow. In the revised version, each functional module is now more clearly aligned with the main processing steps described in Section 2, and the updated flowchart has been replaced in the corresponding position in the manuscript.

In addition, regarding the parameter α(u,v) involved in the non-blind restoration process, we have further supplemented and clarified the fusion mechanism of the WCOB filter in the revised manuscript. The Wiener filter and the Cosine filter exhibit complementary behaviors in terms of detail preservation and ringing-artifact suppression. Specifically, when the image signal-to-noise ratio (SNR) is relatively high, a larger α(u,v) is adopted to strengthen the dominant role of Wiener filtering during deconvolution, thereby enabling more effective recovery of image structures and fine details. Conversely, when the image SNR is low and the risk of noise amplification and ringing artifacts increases, α(u,v) is correspondingly reduced to assign a higher weight to the Cosine filter, which helps suppress high-frequency noise and ringing artifacts introduced during deconvolution and improves the overall stability and robustness of the restoration results. The corresponding revisions have been incorporated around Eq.(34) in the revised manuscript, which facilitates a clearer understanding of the collaborative mechanism between the two filters.

The modification can be found in the revised manuscript on Page 4; Page 12, Paragraph 2.3, Line  334-350.

 

Comments 3: Based on the convolution theorem, the frequency-domain filter iHH(u,v) (defined in Equations (29)-(32)) should be multiplied directly with the Fourier transform of the image. However, Equation (33) appears to incorrectly apply the Fourier transform of iHH itself, suggesting a possible implementation error. The authors are strongly advised to verify and correct this issue. In addition, the variable "I" in this context should be clearly defined—whether it refers to the de-streaked image I or the original observed image B.

Responses 3: Thank you for pointing this out. We agree with this comment. Therefore, we have carefully checked Equation (33) and confirmed that it indeed contained an imprecise expression. Specifically, the variable (“I”) in the original formulation was ambiguous. In the revised manuscript, it has been clearly redefined as the destriped image and replaced with (“ I_destriped”) to accurately reflect the actual input–output relationship of the model. The above modifications and explanations have been provided and marked in the corresponding sections of the paper. We once again appreciate your valuable suggestions. This correction will help improve the accuracy and readability of the paper.

The modification can be found in the revised manuscript on Page 12, Paragraph 2.3,  Line 351-359.

 

Comments 4: Table 6 specifies that two deep learning methods, DestripeGAN+PGDN (De_GANv2) and SLDR+NeRf, were executed on CPUs. However, the computing platform remains unspecified for all other methods, including the authors' proposed approach. The computing environment for the proposed method must be clarified, specifically, whether a GPU or any parallel computing architecture was utilized. Without this clarification, the runtime comparison with CPU-based deep learning methods lacks validity. If providing these details is not feasible, we recommend re-evaluating all methods under consistent hardware conditions to ensure a fair performance comparison.

Responses 4: Thank you for pointing this out. We agree with this comment. Therefore, we would like to clarify the implementation and computational settings of the compared methods. The deep-learning-based methods used for comparison (e.g., DestripeGAN+PGDN and SLDR+NeRF) rely on GPU acceleration, and their inference processes are executed on GPU platforms. In contrast, the traditional methods, including the proposed joint deblurring and destriping approach, are based on variational optimization frameworks, fully implemented in MATLAB and executed in a CPU environment without using GPU acceleration or any parallel computing architecture. Due to the fundamental differences in the underlying implementation paradigms between deep learning methods and traditional optimization-based methods, conducting a strictly identical hardware-level comparison is not practically feasible. To address this concern and improve transparency, we have added a clear description of the computational environment in the first paragraph of Section 3.1.1 explicitly stating that the deep learning methods are implemented on GPU platforms, while the proposed method is entirely executed on CPU. This clarification ensures a more transparent and reasonable comparison setting.

The modification can be found in the revised manuscript on Page 14, Paragraph 3.1.1,  Line 394-402.

