An underwater image super-resolution reconstruction method based on dynamic window and state space cooperation

Abstract

The application is an underwater image super-resolution reconstruction method based on dynamic window and state space cooperation. The method comprises a shallow feature extraction module, a deep feature extraction module and an image reconstruction module, through the construction of a double-path feature interaction mechanism, the alternative learning of a dynamic window attention module and a visual state space module is adopted, the cooperative optimization of local detail enhancement and global context modeling is realized. A dynamic channel gate feedforward network is innovatively designed, through a channel weight dynamic adjustment mechanism guided by a global average pooling, adaptive fusion of multi-scale features is realized. The method effectively solves the problems of blur and contrast reduction of underwater images caused by scattering effect, the PSNR is 28.21dB in the x3 task of the USR-248 dataset, which is 0.35dB higher than that of SRFormer, the experimental results show that the model has excellent performance in objective evaluation index and subjective evaluation, and remarkable effect is achieved in detail recovery, which has significant technical effect and industrial application value.

Description

Technical Field

[0001] This invention relates to the field of super-resolution technology, and is a method for underwater image super-resolution reconstruction based on dynamic window and state space collaboration, which converts low-resolution underwater images into high-resolution underwater images. Background Technology

[0002] With the deepening of marine exploration and the steady advancement of the national marine strategy, underwater imaging technology [1] plays an important role in marine resource development, environmental monitoring and biological research. However, due to the complexity of the underwater environment and the limitations of imaging equipment performance, the acquired underwater images often suffer from problems such as low resolution, color distortion and blurred details. The root cause of these problems is the scattering and absorption effect of underwater light, which leads to a significant decrease in image quality and seriously affects subsequent analysis and application. Therefore, how to effectively improve the resolution and quality of underwater images has become a hot research topic.

[0003] Single-image super-resolution (SISR) techniques aim to reconstruct high-resolution images from low-resolution images and have made significant progress in natural scene image processing in recent years. However, the unique characteristics of the underwater environment make traditional SISR methods difficult to apply directly: on the one hand, underwater images exhibit complex degradation patterns, including light attenuation, color shift, and noise interference; on the other hand, existing methods struggle to effectively balance global context modeling and local detail restoration, resulting in suboptimal reconstruction results. Although methods based on convolutional neural networks (CNNs) and Transformers perform well in image super-resolution tasks, they still face challenges in underwater scenes: CNN methods suffer from insufficient global color correction capabilities due to their difficulty in modeling long-range dependencies; while details can be enhanced by stacking residual blocks, the high-frequency texture restoration effect is limited by the local receptive field of the convolutional kernel. Summary of the Invention

[0004] This invention addresses the image blurring and contrast reduction caused by underwater image scattering effects by proposing an underwater image super-resolution method based on dynamic window and state space collaboration. It comprises three modules: shallow feature extraction, deep feature extraction, and image reconstruction. A dual-path feature interaction mechanism is constructed, achieving a balance between local detail enhancement and global context modeling through alternating learning of dynamic window attention and state space models. Simultaneously, an adaptive residual correction strategy is introduced, utilizing learnable weights to dynamically fuse multi-scale features.

[0005] A method for super-resolution reconstruction of underwater images based on dynamic window and state space coordination, the method comprising the following steps:

[0006] Step 1: Acquire low-resolution underwater images using underwater imaging equipment, and perform degradation feature analysis using Gaussian blur and bicubic downsampling;

[0007] Step 2: Construct a super-resolution network that includes a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module;

[0008] Step 3: Construct a dual-path feature interaction mechanism in the deep feature extraction module, and achieve local detail enhancement and global context modeling through the alternating learning of the dynamic window attention module and the state space module;

[0009] Step 4: In the feature fusion stage, a channel-gated feedforward network is introduced to dynamically fuse multi-scale features using learnable weights, thereby realizing an adaptive residual correction strategy.

[0010] Step 5: Generate high-resolution images using a multi-scale reconstruction module combined with Pixel Shuffle operation;

[0011] Step 6: Improve the mixing loss function by weighted chromaticity loss.

