A method for enhancing underwater image details based on dual-branch optical feature fusion

Abstract

This invention discloses an underwater image detail enhancement method based on dual-branch optical feature fusion. It reduces computation by performing feature preprocessing on the input image through pixel inverse rearrangement; it extracts channel-dimensional optical features and spatial-dimensional detail features using a global color branch and a local detail branch respectively, achieving task decoupling; it employs an optically perceptual detail enhancement module, sequentially performing optical restoration, multi-scale denoising, image restoration, and color enhancement operations via a serial link to address degradation issues in stages; it fuses the enhanced global and local detail features element-wise and rearranges pixels to restore the image size; finally, it achieves adaptive enhancement and normalization through global-local multi-granularity contrast constraints, outputting a high-quality enhanced image. This method effectively solves multi-dimensional degradation in underwater images, balancing color correction, detail preservation, and real-time performance, and is suitable for complex underwater scenarios.

Description

Technical Field

[0001] This invention pertains to image data processing, specifically to a method for enhancing underwater image details based on bi-branch optical feature fusion. Background Technology

[0002] The underwater environment is complex and variable, and the acquisition of high-quality underwater visual information faces severe challenges due to light attenuation and the scattering effect of seawater. This is mainly manifested in phenomena such as underwater scene turbidity, image scattering degradation, and blurred textures. Underwater image processing, as one of the core technologies in marine resource exploration, underwater mechanical navigation, and underwater archaeology, is undeniably important. Therefore, effectively enhancing the quality of underwater images has become a major technical challenge for advancing marine research and exploration missions.

[0003] Currently, underwater image processing methods are mainly divided into two categories: traditional non-deep learning methods and deep learning methods. Traditional methods rely on physical model imaging or prior image knowledge, directly correcting degradation models through mathematical analysis and physical modeling. However, due to the diversity and complexity of the underwater environment, these methods tailored to specific physical scenarios are difficult to adapt to other diverse scenarios, resulting in limited generalization ability. On the other hand, deep learning methods adopt a data-driven approach, utilizing network architectures such as CNNs and Transformers to learn the mapping relationship from degraded images to clear images. This method has shown significant advantages in improving the efficiency and quality of underwater image optimization and has become the mainstream of current research. Nevertheless, these methods still bring new challenges to the field of computer vision in areas such as degradation co-optimization, real-time processing and transmission, and lightweighting of complex image data.

[0004] Despite the progress made by the aforementioned methods, underwater image enhancement still faces several critical unresolved issues. First, the unique scattering and blurring effects underwater cause optical color distortion, introduce noise interference, and result in loss of detail and texture, as well as color aberrations. These problems are often intertwined. Many existing solutions tend to employ parallel processing mechanisms or focus on performance optimization on specific datasets, which can lead to task conflicts, model overfitting, and weak generalization, thus affecting the final enhancement effect. Second, while super-resolution reconstruction techniques can obtain high-resolution images from low-resolution images, they cannot effectively eliminate severe color distortion and semantic loss in underwater images. Furthermore, there may be mutual interference between color correction and detail preservation, affecting the visual naturalness of the image. Additionally, traditional denoising modules may blur edge details while suppressing noise, and color enhancement modules using only Sigmoid or ReLU activation functions may encounter gradient vanishing or hard truncation problems, making it difficult to achieve smooth color transitions. Finally, existing fusion strategies are not balanced in weight allocation, easily leading to over-enhancement or loss of original information, resulting in an unrealistic enhanced image.

[0005] The existence of these problems not only limits the practical application of underwater image enhancement technology, but also highlights the necessity of developing a method that can comprehensively consider color correction, detail preservation, and avoid over-enhancement. Summary of the Invention

[0006] To address the technical problems of existing underwater image enhancement methods, such as difficulty in resolving multidimensional degradation coupling, insufficient generalization ability, and difficulty in simultaneously preserving detail and color, this invention proposes an underwater image detail enhancement method based on dual-branch optical feature fusion. This method designs an optically-sensing detail enhancement module, which sequentially performs optical restoration, multi-scale denoising, and image enhancement via a serial link.

[0007] The restoration and color enhancement operations simulate the gradual repair process of human vision, ensuring that each stage focuses on a single type of degradation and effectively avoiding interference between tasks. For color decoupling and detail processing, this invention constructs a dual-branch architecture. The global branch achieves color balance and optical distortion repair through an optically perceptual detail enhancement module, while the local branch focuses on detail preservation and combines it with a global contrast-weighted correction mechanism to achieve synergistic integration. Addressing the shortcomings of traditional denoising methods that easily blur edges and color enhancement methods that easily produce unnatural transitions, this invention designs a multi-scale denoising and channel-by-channel edge preservation collaborative mechanism, as well as an adaptive color enhancement module based on the Softplus activation function, effectively alleviating edge blurring and color banding problems. Simultaneously, by optimizing the contrast-constrained fusion strategy and the hierarchical deep feature fusion link, the balance of the enhancement effect and the integrity of the original features are further ensured.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] An underwater image detail enhancement method based on dual-branch optical feature fusion includes the following steps:

[0010] S1. Visual feature preprocessing: The input raw underwater image is preprocessed by pixel inverse rearrangement. Block tensor features are obtained by pixel splitting, which reduces the amount of computation while preserving the local feature structure, laying the foundation for the decoupling of global color and local detail features in the two branches.

[0011] S2. Dual-branch feature extraction: Construct a global color branch and a local detail branch. The global color branch extracts channel-dimensional optical features (brightness and / or color) from the block tensor features through 1×1 convolution to obtain channel-dimensional global color correction features. The local detail branch extracts spatial-dimensional detail features (edges and / or textures) from the block tensor features through 3×3 convolution to obtain channel-dimensional local detail features, thus achieving decoupling of multi-branch tasks.

[0012] S3, Optical Detail Enhancement: The global color correction features of the channel dimension are input into the Optical Detail Enhancement (OADEM) module. Optical restoration, multi-scale denoising, image restoration and color enhancement operations are executed sequentially through a serial link to obtain the final enhancement features of the OADEM module. This achieves a phased solution to the single degradation problem and avoids conflicts in image optimization tasks.

[0013] S4. Dual-branch feature fusion and size restoration: The final enhanced features of the OADEM module are fused element-wise with the local detail features of the channel dimension, and the pixel rearrangement operation is performed on the fusion result to restore it to the same spatial size as the original underwater image, thus obtaining the fused enhanced features.

[0014] S5. Global-Local Multi-Granularity Contrast Constraint: Based on the original underwater image, a global contrast weighting module generates global contrast weights, which are then used to adjust the weighted enhancement features. Local adaptive normalization is then performed to constrain the pixel value range and output the final enhanced image.

