MambaSCI: Efficient Mamba-UNet for Quad-Bayer Patterned Video Snapshot Compressive Imaging

Response to Reviewer yWFS

Thanks for your valuable comments.

Q1: Clarification of high resolution images reconstruction.

A: As stated in the main manuscript (lines 233-234), we also evaluated our method on a large-scale simulated dataset. The metrics comparison and visual comparison between our method and the comparison methods for reconstructing 1920x1080 resolution video are presented in Table 4 and Figure 6 in the main manuscript.

Q2: Noise modeling analysis of simulated datasets.

A2: We have explained how the simulated dataset was created and provided an analysis of our method's effectiveness on a real dataset:

• The middle-scale simulated dataset is created by extracting and cropping frames from high-quality open-source videos available online to obtain the groundtruth (GT), while the large-scale dataset uses videos selected from YouTube as the GT. Both datasets include noise of real-world scenes.
• Given that we are tackling a completely new task, our focus is primarily on addressing the degradation caused by SCI compression and developing effective demosaicing techniques to achieve color fidelity and eliminate artifacts for high-quality reconstruction. The aspect of noise modeling has not yet been fully considered.
• Through our studies on various Bayer-array-based methods, their performance on simulated and real datasets is positively correlated, thus I believe our method can still achieve significant advantages on real data.

Q3: Time required to reconstruct each frame.

A: In the table below, we present a comparison of the time required to reconstruct each frame, along with the PSNR and SSIM metrics for the proposed method and other competing methods. Even though our method takes longer than EfficientSCI to reconstruct each frame, it delivers significantly superior performance.

Q4: Clarification of how the masks are generated.

A: In the SCI domain, the mask typically has a specific physical significance. The basic principle involves modulating different frames within the video cube with varying weights and then integrating the light into the sensor to create a 2D measurement [1-3]. Specifically the masks are the 0-1 matrices that are different from each other in the T-frame. Also this mask is customizable subject to its physical significance and has no effect on the reconstruction results. During training, we simulate the masks of different SCI systems by randomly generating 0-1 matrices to improve the robustness of the system.

Q5: Clarification of the relationship between T and the amount of motion between consecutive frames.

A: In SCI tasks, $T$ typically represents the compression ratio ($Cr$) and is not directly related to the motion between consecutive frames. A higher $T$ indicates a higher $Cr$ and increased reconstruction difficulty. As highlighted in Table 1 in the global rebuttal, both reconstruction difficulty and the required FLOPS escalate with increasing $T$. Nonetheless, our method delivers high-quality reconstruction with superior computational efficiency.

We greatly appreciate the suggestion to use smaller compression ratios for fast motion scenes and larger ones for slow motion scenes to improve efficiency. This insight has inspired us to consider dynamically adjusting the compression ratio based on object movement speed. Thank you for this suggestion, which opens a new avenue for exploration.

Q6: Clarification of why there is no tiling done during reconstruction.

A: In the reconstruction process, we first address the SCI compression degradation in the raw domain and then perform demosaicing to convert from the raw domain to RGB. In the raw domain, two common tiling methods involves divideing into non-overlapping patches and splitting the image into four sub-parts: R, G1, G2, and B. We specify why we did not employ either of these methods:

• Dividing into Non-overlapping Patches. Dividing an image into non-overlapping patches will lead to localized detail loss. To address this, we employ convolution for shallow feature extraction, which enables us to train effectively on low-resolution images (using a resolution of 128x128 for training and 256x256 for fine-tuning) and process on high resolution images.
• Dividing into Four Sub-parts. This method is efficient for Bayer arrays due to their physical arrangement. However, the quad-Bayer's unique layout poses challenges. Splitting the image into R, G1, G2, and B components requires a more intricate process because each color forms a 2x2 square, which must be considered as a whole rather than isolated pixel by pixel. We attempted this approach, but found it significantly increased both training and inference times.

Thank you for your question, and we will continue exploring efficient tiling methods to achieve lightweight reconstruction for higher-resolution videos.

[1] Yuan X, Liu Y, Suo J, et al. "Plug-and-play algorithms for large-scale snapshot compressive imaging." CVPR 2020.

[2] Yuan X, Liu Y, Suo J, et al. "Plug-and-play algorithms for video snapshot compressive imaging." IEEE TPAMI 2021.

[3] Yuan X, Brady D J, Katsaggelos A K. "Snapshot compressive imaging: Theory, algorithms, and applications." IEEE SP 2021.

