Lightweight underwater image enhancement method based on adaptive feature fusion

Abstract

The invention provides a lightweight underwater image enhancement method based on adaptive feature fusion, and the method comprises the steps: decomposing an underwater image enhancement task into a strong prior enhancement task and a fine-grained enhancement task through employing a lightweight underwater image enhancement network based on adaptive feature fusion, so as to make full use of the multi-scale information in an image; the strong prior enhancement task is used for extracting global and low-layer features of the image; the fine-grained enhancement task is used for improving local details and color recovery of the image so as to improve the visual quality; the lightweight underwater image enhancement network comprises a multi-scale feature aggregation module and a global channel attention module, the multi-scale feature aggregation module is composed of an adaptive feature fusion module and a segmented convolution feed-forward network, and the lightweight underwater image enhancement network is used for combining global and local information of features and adaptively adjusting importance of each channel through information association between channels.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular to a lightweight underwater image enhancement method based on adaptive feature fusion. Background Art

[0002] Underwater image enhancement is a key image processing task in underwater operations, which aims to solve the problems of low image contrast, color distortion and blur caused by light absorption and scattering. Traditional image enhancement methods based on physical models and statistical characteristics are often limited in complex and dynamic underwater environments. In recent years, the rapid development of deep learning methods has enabled neural networks to be gradually applied in underwater image enhancement. They can autonomously learn complex image enhancement patterns. Convolutional neural networks are widely used in underwater image enhancement tasks due to their visual feature learning capabilities. However, existing deep learning models are often constrained by a large number of parameters and high computational complexity on resource-constrained embedded platforms.

[0003] In order to solve the problem of limited computing resources of embedded platforms in underwater environments, existing technologies have proposed some lightweight underwater image enhancement models that use less computing resources than other methods. For example, Five A+ [1] The network adopts a two-stage enhancement structure. The first stage aims to decompose the challenging underwater degradation into sub-problems, while the second stage contains a multi-branch color enhancement module and a pixel attention module to amplify the network's perception of details. Song et al. [2] An efficient residual dense block search algorithm with multiple objectives is proposed to identify fast, lightweight and accurate networks for image super-resolution. The method combines residual dense blocks in the neural architecture search process and optimizes network performance using a reinforcement learning strategy. Guo et al. [4] We propose an efficient low-light image enhancement method that estimates the illumination of each pixel individually and improves it using a structural prior. [3] The network constructs a shallow network by introducing lightweight network components and residual convolution blocks. [5] A Ranker Loss method is proposed, which is specifically used for underwater image enhancement. This method uses learning to judge the quality of image enhancement and optimize the enhancement process to improve the quality of underwater images. These methods all keep the model lightweight.

[0004] However, underwater images suffer from reduced contrast, color distortion, and image blur due to light absorption and scattering, which has always seriously affected the visual effect of underwater operations and the accuracy of subsequent processing. Figure 1 The main technical problems of underwater image enhancement at present include:

[0005] Low contrast: There are limited light sources in underwater environments, and short-wave light is quickly absorbed by water, resulting in low image contrast and loss of detail information.

[0006] Color distortion: Because water absorbs light of different wavelengths at different rates, especially red light, underwater images often appear blue or green, which not only affects the visual effect, but also the accuracy of subsequent image processing tasks.

[0007] Image blur: Factors such as suspended particles in the water and light scattering cause underwater images to be blurred, which poses a challenge to enhancement processing, especially when it is difficult to achieve ideal results when enhancing details and edges.

[0008] Real-time requirements: In embedded application scenarios such as underwater robots, computing resources are limited and images need to be processed in real time. However, existing deep learning models often have large parameters and high computational complexity, making it difficult to run efficiently on resource-constrained platforms.

[0009] Insufficient utilization of multi-scale information: The underwater environment is dynamic and complex. Traditional image enhancement methods or deep learning models often cannot fully utilize the multi-scale information in the image, resulting in poor enhancement effects, especially when it is necessary to process global information and local details at the same time.

[0010] To address these technical issues, it is necessary to develop a lightweight underwater image enhancement network that can process in real time at low computational cost and can effectively process multi-scale features to improve image quality, enhance details, and adapt to changes in complex underwater environments. Li et al. [6] proposed a multi-color spatial embedding network for underwater image enhancement guided by dielectric transfer, called Ucolor. Huo et al. [7] Wavelet reinforcement learning units are used to decompose hierarchical features into high-frequency and low-frequency components, and then normalization and attention mechanisms are used to strengthen them. Although this method has shown excellent visual results, its extensive network parameters and computational requirements make it unsuitable for existing underwater equipment. In addition, it cannot effectively solve the problem of color distortion.

[0011] References involved in the present invention include:

[0012] [1]Jingxia Jiang,Tian Ye,Jinbin Bai,Sixiang Chen,Wenhao Chai,Shi Jun,Yun Liu,and Erkang Chen.Five A+Network:You only need 9k parameters forunderwater image enhancement.In British Machine Vision Conference,pages 149–151,2023.