[0015] Furthermore, in step S1, the specific process of visual feature preprocessing is as follows:

[0016] The underwater image enhancement target is defined as: the original underwater image (R represents an RGB image, and 3 represents the number of channels) (representing image size), through a network mapping function To generate enhanced images ,Should It can approximate the real, clear underwater images in the calibration standard file of the dataset in terms of visual effects (detail, color, contrast) and quantitative indicators. The enhancement process formula is as follows:

[0017] Formula 1: ,

[0018] in, Indicates an enhanced image. This represents a true and clear underwater image. This represents the learnable parameters of the network. Represents the network mapping function. Represents the model mapping function. loss function This indicates that minimization optimization is performed on all learnable parameters θ in the network, where I represents the original input image. R represents a GRB image, 3 represents the number of channels, and H×W represents the image size;

[0019] To adapt to different real-time requirements, the pixel rearrangement coefficient m is set for the original input image. The feature extraction process is as follows:

[0020] The visual encoder and decoder obtain the converted text and latent representation of the image. The specific process and formulas are as follows:

[0021] Formula 2: ,

[0022] in, Represents the block tensor features, PixelUnshuffle represents the pixel unshuffle function, and m represents the pixel rearrangement coefficients. Indicates the number of channels. This represents the image size. The original input image is processed by PixelUnshuffle to obtain block tensor features. The number of channels has been expanded from 3 to The image size was reduced from H×W to The computational workload is reduced to the original computational workload. This significantly improves the model inference speed and ensures real-time performance in the preprocessing stage.

[0023] Furthermore, in step S2, the global color branch's task is to extract optical features in the channel dimension and correct color shift and brightness unevenness. Therefore, a 1×1 convolution is used to fuse channel information. The 1×1 convolution's lack of spatial receptive field and lightweight nature ensures that the global dimension is not affected by edge texture spatial details and further reduces the model's computational cost. The Tanh constraint on the feature range [-1,1] avoids gradient explosion while extracting color deviations. The specific process and formulas are as follows:

[0024] Formula 3: ,

[0025] Formula 4: ,

[0026] in, Tanh represents the global feature of the channel dimension, and it represents the hyperbolic tangent function. Represents a 1×1 convolution. Represents the global color correction feature in the channel dimension, abbreviated as ReLU represents the modified linear unit activation function.

[0027] Block tensor features Global channel feature extraction is performed using 1×1 convolutions, and bidirectional color bias is captured using hyperbolic tangent function Tanh constraints to obtain global features in the channel dimension. ;Will The ReLU activation function is applied to suppress negative noise and filter color correction features, avoiding additional noise interference from Tanh capture in subsequent enhancement. Finally, a quadratic 1×1 convolution is used to recover the effective features filtered by the ReLU activation function. Through effective weight adjustment, the abstract channel-dimensional features are transformed into channel-dimensional global color correction features that can be directly color corrected. (hereinafter referred to as) It can also ensure the fusion of subsequent module enhancements and local detail features.

[0028] Further, in step S2, the local detail branch: the task is to extract spatial texture features and preserve image edges and detail structures. Therefore, a 3×3 convolution is used to cover local pixel associations. The local pixel fusion features and local spatial receptive field of the 3×3 convolution can adapt to the spatial distribution features of underwater image details, i.e., the gray-level and color abrupt changes in the local pixel neighborhood, and can initially smooth the random noise of underwater images. It can also form a strict bidirectional decoupling of color details with the 1×1 convolution. Edge replication filling is used to protect the edges, and the specific formula is as follows:

[0029] Formula 5: ,

[0030] Formula 6: ,

[0031] in, For channel-dimensional local features, Local detail features in the channel dimension, abbreviated as , It is a 3×3 convolution.

[0032] Block tensor features Local channel feature extraction is performed using 3×3 convolution and Tanh constraints are applied to obtain local features in the channel dimension. The Tanh constraint and ReLU activation are still used because the combination of the two functions can still filter positive detail features and suppress noise to a certain extent. In local details, the ReLU function can also add non-linear mapping to the features, allowing the model to learn the underwater non-linear degradation law, namely the difference in the blurring degree of texture under different turbidity and the difference in the contrast attenuation of edges under different depths. 3×3 convolution is the optimal balance between efficiency and feature representation in convolutional neural networks. Using a second 3×3 convolution to recover effective features compensates for the possible loss of weak details after ReLU filtering, and restores the feature dimension to obtain local detail features in the channel dimension. (hereinafter referred to as) And ensure that it can be integrated with the global optimized color features in the future.

[0033] Furthermore, in step S3, the optical perception detail enhancement module can simulate the human visual restoration process, first addressing basic degradation—optical restoration, then optimizing detail effects—multi-scale denoising and image restoration, and finally enhancing visual effects—image color enhancement. Each stage focuses on a single degradation type to achieve precise image restoration. The specific process includes the following four sub-steps:

[0034] (a) Optical restoration: The global color correction features in the channel dimension are processed by convolution, LeakyReLU activation, batch normalization and layer normalization, and the optical restoration features are output through residual connections. The goal of this step is to address the optical distortion caused by underwater light scattering and absorption. This is achieved by capturing optical features through high-dimensional mapping, co-optimizing the feature distribution formula using double normalization, and then preserving the original features through residual connections after low-dimensional reduction. The specific process and formulas are as follows:

[0035] Formula 7: ,

[0036] Formula 8: ,

[0037] Formula 9: ,

[0038] in, For the global activation feature tensor, For globally normalized features, For batch normalization function, For layer normalization function, For a linear correction unit with leakage, This is an optical recovery feature.

[0039] Using a 3×3 convolutional neural network To perform dimensionality increase, the LeakyReLu function allows a small number of negative gradients to pass through, preventing the emergence of dead neurons, and outputs a global activation feature tensor. BatchNorm Batch normalization is performed to reduce internal covariate bias. LayerNorm performs layer normalization to balance the feature distribution of each channel and adapt to the color imbalance between underwater image channels to obtain globally normalized features. Finally, 3×3 convolution dimensionality reduction is performed, and residual connections are used to... Fusion, output optical restoration features .

[0040] (b) Multi-scale denoising: Downsampling, upsampling, and edge detection operations are performed on the optically restored features, and denoised edge enhancement features are output through residual calculation. The goal of this step is to remove underwater noise pollution and resolve the contradiction with blurred edge details. A pooling denoising strategy is performed through downsampling mechanism and bilinear interpolation restoration. For blurred edges, channel-by-channel edge detection is used, and finally residual connection is performed to achieve multi-scale denoising. The specific process and formulas are as follows:

[0041] Formula 10: ,

[0042] Formula 11: ,

[0043] Formula 12: ,

[0044] Formula 13: ,

[0045] in, MaxPool represents coarse-grained denoising features, while MaxPool represents downsampling. Upsample represents upsampling to restore features, and Upsample represents bilinear interpolation. This represents edge enhancement features, and Laplacian_in_channels represents channel-by-channel edge detection. This represents the noise reduction and edge enhancement features.