Response to Reviewer xnVL

Thanks for your valuable comments.

Q1: Clarification of the novelty.

A: We have elaborated on the novelty in lines 160-167 and 289-296 of the manuscript. Rather than merely adapting ST-Mamba, we introduced the following innovations in video SCI reconstruction:

• High Performance Lightweight Network Design: Although STMamba is an effective lightweight global attention mechanism approach, it still struggles with the high demands of SCI video reconstruction compared to Vivim's medical image segmentation. Our framework enables joint global-local reconstruction, addressing these challenges.
• Combined Global-Local Features. DWConv efficiently extracts local features, avoiding the parameter and complexity issues of common 3D convolutions. Meanwhile, it is not a straightforward incremental combination with STMamba. By integrating linear transformations with DWConv features, the network integrates local and global information, enhancing the global features extracted by STMamba. This fusion is crucial for accurately capturing complex edge structures.
• Enhanced Channel Interaction: Many Mamba-based frameworks overlook channel interaction limitations, leading to channel isolation and suboptimal information fusion, which adversely affects reconstruction quality. To address this, we introduce a lightweight channel attention mechanism to improve multi-channel perception and feature representation, refining reconstruction quality.

Q2: In-depth description of EDR module.

A: We have added a detailed analysis of the EDR module's design principles in lines 184-192, explaining how it enhances video reconstruction by improving the representation of depth features. To better illustrate the EDR module's operation, we redraw the relevant section of Figure 4 in the manuscript. Additionally, further analysis of edge detail reconstruction attributed to the EDR module has been added to the ablation experiment section, specifically in lines 297-299. We additionally provide a visual comparison of the edge detail effect with and without the EDR module on the final reconstruction result (Figure 3 on pdf).

Specifically, key improvements with EDR include:

• DWConv : This technique performs spatial convolution independently on each channel and then applies pointwise (1x1) convolution across channels. It effectively captures localized spatial features, enhancing edge detail perception.
• GELU : Compared to ReLU, GELU handles negative inputs more naturally, improving the model's ability to represent fine details.
• Adaptive Weight Initialization: Truncated normal initialization helps the model capture details more effectively during the early training stages.
• Multi-scale Feature Fusion: Combining linear transformations with DWConv features allows the network to extract global information from local features. This simultaneous processing of global and local information is crucial for understanding complex edge structures.

Q3: Clarification the reasons for inconsistent improvement effects on large-scale datasets.

A: The inconsistencies may be attributed to:

1. Evaluation in Raw Domain: As described in lines 494-500 of the Supplementary Material, we use additional demosaicing model to show RGB images reconstructed by GAP-TV, PnP-FFDnet, and PnP-FastDvDnet, which ultimately reconstruct videos in the raw domain. In contrast, three E2E methods reconstruct RGB video directly. For a fair comparison and excluding the influence of demosaicing, we assess performance in the raw domain. Thus, the RGB video from E2E methods is converted back to the raw domain using a quad-Bayer mask. Figure 6 in manuscript shows that while iterative methods perform better in metrics, they exhibit more visual artifacts.
2. Training Resolution: The training resolution is 128x128 and fine-tuned at 256x256, which is a gap with large-scale datasets.
3. High-Speed Frame Reconstruction: The effect of high-speed moving frame reconstruction is to be strengthened under the limitation of the number of parameters and computational complexity. The number of parameters and flops of the model are as follows. Even with the limited FLOPs, we still achieve better performance.

Q4: Clarification of why Mamba and not self-attention.

A: As stated in our response to Reviewer M9Ao's Q1, from both the technical perspective, including performance metrics and comparisons of Params and FLOPs as shown in Table 2 of main manuscript, and the perspective of pattern processing, where previous methods based on Transformers and CNNs have struggled with artifacts and color distortions when dealing with quad-Bayer arrays. Mamba has proven to be a more suitable choice for addressing the new challenges we face.

Q5: Clarification of why quad-Bayer pattern is better than a single Bayer-patterned measurement?

A: As noted in response to Reviewer M9Ao's Q2,exploring quad-Bayer-based SCI tasks is both necessary and advantageous. Due to the quad-Bayer's physical structure, where each color is represented by a 2x2 grid, it offers higher resolution, better light intake in low-light conditions, and supports HDR video through different exposure settings.

Q6: Comparison on long videos.