[0013] [2]Dehua Song,Chang Xu,Xu Jia,Yiyi Chen,Chunjing Xu,and YunheWang.Efficient residual dense block search for image super-resolution.InProceedings ofthe AAAI Conference on Artificial Intelligence,volume 34,pages12007–12014,2020.

[0014] [3]Ankita Naik,Apurva Swarnakar,and Kartik Mittal.Shallow-uwnet:Compressed model for underwater image enhancement(student abstract).InProceedings ofthe AAAI Conference on Artificial Intelligence,volume35,pages15853–15854,2021.

[0015] [4]Xiaojie Guo,Yu Li,and Haibin Ling.Lime:Low-light image enhancementvia illumi-nation map estimation.IEEE Transactions on Image Processing,26(2):982–993,2017.

[0016] [5]Chunle Guo,Ruiqi Wu,Xin Jin,Linghao Han,Zhi Chai,Weidong Zhang,andChongyi Li.Underwater ranker:Learn which is better and how to be better.InProceedings oftheAAAI Conference onArtificial Intelligence,2023.

[0017] [6]Chongyi Li,Saeed Anwar,Junhui Hou,Runmin Cong,Chunle Guo,and WenqiRen.Underwater image enhancement via medium transmission-guided multi-colorspace embedding.IEEE Transactions on Image Processing,30:4985–5000,2021.

[0018] [7]Fushuo Huo,Bingheng Li,and Xuegui Zhu.Efficient wavelet boostlearning-based multi-stage progressive refinement network for underwaterimage enhancement.In Proceedings of the IEEE / CVF International Conference onComputer Vision,pages 1944–1952,2021.

[0019] [8]Dosovitskiy,Alexey,et al."An image is worth 16x16 words:Transformers for image recognition at scale."arXiv preprint arXiv:2010.11929(2020).

[0020] [9]Jari Korhonen and Junyong You.Peak signal-to-noise ratiorevisited:Is simple beautiful?In 2012 Fourth International Workshop onQuality of Multimedia Experience,pages 37–38,2012.

[0021]

[10] Zhou Wang,Alan C Bovik,Hamid R Sheikh,and Eero P Simoncelli.Imagequality assessment:from error visibility to structural similarity.IEEEtransactions on image processing,13(4):600–612,2004.

[0022]

[11] Miao Yang and Arcot Sowmya.An underwater color image qualityevaluation metric.IEEE Transactions on Image Processing,24(12):6062–6071,2015.

[0023]

[12] Karen Panetta,Chen Gao,and Sos Agaian.Human-visual-system-inspired underwater image quality measures.IEEE Journal of OceanicEngineering,41(3):541–551,2016.

[0024]

[13] Chongyi Li,Chunle Guo,Wenqi Ren,Runmin Cong,Junhui Hou,Sam Kwong,and Dacheng Tao.An underwater image enhancement benchmark dataset andbeyond.IEEE Transactions on Image Processing,29:4376–4389,2019.

[0025]

[14] Zhenqi Fu, Wu Wang, Yue Huang, Xinghao Ding, and Kai-KuangMa. Uncertainty inspired underwater image enhancement. In Computer Vision–ECCV2022:17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVIII, pages 465–482. Springer, 2022. Summary of the invention

[0026] In view of the defects and shortcomings of the above-mentioned prior art, the present invention proposes a new underwater image enhancement network, based on the lightweight underwater image enhancement network (LAFF-Net) with adaptive feature fusion, which decomposes the underwater image enhancement task into two branch tasks---strong prior enhancement task and fine-grained enhancement task, so as to make full use of the multi-scale information in the image, minimize the parameters and computational complexity, and provide excellent image enhancement performance in complex underwater scenes. The present invention introduces a multi-scale feature aggregation module and a global channel attention module. The adaptive feature fusion module and the segmented convolution feedforward network constitute a multi-scale feature aggregation module. These technologies combine the global and local information of the features, and adaptively adjust the importance of each channel through the information association between channels, so as to achieve efficient image enhancement on the embedded platform while maintaining a low parameter amount, so as to be more suitable for real-time applications in embedded scenarios. Specifically, the proposed algorithm first provides global features and low-level features for the network in the strong prior enhancement task, and builds the foundation of the entire network. In the fine-grained enhancement task, it focuses on enhancing the local details and colors of the image, thereby improving the visual quality. It aims to effectively extract complex multi-scale features of underwater images while keeping the model's parameters and computational complexity low on an embedded platform. This technology is widely used in practical scenarios such as underwater robots.

