Depth-sensing underwater image enhancement methods and systems for high-resolution deployment

Abstract

This invention discloses a depth-sensing underwater image enhancement method and system for high-resolution deployment, belonging to the field of underwater image processing technology. The method includes: acquiring the underwater image to be enhanced and its corresponding depth information, performing YCbCr preprocessing and brightness gradient extraction to obtain brightness channel, chroma channel, and brightness gradient features; predicting image-level lookup table fusion weights and pixel-level color degradation indexes based on a dual-branch degradation encoder; performing depth-sensing four-dimensional lookup table mapping to obtain preliminary enhancement results in the YCbCr space and locally refining them to obtain refined YCbCr enhancement results; converting these results back to the RGB space; and outputting the final enhanced underwater image. This invention combines depth information with the YCbCr space, achieving coordinated correction of brightness attenuation and color distortion through a depth-sensing four-dimensional lookup table, meeting the requirements of high-resolution real-time deployment while maintaining high enhancement quality.

Description

Technical Field

[0001] This invention belongs to the field of underwater image processing technology, specifically relating to a depth-sensing underwater image enhancement method and system for high-resolution deployment. Background Technology

[0002] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art.

[0003] Currently, underwater vision technology has been widely applied in various fields such as marine exploration, autonomous navigation of underwater robots, ecological monitoring, underwater surveying, underwater operations, and target recognition. With the continuous development of underwater intelligent equipment and marine information applications, underwater images, as a crucial information carrier for environmental perception, target analysis, and intelligent decision-making, directly impact the performance of downstream tasks such as target detection, feature matching, localization and mapping, and autonomous control. Simultaneously, practical applications typically require image processing methods to adapt to high-resolution input and possess high real-time inference capabilities to meet the processing efficiency and deployment performance requirements of embedded platforms, edge devices, and high-throughput operating systems.

[0004] Currently, research on underwater image enhancement mainly focuses on physical model-based methods and deep learning-based methods, while some methods also attempt to introduce lookup table mechanisms to improve inference efficiency. In recent years, deep learning-based methods (such as complex networks that integrate multiple color spaces and multi-scale features) have made some progress in enhancing image quality. However, these methods typically suffer from problems such as complex network structures, large number of model parameters, high computational cost, and long inference latency, making them difficult to adapt to underwater robots and embedded processing platforms with high real-time requirements.

[0005] While existing lookup table methods offer high inference efficiency, their ability to represent complex underwater degradation is limited. Most current lookup table enhancement methods primarily employ 3D lookup table modeling in RGB space. However, because brightness and color information are coupled in RGB space, it's difficult to effectively distinguish between brightness attenuation caused by propagation distance changes and color distortion caused by wavelength differences. Furthermore, existing methods generally lack the ability to co-model the two degradation factors of "brightness attenuation caused by propagation distance" and "color shift caused by wavelength-selective absorption," and there is also a lack of effective coordination between depth information and efficient enhancement mechanisms. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a depth-sensing underwater image enhancement method and system for high-resolution deployment.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the technical solution of the present invention provides a depth-sensing underwater image enhancement method for high-resolution deployment, comprising: Acquire the underwater image to be enhanced and its corresponding depth information; Underwater images were preprocessed with YCbCr and brightness gradient extracted to obtain brightness channel, chroma channel and brightness gradient features; The luminance channel, chrominance channel, depth information, and luminance gradient features are input into the dual-branch degradation encoder to predict image-level lookup table fusion weights and pixel-level color degradation indexes. Based on the brightness value, depth value and color degradation index, a depth-aware four-dimensional lookup table mapping is performed to obtain the preliminary enhancement results in the YCbCr space; The preliminary enhancement results are locally refined to obtain the refined YCbCr enhancement results; The refined YCbCr enhancement result is converted back to RGB space to output the final enhanced underwater image.

[0008] In at least one embodiment, underwater images undergo YCbCr preprocessing and brightness gradient extraction, specifically including: The acquired underwater RGB image to be enhanced is converted to the YCbCr color space to generate the YCbCr image of the underwater RGB image to be enhanced, thus obtaining the luminance channel and chrominance channel of the YCbCr image. Gradient features are obtained by extracting gradients from the brightness channel using the Sobel operator.