A: We have conducted new experiments to further validate our method's effectiveness with long sequence videos and also provided corresponding visual comparison images (Figure 4 on PDF) and a performance comparison table (Table 1 in global rebuttal). The default $T$ is set to 8, meaning reconstruction is performed after compressing 8 frames of video into 2D observations. To test our method's performance with longer sequences, we also evaluated it with $T = 16$ and $T = 32$.

Thanks for your valuable comments.

Q1: Clarification about comparison on long videos.

A1: Regarding your questions about run-time and performance, we'll address them as follows:

• Performance: As the model's FLOPs increase with $T$, our method maintains only 9.8% of the FLOPs required by EfficientSCI. While achieving comparable PSNR to EfficientSCI, our method significantly outperforms it in SSIM metrics. Additionally, as shown in the visual comparison in Figure 4 of the submitted PDF, our method provides superior visualization compared to EfficientSCI.
• Run-time: The reason why our method's runtime is not as competitive as EfficientSCI's can be attributed to two key factors:
  • Mamba Characteristics. Mamba's acceleration is tied to the GPU's hardware characteristics and the underlying CUDA architecture, limiting our ability to enhance its speed directly.
  • Network Framework Optimization. To achieve high-quality reconstruction, our framework currently uses 3D convolution for channel feature interactions at the end of each Residual-Mamba-Block. However, this process becomes more time-intensive as $T$ increases. We plan to optimize the framework further in the future to develop a lightweight, high-quality, and faster reconstruction algorithm.

Q2: Clarification of the innovation of DWConv.

A2: The depth-wise convolution used by Vivim [1] is essentially a 3D depth-wise convolutional layer, which focuses solely on spatial feature extraction. It lacks a pointwise convolution component, meaning it doesn't account for the relationships between channels. In contrast, the DWConv we designed integrates both depthwise and pointwise convolution. First, it processes spatial information through depthwise convolution, then it combines the relationships between channels via pointwise convolution to produce the final output features. This design allows us to effectively extract local detailed features and combinations between channels while reducing computational costs and the number of parameters.

Thank you for your insightful feedback. We will update the DWConv section of Figure 4 in the main manuscript accordingly and provide a more detailed explanation of DWConv in lines 186-187. Additionally, we will include further analysis in the ablation study in Section 4.5 to enhance the clarity and robustness of our findings.

[1] Yang Y, Xing Z, Zhu L. Vivim: a video vision mamba for medical video object segmentation[J]. arXiv preprint arXiv:2401.14168, 2024.

Response to Reviewer zshs

Thanks for your valuable comments.

Q1: Differences between the proposed method and the video enhancement-based ones.

A: MambaSCI significantly differs from video enhancement technology in the nature of the task, input differences, and use of prior knowledge, as detailed below:

• Nature of the Task: Video enhancement tasks typically improve RGB video quality by addressing noise or missing frames. In contrast, the MambaSCI network addresses the more complex SCI task, which involves not only overcoming compression degradation but also handling the additional effects of the quad-Bayer array mask. Thus, the MambaSCI network has a dual mission: accurate reconstruction of compressed data and conversion and demosaicing from the raw domain to standard RGB format, requiring highly flexible and capable of managing the unique complexities of SCI technology.
• Huge Gap Between the Input of Two Tasks: Same to previous models like STFormer and EfficientSCI, our MambaSCI reconstruction network also requires an initialization block for up-sampling, which is a crucial component of the SCI reconstruction task. However, the $\mathbf{x_{in}}$ obtained after initialization differs significantly from traditional video enhancement inputs. We present $\mathbf{x_{in}}$ after initialization in Figure 1 in supplemental PDF, where it is evident that the degradation is much more severe than what common video enhancement networks can handle.
• Utilization of Different Prior Knowledge: In the reconstruction process, MambaSCI leverages the unique priori knowledge of compression and quad-Bayer arrays, contrasting sharply with traditional video enhancement’s traditional focus on smoothness and edge preservation.

Q2: Clarification of the quad-Bayer pattern is not characterized in the same way as in previous work.