[0027] The present invention specifically adopts the following technical solutions:

[0028] A lightweight underwater image enhancement method based on adaptive feature fusion: a lightweight underwater image enhancement network based on adaptive feature fusion is adopted to decompose the underwater image enhancement task into a strong prior enhancement task and a fine-grained enhancement task to make full use of the multi-scale information in the image; the strong prior enhancement task is used to extract the global and low-level features of the image; the fine-grained enhancement task is used to enhance the local details and color restoration of the image, thereby improving the visual quality; the lightweight underwater image enhancement network includes a multi-scale feature aggregation module composed of an adaptive feature fusion module and a segmented convolution feedforward network, and a global channel attention module, which is used to combine the global and local information of the features, and adaptively adjust the importance of each channel through the information association between channels.

[0029] Lightweight convolution is used in each module of the lightweight underwater image enhancement network, and channel dimensions are controlled during feature aggregation to achieve precise control of low parameters.

[0030] Furthermore, in the lightweight underwater image enhancement network, the input image first performs the strong prior enhancement task after point-by-point convolution. In the strong prior enhancement task, one branch passes through a multi-branch color enhancement module, a multi-scale feature aggregation module and a global channel attention module, and the other branch passes through a multi-scale enhancement network and then jointly inputs into a mixed domain feature interaction module for fusion and then enters into the fine-grained enhancement task; in the fine-grained enhancement task, it first undergoes point-by-point convolution, and then sequentially inputs into a multi-scale feature aggregation module, a global channel attention module, a multi-branch color enhancement module and a channel attention layer.

[0031] The global channel attention module improves the representation ability of important channels in the feature map by calculating channel weights, and the multi-scale feature aggregation module enhances the network's ability to process global and local information by extracting and fusing features of global and local information.

[0032] Furthermore, the multi-scale feature aggregation module combines an adaptive feature fusion module and a piecewise convolutional feedforward network; first, the adaptive feature fusion module is used to perform spatial enhancement of features, extract global features and enhance local details; then, the piecewise convolutional feedforward network is used to retain some original features while enhancing some features; and feature normalization is applied to the feature input of each step to ensure that the input features remain numerically stable.

[0033] Furthermore, the adaptive feature fusion module combines two branches: self-attention approximation and local detail estimation; the self-attention approximation is used to capture global features through an adaptive pooling method to reduce computational complexity; the local detail estimation is used to enhance local detail processing capabilities and achieve detail recovery through 1×1 convolution:

[0034] The input features are channel-added through the convolution layer, and then split into two parts in the channel dimension to obtain feature maps X1 and X2 respectively;

[0035] For feature map X1, adaptive maximum pooling is first performed to generate a downsampled feature map to capture spatial context information; then, the pooling result is further processed by deep convolution to enhance local detail information; and the variance of feature map X1 in the spatial dimension is calculated to extract global statistical information and obtain the global context of the feature map; the local pooled features are multiplied by a learnable weight parameter alpha to amplify local feature information; the global variance features are multiplied by another learnable parameter belt to enhance global feature information; the local and global features are weighted summed to fuse information of different scales; the fused features are nonlinearly processed by the GELU activation function, and the fused features are further compressed and adjusted by 1x1 convolution; finally, the nearest neighbor interpolation is used to restore the fused features from the downsampled size to the original size;

[0036] The feature map X2 is locally enhanced through a depthwise separable convolution layer, a pointwise convolution layer, and a GELU activation function; the depthwise separable convolution layer is composed of a combination of depthwise convolution and pointwise convolution.

[0037] Furthermore, the piecewise convolutional feedforward network first performs dimensionality transformation through 1x1 convolution, and then uses the activation function GELU to enhance nonlinearity and expand the input feature dimension; then, in the training stage, the segmentation operation is used to divide the hidden layer features into two parts in the channel dimension: one part learns local features through 3x3 convolution, and uses the activation function GELU to enhance the nonlinear feature expression; the other part keeps the original features unchanged; these features are processed through different convolutional layers, and then re-spliced, and dimensionality transformation is performed using 1x1 convolution; in the inference stage, the local features are directly convolved, and no segmentation operation is used to simplify the inference process.

[0038] Furthermore, the global channel attention module performs weighted processing on the input feature map through a channel-level attention mechanism to enhance the representation capability of important features; and uses an activation function Softsign to balance performance and computational complexity to obtain a weight coefficient for each channel;

[0039] The global channel attention module first performs global average pooling on each channel and calculates the weight, expressed as:

[0040]

[0041] z f =Flatten(z m ),

[0042] y1=S(W1zf ),

[0043] y2=S(W2y1),

[0044] Where X m represents the mth feature map that has been extracted; z m represents the average global pooling of the mth channel; z f It means that the tensor z m The process of flattening to a one-dimensional vector, W1 and W2 are weight matrices used to calculate z f Perform a linear transformation and use W1 to transform the one-dimensional vector z f Channel attenuation is performed to reduce the number of channels from the original to 1 / 4, and then W2 is used to increase the dimension of the one-dimensional vector y1 to obtain the weight of each channel; the symbol S represents the Softsign activation function, and y1 is activated by using the weight matrix z f After dimensionality reduction, y2 is obtained as the initial result of the channel attention weight after dimension recovery, which is a one-dimensional vector;

[0045] Then resize and reweight the channels:

[0046] A e = expand(y2),

[0047] X′=X×A e ,

[0048] Among them A e Indicates adjusting the dimension of y2 to match the dimension of the input feature map X;

[0049] Finally, the weight A e Multiply it with the input feature map X to get a new feature map X'.