[0009] In at least one embodiment, the dual-branch degradation encoder specifically includes a feature splicing module, a shared feature extraction module, a weight branch, and a color degradation branch; wherein, the weight branch and the color degradation branch are two parallel branches, both of which take the output of the shared feature extraction module as input features; the weight branch is used to predict the image-level lookup table fusion weights that characterize the overall degradation state of the image, and the color degradation branch is used to predict the pixel-level color degradation index that characterizes the local color degradation state of a pixel.

[0010] In at least one embodiment, predicting the image-level lookup table fusion weights and the pixel-level color degradation index specifically includes: The feature stitching module stitches together the luminance channel, chroma channel, depth map, and gradient features to generate the input features for the degradation encoder. Based on the shared feature extraction module, feature extraction is performed on the input features of the degenerate encoder to obtain shared degenerate features; Based on shared degradation features, image-level lookup table fusion weights are predicted through weighted branching. Based on shared degradation features, pixel-level color degradation indices are predicted through color degradation branches.

[0011] In at least one embodiment, based on shared degradation features, image-level lookup table fusion weights are predicted by weighted branching, specifically including: first, global aggregation of shared degradation features, then outputting image-level lookup table fusion weights through a weight prediction head, and normalizing the image-level lookup table fusion weights through a Softmax function; wherein, the sum of all image-level lookup table fusion weights is equal to 1.

[0012] In at least one embodiment, based on shared degradation features, pixel-level color degradation indexes are predicted via a color degradation branch, specifically including: The shared degradation features are processed pixel by pixel. The color degradation prediction head outputs the first pixel-level color degradation index and the second pixel-level color degradation index. The Sigmoid function is used to constrain the range of the first pixel-level color degradation index and the second pixel-level color degradation index.

[0013] In at least one embodiment, the depth-aware four-dimensional lookup table mapping specifically includes: Construct multiple learnable four-dimensional lookup tables; For any four-dimensional lookup table, the lookup is performed using the brightness value, depth value, first color degradation index, and second color degradation index as four lookup dimensions, and the output results of each four-dimensional lookup table are obtained through four-linear interpolation. The output results of each four-dimensional lookup table are weighted and fused based on the image-level lookup table fusion weight to obtain the lookup table residual results in the YCbCr space. The lookup table residuals in the YCbCr space are added to the original YCbCr image to obtain the preliminary enhancement result.

[0014] In at least one embodiment, the preliminary enhancement result is locally refined to obtain a refined YCbCr enhancement result, specifically including: The preliminary enhancement result is stitched together with the original YCbCr image, and the thinning residual is output through the local thinning module. The thinning residual is added to the preliminary enhancement result to obtain the thinned YCbCr enhancement result.

[0015] In at least one embodiment, an adaptive resolution inference step is further included, specifically comprising: The input sample is subjected to resolution detection. If the resolution is between 1080P and 4K, the subsequent image enhancement steps are performed at 0.5 times the original resolution. If the resolution is not lower than 4K, the subsequent image enhancement steps are performed at 0.25 times the original resolution.

[0016] Secondly, the technical solution of the present invention also provides a depth-sensing underwater image enhancement system for high-resolution deployment, comprising: The image acquisition module is configured to acquire the underwater image to be enhanced and its corresponding depth information. The feature extraction module is configured to perform YCbCr preprocessing and brightness gradient extraction on the underwater image to obtain brightness channel, chroma channel and brightness gradient features; The dual-branch prediction module is configured to: input the luminance channel, chrominance channel, depth information and luminance gradient features into the dual-branch degradation encoder, and predict the image-level lookup table fusion weights and pixel-level color degradation index; The initial enhancement module is configured to perform a depth-aware four-dimensional lookup table mapping based on luminance values, depth values, and color degradation indexes to obtain the initial enhancement results in the YCbCr space. The refinement and enhancement module is configured to: refine the preliminary enhancement results locally to obtain the refined YCbCr enhancement results; The output module is configured to convert the refined YCbCr enhancement result back to RGB space and output the final enhanced underwater image.