[0012] Preferably, step 1, the data acquisition stage, specifically comprises:

[0013] (1) Using an integrated underwater vehicle, the camera was securely fixed on a suitable underwater platform and fitted with a polarizing filter system. This system ensured that the camera could shoot under different polarization directions. To fully describe the polarization state of light, three representative polarization directions were selected: 0°, 45°, and 90°. Before the actual shooting, a uniformly lit environment was carefully selected to minimize interference with polarization information. During the shooting process, the polarizing filter was rotated sequentially to the predetermined polarization directions of 0°, 45°, and 90° to shoot the same underwater scene, thereby obtaining multiple images with different polarization directions. Finally, by comprehensively processing these images, the underwater image closest to the real situation was obtained.

[0014] (2) Gaussian blur and bicubic downsampling degraded image features were applied to the acquired underwater images to obtain a low-resolution underwater image dataset. Gaussian blur is performed by weighted averaging of the images using a normal distribution convolution kernel. The specific formula is as follows:

[0015]

[0016] in The standard deviation is used to control the blur intensity, and x and y represent the coordinates of each element in the convolution kernel relative to the center point, used to calculate the weight value at that location. The pixel value for each sampling point in the image is constructed by calculating the weighted average of the surrounding 5×5 pixels, with the weights decreasing exponentially with distance from the center point.

[0017] (3) The bicubic downsampling process is implemented using a bicubic interpolation algorithm. Interpolation is performed on the 16 nearest neighbor pixels of the point to be sampled using a bicubic polynomial. The formula for bicubic interpolation is as follows:

[0018]

[0019]

[0020] Where t represents the normalized distance t ), is the normalized distance between the target interpolation point and the reference pixel, representing the position offset of the interpolation point relative to the reference pixel grid. 'a' represents the curve sharpness control parameter, a key parameter that adjusts the shape of the interpolation curve and controls the smoothness of the weight distribution.

[0021] Preferably, step 2, which constructs a super-resolution network comprising a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module, specifically involves:

[0022] (1) In the shallow feature extraction stage, the input image is first... The input is mapped to a high-dimensional feature space by performing a 3×3 convolution followed by a tensor reshaping operation, where B represents the batch, C represents the number of channels, and H and W represent the image height and width, respectively.

[0023] (2) The deep feature extraction stage involves stacking N residual groups, with dynamic spatial attention blocks and Mamba state space modules alternately deployed in each residual group;

[0024] (3) In the image reconstruction stage, 3×3 convolution and Pixel Shuffle multi-scale reconstruction are used to achieve super-resolution reconstruction of the target image.

[0025] Preferably, step 2, which constructs a super-resolution network comprising a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module, specifically involves:

[0026] In the shallow feature extraction stage, the input image is first... The input is mapped to a high-dimensional feature space by 3×3 convolution followed by tensor reshaping, where B represents the batch size, C represents the number of channels, and H and W represent the image height and width, respectively. In the deep feature extraction stage, N residual groups are stacked, and dynamic spatial attention blocks and Mamba state space modules are alternately deployed in each residual group. In the image reconstruction stage, 3×3 convolution and pixel shuffle multi-scale reconstruction are used to achieve super-resolution reconstruction of the target image.

[0027] Preferably, the specific implementation of step 3 in the deep feature extraction module to construct the dual-path feature interaction mechanism is as follows:

[0028] (1) Construct a dynamic window attention module, using dynamic window partitioning units, and alternately employing horizontal and vertical window segmentation strategies based on the hierarchical index, dynamically selecting the window direction based on the hierarchical index. When the residual group index is even, the input feature map is... Divided into A horizontal window, switched to an odd index. The vertical window alternately captures horizontal and vertical texture features, and then expands the local receptive field through cyclic displacement operations.

[0029] (2) The position-aware attention calculation stage adopts dual-path parallel processing, and the window features are divided for each window block. During the calculation query ( ),key( ) and value ( ) matrix, and inject dynamic position bias. The formula is shown below:

[0030]

[0031]

[0032] in , .

[0033] (3) During feature fusion, the horizontal window attention output is... and vertical window results Stitching along the channel dimension:

[0034]

[0035] (4) Then add parallel depthwise convolution paths to compensate for high-frequency details, as shown in the following formula:

[0036]

[0037] (5) Perform dual-path attention collaborative computation, as follows:

[0038]

[0039] Among them, the gating weight Generated by global average pooling.