[0050] Furthermore, the lightweight underwater image enhancement network is optimized by combining loss terms; the training objective is divided into the perceptual loss L vgg , SSIM loss, Charbonnier Loss, combined in a weighted manner:

[0051] L total =λ1L vgg1 +λ2L ssim +λ3L c +λ4L vgg2

[0052] Among them, λ1,λ2,λ3,λ4 are all weight parameters;

[0053] The formula for Charbonnier Loss is as follows:

[0054]

[0055] Where X is the predicted value and Y is the target value; ∈ is used to prevent numerical instability when the sum of squares approaches zero.

[0056] A lightweight underwater image enhancement system based on adaptive feature fusion comprises: a lightweight underwater image enhancement network based on adaptive feature fusion, comprising a strong prior enhancement task module and a fine-grained enhancement task module, so as to make full use of multi-scale information in an image; the strong prior enhancement task module is used to extract global and low-level features of an image; the fine-grained enhancement task module is used to enhance local details and color restoration of an image, thereby improving visual quality; the lightweight underwater image enhancement network comprises a multi-scale feature aggregation module and a global channel attention module composed of an adaptive feature fusion module and a segmented convolution feedforward network, so as to combine global and local information of features, and adaptively adjust the importance of each channel through information association between channels.

[0057] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the lightweight underwater image enhancement method based on adaptive feature fusion as described above are implemented.

[0058] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the lightweight underwater image enhancement method based on adaptive feature fusion as described above.

[0059] Compared with the prior art, the present invention and its preferred embodiment have the following advantages:

[0060] The network of the present invention enhances the expressiveness of features compared to existing solutions, especially in complex underwater environments. It can capture richer details and global information, ensuring that when processing blurred and low-contrast images, it can maintain global consistency and accurately enhance local details. This is mainly reflected in the following points:

[0061] Better feature expression: The global channel attention module adaptively adjusts the channel attention so that the network can focus on more representative channels, avoiding redundancy or loss of channel information, and the computational complexity is very low. The multi-scale feature aggregation module processes global and local features through a multi-path structure, making the feature expression more comprehensive and accurate, and adopts residual links to ensure that feature information is not lost.

[0062] Enhanced detail capture capability: Underwater images are often blurred and have detail loss due to poor lighting conditions, physical obstacles, and noise. The multi-scale feature aggregation module, through its unique local enhancement mechanism, can enhance the expression of local details while maintaining global information, helping to capture more effective information, and obtain global statistical information by calculating variance, thereby further maintaining the integrity of global features.

[0063] Strong adaptability: The complexity of the underwater environment requires the network to have strong adaptability. The global channel attention module and the multi-scale feature aggregation module introduce learnable weights in the training process (such as the adaptive channel attention weights in GCAM and the alpha and belt parameters in MFA), so that the network can adaptively adjust according to the input features, thereby processing different types of underwater images more flexibly. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0065] Figure 1 Schematic diagram of underwater imaging;

[0066] Figure 2 Schematic diagram of the network model architecture of an embodiment of the present invention;

[0067] Figure 3 A framework diagram of an adaptive feature fusion module according to an embodiment of the present invention;

[0068] Figure 4 The segmented convolution feedforward network diagram of the embodiment of the present invention;

[0069] Figure 5 Visualization results of different underwater image enhancement networks according to embodiments of the present invention. DETAILED DESCRIPTION

[0070] In order to make the features and advantages of this patent more obvious and easy to understand, the following embodiments are specifically described in detail with the accompanying drawings as follows:

[0071] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0072] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0073] In view of the defects and shortcomings of the above-mentioned prior art, the present invention proposes a new underwater image enhancement network, based on the lightweight underwater image enhancement network (LAFF-Net) with adaptive feature fusion, which decomposes the underwater image enhancement task into two branch tasks---strong prior enhancement task and fine-grained enhancement task, so as to make full use of the multi-scale information in the image, minimize the parameters and computational complexity, and provide excellent image enhancement performance in complex underwater scenes. The present invention introduces a multi-scale feature aggregation module and a global channel attention module. The adaptive feature fusion module and the segmented convolution feedforward network constitute a multi-scale feature aggregation module. These technologies combine the global and local information of the features, and adaptively adjust the importance of each channel through the information association between channels, so as to achieve efficient image enhancement on the embedded platform while maintaining a low parameter amount, so as to be more suitable for real-time applications in embedded scenarios. Specifically, the proposed algorithm first provides global features and low-level features for the network in the strong prior enhancement task, and builds the foundation of the entire network. In the fine-grained enhancement task, it focuses on enhancing the local details and colors of the image, thereby improving the visual quality. It aims to effectively extract complex multi-scale features of underwater images while keeping the model's parameters and computational complexity low on an embedded platform. This technology is widely used in practical scenarios such as underwater robots.