[0017] The beneficial effects of the above-described technical solution of the present invention are as follows: 1) The depth-sensing underwater image enhancement method for high-resolution deployment of the present invention can combine scene depth information with YCbCr color space to perform collaborative modeling and enhancement correction of brightness attenuation and color distortion in underwater images, thereby achieving efficient enhancement processing of underwater images, improving the image visibility quality in complex underwater environments and the applicability of subsequent downstream visual tasks, and meeting the real-time deployment requirements in high-resolution scenarios.

[0018] 2) This invention combines scene depth information with YCbCr color space representation to achieve collaborative modeling of depth-related brightness degradation and wavelength-related color distortion in underwater images; by predicting image-level lookup table fusion weights and pixel-level color degradation indexes through a dual-branch degradation encoder, global degradation information and local degradation information jointly participate in enhancement mapping; and by cooperating with a depth-aware four-dimensional lookup table and a local refinement module, efficient enhancement and local consistency correction are achieved.

[0019] 3) By proposing an adaptive resolution inference strategy to improve processing efficiency in high-resolution input scenarios, the depth-sensing underwater image enhancement method of this invention for high-resolution deployment can be applied to real-time high-resolution underwater image enhancement and related engineering deployment scenarios. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a schematic diagram of the depth-sensing underwater image enhancement method for high-resolution deployment disclosed in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the depth-sensing underwater image enhancement method for high-resolution deployment disclosed in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of YCbCr pretreatment and brightness gradient extraction disclosed in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the bi-branch prediction disclosed in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the depth-sensing four-dimensional lookup table mapping disclosed in Embodiment 1 of the present invention. Detailed Implementation

[0022] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example 1 As introduced in the background section, current research on underwater image enhancement mainly focuses on physical model-based methods and deep learning-based methods. Some methods also attempt to introduce lookup table mechanisms to improve inference efficiency. However, existing technologies still face the following significant technical bottlenecks and practical problems in real-world applications: (1) Insufficient ability to express degradation mechanisms makes it difficult to unify the modeling of depth-dependent brightness degradation and wavelength-dependent color distortion. Underwater image degradation is not only related to the propagation distance of light in water, but also closely related to the absorption differences of different wavelengths of light in the medium. Most existing methods only deal with color correction, contrast enhancement or overall brightness restoration, and lack the ability to coordinate the modeling of the two types of degradation factors, namely "brightness attenuation caused by propagation distance" and "color shift caused by wavelength selective absorption". Therefore, it is difficult to achieve stable and accurate enhancement effects in complex underwater scenes.

[0024] (2) While existing lookup table methods have high inference efficiency, their ability to represent complex underwater degradation is limited. Lookup table methods typically possess good inference efficiency and deployment advantages due to their offline learning and online lookup capabilities. However, most existing lookup table enhancement methods primarily employ 3D lookup table modeling in RGB space. Because brightness and color information are coupled in RGB space, it is difficult to effectively distinguish between brightness attenuation caused by changes in propagation distance and color distortion caused by wavelength differences, thus limiting the ability of lookup table methods to accurately represent underwater degradation processes.

[0025] (3) Existing methods struggle to balance enhancement quality, modeling capability, and efficient deployment performance. In practical applications, underwater image enhancement methods not only need to improve image visual quality but also require strong generalization ability and low computational complexity. In existing technologies, some methods focus on enhancement effects but sacrifice model efficiency; others focus on inference speed but struggle to effectively handle complex and non-uniform underwater degradation problems. Therefore, existing methods generally struggle to achieve a good balance between enhancement quality, degradation mechanism representation capability, and high-resolution real-time deployment performance.

[0026] (4) Lack of effective synergy between depth information and efficient enhancement mechanisms. Although some existing methods have begun to attempt to introduce depth information into underwater image restoration or enhancement to assist in estimating the degree of degradation or improve restoration accuracy, the relevant schemes usually fail to form a unified design with efficient inference mechanisms. As a result, the introduction of depth information remains more of an auxiliary function and is difficult to truly transform into efficient enhancement capabilities for real-time deployment. In complex underwater scenarios, without the collaborative modeling between depth information and color representation, it is still difficult to accurately describe the spatial non-uniform degradation process.