A: Both STFormer and EfficientSCI were designed as Bayer array-based video compression reconstruction methods, only considered Bayer pattern features in their initialization blocks. In contrast, we have developed a specialized initialization block to address the unique characteristics of quad-Bayer physical arrays, and applied it specifically to GAP-TV, PnP-FFDnet, and PnP-FastDVDnet. Details are as follows:

• GAP-TV, PnP-FFDnet, and PnP-FastDVDnet use iterative optimization by dividing the raw domain into R, G1, G2, and B blocks for independent reconstruction. Similarly, we divided the quad-Bayer array into R, G1, G2, and B sub-blocks as shown in lines 103-107 in the main manuscript. However, due to the quad-Bayer pattern's 2x2 grid representation for each color, processing is significantly more complex than standard Bayer pattern, resulting in an increased processing time from 0.0015s to 7.3312s at 512x512 resolution compared to the SCI generic initialization method.
• Leveraging the powerful modeling capabilities of the E2E model and considering the need for a lightweight design due to the limited computing power of future handsets, we selected the SCI generic initialization block for STFormer, EfficientSCI, and MambaSCI. Our experiments confirmed nearly no performance loss with this choice.

Figure 1 of our PDF submission provides a visual comparison of $\mathbf{x_{in}}$ obtained using the SCI generic initialization block versus an initialization block designed specifically for the physical characteristics of a quad-Bayer array.

Q3: Enhancing in-depth analysis of certain modules.

A: We have added a deeper analysis of the EDR module, bottleneck layer, and the decoding layer in the following areas:

• More Detailed Analysis: We have thoroughly analyzed the EDR module, detailing its inner workings and its critical role in our framework. Additionally, we have improved the presentation and analysis of the bottleneck and decoding layers.
• Figure-Text Interaction: We have redrawn Figures 3 and 4 in the main manuscript and improved their integration with the text to facilitate better understanding and clarity for the reader.
• Corresponding to the Ablation Experiment: The module description is linked to the associated ablation experiments, allowing the reader to intuitively grasp the module's important role in the network.

Q4: Lack of real experiments on videos captured by quad-Bayer patterns.

A: The reasons for the lack of real experiments are as follows:

• Since we are exploring a new task and a new discovery that is not currently being investigated, there are no publicly available datasets for comparison, as stated by reviewer yWFS (there is no real SCI dataset based on quad-Bayer pattern).
• To obtain the real-world datasets, a corresponding optical coding system must be assembled. We are working on building this system, but it will take six months to one year to complete, even with experienced personnel. Additionally, we need to upgrade the existing system by replacing the current modules with four-layer modules.
• To demonstrate the advantages of quad-Bayer over Bayer, we plan to acquire more datasets from extreme scenes and collect HDR data. This will significantly expand the current research field of SCI.

Q5: Analysis of artifacts and color distortion in the lower left corner of "Swinger".

A: Thanks to your careful observation, we have added the visual comparison in Figure 2 of the submitted PDF, showing different methods to locally zoom in on the lower left corner of the "swinger." The intricate ropes in this area, as depicted in the GT, present a significant challenge due to their dense packing and rapid movement between frames, leading to ghosting artifacts in the reconstruction.

To address this limitation, we plan to refine our model further. In future work, we will focus on improving the model's ability to handle fast motion between frames, minimizing artifacts and achieving more accurate reconstructions in complex and dynamic scenes.

Thanks for your valuable comments.

Q1: Clarification on the lack of real experiments using videos captured by quad-Bayer patterns.

A1: As we are pioneering a new task in the field of video SCI, which is still in its exploratory stage, no public dataset currently exists for this class. Meanwhile, our main innovations include:

• New task: Quad-Bayer offers significant advantages over traditional Bayer sensors, such as higher resolution and HDR capabilities, and is widely used in smartphone cameras. However, there have been no studies on its application in video SCI. We are the first to introduce quad-Bayer into video SCI, aiming to expand the field and achieve higher quality reconstruction.
• New method: To enable lightweight reconstruction, we are the first to utilize an asymmetric U-shaped architecture, employing customized Residual-Mamba-Blocks modules to achieve efficient, high-quality video reconstruction.

Furthermore, constructing the corresponding SCI encoding system requires specialized optical encoding hardware and significant time, which is often beyond the capabilities of researchers focused on decoding algorithms.

We are working on building and refining this system to fully leverage the advantages of quad-Bayer over Bayer, such as capturing video in low-light environments and compressing HDR video SCI by adjusting pixel exposure within the 2x2 color grid.

Q2: Clarification of swinger's visual comparison.