[0040] (4) The visual state space module adopts a dual-channel feature processing architecture. First, the input tensor X is projected in two ways, and then it is divided into two sub-tensors x and y equally along the channel dimension. x is processed independently by 3x3 grouped convolution depthwise separable convolution, and a non-linear SiLU activation function is introduced to enhance the feature expression capability.

[0041] (5) A 2D selective scanning module is used to transform the two-dimensional image into a multi-directional one-dimensional feature sequence through spatial scanning in four symmetrical directions. Each pixel is acquired along the diagonal and horizontal and vertical directions, and four sets of direction-sensitive feature streams are formed by combining contextual information.

[0042] (6) The feature flow is used as the input of the S6 block to perform multi-directional feature fusion. The processed feature flow is reshaped and merged to form the output graph.

[0043] Preferably, step 4 introduces a channel-gated feedforward network in the feature fusion stage, specifically as follows:

[0044] Dynamic channel weights are generated using Global Average Pooling (GAP). The calculation formula is:

[0045]

[0046] Among them, feature map B represents the batch size, N represents the product of the image height and width, and C represents the number of channels. The matrix is ​​a learnable linear weight matrix, with sigmoid as the activation function and MLP as the multilayer perceptron. The weights dynamically adjust the importance of each channel based on the global statistical properties of the input features. Hadamard products are used to weight the feature maps channel-by-channel, achieving data-driven feature selection.

[0047] Preferably, step 5 employs a 3×3 convolution and a pixel reconstruction module to convert the channel dimension into a spatial dimension, thereby achieving upsampling of the feature map from low resolution to high resolution. The specific steps are as follows:

[0048] (1) Decompose the number of channels: Decompose the number of channels C of the input feature map into Where C is the new number of channels and r is the magnification.

[0049] (2) Increase spatial dimension: Change the dimension of the input feature map from C×H×W to .

[0050] (3) Rearrange the feature map: Move the r×r blocks to the spatial dimension to form a feature map of size r×r. High-resolution feature maps.

[0051] Preferably, step 6 optimizes the network parameters using a hybrid loss function, specifically as follows:

[0052] (1) The L1 absolute error loss function provides a smoother optimization direction in the early stage of training and has a strong ability to correct the blue-green bias in underwater images. The specific formula is as follows:

[0053]

[0054] H and W represent the number of channels, height, and width of the image, respectively. IHR represents the true high-resolution image tensor, and ISR represents the high-resolution image tensor generated by the network.

[0055] (2) The L2 mean squared error loss function is used to measure the difference between the predicted value and the true value. The specific formula is as follows:

[0056]

[0057] Let i be the true value of the i-th sample. is the corresponding predicted value, and N is the total number of samples.

[0058] (3) Improved mixing loss function by weighted chromaticity loss L1 loss is applied to the luminance channel Y. =0.7), and L2 loss is applied to the chroma channel UV ( =0.3) The specific formula is as follows:

[0059] Attached Figure Description

[0060] Figure 1 This is the overall flowchart of the model of the present invention;

[0061] Figure 2 This is a diagram of the overall network structure of the present invention;

[0062] Figure 3 This is a structural diagram of the dynamic spatial attention mechanism of the present invention;

[0063] Figure 4 This is a structural diagram of the 2D selective scanning module of the present invention;

[0064] Figure 5 This is a diagram of the channel-gated feedforward network structure of the present invention; Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0066] The following detailed description of an underwater image super-resolution reconstruction method based on dynamic window and state space collaboration, in conjunction with specific embodiments, provides an explanation of the present invention.

[0067] The method is specifically implemented by the following steps.

[0068] Step 1: Acquire low-resolution underwater images using underwater imaging equipment, and perform degradation feature analysis using Gaussian blur and bicubic downsampling;

[0069] An integrated underwater vehicle was used to securely mount the camera on a suitable underwater platform, and a polarizing filter system was added. This system ensured that the camera could capture images under different polarization directions. To fully describe the polarization state of light, three representative polarization directions—0°, 45°, and 90°—were selected. Before the actual shooting, a well-lit environment was carefully chosen to minimize interference with polarization information. During shooting, the polarizing filter was rotated sequentially to the predetermined 0°, 45°, and 90° polarization directions, and images of the same underwater scene were captured, resulting in multiple images with different polarization directions. Finally, these images were processed together to obtain the underwater image that most closely resembled reality.