[0046] 3×3 convolution extracts upper-layer input optical restoration features After local feature extraction, MaxPool downsampling reduces the feature size, suppresses high-frequency noise, and outputs coarse-grained denoised features. Subsequently, a 3×3 convolution is used to perform feature restoration, and bilinear interpolation upsampling is used to restore the features to their original size, generating upsampled restored features. For edge enhancement, a Laplacian edge detection kernel is used, employing optimal weight parameters instead of dynamic parameters to reduce computational cost. This is achieved through grouped convolution pairs. Perform channel-wise edge detection using Laplacian in-channels to highlight detailed edge features in the image and output edge enhancement features. Finally, noise suppression is achieved through channel residual calculation. Simultaneously, a 0.15x edge feature overlay strategy is employed to balance denoising and edge preservation, outputting denoised edge enhancement features. .

[0047] (c) Image Restoration: An encoder-decoder structure is used to process the denoised edge enhancement features, outputting detailed restoration features.

[0048] The purpose of this step is to address the residual blurring of details after denoising. A lightweight encoder-decoder structure is used, employing only 4 convolutional layers (without pooling). This approach controls the number of parameters while ensuring detail restoration capabilities. Image details are restored by expanding channel activation and shrinking channel restoration. The specific process and formulas are as follows:

[0049] Formula 14: ,

[0050] Formula 15: ,

[0051] in, This represents the high-order encoded features. Indicates detailed restoration features.

[0052] During the encoding stage, the number of feature channels is gradually increased through two layers of 3×3 convolutions to extract... ReLU introduces non-linear representations to output high-bit encoded features, providing deeper, more detailed information. The decoding stage uses two layers of 3×3 convolutions to progressively reduce the number of feature channels, thereby mapping the high-dimensional encoding back to the original dimension and restoring the detailed structure, outputting detailed restored features. This scheme abandons complex attention mechanisms and dense connections, and uses a simple encoder-decoder structure to achieve detailed restoration.

[0053] (d) Color Enhancement: Convolution and Softplus activation are performed on the detail restoration features to output color enhancement features. These color enhancement features are then fused with channel-dimensional global color correction features to obtain the final enhanced features of the OADEM module. The purpose of this step is to correct color shifts and resolve unnatural color transitions. Smooth color enhancement is achieved through high-dimensional unfolding, Softplus smoothing activation, and low-dimensional restoration. The specific process and formulas are as follows:

[0054] Formula 16: ,

[0055] in, Softplus represents the color enhancement feature, and Softplus represents the activation function.

[0056] Two layers of 3×3 convolutions and a ReLU activation function are used to construct a feature refinement chain to extract detailed and restored features. Color features; finally, the Softplus activation function is used: It replaces the traditional Sigmoid and ReLU, and its output range is Compared to other activation functions, this approach is more flexible and smoother, enabling smooth color enhancement and avoiding color banding caused by hard stages, thus outputting enhanced color features. Global color correction features across channel dimensions. and Residual fusion is performed to preserve the basic structure of the original color information while overlaying enhanced color details, avoiding feature distortion during the enhancement process, and outputting the final enhanced features of the OADEM module. As shown in Formula 17:

[0057] Formula 17: ,

[0058] in, This is the final enhancement feature for the OADEM module.

[0059] Furthermore, in step S4, the specific process and formulas for bi-branch feature fusion and size restoration are as follows:

[0060] Formula 18: ,

[0061] in, This represents the fusion enhancement feature, and PixelShuffle represents the pixel rearrangement function. This indicates element-wise multiplication.

[0062] OADEM module final enhancement features Local detail features of the channel dimension Element-wise multiplication guides the interplay between color enhancement and detail preservation, achieving precise matching and optimization. Then, pixel rearrangement... The operation maps the fused features back to the original spatial dimension and performs residual fusion with the original image (m is used as the pixel rearrangement coefficient to control the model parameter scale), which can avoid over-enhancement and loss of original structure, and output fused enhanced features. This is to prepare for subsequent contrast constraint adjustments.

[0063] Furthermore, in step S5, global contrast weighting and local adaptive normalization are integrated, performed in the order of global adjustment and local normalization, balancing the enhancement effect and visual naturalness. The purpose of global contrast weighting is to generate adaptive global weights, increasing the brightness of low-contrast areas and suppressing overexposure in high-contrast areas. The specific formula for global contrast weighting is as follows:

[0064] Formula 19: ,

[0065] Formula 20: ,

[0066] in, Indicates the difference between light and dark values Represents the flattening function. This indicates 8×8 average pooling. Represents 1×10 -4 Indicates the global contrast weight;

[0067] Use 8×8 average pooling Used to extract the region brightness of the original input I, the one-dimensional flattening facilitates the calculation of grid brightness extrema, and the brightness difference is calculated for each grid. 1e-4 avoids division by zero error; Formula (20) uses a non-linear formula to convert the brightness difference into a global contrast weight of [0.5,1]. Used for weighted feature enhancement The weights are used as shown in Formula 21.

[0068] Formula 21: ,

[0069] in, This represents the global contrast-weighted feature.

[0070] Fusion Enhancement Features Multiplied by global contrast weight It can brighten dark areas, suppress exposure, and overlay the original image at 0.2x magnification. It can preserve the original image scene structure and output global contrast-weighted features. .

[0071] After global contrast-weighted adjustment, some pixel values ​​in the image may exceed the normal image range, resulting in a glaring image and color loss. If the pixel value exceeds [0,1], local adaptive normalization is used to scale the pixel value. Finally, the original features are fused to preserve details. The specific calculation process and formula are as follows:

[0072] Formula 22: ,

[0073] Formula 23: ,

[0074] Formula 24: ,

[0075] in, 'Indicates the difference in local contrast. This indicates 6×6 average pooling. This represents the locally normalized feature, and Clamp represents the cutoff function. express ;

[0076] Global contrast-weighted features Perform 6×6 average pooling Extract local brightness and calculate the local contrast difference of the pooled data. A constant of 0.01 is used to prevent division by zero; then, a local normalization operation is performed on each grid to obtain the local normalized feature by stretching or compressing the brightness range of each grid to [0,1]. Finally, the Clamp function is used to forcibly truncate images exceeding a certain pixel limit. A 7:3 residual fusion strategy is employed, using 70% of the normalized natural image and 30% of the globally contrast-weighted enhanced features. This approach preserves enhanced details while maintaining a natural image quality, resulting in the final enhanced image. .