[0027] Therefore, existing underwater image enhancement techniques generally suffer from problems such as insufficient expression of degradation mechanisms, high computational overhead, limited lookup table modeling capabilities, and difficulty in achieving high-resolution real-time deployment. They cannot effectively meet the comprehensive requirements for enhancement quality, processing efficiency, and engineering practicality in complex underwater environments. Therefore, it is necessary to propose a new underwater image enhancement method to address the aforementioned problems in existing technologies.

[0028] To overcome the shortcomings of the prior art, in a typical embodiment of the present invention, such as... Figures 1 to 5 As shown, this embodiment discloses a depth-sensing underwater image enhancement method for high-resolution deployment, specifically including the following steps: S1. Obtain the underwater image to be enhanced and its corresponding depth information; S2. Perform YCbCr preprocessing and brightness gradient extraction on the underwater image to obtain brightness channel, chroma channel and brightness gradient features; S3. Input the luminance channel, chrominance channel, depth information and luminance gradient features into the dual-branch degradation encoder to predict the image-level lookup table fusion weights and pixel-level color degradation index; S4. Perform depth-aware four-dimensional lookup table mapping based on luminance value, depth value and color degradation index to obtain preliminary enhancement results in YCbCr space; S5. The preliminary enhancement results are locally refined to obtain the refined YCbCr enhancement results; S6. Convert the refined YCbCr enhancement result back to RGB space and output the final enhanced underwater image.

[0029] The following describes in detail the depth-sensing underwater image enhancement method for high-resolution deployment, with reference to specific implementation methods.

[0030] S1. Obtain the underwater image to be enhanced and its corresponding depth information.

[0031] In order to obtain the original image data and propagation distance-related prior information required for subsequent enhancement processing, this step obtains the underwater RGB image to be enhanced and the depth map corresponding to the underwater RGB image.

[0032] Specifically, the acquired underwater RGB image to be enhanced is denoted as... ,and The corresponding depth map is denoted as Depth map Specifically, this is a scene depth map, used to characterize the distance between target points in the scene and the imaging sensor, reflecting the propagation distance of reflected light in the water. Based on the acquired underwater image to be enhanced... and its corresponding depth map Construct the input sample, denoted as .

[0033] As a further implementation, the underwater RGB image to be enhanced It can be acquired through underwater cameras, AUV / ROV payload cameras, fixed underwater observation equipment, or other imaging devices. (Compared with the underwater RGB image to be enhanced.) Corresponding depth map It can be obtained from a monocular depth estimation model, or from a binocular vision system, depth sensor, or other imaging ranging device.

[0034] S2. Perform YCbCr preprocessing and brightness gradient extraction on the underwater image to obtain brightness channel, chromaticity channel and brightness gradient features.

[0035] To separate and represent luminance, chromaticity, and local structure information, this step involves processing the acquired underwater RGB image to be enhanced. Perform YCbCr color space transformation and extract gradient features based on the luminance channel.

[0036] Specifically, such as Figure 3 As shown, the underwater RGB image to be enhanced is first obtained. Convert to YCbCr color space to obtain the underwater RGB image to be enhanced. YCbCr image Specifically, it can be expressed as:

[0037] In the formula, This is the luminance channel, i.e., the luminance value; and For chroma channel. For luminance channel. and chroma channels and can be The result of the conversion is as follows:

[0038] Then, the Sobel operator is used on the luminance channel. Gradient extraction is performed to obtain gradient features, which are specifically represented as follows:

[0039] In the formula, The extracted gradient features are used to characterize the edge, texture, and local structural changes in the image, thereby providing auxiliary information for subsequent degradation feature encoding and enhancement mapping.

[0040] S3. Input the luminance channel, chrominance channel, depth information and luminance gradient features into the dual-branch degradation encoder to predict the image-level lookup table fusion weights and the pixel-level color degradation index.