[0070] Gaussian blur and bicubic downsampling degrade image features were applied to the acquired underwater images to obtain a low-resolution underwater image dataset. The Gaussian blur was performed by applying a weighted average to the images using a normal convolution kernel; the specific formula is as follows:

[0071]

[0072] in The standard deviation is used to control the blur intensity, and x and y represent the coordinates of each element in the convolution kernel relative to the center point, used to calculate the weight value at that location. The pixel value for each sampling point in the image is constructed by calculating the weighted average of the surrounding 5×5 pixels, with the weights decreasing exponentially with distance from the center point.

[0073] Secondly, bicubic downsampling is employed. Interpolation is performed on the 16 nearest neighbor pixels of the point to be sampled using a bicubic polynomial. The formula for bicubic interpolation is as follows:

[0074]

[0075]

[0076] Where t represents the normalized distance t ), is the normalized distance between the target interpolation point and the reference pixel, representing the position offset of the interpolation point relative to the reference pixel grid. 'a' represents the curve sharpness control parameter, a key parameter that adjusts the shape of the interpolation curve and controls the smoothness of the weight distribution.

[0077] Step 2: Construct a super-resolution network that includes a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module, such as... Figure 2 As shown;

[0078] In the shallow feature extraction stage, the input image is first... The input is mapped to a high-dimensional feature space by 3×3 convolution followed by tensor reshaping, where B represents the batch size, C represents the number of channels, and H and W represent the image height and width, respectively. In the deep feature extraction stage, N residual groups are stacked, and dynamic spatial attention blocks (DSAB) and Mamba state space modules (MAB) are alternately deployed in each residual group. In the image reconstruction stage, 3×3 convolution and pixel shuffle multi-scale reconstruction are used to achieve super-resolution reconstruction of the target image.

[0079] Step 3: Construct a dual-path feature interaction mechanism, achieving local detail enhancement and global context modeling through alternating learning of the dynamic window attention module and the state space module, such as... Figure 3 As shown;

[0080] A dynamic window attention module is constructed, using dynamic window partitioning units. Horizontal and vertical window segmentation strategies are alternately employed based on the hierarchical index, and the window orientation is dynamically selected according to the hierarchical index. When the residual group index is even, the input feature map is... Divided into A horizontal window, switched to an odd index. The vertical window alternately captures horizontal and vertical texture features, and then expands the local receptive field through cyclic displacement operations.

[0081] The location-aware attention computation stage employs dual-path parallel processing, dividing each window block into window features. During the calculation query ( ),key( ) and value ( ) matrix, and inject dynamic position bias. The formula is shown below:

[0082]

[0083]

[0084] in , .

[0085] The horizontal window attention output is used during feature fusion. and vertical window results Stitching along the channel dimension:

[0086]

[0087] Then, a parallel depthwise convolution path is added to compensate for high-frequency details, as shown in the following formula:

[0088]

[0089] Next, we perform dual-path attention collaborative computation, as shown in the following formula:

[0090]

[0091] Among them, the gating weight Generated by global average pooling.

[0092] Subsequently, the visual state space module adopts a dual-channel feature processing architecture. First, the input tensor X is projected in two paths, and then it is equally divided into two sub-tensors x and y along the channel dimension. x is processed independently by 3x3 grouped depthwise separable convolution, and a non-linear SiLU activation function is introduced to enhance the feature representation capability.

[0093] Feature extraction employs a 2D selective scanning module, such as... Figure 4 As shown, a two-dimensional image is transformed into a multi-directional one-dimensional feature sequence through spatial scanning in four symmetrical directions. Each pixel is acquired along the diagonal and horizontal / vertical directions, and contextual information is incorporated to form four sets of direction-sensitive feature flows.

[0094] The feature stream is used as input to the S6 block for multi-directional feature fusion. The processed feature stream is then reshaped and merged to form the output map.