[0077] The beneficial effects of this invention are as follows:

[0078] (1) By introducing a dual-branch feature extraction architecture, the present invention decouples the global color correction and local detail preservation tasks. Combined with the optical perception detail enhancement module, it effectively solves the multi-dimensional degradation problems such as color distortion, noise interference and texture blur in underwater images, and avoids task conflicts in parallel optimization.

[0079] (2) Compared with traditional single-line processing or simple multi-task parallel methods, the present invention performs optical restoration, multi-scale denoising, detail restoration and color enhancement operations sequentially through a serial link, simulating the human visual restoration logic. It can establish a better synergistic relationship between color balance and detail preservation, and significantly improve the visual naturalness and detail restoration realism of the enhancement results.

[0080] (3) This invention not only focuses on global differences at the pixel level, but also imposes detailed constraints on optical degradation areas, so that it can maintain a robust enhancement effect in harsh underwater noise scenarios such as low light and high turbidity, and the output image performs well in both visual quality and quantitative indicators.

[0081] (4) The present invention reduces computational complexity through pixel rearrangement mechanism. The model has good lightweight characteristics and generalization ability, and can adapt to the real-time processing needs of underwater images with different resolutions such as 720p and 1080p, providing an efficient and feasible solution for underwater image enhancement. Attached Figure Description

[0082] Figure 1 This is a detailed framework diagram of the model of the present invention;

[0083] Figure 2 This is a schematic diagram of the data flow processing of the optical sensing enhancement module in step S3 of the present invention.

[0084] Figure 3 This is a schematic diagram illustrating the working principle of optical repair in step S3 of the present invention.

[0085] Figure 4 This is a schematic diagram illustrating the working principle of multi-scale noise reduction in step S3 of the present invention.

[0086] Figure 5 This is a schematic diagram illustrating the working principle of image restoration in step S3 of the present invention.

[0087] Figure 6 This is a schematic diagram illustrating the working principle of color enhancement in step S3 of the present invention.

[0088] Figure 7 A comparison of the FPS (frames per second) performance of various image enhancement models at different resolutions;

[0089] Figure 8 Comparison of computational cost (FLOPs) for various image enhancement models at different resolutions;

[0090] Figure 9 The underwater image enhancement results of different methods on the UIEB and EUVP datasets are shown, where (a) GT standard optimized set, (b) original image input set, (c) the DualOademNet model of this invention, (d) FUnIE-GAN model, (e) LiteEnhanceNet model, (f) Shallow-UWnet model, (g) Dnnet model, (h) GC model, and (i) LU2net model.

[0091] Figure 10The results of different methods for deep-sea image enhancement on the LSUI dataset are shown, where (a) GT standard optimized set, (b) original image input set, (c) the DualOademNet model of this invention, (d) FUnIE-GAN model, (e) LiteEnhanceNet model, (f) Shallow-UWnet model, (g) Dnnet model, (h) GC model, and (i) LU2net model.

[0092] Figure 11 A radar chart for quantitative evaluation of the model is provided, where MSE represents mean squared error, PSNR represents peak signal-to-noise ratio, SSIM represents structural similarity index, UIQM represents underwater image quality metric, FPS represents frame rate, DualOademNet represents the model of this invention, GC represents the mathematical technique in the prior art that adjusts image brightness and contrast through nonlinear pixel value transformation, LU2net represents the U-shaped network in the prior art specifically designed for real-time underwater image enhancement, LiteEnhanceNet represents the model in the prior art for single underwater image enhancement, FunIE-GAN represents the underwater image enhancement technique based on visual perception fusion in the prior art, Shallow-Uwnet represents the hybrid architecture model in the prior art that adopts a fully convolutional network + dense connection + residual mechanism, and DNnet represents the lightweight neural network in the prior art that uses high-resolution underwater image real-time enhancement. Detailed Implementation

[0093] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited thereto.

[0094] Detailed framework diagram of the present invention is as follows: Figure 1 As shown. This invention provides an underwater image detail enhancement method based on dual-branch optical feature fusion, which specifically includes the following steps:

[0095] In step S1, visual feature preprocessing involves performing pixel splitting on the input image through pixel inverse rearrangement. This reduces computational cost while preserving local feature structures, laying the foundation for decoupling global color and local detail features in the dual-branch process. Specifically:

[0096] Underwater image enhancement targets can be defined as: the original input image (low-quality underwater image). Through network mapping function To generate enhanced images ,Should It can approximate the real, clear underwater images in the calibration standard file of the dataset in terms of visual effects (detail, color, contrast) and quantitative indicators. The enhancement process formula is as follows:

[0097] Formula 1: ,

[0098] in As learnable parameters of the network ( ) represents the loss function.

[0099] To adapt to different real-time requirements, the pixel rearrangement coefficient m is set for the original input image. The feature extraction process is as follows:

[0100] The visual encoder and decoder obtain the converted text and latent representation of the image. The specific process and formulas are as follows:

[0101] Formula 2: ,

[0102] Where I represents the original input image, which is obtained as a block tensor feature after pixel unshuffle. The number of channels has been expanded from 3 to... The space dimensions were reduced from H×W to The computational workload is reduced to the original computational workload. This significantly improves the model inference speed and ensures real-time performance in the preprocessing stage.

[0103] In step S2, the dual-branch feature extraction involves a global color branch that extracts optical features (brightness and / or color) in the channel dimension through a 1×1 convolution, and a local branch that extracts spatial detail features (edges and / or texture) through a 3×3 convolution. This decouples the multi-branch tasks. Specifically:

[0104] Global channel color branch: The task is to extract optical features in the channel dimension and correct color shift and brightness unevenness. Therefore, 1×1 convolution is used to fuse channel information. The 1×1 convolution's lack of spatial receptive field and lightweight nature ensures that the global dimension is not affected by edge texture spatial details, and also further reduces the model's computational cost. The Tanh constraint on the feature range [-1,1] can avoid gradient explosion while extracting color deviations. The specific formula is as follows:

[0105] Formula 3: ,

[0106] Formula 4: ,

[0107] Block tensor features Global channel feature extraction is performed using 1×1 convolutions, and bidirectional color bias is captured using hyperbolic tangent function Tanh constraints to obtain global features in the channel dimension. ;Will The ReLU activation function is applied after the linear unit activation function to suppress negative noise and filter color correction features, avoiding additional noise interference from Tanh capture in subsequent enhancement. Finally, a quadratic 1×1 convolution is used to recover the effective features filtered by the ReLU activation function. Through effective weight adjustment, the abstract channel-dimensional features are transformed into channel-dimensional global color correction features that can be directly color corrected. It can also ensure the fusion of subsequent module enhancements and local detail features.