[0074] The overall framework of the algorithm of the embodiment of the present invention is as follows:

[0075] 1. Model optimization and lightweight design

[0076] The underwater image enhancement network of the present invention adopts a lightweight design. To ensure the real-time performance of the network on the embedded platform, the network structure is designed to reduce the amount of calculation and the number of parameters as much as possible. Lightweight convolution (such as 1×1 convolution) is introduced in each module, and the channel dimension is strictly controlled during feature aggregation to achieve precise control of low parameters. The network uses only about 14k parameters, which can achieve real-time processing with low computational overhead in underwater scenes.

[0077] 2 Network Structure Design

[0078] like Figure 2As shown in the figure, the overall design of the algorithm of the present invention is divided into two parts: the strong prior and fine-grained stages: the strong prior stage is used to extract the global and low-level features of the image and provide the network with basic information about the image. The fine-grained stage focuses on improving the local details and color restoration of the image and enhancing the visual quality of the image. The following are the functions of each module of the network:

[0079] The Multiscale Enhancement Network (MEN) in this scheme realizes multi-scale feature extraction and fusion, improving the details and visual effects of the image. It uses different convolution, pooling and upsampling operations to perform feature fusion in the channel dimension.

[0080] The multi-branch color enhancement block (MCEBlock, from the MCEM module in the paper Jingxia Jiang, TianYe, Jinbin Bai, Sixiang Chen, Wenhao Chai, Shi Jun, Yun Liu, and Erkang Chen. Five A+ Network: You only need 9k parameters for underwater image enhancement. In British Machine Vision Conference, pages 149–151, 2023) in this scheme achieves efficient color restoration by processing the red, green, blue and remaining channels of the input feature map separately. This design ensures that the color information is fully extracted and enhanced in the channel dimension, and effectively avoids the color confusion problem by independently processing each channel. At the same time, the fourth branch retains the original input information to enhance the integrity of the overall feature expression and achieve high-precision color restoration.

[0081] The mixed domain feature interaction module (MDFIM, from the SFDIM module in the paper Jingxia Jiang, TianYe, Jinbin Bai, Sixiang Chen, Wenhao Chai, Shi Jun, Yun Liu, and Erkang Chen. Five A+Network: You only need 9k parameters for underwater image enhancement. In British Machine Vision Conference, pages 149–151, 2023.) in this scheme makes full use of the frequency information of the image to enhance the detail expression by performing feature interaction between the frequency domain and the spatial domain.

[0082] In order to further improve the network's feature representation ability and information fusion effect, this paper introduces two improved key modules: the global channel attention module (GCAM) and the multi-scale feature aggregation module (MFA). The global channel attention module improves the representation ability of important channels in the feature map by calculating channel weights, while the multi-scale feature aggregation module enhances the network's ability to process global and local information through feature extraction and fusion of global and local information.

[0083] Introduction to the overall process of the network:

[0084] like Figure 2 As shown, the network of the present invention first passes the input image Input (3-channel RGB image) through a convolution layer (Conv1) to convert it into a feature mapping space, and the number of channels of the feature map is 16. Then, the feature map is processed using a multi-scale enhancement network, and an enhanced image is obtained by utilizing a feature pyramid. The feature map obtains multi-scale information through average pooling operations of different scales (pooling of 128, 64, and 32 sizes). After the information of each scale is processed by the convolution layer, it is upsampled to the original size and spliced ​​with the initial feature map. This step can capture the detailed features of different scales in the image and further improve the clarity and detail recovery ability of the image.

[0085] At the same time, the network uses a multi-branch color enhancement module to process different color channels separately. The input feature map first divides the channel into 4 parts (R, G, B and the remaining channels), and processes the single image pixel through 1x1 convolution and compresses the channel to reduce the computational complexity. Then, for each channel, the instance normalization operation is used to remove the statistical deviation in the color features. The normalized features are restored to the initial number of channels again through 1x1 convolution, and then the outputs of the three channels are added and fused with the original features of the fourth branch, and finally the processed outputs are integrated through the splicing operation. After that, the multi-scale feature aggregation module and the global channel attention module work together to enhance the global detail processing and information fusion capabilities of the image.