[0095] Step 4: In the feature fusion stage, a channel-gated feedforward network is introduced. Learnable weights are used to dynamically fuse multi-scale features, achieving an adaptive residual correction strategy, such as... Figure 5 As shown;

[0096] Dynamic channel weights are generated using Global Average Pooling (GAP). The calculation formula is:

[0097]

[0098] Among them, feature map B represents the batch size, N represents the product of the image height and width, and C represents the number of channels. The matrix is ​​a learnable linear weight matrix, with sigmoid as the activation function and MLP as the multilayer perceptron. The weights dynamically adjust the importance of each channel based on the global statistical properties of the input features. Hadamard products are used to weight the feature maps channel-by-channel, achieving data-driven feature selection.

[0099] Step 5: Using 3×3 convolution and a pixel reconstruction module, the feature map is upsampled from low resolution to high resolution by converting the channel dimension to the spatial dimension.

[0100] Decompose the number of channels C of the input feature map into Where C is the new number of channels and r is the magnification factor. The dimensions of the input feature map are changed from C×H×W to... Add spatial dimension. Then rearrange the feature map: move the r×r blocks to the spatial dimension to form a new feature map of size r×r. High-resolution feature maps.

[0101] Step 6: Set the hybrid loss function to optimize network parameters:

[0102] First, the L1 absolute error loss function is used to provide a smoother optimization direction in the early stages of training, and it has a strong ability to correct the blue-green bias in underwater images. The specific formula is as follows:

[0103]

[0104] H and W represent the number of channels, height, and width of the image, respectively. IHR represents the true high-resolution image tensor, and ISR represents the high-resolution image tensor generated by the network.

[0105] Secondly, the L2 mean squared error loss function is used to measure the difference between the predicted value and the true value. The specific formula is as follows:

[0106]

[0107] Let i be the true value of the i-th sample. is the corresponding predicted value, and N is the total number of samples.

[0108] The improved mixing loss function is achieved through weighted chromaticity loss. L1 loss is applied to the luminance channel Y. =0.7), and L2 loss is applied to the chroma channel UV ( =0.3) The specific formula is as follows:

[0109]

[0110] The above description is merely a preferred embodiment of an underwater image super-resolution reconstruction method based on dynamic window and state space collaboration. The scope of protection for this underwater image super-resolution reconstruction method is not limited to the above embodiments; all technical solutions falling within this conceptual framework are within the protection scope of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the protection scope of this invention.

Claims

1. A method for underwater image super-resolution reconstruction based on dynamic window and state space coordination, characterized in that, Includes the following steps: Step 1: Acquire low-resolution underwater images using underwater imaging equipment, and generate low-resolution images by applying Gaussian blur and bicubic downsampling; Step 2: Construct a super-resolution network that includes a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module; Step 3: Construct a dual-path feature interaction mechanism in the deep feature extraction module, and achieve local detail enhancement and global context modeling through the alternating learning of the dynamic window attention module and the state space module; Step 4: In the feature fusion stage, a channel-gated feedforward network is introduced to dynamically fuse multi-scale features using learnable weights, thereby realizing an adaptive residual correction strategy. Step 5: Generate high-resolution images using a multi-scale reconstruction module combined with Pixel Shuffle operation; Step 6: Improve the mixing loss function using weighted chroma loss.

2. The method according to claim 1, characterized in that, The underwater imaging device used in step 1 is specifically: An integrated underwater vehicle was used to securely mount the camera on a suitable underwater platform, and a polarizing filter system was added. This system ensured that the camera could capture images under different polarization directions. To fully describe the polarization state of light, three representative polarization directions—0°, 45°, and 90°—were selected. Before the actual shooting, a well-lit environment was carefully chosen to minimize interference with polarization information. During shooting, the polarizing filter was rotated sequentially to the predetermined 0°, 45°, and 90° polarization directions, and images of the same underwater scene were captured, resulting in multiple images with different polarization directions. Finally, these images were processed together to obtain the underwater image that most closely resembled reality.