[0108] Local channel detail branch: The task is to extract spatial texture features and preserve image edges and detailed structures. Therefore, a 3×3 convolution is used to cover local pixel associations. The local pixel fusion features and local spatial receptive field of the 3×3 convolution can adapt to the spatial distribution features of underwater image details, i.e., the gray-level and color abrupt changes in the local pixel neighborhood, and can initially smooth the random noise in underwater images. It can also form a strict bidirectional decoupling of color details with the 1×1 convolution. Edge replication is used to fill and protect the edges. The specific formula is as follows:

[0109] Formula 5: ,

[0110] Formula 6: ,

[0111] Block tensor features Local channel feature extraction is performed using 3×3 convolutions and Tanh constraints are applied to obtain local features in the channel dimension. The Tanh constraint and ReLU activation are still used because the combination of the two functions can still filter positive detail features and suppress noise to a certain extent. In local details, the ReLU function can also add non-linear mapping to the features, allowing the model to learn the underwater non-linear degradation law, namely the difference in the blurring degree of texture under different turbidity and the difference in the contrast attenuation of edges under different depths. 3×3 convolution is the optimal balance between efficiency and feature representation in convolutional neural networks. Using a second 3×3 convolution to recover effective features compensates for the possible loss of weak details after ReLU filtering, and restores the feature dimension to obtain local detail features in the channel dimension. And ensure that it can be integrated with the global optimized color features in the future.

[0112] Step S3 uses the optically perceived detail enhancement module to sequentially perform optical inpainting, multi-scale denoising, image restoration, and color enhancement operations on global color branch features via a serial link. This addresses single degradation issues in stages and avoids conflicts in image optimization tasks. Specifically:

[0113] The optical perception detail enhancement module simulates the human visual restoration process, first addressing basic degradation—optical restoration—then optimizing detail effects—multi-scale denoising and image restoration—and finally enhancing visual effects—image color enhancement. Each stage focuses on a single type of degradation, achieving precise image restoration. Figure 2 A schematic diagram illustrating the data flow of the optical sensing detail enhancement module of this invention is shown. The specific process is as follows:

[0114] 1. Optical repair layer

[0115] The core objective of this layer is to address the optical distortion caused by underwater light scattering and absorption. This is achieved by capturing optical features through high-dimensional mapping, co-optimizing the feature distribution formula using dual normalization, and preserving the original features through residual connections after low-dimensional reduction. Figure 3 The working principle diagram of the optical repair layer of the present invention is shown below, and the specific process and formulas are as follows:

[0116] Formula 7: ,

[0117] Formula 8: ,

[0118] Formula 9: ,

[0119] and For batch normalization and layer normalization functions, For linear correction units with leakage, a 3×3 convolutional neural network is used. To perform dimensionality increase, the LeakyReLU activation function allows a small number of negative gradients to pass through, preventing the emergence of dead neurons, and outputs a global activation feature tensor. BatchNorm Batch normalization is performed to reduce internal covariate bias. LayerNorm performs layer normalization to balance the feature distribution of each channel and adapt to the color imbalance between underwater image channels to obtain globally normalized features. Finally, 3×3 convolution dimensionality reduction is performed, and residual connections are used to... Fusion, output optical restoration features .

[0120] 2. Multi-scale denoising layer

[0121] The core objective of this layer is to remove underwater noise pollution and resolve the contradiction with blurred edge details. A pooling denoising strategy is implemented through downsampling and bilinear interpolation. For blurred edges, channel-by-channel edge detection is used, and finally, residual connections are performed to achieve multi-scale denoising. Figure 4 The working principle diagram of the multi-scale denoising layer of this invention is shown in the figure. The specific process and formulas are as follows:

[0122] Formula 10: ,

[0123] Formula 11: ,

[0124] Formula 12: ,

[0125] Formula 13: ,

[0126] 3×3 convolution extracts upper-layer input optical restoration features After local feature extraction, MaxPool downsampling reduces the feature size, suppresses high-frequency noise, and outputs coarse-grained denoised features. Subsequently, a 3×3 convolution is used to perform feature restoration, and bilinear interpolation upsampling is used to restore the features to their original size, generating upsampled restored features. For edge enhancement, a Laplacian edge detection kernel is used, employing optimal weight parameters instead of dynamic parameters to reduce computational cost. This is achieved through grouped convolution pairs. Perform channel-wise edge detection using Laplacian in-channels to highlight detailed edge features in the image and output edge enhancement features. Finally, noise suppression is achieved through channel residual calculation. Simultaneously, a 0.15x edge feature overlay strategy is employed to balance denoising and edge preservation, outputting denoised edge enhancement features. .

[0127] 3. Image detail restoration layer

[0128] The core objective of this layer is to address the residual detail blurring after denoising. A lightweight encoder-decoder structure is employed, using only four convolutional layers (without pooling). This approach controls the number of parameters while maintaining detail restoration capabilities, restoring image details through channel expansion for activation and channel contraction for restoration. Figure 5 The working principle diagram of the image detail restoration layer of this invention is shown below. The specific process and formulas are as follows:

[0129] Formula 14: ,

[0130] Formula 15: ,

[0131] During the encoding stage, the number of feature channels is gradually increased through two layers of 3×3 convolutions to extract... ReLU introduces non-linear representations to output high-bit encoded features, providing deeper, more detailed information. The decoding stage uses two layers of 3×3 convolutions to progressively reduce the number of feature channels, thereby mapping the high-dimensional encoding back to the original dimension and restoring the detailed structure, outputting detailed restored features. This scheme abandons complex attention mechanisms and dense connections, and uses a simple encoder-decoder structure to achieve detailed restoration.

[0132] 4. Color Enhancement Layer

[0133] The core objective of this layer is to correct color shifts and resolve unnatural color transitions. This is achieved through high-dimensional expansion, activation using the Softplus smoothing function, and low-dimensional restoration to smooth color enhancement. Figure 6 The working principle diagram of the color enhancement layer of this invention is shown, and the specific process and formulas are as follows:

[0134] Formula 16: ,

[0135] Two layers of 3×3 convolutions and a ReLU activation function are used to construct a feature refinement chain to extract detailed and restored features. Color features; finally, the Softplus activation function is used: It replaces the traditional Sigmoid and ReLU, and its output range is Compared to other activation functions, this approach is more flexible and smoother, enabling smooth color enhancement and avoiding color banding caused by hard stages, thus outputting enhanced color features. .

[0136] The final output of the optical perception detail enhancement module is the residual fusion result, as shown in Equation 17, which integrates the original global color features. and Residual fusion preserves the basic structure of the original color information while superimposing enhanced color details, avoiding feature distortion during the enhancement process, and outputting the final enhanced features of the OADEM module. ,

[0137] Formula 17: ,

[0138] In step S4, the dual-branch feature fusion and size restoration involves fusing the enhanced global color channel features and local detail features element by element, and performing pixel rearrangement on the fused features to restore them to the original image size.