[0086] The hybrid domain feature interaction module further fuses the outputs of the global channel attention module and the multi-scale enhancement network to enhance the feature characterization capability. In the fusion step, the output feature maps of the global channel attention module and the multi-scale enhancement network are added element by element, and then the feature maps are converted to the frequency domain for processing. The input features are decomposed into amplitude and phase features by fast Fourier transform (FFT), and then these amplitude and phase information are convolved to enhance their detail characteristics. Subsequently, the frequency domain features are restored to spatial domain features by complex operations and inverse Fourier transforms. The frequency domain information is then fused with the original spatial domain information in a linear weighted manner to enhance the expressive power of the features, wherein the hyperparameter α is used to control the fusion ratio of spatial and frequency domain information, and α in the present embodiment is preferably 0.1. This frequency domain processing can further enhance the expressive power of the features, especially for capturing low-frequency information and global features.

[0087] Based on the fused feature map, the network further enhances the image details through a multi-scale feature aggregation module and a global channel attention module. Then a multi-branch color enhancement module is used to process different color channels separately, and further color correction and restoration are performed. As a more optimized processing method, this embodiment further uses a simple channel attention layer SCAL (from the paper Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Enhua Wu. "Squeeze-and-Excitation Networks." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. The proposed module) performs channel attention weighting on the color correction of the multi-branch color enhancement module to better fuse the details of important channels to restore the color. As a preferred method, the channel attenuation of the channel attention layer is set to 4, which can enhance the feature representation ability and generalization ability of the model while retaining more channel information, and adapt to complex tasks. The simple channel attention layer uses average pooling and then passes through two convolutional layers for channel attenuation and restoration, and then uses the Sigmoid function to map the output to a value between [0,1] to weight the channel importance. Finally, a simple spatial attention module (preferably the spatial attention module proposed in the paper Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Enhua Wu. "Squeeze-and-Excitation Networks." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.) is used to further enhance the key details of the image. This module generates a spatial attention map through a simple convolutional network (1x1 convolution), and then multiplies the attention map with the input feature map element by element, thereby achieving spatial weighting of the input features.

[0088] At the same time, after processing in each module, the feature map is gradually refined through the convolutional layers Conv2 (after MDFIM), Conv3 (used to convert back to a 3-channel RGB image at the end) and Conv4 (after MDFIM), and the restored image is finally output. At the same time, Conv4 is used as an intermediate output to provide phased feature extraction results to ensure the stability of the network during the training process.

[0089] Introduction of multi-scale feature aggregation module and global channel attention module:

[0090] The main innovation of this solution lies in the structural improvement of the multi-scale feature aggregation module and the improvement of the global channel attention module. By designing a more efficient feature fusion mechanism, it works together with the existing MEN, MCEBlock, and MDFIM modules to achieve deep fusion of global and local information. The multi-scale feature aggregation module and the global channel attention module are introduced below:

[0091] Multi-scale feature aggregation module such as Figure 2 As shown in the figure, the design of this module transmits and fuses the features of the output of the adaptive feature fusion module (AFFM) and the segmented convolutional feedforward network (SCFN) through residual connections, and uses Normalize to enhance nonlinearity after the input and adaptive feature fusion modules, thereby enhancing the network's multi-scale information processing capabilities. The introduction of residual connections preserves the original input information, effectively alleviates the gradient vanishing problem, and accelerates the convergence process of the network. The multi-scale feature aggregation module allows the network to focus on the overall and detailed information of the image at the same time, which is particularly suitable for processing complex underwater scenes. The following is a step-by-step introduction to its processing details.

[0092] The multi-scale feature aggregation module is composed of an adaptive feature fusion module and a segmented convolutional feedforward network. The module first performs spatial enhancement of features through the adaptive feature fusion module, extracts global features and enhances local details. Next, the segmented convolutional feedforward network retains some of the original features while enhancing some of the features. Feature normalization is applied to the feature input of each step to ensure that the input features remain numerically stable, which helps to improve the stability of the model training process. By combining the adaptive feature fusion module and the segmented convolutional feedforward network, both local features can be enhanced and global feature information can be maintained.

[0093] Adaptive feature fusion module such as Figure 3 As shown in the figure, this module combines two branches: self-attention approximation and local detail estimation. The former is used to capture global features through adaptive pooling methods to reduce computational complexity; the latter enhances local detail processing capabilities and achieves detail recovery through 1×1 convolution. The dual branches designed in this way can effectively balance global information and detail information. In addition, the variance is calculated to balance global statistical information and improve the robustness of the network in multi-scale environments.

[0094] The process of the adaptive feature fusion module in the overall network is to increase the channel of the input feature through the convolution layer, and then split it into two parts in the channel dimension to obtain feature maps X1 and X2 respectively. Among them, X1 is further processed, firstly adaptive maximum pooling is performed to generate a downsampled feature map to capture spatial context information. Subsequently, the pooling result is further processed by deep convolution to enhance local detail information. And the variance of feature X1 in the spatial dimension (i.e. in the height and width dimensions) is calculated to extract global statistical information. This step provides the global context of the feature map. The local pooled features are multiplied by the learnable weight parameter alpha to amplify the local feature information. The global variance feature is multiplied by another learnable parameter belt to enhance the global feature information. The local and global features are weighted summed to fuse information of different scales. These parameters are obtained through training. The fused features are processed nonlinearly by the GELU activation function to enhance the feature expression ability. The fused features are further compressed and adjusted by 1x1 convolution. The fused features are restored from the downsampled size to the original size using nearest neighbor interpolation. The model can learn multi-scale information from feature maps of different resolutions.