 

Comments 5: The experimental results section currently lacks statistical significance testing. It is important to note that the provided tables report only average performance metrics, such as the mean PSNR. However, comparing mean values alone is insufficient to establish the superiority of one method over another scientifically. To robustly demonstrate that the observed performance improvements are not due to random chance, it is essential to incorporate statistical significance tests. This would significantly strengthen the validity of the conclusions drawn from the experimental data.

Responses 5: Thank you for pointing this out. We agree with this comment, as it is important for further validating the reliability and statistical significance of the conclusions. Therefore, we have reintroduced statistical significance testing into the revised manuscript. To evaluate whether the performance differences among different methods are statistically significant across multiple evaluation metrics, independent significance analyses were conducted on a per-image basis over the test dataset. Specifically, for metrics such as PSNR and SSIM, which approximately satisfy the normality assumption, paired t-tests were employed to compute the corresponding p-values. For metrics such as NIQE, ICV, and MRD, which do not follow a normal distribution or whose distributions are unknown, the Wilcoxon signed-rank test was adopted to obtain p-values. All significance tests were performed based on paired observations for each image, and the p-values were computed using MATLAB’s built-in statistical functions. The statistical significance thresholds were set as p < 0.05 for statistically significant differences and p < 0.01 for highly significant differences. The significance test results for flat scenes and complex scenes have been organized into separate tables and are presented at the corresponding locations in the revised manuscript. In addition, a detailed description of the significance testing methodology and implementation has been added to Section 3.1.2, ensuring that the experimental design is transparent, rigorous, and reproducible.

The modification can be found in the revised manuscript on Page 14, Paragraph 3.1.2,  Line 404-458; Page 23, Paragraph 3.2, Line 567-577; Page 26, Paragraph 3.3,  Line 623-630.

 

Comments 6: The use of bold formatting in Tables 2-7 is inconsistent and should be improved. We recommend establishing a uniform rule for emphasizing the best-performing result in each metric column, and clearly stating this rule either in the main text or in each table's caption. Specifically, for metrics where a lower value indicates better performance (e.g., NIQE, MRD), the minimum value should be bolded; and for metrics where a higher value reflects superior performance (e.g., SSIM, PSNR, ICV), the maximum value should be bolded.

Responses 6: Thank you for pointing this out. We agree with this comment. Therefore, we have standardized the boldface formatting in Tables 2–7 in the revised manuscript to ensure consistency and clarity. Specifically, for evaluation metrics where lower values indicate better performance (e.g., NIQE and MRD), the minimum values are highlighted in bold. Conversely, for metrics where higher values correspond to better performance (e.g., SSIM, PSNR, and ICV), the maximum values are emphasized in bold. In addition, we have provided a systematic description of the definitions and calculation methods of all evaluation metrics in Section 3.1.2 of the revised manuscript, explicitly clarifying the relationship between metric values and performance quality. Moreover, in the paragraphs below Figures 9 and 15, where these comparative metrics are first discussed, we have added explicit explanations of the meaning of the boldface notation to improve consistency and readability when interpreting the tables. For Table 6, which compares the running time of different methods under varying image sizes, the shortest execution time is highlighted in bold. A corresponding explanation of this formatting convention has also been added to the paragraph below Figure 15. It is worth noting that the proposed method achieves the shortest running time among all traditional methods, and this result has been explicitly emphasized at the corresponding locations in the revised manuscript.

The modification can be found in the revised manuscript on Page 17, Paragraph 3.2,  Line 468-478; Page 23, Paragraph 3.3, Line 586-596.

 

Comments 7: The visual clarity of figures requires improvement (g., Figures 4, Figures 6, Figures 9, Figures 11-13). The small font size of labels and fine line weights make the textual and graphical elements difficult to distinguish, even when viewed at high resolution. As a result, the figure's current presentation compromises its readability and effectiveness in conveying comparative results. It is recommended that the figure be reformatted with larger fonts and more distinct line styles to ensure all components remain clearly legible at its intended publication size.