[0139] By fusing the dual-branch features, the original image size is restored, laying the foundation for subsequent contrast constraints. The specific formula is as follows:

[0140] Formula 18: ,

[0141] Among them, the optical perception detail enhancement module ultimately enhances the features. Local detail branch features Element-wise multiplication guides the interplay between color enhancement and detail preservation, achieving precise matching and optimization. Then, pixel rearrangement... The operation maps the fused features back to the original spatial dimension and performs residual fusion with the original image (m is used as the pixel rearrangement coefficient to control the model parameter scale), which can avoid over-enhancement and loss of original structure, and output fused enhanced features. This is to prepare for subsequent contrast constraint adjustments.

[0142] Step S5 uses global-local multi-granularity contrast constraints: A global contrast weighting module generates global contrast weights to adaptively enhance the fused features. Then, a feature adaptive normalization module performs final local normalization to constrain the pixel value range, outputting the final enhanced image. Specifically:

[0143] Integrating global contrast weighting and local adaptive normalization, performed in the order of global adjustment → local normalization, balances enhancement effects with visual naturalness. The purpose of global contrast weighting is to generate adaptive global weights. To increase brightness in low-contrast areas and suppress overexposure in high-contrast areas, the specific formula for global contrast weighting is as follows:

[0144] Formula 19: ,

[0145] Formula 20: ,

[0146] Use 8×8 average pooling Used to extract the region brightness of the original input I, the one-dimensional flattening facilitates the calculation of grid brightness extrema, and the brightness difference is calculated for each grid. 1e-4 avoids division by zero error; Formula (20) uses a non-linear formula to convert the brightness difference into a global contrast weight of [0.5,1]. Used for weighted feature enhancement The weights are used as shown in Formula 21.

[0147] Formula 21: ,

[0148] in, It is the fusion and enhancement feature of the previous modules, multiplied by the weights. It can brighten dark areas, suppress exposure, and overlay the original image at 0.2x magnification. It can preserve the original image scene structure and output global contrast-weighted features. .

[0149] After global contrast-weighted adjustment, some pixel values ​​in the image may exceed the normal image range, resulting in a glaring image and color loss. If the pixel value exceeds [0,1], local adaptive normalization is used to scale the pixel value. Finally, the original features are fused to preserve details. The specific calculation process and formula are as follows:

[0150] Formula 22: ,

[0151] Formula 23: ,

[0152] Formula 24: ,

[0153] Among them, for Perform 6×6 average pooling Extract local brightness and calculate the local contrast difference of the pooled data. A constant of 0.01 is used to prevent division by zero; then, a local normalization operation is performed on each grid to obtain the local normalized feature by stretching or compressing the brightness range of each grid to [0,1]. Finally, the Clamp function is used to forcibly truncate images exceeding the pixel limit. A 7:3 residual fusion strategy is employed, using 70% of the normalized natural image and 30% of the globally contrast-weighted enhanced features. This preserves and enhances details while maintaining a natural image quality. .

[0154] The following is the experimental setup and results analysis of this invention:

[0155] This invention uses an NVIDIA GeForce RTX 4070Ti SUPER GPU to train the model. The image size used for training is forcibly adjusted to 256×256. Based on the evaluation results of three datasets, the model performance is evaluated during training using mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and underwater image quality metric (UIQM), and the optimal model weights for each metric are saved. The following is the test dataset:

[0156] (1) UIEB dataset: contains 950 real underwater images, covering diverse underwater environments such as oceans, lakes, and artificial pools, and showcasing a rich variety of target categories such as coral reefs, fish, underwater structures, and aquatic plants.

[0157] (2) EUVP dataset: A multifunctional underwater image dataset containing approximately 11,000 synthetic subset images. The dataset is designed based on an underwater optical physics model and can simulate degradation under different "water types" (clear water, turbid water, and extremely turbid water), shooting distances, and lighting conditions.

[0158] (3) LSUI dataset: Low light underwater dataset. LSUI contains 5,004 pairs of images, making it the largest of the three datasets. All images were collected under real low light conditions, with a particular focus on low light environments such as deep sea, nighttime, and turbid waters.

[0159] To evaluate the model's performance, the model of this invention, abbreviated as DualOademNet, will be compared quantitatively and qualitatively with six current state-of-the-art methods:

[0160] GC: A mathematical technique that adjusts the brightness and contrast of an image through nonlinear pixel value transformation. It is a commonly used basic module in current underwater image enhancement models and can solve the problems of loss of details in dark areas and partial contrast imbalance in underwater images.

[0161] LU2net: A novel U-shaped network specifically designed for real-time underwater image enhancement. The proposed model combines axial depth convolution and channel attention modules, which significantly reduces computational requirements and model parameters, thereby improving processing speed.

[0162] LiteEnhanceNet: A model for single-pass underwater image enhancement that employs depthwise separable convolutions as its primary building blocks to reduce computational complexity. Single-pass aggregation connections are used to efficiently extract features from low-level and intermediate layers. Furthermore, appropriate activation functions and squeeze-activation modules are integrated at suitable locations within the network to further reduce computational complexity.

[0163] FunIE-GAN: An underwater image enhancement technique based on visual perception fusion. The method consists of three stages: color correction, contrast enhancement, and multi-task fusion.

[0164] Shallow-Uwnet is a hybrid architecture model that uses a fully convolutional network, dense connections, and residual mechanisms. It optimizes feature utilization efficiency while avoiding redundant computation. In addition, the lightweight design of the model mechanism greatly reduces the computational parameters.

[0165] DNnet is a lightweight neural network that uses high-resolution underwater images for real-time enhancement. It achieves a balance between high-resolution underwater images and real-time enhancement effects by using pixel rearrangement, FAN normalization, and CDR dynamic coordination to improve image optimization efficiency, suppress over-enhancement, and balance image brightness distribution.

[0166] The results of the comparative evaluation experiments on the UIEB dataset are shown in Table 1:

[0167] Table 1 Quantitative Comparative Evaluation of UIEB Datasets

[0168]

[0169] The results of the comparative evaluation experiments on the EUVP dataset are shown in Table 2:

[0170] Table 2 Quantitative Comparative Evaluation of EUVP Datasets

[0171]

[0172] The results of the comparative evaluation experiments on the LSUI dataset are shown in Table 3:

[0173] Table 3 Quantitative Comparison and Evaluation of the LSUI Dataset

[0174]

[0175] Real-time performance metrics on the UIEB classic underwater dataset are as follows: Figure 7 and Figure 8 As shown in the line graph. Figure 7 and Figure 8 It is known that the higher the FPS, the lower the FLOPs, the stronger the real-time performance, and the lower the required computing power. It is generally accepted that image enhancement can process images in real time at a frame rate higher than 30 FPS.