3. The method according to claim 1, characterized in that, The process of generating a low-resolution image using Gaussian blurring and bicubic downsampling in step 1 is as follows: Gaussian blurring and bicubic downsampling were applied to the acquired underwater images to degrade image features and obtain a low-resolution underwater image dataset. Gaussian blurring involved weighted averaging of the image using a normally distributed convolution kernel. The pixel value of each sampling point was calculated by averaging the surrounding 5×5 pixels, with the weights decreasing exponentially with distance from the center point. The bicubic downsampling process was implemented using a bicubic interpolation algorithm, performing interpolation calculations based on a bicubic polynomial over the 16 nearest neighbor pixels of the sampling point.

4. The method according to claim 1, characterized in that, Step 2 includes a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module, specifically: In the shallow feature extraction stage, the input image X is first processed by 3×3 convolution, and then tensor reshaping is performed to map the input to a high-dimensional feature space, where B represents the batch, C represents the number of channels, and H and W represent the image height and width, respectively. In the deep feature extraction stage, N residual groups are stacked, and dynamic spatial attention blocks and Mamba state space modules are alternately deployed in each residual group. In the image reconstruction stage, 3×3 convolution and pixel shuffle multi-scale reconstruction are used to achieve super-resolution reconstruction of the target image.

5. The method according to claim 1, characterized in that, The dynamic window attention module in step 3 specifically refers to: The dynamic window partitioning unit alternates between horizontal and vertical window segmentation strategies based on the hierarchical index. When the residual group index is even, the input feature map is divided into horizontal windows; when the index is odd, it switches to vertical windows, alternately capturing horizontal and vertical texture features. Then, the local receptive field is expanded through cyclic shift operations. The cyclic shift operation unit expands the local receptive field while maintaining the continuity of the neighborhood structure. The dual-path attention calculation unit integrates spatial attention mechanisms with deep convolutional paths. The gated collaboration unit generates channel weights through global average pooling to achieve dynamic feature fusion.

6. The method according to claim 1, characterized in that, The state space module in step 3: The input image tensor X is subjected to dual-path projection and then equally divided into two sub-tensors x and y along the channel dimension. x is processed independently by 3x3 grouped depthwise separable convolution, and a nonlinear SiLU activation function is introduced to enhance feature representation. y is scanned in two dimensions using selective scanning along the horizontal, vertical, forward, and reverse directions to transform the two-dimensional image into a multi-directional one-dimensional feature sequence. To support end-to-end gradient propagation and efficiently cooperate with deep learning frameworks, the continuous one-dimensional feature sequence system is discretized to adapt to discrete inputs in deep learning. Therefore, a zero-order hold is used to convert the continuous parameters A and B into discrete parameters.

7. The method according to claim 1, characterized in that, The specific implementation of the channel-gated feedforward network in step 4 is as follows: Dynamic channel weights α are generated by Global Average Pooling (GAP). These weights dynamically adjust the importance of each channel based on the global statistical characteristics of the input features. The feature map is then weighted channel by channel using the Hadamard product to achieve data-driven feature selection.

8. The method according to claim 1, characterized in that, The image reconstruction module in step 5 is specifically implemented as follows: By fusing shallow high-frequency details, mid-level structural features, and deep semantic information through 3×3 convolution, a representation containing multi-scale features is formed. Then, the channel dimension is reorganized into a spatial dimension through the Pixel Shuffle operation.

9. The method according to claim 1, characterized in that, The hybrid loss function in step 6 is specifically as follows: The L1 absolute error loss function provides a smoother optimization direction in the early stages of training and has a strong ability to correct the blue-green bias in underwater images. The L2 mean square error loss function is used to measure the difference between the predicted value and the true value. The hybrid loss function is improved by weighted chroma loss. L1 loss (0.7) is used for the luminance channel Y and L2 loss (0.3) is used for the chroma channel UV.

10. A method for super-resolution reconstruction of underwater images based on dynamic window and state space coordination, characterized in that: The data acquisition module is configured to acquire underwater low-resolution images and corresponding degradation parameters; the feature extraction module includes the dynamic window attention module as described in claim 5 and the state space module as described in claim 6; the learnable weights dynamically fuse multi-scale features, involving the channel-gated feedforward network as described in claim 7 to implement an adaptive residual correction strategy. The reconstruction output module generates a high-resolution image and outputs the color correction results.

Images