A2: As detailed in lines 494-500 of the Supplementary Material of main manuscript, we utilized the high-performance BJDD [1] model for joint denoising and demosaicing to display the RGB images reconstructed by PnP-FastDvDnet. Specifically, PnP-FastDvDnet only handles Raw domain reconstruction, while BJDD performs further denoising and demosaicing to obtain the final RGB video. In contrast, our model integrates both Raw domain reconstruction and demosaicing. As a result, the visual quality of PnP-FastDvDnet may appear superior due to the BJDD network's influence. Therefore, as described in the main text, to ensure a fair comparison, we converted the RGB video output from MambaSCI back to the Raw domain and calculated metrics such as PSNR and SSIM in the Raw domain. The performance metrics of MambaSCI vs. PnP-FastDvDnet on swinger in Raw domain are compared as follows:

As seen in the table, our method outperforms PnP-FastDVDnet in both PSNR and SSIM in the Raw domain. Additionally, as shown in Figure 6 and Figure 15 of the main manuscript, our approach achieves superior reconstruction in other detailed areas compared to PnP-FastDVDnet.

[1] A Sharif S M, Naqvi R A, Biswas M. "Beyond joint demosaicking and denoising: An image processing pipeline for a pixel-bin image sensor." CVPR 2021.

Thanks for your valuable comments.

Q1: Real System.

A1: Since we are exploring a new task in the SCI field, our focus has been on addressing issues like artifacts and color distortion when dealing with quad-Bayer array video. We appreciate your insight into the potential challenges that might arise in real-world scenarios. We will take your suggestions into account and work on developing a quad-Bayer-based coding system as soon as possible.

Q2: Reconstruction Quality.

A2: Our proposed E2E network strikes a balance between performance and efficiency. As demonstrated in Figure 5 of the main manuscript, at a resolution of 512x512, our method produces superior visual results compared to the cascaded "PnP-FastDvDnet + BJDD" combination. Our approach maintains high color fidelity without introducing additional artifacts, whereas the "PnP-FastDvDnet + BJDD" combination suffers from severe artifacts and shape distortions across multiple datasets. Also, we provide a comparison of the corresponding performance metrics as well as a comparison of the reconstruction times in the table below:

Table 1: A comparison of performance metrics (PSNR (dB), SSIM) and reconstruction time between PnP-FastDVDnet and MambaSCI. (Note: The time for PnP-FastDvDnet does not include the additional time required for the demosaicing process performed by BJDD.)

Table 1 clearly highlights the superior performance of our approach compared to PnP-FastDVDnet, along with a significant advantage in the time required for reconstruction.

Additionally, we are actively exploring improvements to the pipeline and have identified the following areas for enhancement:

• Raw Domain Reconstruction Performance: While our method performs well on middle-scale datasets, it does not achieve superior metrics on large-scale datasets, primarily due to the significant difference in resolution sizes between training (128*128, and 256*256 for finetuning) and testing (1920*1080). We are trying to further investigate multi-scale generalization approaches to address this issue.
• Improved Demosaicing Algorithms: Similar to previous Bayer-based reconstruction methods, and for lightweight considerations, our current approach utilizes 3D convolution for the final demosaicing operation. However, the distinct structure of the quad-Bayer pattern makes it more challenging to demosaic compared to traditional Bayer arrays. This complexity can often result in color confusion and artifacts during the demosaicing process. Recognizing this limitation, we have designed a lightweight demosaicing network that has already shown promising results in preliminary experiments. We will continue to refine and explore this approach further.

Response to Reviewer M9Ao

Thank for your valuable comments.

Q1. Clarification of why Mamba and not transformers or CNNs.

A: We have added further discussion of this topic in lines 59-61 of main manuscript, with a more detailed analysis and experimental validation as follows:

Transformer and CNN-based Video Snapshot Compressive Imaging methods have been proposed such as transformer-based method STFormer and the hybrid CNN- and transformer-based method EfficientSCI. However, existing methods still face two issues:

1. Technique perspective:
   • The high computational complexity of Transformers and the lack of a global attention mechanism in CNNs hinder their extension into modern, lightweight, and efficient network architectures. Therefore, we focus on exploring multi-scale reconstruction of Mamba-Unet into the Quad-Bayer pattern to achieve lightweight design and enable deployment on mobile devices.
   • As shown in Table 2 of main manuscript, MambaSCI-B outperforms STFormer and EfficientSCI, using only 31% of STFormer’s parameters and 4.5% of its FLOPS, and 69% of EfficientSCI's parameters and 9.8% of its FLOPS.
2. Pattern perspective: All existing video SCI methods are designed based on the traditional Bayer pattern. When applied to videos captured by quad-Bayer cameras, these methods often result in color distortion and ineffective demosaicing, rendering them impractical for primary equipment. Thus, we aim to solve these issues with two key contributions: new task and new method. That is, efficient and qualitative recovery of quad-Bayer arrays through the asymmetric Mamba-Unet framework. To the best of our knowledge, we are the first one to develop the Quad-Bayer Patterned Video Snapshot Compressive Imaging task and the first algorithm to elaborate designs of Mamba-UNet to this task.