[0041] To simultaneously extract both overall and local degradation information of the image, a dual-branch degradation encoder is constructed in this step. Shared feature extraction and branch prediction are performed on the input features to obtain control parameters for subsequent enhancement mapping.

[0042] In this embodiment, the dual-branch degradation encoder specifically includes a feature concatenation module, a shared feature extraction module, a weight branch, and a color degradation branch. The shared feature extraction module extracts common degradation features from the input image and depth information. The weight branch and color degradation branch are two parallel branches, both using the output of the shared feature extraction module as their input features. The weight branch predicts the image-level lookup table fusion weights representing the overall degradation state of the image, while the color degradation branch predicts the pixel-level color degradation index representing the local color degradation state of pixels, enabling global and local degradation information to jointly participate in the subsequent lookup table enhancement mapping process.

[0043] Specifically, such as Figure 4 As shown, first, the brightness channel... chroma channel and Depth map and gradient features The features are concatenated in the feature concatenation module to form the input features of the degenerate encoder, specifically represented as follows:

[0044] In the formula, Input features for the degenerate encoder; This is for splicing operations.

[0045] Then, through the shared feature extraction module Input features of the degenerate encoder Feature extraction is performed to obtain shared degradation features, specifically represented as follows:

[0046] In the formula, To share degradation features.

[0047] Next, based on shared degradation features Image-level lookup table fusion weights are predicted through weighted branching. Specifically, the weighted branch first processes the shared degenerate features. Perform global aggregation to obtain image-level feature representations. Specifically, it is expressed as:

[0048] In the formula, This is a global pooling operation.

[0049] Then, the weights are fused using an image-level lookup table output by the weight prediction head. The image-level lookup table fusion weights are normalized using the Softmax function, which can be specifically expressed as follows:

[0050] In the formula, Indicates the weighted prediction head; Used to normalize the output into a fusion weight between multiple lookup tables.

[0051] Let the number of learnable four-dimensional lookup tables be . The image-level lookup table fusion weights can be specifically expressed as:

[0052] and The sum of the fusion weights of the image-level lookup tables equals 1, specifically expressed as:

[0053] Image-level lookup table fusion weights It is used to characterize the overall degradation state of the input image and to control the weighted fusion ratio of multiple four-dimensional lookup table outputs.

[0054] Finally, based on shared degradation features Predict pixel-level color degradation index through color degradation branch and Specifically, the color degradation branch shares degradation features. Perform pixel-by-pixel processing and output two pixel-level color degradation indices. and And the pixel-level color degradation index is obtained through the Sigmoid function. and Applying range constraints can be represented as follows:

[0055] In the formula, The first color degradation index; For the second color degradation index; The prediction head representing the color degradation branch; This represents the Sigmoid activation function, used to constrain the output within a predetermined range. Pixel-level color degradation index. and Used to characterize the color degradation state of local regions of an image and as the lookup dimension for subsequent four-dimensional lookup table mapping.

[0056] S4. Perform depth-aware four-dimensional lookup table mapping based on luminance value, depth value and color degradation index to obtain preliminary enhancement results in YCbCr space.

[0057] In this step, multiple learnable four-dimensional lookup tables are constructed, and weighted enhancement mapping is completed based on the image-level lookup table fusion weights to achieve joint enhancement mapping based on brightness, depth, and color degradation indexes, thereby achieving collaborative correction of brightness attenuation and color distortion in underwater images.

[0058] Specifically, such as Figure 5 As shown, multiple learnable four-dimensional lookup tables are first constructed. This is used to learn the enhancement mapping relationships under different degradation modes. , This is the sequence number of the four-dimensional lookup table; middle, These represent the number of discrete samples in the four lookup dimensions, which correspond to the brightness values ​​respectively. Depth value First color degradation index Second color degradation index , express , and The residual components of the three channels. Preferably, the number of discrete samples is the same across the four lookup dimensions, i.e. .