Responses 7: Thank you for pointing this out. We agree with this comment. Therefore, we have replaced Figures 4, 6, 9, and 12 with clearer versions in the revised manuscript. In addition, local zoom-in views have been added to Figures 6, 9, and 13 to present image details more intuitively and enhance visual clarity. Regarding Figure 11, the currently retained version already has sufficient resolution, and its layout is intentionally aligned with that of Figure 10 for visual correspondence. Further enlargement may disrupt the overall page layout; therefore, no additional modification has been made at this stage. Nevertheless, if further optimization is deemed necessary, we are fully willing to revise this figure according to the reviewer’s suggestions to further improve the manuscript quality.

The modification can be found in the revised manuscript on Page 10; Page 13; Page 17; Page 20; Page 21.

 

Comments 8: The red annotation boxes in Figures 5, 10, 15, and 16 are not prominent enough. They should be made thicker or changed to a high-contrast color. This will ensure they clearly highlight the relevant areas.

Responses 8: Thank you for pointing this out. We agree with this comment. Therefore, we have comprehensively revised the annotation boxes in Figures 5, 10, 15, and 16, replacing the original red boxes with high-visibility yellow highlights. This adjustment makes the enlarged regions more prominent and allows the relevant details to be emphasized in a clearer and more intuitive manner. Overall, this modification effectively improves the readability and visual guidance of the figures, enabling the key content to be more easily identified and understood by the reader.

The modification can be found in the revised manuscript on Page 12; Page19 ; Page25 ; Page27.

 

Comments 9: The manuscript requires thorough proofreading to correct grammatical errors and spelling mistakes. Additionally, the meaning of the symbol "?" above Figure 9 (line 353) should be clearly explained. Finally, the citation format throughout the references section needs to be made consistent.

Responses 9: Thank you for pointing this out. We agree with this comment. Therefore, we sincerely apologize for this oversight caused by insufficient checking prior to submission. We confirm that the question mark symbol (“?”) appearing above Figure 9 (Line 353) in the original manuscript was due to a missing reference that had not been properly included in the reference list. This reference has now been added and completed accordingly in the revised manuscript. Furthermore, before resubmission, we have carefully rechecked the entire manuscript and all references in a comprehensive and meticulous manner to ensure that no similar omissions remain.

The modification can be found in the revised manuscript on Page 17,  Line 468-478.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper focuses on the critical problem of joint deblurring and destriping in infrared remote sensing images, proposing a three-stage variational framework with clear modular design. The method achieves significant performance improvements on both simulated and real satellite datasets, demonstrating strong practical value. However, the manuscript can be further enhanced in terms of theoretical depth, experimental comprehensiveness, and reproducibility details. With targeted supplements such as strengthened theoretical analysis, expanded comparative experiments, and detailed parameter setting explanations, this work will make a more solid contribution to the field of infrared remote sensing image processing.

1. For the structure-tensor-based adaptive edge-preserving operator, the paper only describes its implementation process but lacks in-depth theoretical analysis of why this operator can effectively distinguish stripe noise from image edges. For example, there is no quantitative analysis of the relationship between eigenvalues (λ₁, λ₂) and edge/stripe characteristics, nor a comparison with other edge-preserving strategies in terms of theoretical advantages.
2. The parameter setting and computational efficiency analysis are insufficient. The model involves regularization parameters (such as ρ₁, ρ₂, β₁, etc.). The parameter values are all determined based on empirical verification, without providing parameter sensitivity analysis (e.g., curves showing the impact of parameter changes on SSIM and PSNR).
3. No discussion on the impact of wavelet decomposition levels. The paper uses multi-scale wavelet transform but does not explain why the decomposition level n>1 is chosen, nor does it briefly verify whether different decomposition levels (e.g., 2-level vs. 3-level) affect stripe noise extraction and edge preservation, leaving room for optimization of this key preprocessing step.
4. In the no-reference image experiments using SDGSAT-1 data, the paper mentions that the ICV value is higher, indicating better restoration performance, but does not explain the definition and calculation method of the ICV metric. It is recommended to supplement the formula or reference source of ICV to ensure the transparency of the evaluation process.
5. The comparative experiments only include traditional model-based methods and a few deep learning-based methods (such as DestripeCycleGAN, SLDR+eNeRf). It is recommended to add comparisons with more state-of-the-art deep learning methods (such as transformer-based or diffusion-model-based image restoration methods) to fully reflect the competitiveness of the proposed method.
6. Insufficient evaluation of restoration performance under extreme imaging conditions. The paper does not test the method’s effectiveness in scenarios such as strong blur (Gaussian blur with variance greater than 2), high-intensity stripe noise (noise amplitude exceeding [-40, 40]), or extremely low-light environments, making it hard to clarify the robustness boundary of the method.