Q2. Clarification of the necessity and advantages for quad-Bayer .

A: Quad-Bayer arrays are prevalent and almost all current flagship smartphones cameras, such as the iPhone 14 Pro/Max, vivo X90 Pro+, Xiaomi 13S Ultra, and OPPO Find X6 Pro, utilize quad-Bayer arrays. We have added a description of the main applications where quad-Bayer is used and its advantages over Bayer in lines 35-38 of the main manuscript. Meanwhile, integrating quad-Bayer array into video SCI not only is necessary but also offers significant advantages over Bayer arrays:

1. Necessity:
   • Widespread Use on Smartphones: Quad-Bayer arrays are common in smartphone cameras [1], which are frequently used for video recording. Existing methods face color distortion and artifacts, which our work aims to provide a customized solution for Quad-Bayer color video SCI tasks to overcome these issues while offering space-saving, high-quality reconstruction through compression and lightweight construction on smartphones.
   • New Industrial Demands and Research Trends: Bayer arrays have been widely used in industry and well-studied in academia due to their long history, performing adequately in bright scenes. However, the imaging capabilities in low-light conditions are limited. Therefore, exploring quad-Bayer arrays is essential to address these deficiencies.
2. Advantage: Quad-Bayer arrays has the following advantages over Bayer arrays:
   • Higher Resolution: Quad-Bayer provides higher resolution than Bayer arrays by realizing higher pixel density on the sensor for better detail reproduction compared to conventional Bayer arrays [2].
   • Low-Light Performance Enhancemente: Quad-Bayer collects more light in low-light environments, improving the sensor's signal-to-noise ratio and reducing noise interference.
   • HDR Capability: By adjusting quad-Bayer sub-pixels and setting different exposure values on different sub-pixels, high dynamic ranging (HDR) images or videos can be captured to improve the detail performance of highlights and shadows [3-4].
   • Color Accuracy: Quad-Bayer pattern captures richer color information by sampling red, green and blue colors in four directions. This fine sampling reduces the likelihood of color distortion and improves the color accuracy and realism of the image [5].

Therefore, we have introduced quad-Bayer arrays to video SCI for the first time, providing a solution for reconstructing higher-quality videos even HDR videos in the future.

Q3. Section 3 lacks clarity.

A: To address this, we have made the following changes to the paper:

• Top-Down Structure: We use a top-down structure to introduce our framework, providing an overview followed by detailed component breakdowns. This approach ensures a clear understanding of the overall algorithm and the specific functions of each individual part.
• Clear Delineation: We clearly delineate Section 3.2 to provide a detailed analysis of each module, including a thorough examination of the internal structure and a clear explanation of each module's function.
• Figure-Text Interaction: Figures 3 and 4 in the main manuscript have been revised to align with the content and more effectively present the network framework and internal modules details.
• Contextual Analysis: We ensured the model description is consistent with Section 4’s experimental setup and interacts with the Supplementary Material’s pseudo-code to aid reproduction.

[1] Madhusudana P C, et al. "Mobile Aware Denoiser Network (MADNet) for Quad Bayer Images" CVPRW 2024.

[2] Zheng B, et al. "Quad Bayer Joint Demosaicing and Denoising Based on Dual Encoder Network with Joint Residual Learning" AAAI 2024.

[3] Kim J, Kim M H. "Joint demosaicing and deghosting of time-varying exposures for single-shot hdr imaging." ICCV 2023.

[4] Wu T, et al. "High Dynamic Range Imaging with Multi-Exposure Binning on Quad Bayer Color Filter Array. " ICIP 2023.

[5] Lee H, et al. "Efficient unified demosaicing for bayer and non-bayer patterned image sensors." ICCV 2023.

We sincerely appreciate your thorough review of our paper, careful consideration of our rebuttals, and raising the score. If you have any further questions or concerns, we are always available to provide additional clarification as necessary.