[0059] Then, with brightness value Depth value First color degradation index Second color degradation index As a lookup dimension, a lookup is performed on each of the four-dimensional lookup tables. Specifically, for any pixel in the input image, the pixel's brightness value is used as the lookup dimension. Depth value First color degradation index Second color degradation index As input for the lookup, multiple four-dimensional lookup tables are queried, and the output results of each four-dimensional lookup table are obtained through quadlinear interpolation. Then, the... The output corresponding to each four-dimensional lookup table can be represented as:

[0060] Next, the weights are fused based on the image-level lookup table. The outputs of each four-dimensional lookup table are weighted and fused to obtain the lookup table residuals in the YCbCr space, which can be specifically represented as follows:

[0061] In the formula, This represents the lookup table residual results in the YCbCr space; Represents the image-level lookup table fusion weights The corresponding number in the middle The weight values ​​of a four-dimensional lookup table.

[0062] Finally, the lookup table residuals in the YCbCr space are added to the original YCbCr image to obtain the preliminary enhancement result, specifically as follows:

[0063] In the formula, This indicates the initial enhancement result after enhancement using a four-dimensional lookup table.

[0064] S5. The preliminary enhancement results are locally refined to obtain the refined YCbCr enhancement results.

[0065] To improve the local discontinuity problem that may be caused by pixel-by-pixel mapping of the lookup table, a local refinement module is set up in this step to further correct the preliminary enhancement results, so as to correct the enhancement results of edge and texture regions, thereby improving the spatial consistency and visual naturalness of the enhanced image in local regions.

[0066] Specifically, the initial enhancement results will be... Compared with the original YCbCr image By concatenating the data, we obtain the input for the local refinement module, specifically represented as follows:

[0067] Then, through local refinement of modules Output refined residuals. In this step, the local refinement module... A lightweight 3×3 neighborhood convolutional residual thinning module is employed to perform local spatial correction after the output of the four-dimensional lookup table, compensating for the local discontinuities that may arise from independent pixel-by-pixel lookups in the LUT. The input to the local thinning module is... Among them, preliminary enhancement results The original YCbCr image has 3 channels. The number of channels is also 3, therefore, the local refinement module The total number of input channels is 6.

[0068] In the local refinement module First, input the 6 channels. The hidden features are mapped to a 16-dimensional hidden feature space using a 3×3 ordinary convolutional layer; followed by a LeakyReLU activation layer with a negative slope of 0.2; then, a depthwise separable convolutional block is used for lightweight local spatial feature extraction, consisting of a 3×3 depthwise convolution and a 1×1 pointwise convolution, followed by another LeakyReLU activation layer with a negative slope of 0.2; finally, a 3×3 output convolutional layer maps the hidden features to a 3-channel YCbCr refined residual. This process can be specifically represented as:

[0069] In the formula, This represents the refined residual output by the local refinement module.

[0070] In this embodiment, the local refinement module The hidden feature dimension is 16, the default number of network layers is 3, and the number of parameters is approximately 1.7K, making it very lightweight overall. The weights of the output convolutional layers are initialized with a small standard deviation, and the bias is initialized to 0, so that the module outputs near-zero residuals in the early stages of training, without significantly damaging the initial enhancement results of the LUT.

[0071] Next, the obtained refined residuals Compared with preliminary enhancement results Adding them together yields the refined YCbCr enhancement result, specifically expressed as follows:

[0072] In the formula, This represents the enhancement result of YCbCr after local refinement.

[0073] This step involves designing a local refinement module that, after the depth-aware four-dimensional lookup table, specifically corrects the local residuals in the YCbCr space of the LUT output, effectively improving the local consistency and visual naturalness of edge and texture regions.

[0074] S6. Convert the refined YCbCr enhancement result back to RGB space and output the final enhanced underwater image.

[0075] In this step, the refined YCbCr enhancement result is converted back to RGB space to obtain the final enhanced underwater image. This conversion process can be represented as follows:

[0076] In the formula, This represents the final enhanced underwater RGB image output.