[0095] X2 uses two convolutional layers, namely a depthwise separable convolutional layer and a pointwise convolutional layer, and an activation function for local enhancement. The first convolutional layer is composed of a combination of depthwise convolution and pointwise convolution. The purpose is to extract more effective features with less computation. The depthwise convolution only performs convolution operations on each channel independently to maintain the feature extraction capability of the spatial dimension. Pointwise convolution (1x1 convolution) is used to mix channel information to integrate the features of each channel. The second convolutional layer is still a 1x1 convolution, and its function is to restore the original number of channels. The activation function uses GELU, which can handle negative values ​​more effectively than ReLU, which helps the model retain more feature details during training. By combining depthwise convolution and pointwise convolution, features can be processed efficiently while maintaining computational efficiency. After processing, the representation capability of features is improved, especially the enhancement of detail information, so that subsequent modules can obtain richer inputs when converting features.

[0096] Segmented Convolutional Feedforward Network (SCFN) Figure 4 As shown in Figure 2, it aims to further optimize the features output from the adaptive feature fusion module. The segmented convolutional feedforward network divides the feature map into two parts, one part extracts new features through convolution, and the other part retains the original features.

[0097] The purpose of the design of the piecewise convolutional feedforward network is to retain some of the original features while learning new feature representations. The overall network process is to first perform dimensional transformation through 1x1 convolution, and then use the activation function GELU to enhance nonlinearity, expanding the input feature dimension to a larger hidden dimension (in this solution, it is twice the original dimension). Then, during the training phase, Split is used to divide the hidden layer features into two parts in the channel dimension: one part learns local features through 3x3 convolution, and uses the activation function GELU to enhance the nonlinear feature expression. The other part keeps the original features unchanged. After processing through different convolutional layers, these features are re-joined and transformed using 1x1 convolution. This design can locally enhance some features while retaining the original features. This partial convolution strategy can effectively prevent feature loss while allowing new features to be learned. In the inference phase, local features are directly convolved, and the segmentation operation is no longer used, further simplifying the inference process.

[0098] In addition, considering that the self-attention mechanism performs well in the field of image processing, it is difficult to adapt to application scenarios with limited computing resources due to its high computational complexity. To this end, the present invention avoids using computationally intensive operations such as self-attention mechanisms and large kernel convolutions, but instead introduces lightweight global channel attention to select more important feature channels. This design greatly reduces the computational overhead and ensures accurate control of model parameters, thereby enabling efficient image enhancement in resource-limited environments such as embedded platforms. Because different image scenes may have different degrees of dependence on different channels, the global channel attention module designed in the present invention can adaptively adjust the channel weights according to the input image, ensuring that the most important channel information is effectively retained and enhanced when processing complex underwater environments (such as low contrast, blur, and noise). By assigning different attention weights to each channel, the global channel attention module can enhance the response of a specific channel, thereby improving feature discrimination.

[0099] The global channel attention module performs weighted processing on the input feature map through the channel-level attention mechanism to enhance the representation ability of important features. Among them, the combination of Relu activation function and Sigmoid is not used, but the activation function Softsign is used to balance performance and computational complexity to obtain the weight coefficient of each channel, because it is simple to calculate, does not rely on complex exponential operations, and has low computational cost. The output of the ReLU function has discontinuity at the zero point, and the output of ReLU is zero when the input is negative, which may cause some neurons to be permanently suppressed during the training process and unable to participate in subsequent learning, which will affect the effect of the channel attention mechanism. Sigmoid relies on exponential operations and is relatively complex to calculate.

[0100] Specifically, the global channel attention module performs global average pooling on each channel and calculates its weight, selects important channel features, and improves the performance of the network in complex tasks. This method can highlight key feature channels, suppress redundant information, and improve the network's ability to capture important features. The following is the formula for global channel attention:

[0101]