[0176] Comparison of the underwater visual quality of the model of this invention with six other models, for example Figure 9 and Figure 10 As shown. The ablation experiments in this study were divided into two aspects. The first was a quantitative analysis and comparison of ablation using a multi-branch module to verify the optimization of image details and colors by the multi-branch strategy. The second was a quantitative analysis and comparison of ablation using the Optical Inpainting Natural Enhancement (OADEM) module to verify the effectiveness of the OADEM natural branching strategy in image optimization.

[0177] The model removes local branch structures, retains global color features for use in the OADEM model, and directly performs multi-granularity global contrast fusion, resulting in a model called GlobalOademNet. Quantitative evaluation results are shown in Table 4.

[0178] Table 4. Quantitative Comparative Evaluation of Local Branch Ablation

[0179]

[0180] The OADEM module is removed, and global color features and local detail features are retained. Multi-granularity global contrast fusion is then performed directly, resulting in a variant called DualNet. The quantitative evaluation results are shown in Table 5.

[0181] Table 5 Comparative Evaluation of Ablation Quantities in OADEM Modules

[0182]

[0183] The complete architecture of the model of this invention is retained compared to six other models. Preprocessing and subsequent processing steps are excluded. In the same experimental environment, images are divided into four resolutions: 720p (1280×720 px), 1080p (1920×1080 px), 2K (2560×1440 px), and 4K (3840×2160 px). Real-time FPS measurements are performed on each model. The results of the model FPS real-time measurement are shown in Table 6.

[0184] Table 6 FPS Real-Time Measurement Results

[0185]

[0186] Radar charts showing the quantitative experimental results of the model of this invention compared to six current advanced models are shown below. Figure 11 As shown.

[0187] Figure 11 The experimental average results of seven models, including the model of this invention, on five major evaluation indicators, namely MSE, PSNR, SSIM, UIQM, and FPS, are presented, with the MSE values ​​displayed in reverse order.

[0188] Experimental Analysis: Quantitative analysis of the comparative experiments on three datasets shows that DualOademNet significantly outperforms mainstream comparison methods such as LiteEnhanceNet, FunIE-GAN, and Shallow-UWnet under different pixel rearrangement coefficients m, and exhibits an overwhelming advantage over GC. This comparison confirms the ability of the dual-branch decoupling and OADEM module to repair multi-dimensional degradation problems in underwater images. Gradient experiments with different m values ​​demonstrate that larger m values ​​reduce computational cost by compressing spatial dimensions without sacrificing core performance enhancement. Furthermore, channel expansion can significantly improve color and detail restoration, providing flexible parameter selection for real-time deployment of underwater mobile devices.

[0189] The real-time performance test results show that the FPS output frame rate of the model in this invention is higher than that of most models, and the FLOPs value does not reach the upper limit of the lightweight model index. It can efficiently perform lightweight underwater image processing tasks at 720p and 1080p. Furthermore, across four resolutions—720p, 1080p, 2K, and 4K—the model in this invention demonstrates a good balance between FPS rate and FLOPs computational load, providing a solid foundation for future lightweight real-time improvements and expansions at 2K and 4K resolutions.

[0190] The quality comparison analysis results from the three datasets show that the model of this invention is significantly better than GC, LiteEnhanceNet, Shallow-Uwnet, and DNnet models in terms of visual quality and restoration quality. Its optimization quality is on par with FUnIE-GAN and LU2net models, indicating that DualOademNet excels in restoring chromatic aberration, light scattering, underwater blur noise, and other aspects, and the output image can be clearly presented.

[0191] Quantitative analysis of the multi-branch module ablation experiment and the OADEM module ablation experiment shows that DualOademNet has the best performance in both ablation experiments. This successfully demonstrates that branch decoupling fusion and global color feature optical perception detail enhancement are the core of the model and are indispensable. They can help the model complete the key feature completion for underwater image restoration and enhancement, laying a solid foundation for the model to achieve excellent performance.

[0192] The above describes the basic principles and specific implementation process of this invention. By introducing a dual-branch feature extraction architecture and an optically perceived detail enhancement module, this invention not only achieves decoupling optimization of global color correction and local detail preservation to solve the multi-dimensional degradation problem of underwater images, but also fully considers the collaborative repair logic and feature hierarchy association weights among optical distortion, noise interference, and color shift. It can comprehensively capture the complex correlation between the optical features and spatial details of underwater images, significantly improving the naturalness, detail restoration, and real-time processing efficiency of image enhancement. This method can establish a better synergistic relationship between color balance and detail preservation, and is suitable for image enhancement tasks in complex underwater scenarios such as marine resource exploration, underwater mechanical navigation, and underwater archaeology. This model integrates both color correction accuracy and detail restoration effect information, making the enhancement effect more comprehensive and accurate, overcoming the over-enhancement and color banding deviations that may result from single-module processing or simple feature stitching. This makes the invention more robust in noisy and complex underwater data environments, adaptable to images of different resolutions such as 720p and 1080p, and able to meet diverse practical underwater image processing needs.

Claims

1. A method for enhancing underwater image details based on dual-branch optical feature fusion, characterized in that, Includes the following steps: S1. Visual feature preprocessing: Perform pixel inverse rearrangement preprocessing on the input raw underwater image to obtain block tensor features; S2. Dual-branch feature extraction: Construct a global color branch and a local detail branch. The global color branch extracts the channel-dimensional optical features from the block tensor features through a 1×1 convolution to obtain the channel-dimensional global color correction features. The local detail branch extracts the spatial-dimensional detail features from the block tensor features through a 3×3 convolution to obtain the channel-dimensional local detail features. S3, Optical Awareness Detail Enhancement: The global color correction features of the channel dimension are input into the Optical Awareness Detail Enhancement module OADEM. Optical restoration, multi-scale denoising, image restoration and color enhancement operations are performed sequentially through a serial link to obtain the final enhanced features of the OADEM module. S4. Dual-branch feature fusion and size restoration: The final enhanced features of the OADEM module are fused element-wise with the local detail features of the channel dimension, and the pixel rearrangement operation is performed on the fusion result to restore it to the same spatial size as the original underwater image, thus obtaining the fused enhanced features. S5. Global-Local Multi-Granularity Contrast Constraint: Based on the original underwater image, a global contrast weighting module generates global contrast weights, which are then used to adjust the weighted enhancement features. Local adaptive normalization is then performed to constrain the pixel value range and output the final enhanced image. In step S3, the color enhancement involves performing convolution and Softplus activation on the detail restoration features output from the image restoration to output color enhancement features. These color enhancement features are then fused with the channel-dimensional global color correction features to obtain the final enhancement features of the OADEM module. The specific process and formulas are as follows: Official 16: , Official 17: , in, Softplus represents the color enhancement feature, and Softplus represents the activation function. This is the final enhancement feature for the OADEM module; Two layers of 3×3 convolutions and a ReLU activation function are used to construct a feature refinement chain to extract detailed and restored features. Color features; finally, the Softplus activation function is used: It replaces the traditional Sigmoid and ReLU, and its output range is Output color enhancement features Global color correction features in the channel dimension and Perform residual fusion to output the final enhanced features of the OADEM module. .

2. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S1, the formula for visual feature preprocessing is as follows: Official 2: , in, Let represent the block tensor features, PixelUnshuffle represent the pixel unshuffle function, I represent the original underwater image, and m represent the pixel rearrangement coefficients. Indicates the number of channels. Indicates the image size.

3. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S2, the specific process and formula for the global color branch are as follows: Official 3: , Official 4: , in, Tanh represents the global feature of the channel dimension, and it represents the hyperbolic tangent function. Represents a 1×1 convolution. Represents the global color correction feature at the channel level, abbreviated as ReLU represents the modified linear unit activation function; Block tensor features Global channel feature extraction is performed using 1×1 convolutions, and bidirectional color bias is captured using hyperbolic tangent function Tanh constraints to obtain global features in the channel dimension. ,Will The ReLU activation function is applied, and finally a 1×1 convolution is used to recover the effective features filtered by the ReLU activation function, thus obtaining the channel-dimensional global color correction features. .

4. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S2, the specific process and formula for the local detail branch are as follows: Official 5: , Official 6: , in, For channel-dimensional local features, Local detail features in the channel dimension, abbreviated as , It is a 3×3 convolution; Block tensor features Local channel feature extraction is performed using 3×3 convolution and constrained by the hyperbolic tangent function Tanh to obtain local features in the channel dimension. ReLU activation is employed, and 3×3 convolution is used to recover effective features, thereby obtaining local detail features in the channel dimension. .

5. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S3, the optical restoration involves processing the global color correction features of the channel dimension through convolution, LeakyReLU activation, batch normalization, and layer normalization, and then outputting the optical restoration features through residual connections. The specific process and formula are as follows: Official 7: , Official 8: , Official 9: , in, For the global activation feature tensor, This is a global color correction feature for the channel dimension. For globally normalized features, For batch normalization function, For layer normalization function, For a linear correction unit with leakage, It is an optical restoration feature; Using a 3×3 convolutional neural network Dimensional upscaling is performed, and the LeakyReLu function is used to output the global activation feature tensor. BatchNorm Perform batch normalization, then LayerNorm performs layer-level normalization to obtain globally normalized features. Finally, 3×3 convolution dimensionality reduction is performed, and residual connections are used to... Fusion, output optical restoration features .

6. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S3, the multi-scale denoising involves performing downsampling, upsampling, and edge detection operations on the optical restoration features output by the optical restoration, and outputting denoised edge enhancement features through residual calculation. The specific process and formula are as follows: Official 10: , Official 11: , Official 12: , Official 13: , in, MaxPool represents coarse-grained denoising features, while MaxPool represents downsampling. Upsample represents upsampling to restore features, and Upsample represents bilinear interpolation. This represents edge enhancement features, and Laplacian_in_channels represents channel-by-channel edge detection. This represents the denoising edge enhancement feature; Optical restoration features obtained from optical restoration were extracted using 3×3 convolution. After local feature extraction, MaxPool downsampling reduces the feature size, suppresses high-frequency noise, and outputs coarse-grained denoised features. Subsequently, a 3×3 convolution is used to perform feature restoration, and bilinear interpolation upsampling is used to restore the features to their original size, generating upsampled restored features. ; through grouped convolution pairs Perform channel-wise edge detection using Laplacian in-channels and output edge enhancement features. Finally, noise suppression is achieved through channel residual calculation, while a 0.15x edge enhancement feature overlay strategy is used to balance denoising and edge preservation effects, outputting denoised edge enhancement features. .

7. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S3, the image restoration involves processing the denoised edge enhancement features of the multi-scale denoised output using an encoder-decoder structure to output detail restoration features. The specific process and formula are as follows: Official 14: , Official 15: , in, This represents the high-order encoded features. This represents the edge enhancement feature for denoising. Indicates detailed restoration features, During the encoding stage, the number of feature channels is gradually increased through two layers of 3×3 convolutions to extract... ReLU introduces non-linear representations to output high-bit encoded features, providing deeper, more detailed information. The decoding stage uses two layers of 3×3 convolutions to progressively reduce the number of feature channels, thereby mapping the high-dimensional encoding back to the original dimension and restoring the detailed structure, outputting detailed restored features. .

8. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S4, the specific process and formulas for bi-branch feature fusion and size restoration are as follows: Official 18: , in, This represents the fusion enhancement feature, and PixelShuffle represents the pixel rearrangement function. This indicates element-wise multiplication, where m represents the pixel rearrangement coefficient; OADEM module final enhancement features Local detail features of the channel dimension Element-wise multiplication, then pixel rearrangement The operation maps the fused features back to the original spatial dimension and performs residual fusion with the original underwater image I, outputting the fused enhanced features. .

9. The underwater image detail enhancement method based on dual-branch optical feature fusion according to claim 1, characterized in that, In step S5, global contrast weighting and local adaptive normalization are integrated, and the process and formulas are as follows: Official 19: , Official 20: , Official 21: , Official 22: , Official 23: , Official 24: , in, Indicates the difference between light and dark values Represents the flattening function. This indicates 8×8 average pooling. Represents 1×10 -4 Represents the nonlinear global contrast weight. This represents the global contrast-weighted feature. 'Indicates the difference in local contrast. This indicates 6×6 average pooling. Represents the local normalized feature, Clamp represents the cutoff function, I t This represents the final enhanced image; Use 8×8 average pooling Extracting the region brightness from the original underwater image I, the Flatten function is used to calculate the grid brightness extrema; fusing and enhancing features. Multiplied by non-linear global contrast weight Achieve optimized effects such as brightening dark areas and suppressing exposure, while overlaying the original underwater image at 0.2x magnification. It can preserve the original image scene structure and output global contrast-weighted features. , Global contrast-weighted features Perform 6×6 average pooling Extract local brightness and calculate the local contrast difference of the pooled data. A constant of 0.01 is used to prevent division by zero; then, each mesh is locally normalized to obtain the local normalized feature by stretching or compressing the brightness range of each mesh to [0,1]. Finally, the Clamp function is used to forcibly truncate images exceeding a certain pixel limit. A 7:3 residual fusion strategy is employed, using 70% of the normalized natural image and 30% of the globally contrast-weighted enhanced features to output the final enhanced image. .

Images