[0077] Deep learning-based underwater image enhancement methods typically suffer from complex network structures, large model parameter counts, high computational costs, and long inference latency. These computational overheads increase further in high-resolution input scenarios, making them unsuitable for underwater robots, edge computing nodes, and embedded processing platforms with high real-time requirements, thus limiting their engineering applications. To overcome the shortcomings of existing high-performance enhancement methods, such as high computational overhead and difficulty in meeting the demands of high-resolution real-time deployment, this embodiment proposes an adaptive resolution inference strategy to improve deployment efficiency in high-resolution input scenarios.

[0078] Specifically, the adaptive resolution inference strategy refers to the strategy of applying the input sample obtained in step S1 to the input sample. Detection is performed at a resolution where the input sample is... If the resolution is lower than 1080P, then proceed with steps S2 to S6 at the original resolution; if the input sample If the input sample resolution is between 1080P and 4K, then steps S2 to S6 are performed at 0.5 times the original resolution. After obtaining the enhancement result, the enhancement result is then upsampled back to the original resolution. If the resolution is not lower than 4K, then steps S2 to S6 are performed at 0.25 times the original resolution. After obtaining the enhancement result, the enhancement result is upsampled back to the original resolution.

[0079] As an example, when input samples resolution is At that time, if the resolution If the resolution is lower than 1080P, it will be based on the original resolution. Execute steps S2 to S6; if the resolution For images between 1080P and 4K resolution, first obtain the high-resolution underwater image to be enhanced. and its corresponding depth Scaling by 0.5x, then... Perform steps S2 to S6 at the specified resolution to obtain... The enhanced RGB image is then upsampled at its original resolution, and finally restored to its original resolution. This serves as the final enhanced underwater image; if the resolution... If the resolution is no lower than 4K, then first process the high-resolution underwater image to be enhanced. and its corresponding depth Scaling by 0.25x, then... Perform steps S2 to S6 at the specified resolution to obtain... The enhanced RGB image is then upsampled at its original resolution, and finally restored to its original resolution. This is the final enhanced underwater image.

[0080] Example 2 In one typical embodiment of the present invention, this embodiment discloses a depth-sensing underwater image enhancement system for high-resolution deployment, comprising: The image acquisition module is configured to acquire the underwater image to be enhanced and its corresponding depth information. The feature extraction module is configured to perform YCbCr preprocessing and brightness gradient extraction on the underwater image to obtain brightness channel, chroma channel and brightness gradient features; The dual-branch prediction module is configured to: input the luminance channel, chrominance channel, depth information and luminance gradient features into the dual-branch degradation encoder, and predict the image-level lookup table fusion weights and pixel-level color degradation index; The initial enhancement module is configured to perform a depth-aware four-dimensional lookup table mapping based on luminance values, depth values, and color degradation indexes to obtain the initial enhancement results in the YCbCr space. The refinement and enhancement module is configured to: refine the preliminary enhancement results locally to obtain the refined YCbCr enhancement results; The output module is configured to convert the refined YCbCr enhancement result back to RGB space and output the final enhanced underwater image.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A depth-sensing underwater image enhancement method for high-resolution deployment, characterized in that, include: Acquire the underwater image to be enhanced and its corresponding depth information; Underwater images were preprocessed with YCbCr and brightness gradient extracted to obtain brightness channel, chroma channel and brightness gradient features; The luminance channel, chrominance channel, depth information, and luminance gradient features are input into the dual-branch degradation encoder to predict image-level lookup table fusion weights and pixel-level color degradation indexes. Based on the brightness value, depth value and color degradation index, a depth-aware four-dimensional lookup table mapping is performed to obtain the preliminary enhancement results in the YCbCr space; The preliminary enhancement results are locally refined to obtain the refined YCbCr enhancement results; The refined YCbCr enhancement result is converted back to RGB space to output the final enhanced underwater image.

2. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 1, characterized in that, Underwater images undergo YCbCr preprocessing and brightness gradient extraction, specifically including: The acquired underwater RGB image to be enhanced is converted to the YCbCr color space to generate the YCbCr image of the underwater RGB image to be enhanced, thus obtaining the luminance channel and chrominance channel of the YCbCr image. Gradient features are obtained by extracting gradients from the brightness channel using the Sobel operator.

3. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 1, characterized in that, The dual-branch degradation encoder specifically includes a feature stitching module, a shared feature extraction module, a weight branch, and a color degradation branch. The weight branch and the color degradation branch are two parallel branches, both of which take the output of the shared feature extraction module as input features. The weight branch is used to predict the image-level lookup table fusion weights that characterize the overall degradation state of the image, while the color degradation branch is used to predict the pixel-level color degradation index that characterizes the local color degradation state of a pixel.

4. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 3, characterized in that, Predicting image-level lookup table fusion weights and pixel-level color degradation indexes, specifically including: The feature stitching module stitches together the luminance channel, chroma channel, depth map, and gradient features to generate the input features for the degradation encoder. Based on the shared feature extraction module, feature extraction is performed on the input features of the degenerate encoder to obtain shared degenerate features; Based on shared degradation features, image-level lookup table fusion weights are predicted through weighted branching. Based on shared degradation features, pixel-level color degradation indices are predicted through color degradation branches.

5. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 4, characterized in that, Based on shared degradation features, image-level lookup table fusion weights are predicted through weighted branching. Specifically, the shared degradation features are first globally aggregated, then the image-level lookup table fusion weights are output through the weight prediction head, and the image-level lookup table fusion weights are normalized using the Softmax function. The sum of all image-level lookup table fusion weights is equal to 1.

6. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 4, characterized in that, Based on shared degradation features, pixel-level color degradation indices are predicted through a color degradation branch, specifically including: The shared degradation features are processed pixel by pixel. The color degradation prediction head outputs the first pixel-level color degradation index and the second pixel-level color degradation index. The Sigmoid function is used to constrain the range of the first pixel-level color degradation index and the second pixel-level color degradation index.

7. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 1, characterized in that, The depth-aware four-dimensional lookup table mapping specifically includes: Construct multiple learnable four-dimensional lookup tables; For any four-dimensional lookup table, the lookup is performed using the brightness value, depth value, first color degradation index, and second color degradation index as four lookup dimensions, and the output results of each four-dimensional lookup table are obtained through four-linear interpolation. The output results of each four-dimensional lookup table are weighted and fused based on the image-level lookup table fusion weight to obtain the lookup table residual results in the YCbCr space. The lookup table residuals in the YCbCr space are added to the original YCbCr image to obtain the preliminary enhancement result.

8. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 1, characterized in that, The preliminary enhancement results are locally refined to obtain the refined YCbCr enhancement results, specifically including: The preliminary enhancement result is stitched together with the original YCbCr image, and the thinning residual is output through the local thinning module. The thinning residual is added to the preliminary enhancement result to obtain the thinned YCbCr enhancement result.

9. The depth-sensing underwater image enhancement method for high-resolution deployment as described in claim 1, characterized in that, It also includes an adaptive resolution inference step, specifically including: The input sample is subjected to resolution detection. If the resolution is between 1080P and 4K, the subsequent image enhancement steps are performed at 0.5 times the original resolution; if the resolution is not lower than 4K, the subsequent image enhancement steps are performed at 0.25 times the original resolution. After obtaining the enhanced result at the scaled resolution, the enhanced result is upsampled to restore the original resolution, which is then used as the final enhanced underwater image.

10. A depth-sensing underwater image enhancement system for high-resolution deployment, characterized in that: include: The image acquisition module is configured to acquire the underwater image to be enhanced and its corresponding depth information. The feature extraction module is configured to perform YCbCr preprocessing and brightness gradient extraction on the underwater image to obtain brightness channel, chroma channel and brightness gradient features; The dual-branch prediction module is configured to: input the luminance channel, chrominance channel, depth information and luminance gradient features into the dual-branch degradation encoder, and predict the image-level lookup table fusion weights and pixel-level color degradation index; The initial enhancement module is configured to perform a depth-aware four-dimensional lookup table mapping based on luminance values, depth values, and color degradation indexes to obtain the initial enhancement results in the YCbCr space. The refinement and enhancement module is configured to: refine the preliminary enhancement results locally to obtain the refined YCbCr enhancement results; The output module is configured to convert the refined YCbCr enhancement result back to RGB space and output the final enhanced underwater image.

Images