Author Response

Comments 1: For the structure-tensor-based adaptive edge-preserving operator, the paper only describes its implementation process but lacks in-depth theoretical analysis of why this operator can effectively distinguish stripe noise from image edges. For example, there is no quantitative analysis of the relationship between eigenvalues (λ₁, λ₂) and edge/stripe characteristics, nor a comparison with other edge-preserving strategies in terms of theoretical advantages.

Responses 1: Thank you for pointing this out. We agree with this comment. Therefore, we have revised the manuscript by adding a more detailed theoretical analysis below the revised manuscript Eq. (10) to explain why the proposed operator can effectively distinguish true image edges from stripe noise, thereby ensuring that edge information is not inadvertently removed during the restoration process. Specifically, we have supplemented the theoretical explanation to clarify the role of the structure tensor in differentiating edge structures and stripe artifacts. In addition, we have compared the proposed method with several commonly used edge-preserving strategies, and the differences in edge-preservation performance are visually demonstrated through representative images and quantitatively illustrated using EPI-based evaluations.

The modification can be found in the revised manuscript on Page 7, Paragraph 2.1,  Line 214-253; Page 8, Figure 3.

 

Comments 2: The parameter setting and computational efficiency analysis are insufficient. The model involves regularization parameters (such as ρ₁, ρ₂, β₁, etc.). The parameter values are all determined based on empirical verification, without providing parameter sensitivity analysis (e.g., curves showing the impact of parameter changes on SSIM and PSNR).

Responses 2: Thank you for pointing this out. We agree with this comment. Therefore, we have carefully examined the issues related to parameter selection and computational efficiency. Regarding the regularization parameters (e.g., ρ₁, ρ₂, β₁, etc.), we conducted extensive experimental investigations during the study. Specifically, after determining an approximate effective range for each parameter, a fine-grained validation was performed using small step sizes (e.g., 0.x or 0.0x). By systematically comparing the image restoration results under different parameter combinations, the parameter configuration yielding superior performance was ultimately selected. In addition, Section 3.1.1 of the manuscript provides a dedicated explanation of the parameter selection strategy. Due to space limitations, a complete set of parameter sensitivity curves could not be included in the main text. If the reviewer considers this analysis necessary for the evaluation process, we would be pleased to provide the corresponding curves and further discussion as supplementary material.

The modification can be found in the revised manuscript on Page 14, Paragraph 3.1.1,  Line 387-396.

 

Comments 3: No discussion on the impact of wavelet decomposition levels. The paper uses multi-scale wavelet transform but does not explain why the decomposition level n>1 is chosen, nor does it briefly verify whether different decomposition levels (e.g., 2-level vs. 3-level) affect stripe noise extraction and edge preservation, leaving room for optimization of this key preprocessing step.

Responses 3: Thank you for pointing this out. We agree with the reviewer’s comment. Therefore, we carefully re-examined the manuscript and confirmed that there was indeed a typographical error, in which “wavelet transform” was mistakenly written as “multi-scale wavelet variation.” This error has now been corrected. In addition, we acknowledge that the original description of the wavelet transform may have caused some confusion for readers. To address this issue, we have explicitly clarified at the end of the relevant paragraph that the proposed method ultimately adopts a single-level wavelet transform, rather than a multi-level wavelet decomposition. Furthermore, we have reorganized and revised all wavelet-related descriptions in the manuscript to improve clarity and logical consistency. Figure 2 now explicitly illustrates the subband components obtained from the single-level wavelet transform, and the subsequent text has been supplemented with a theoretical discussion on the effects of second- and third-level wavelet decompositions on the processing performance. This additional explanation is intended to help readers better understand the rationale for selecting single-level decomposition in this study.

The modification can be found in the revised manuscript on Page 5, Paragraph 2.1,  Line 178-202.

 

Comments 4: In the no-reference image experiments using SDGSAT-1 data, the paper mentions that the ICV value is higher, indicating better restoration performance, but does not explain the definition and calculation method of the ICV metric. It is recommended to supplement the formula or reference source of ICV to ensure the transparency of the evaluation process.

Responses 4: Thank you for pointing this out. We agree with the reviewer’s comment. Therefore, we have revised the manuscript by adding supplementary explanations in Section 3.1.2 of the revised manuscript, where the computation methods, definitions, and relevant references of the experimental evaluation metrics, including ICV, are clearly described. These additions make the overall evaluation procedure more transparent and easier for readers to understand, and they clarify the rationale for adopting each metric in the experimental analysis.

The modification can be found in the revised manuscript on Page 14, Paragraph 3.1.2,  Line 404-432.

 

Comments 5: The comparative experiments only include traditional model-based methods and a few deep learning-based methods (such as DestripeCycleGAN, SLDR+eNeRf). It is recommended to add comparisons with more state-of-the-art deep learning methods (such as transformer-based or diffusion-model-based image restoration methods) to fully reflect the competitiveness of the proposed method.

Responses 5: Thank you for pointing this out. We agree with the reviewer’s comment. Therefore, based on a thorough review of recent related studies, we have added an additional comparison method composed of deep learning techniques. Specifically, a combined strategy of CNN-based denoising followed by GAN-based deblurring has been incorporated into the experimental comparisons. This combined approach is implemented with reference to two recent studies: “Self-BSR: Self-Supervised Image Denoising and Destriping Based on Blind-Spot Regularization” and “Blind Image Deconvolution by Generative-Based Kernel Prior and Initializer via Latent Encoding.” A systematic comparison with this newly added method has been conducted in the experimental section, and the corresponding experimental results have been updated in the relevant tables. In the comparative framework, the RBDS method represents a hybrid strategy that integrates traditional model-based techniques with deep learning approaches. As a result, the overall comparison now achieves a more balanced composition, including five traditional methods and four deep learning-based methods. This experimental design enables a more comprehensive evaluation of performance differences across different categories of methods, further demonstrating the competitiveness of the proposed approach and providing readers with more intuitive and reliable comparative results.

The modification can be found in the revised manuscript on Page 17, Paragraph 3.2,  Line 490-492; Page 19, 20 and other related experimental diagrams.

 

Comments 6: Insufficient evaluation of restoration performance under extreme imaging conditions. The paper does not test the method’s effectiveness in scenarios such as strong blur (Gaussian blur with variance greater than 2), high-intensity stripe noise (noise amplitude exceeding [-40, 40]), or extremely low-light environments, making it hard to clarify the robustness boundary of the method.

Responses 6: Thank you for pointing this out. We agree with the reviewer’s comment. Therefore, to further evaluate the performance of the proposed algorithm under extreme imaging conditions, we conducted additional experiments using multiple sets of test cases with varying degrees of degradation. Specifically, test images with different blur strengths (Gaussian blur variances of 2, 4, 6, and 9) and different stripe noise amplitudes were constructed. Under a fixed parameter configuration, systematic experiments were performed on these severely degraded images. The corresponding results are presented and analyzed in the revised manuscript in both visual form (Figure 18) and quantitative metrics (Table 12). The experimental results demonstrate that, although image quality inevitably degrades under highly adverse conditions, the proposed method is still able to effectively suppress noise while preserving the main structural information, indicating a certain degree of robustness.

The modification can be found in the revised manuscript on Page 28, Paragraph 3.4,  Line 670-685.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors Accept in present form Comments on the Quality of English Language Accept in present form