[0102] z f =Flatten(z m ),

[0103] y1=S(W1z f ),

[0104] y2=S(W2y1),

[0105] Where X m Represents the mth feature map that has been extracted. m represents the average global pooling of the mth channel. f It means that the tensor z m The process of flattening to a one-dimensional vector, and W1 and W2 are the two weight matrices used in this operation to adjust z f Perform linear transformation to capture important features and reduce the amount of calculation. The channel attenuation used in the present invention is 4, that is, using W1 to transform the one-dimensional vector z f Channel attenuation is performed to reduce the number of channels from the original to 1 / 4, and then W2 is used to increase the dimension of the one-dimensional vector y1 to obtain the weight of each channel. The symbol S represents the Softsign activation function, whose output is between (-1,1) and is used to smooth the output. In addition, y1 is obtained by using the weight matrix z f What is obtained by dimensionality reduction is the core part of the information, ignoring the noise or unimportant parts. And y2, as the initial result of the channel attention weight after dimensionality recovery, is a one-dimensional vector. These weights reflect the importance of each channel to the final output. The present invention uses a linear layer to reduce and then expand the dimension of the channel, and can extract and process information in a lower dimensional space to reduce the computational burden and improve the flexibility of feature selection. Then, the formula for re-adjusting the size and weight of the channel of the present invention is as follows:

[0106] A e = expand(y2),

[0107] X′=X×A e ,

[0108] Among them A e Represents adjusting the dimension of y2 to match the dimension of the input feature map X. Finally, the weight A eMultiply it with the input feature map X to obtain a new feature map X'. In this way, the present invention can obtain an adaptive channel attention.

[0109] Compared with the existing solutions, the advantages of the network of the present invention are:

[0110] The network of the present invention enhances the expressiveness of features compared to existing solutions, especially in complex underwater environments. It can capture richer details and global information, ensuring that when processing blurred and low-contrast images, it can maintain global consistency and accurately enhance local details. This is mainly reflected in the following points:

[0111] Better feature expression: The global channel attention module adaptively adjusts the channel attention so that the network can focus on more representative channels, avoiding redundancy or loss of channel information, and the computational complexity is very low. The multi-scale feature aggregation module processes global and local features through a multi-path structure, making the feature expression more comprehensive and accurate, and adopts residual links to ensure that feature information is not lost.

[0112] Enhanced detail capture capability: Underwater images are often blurred and have detail loss due to poor lighting conditions, physical obstacles, and noise. The multi-scale feature aggregation module, through its unique local enhancement mechanism, can enhance the expression of local details while maintaining global information, helping to capture more effective information, and obtain global statistical information by calculating variance, thereby further maintaining the integrity of global features.

[0113] Strong adaptability: The complexity of the underwater environment requires the network to have strong adaptability. The global channel attention module and the multi-scale feature aggregation module introduce learnable weights in the training process (such as the adaptive channel attention weights in GCAM and the alpha and belt parameters in MFA), so that the network can adaptively adjust according to the input features, thereby processing different types of underwater images more flexibly.

[0114] 3. Training Goals

[0115] The proposed lightweight underwater image enhancement network based on adaptive feature fusion is an end-to-end network that can be optimized by combining loss terms. The entire training objective can be divided into perceptual loss L vgg , SSIM loss, Charbonnier Loss, combined in a weighted manner:

[0116] L total =λ1L vgg1 +λ2L ssim +λ3L c +λ4L vgg2

[0117] Among them, λ1, λ2, λ3, and λ4 are all weight parameters. In the present invention, λ1 is set to 0.2, λ2 is set to 0.5, λ3 is set to 1, and λ4 is set to 0.2.

[0118] The formula for Charbonnier Loss is as follows:

[0119]

[0120] Where X is the predicted value and Y is the target value. ∈ is a small constant used to prevent numerical instability when the sum of squares is close to zero. This loss function can be regarded as a smooth L1 loss, avoiding the limitations of L1 loss at gradient discontinuities.

[0121] The embodiment of the present invention compares the proposed LAFF-Net with several classic underwater image enhancement methods, including Ucolor [6] 、Water-Net

[13] 、Shallow-UWnet [3] PUIE-Net

[14] NU2Net [5] and FiveA+ [1] Table 2 gives a comparison of these methods based on objective indicators. The role of each module is explored on the U45 and T90 datasets. The algorithm of the present invention adopts peak signal-to-noise ratio (PSNR) [9] , Structural Similarity Index (SSIM)

[10] , Underwater Color Image Quality Evaluation (UCIQE)

[11] and Underwater Image Quality Metric (UIQM)

[12] As a performance indicator of image quality.

[0122] 1 Parameter selection

[0123] During the training process, the number of epochs was set to 500, and the batch size for training and validation was 1. The optimizer used was Adam, and the initial learning rate was set to 4×10 -4 , the default values ​​of β1 and β2 are 0.5 and 0.999 respectively. To dynamically adjust the learning rate, the present invention uses the CyclicLR scheduler, and the initial momentum is set to 0.9 and 0.999. The data augmentation strategy includes horizontal flipping and random rotation of 90 degrees, 180 degrees and 270 degrees.

[0124] Table 1 Ablation experiment

[0125]

[0126]

[0127] In order to deeply understand the contribution of each module and its limitations, the present invention designed an ablation experiment, as shown in Table 1, to evaluate the impact of the multi-scale feature aggregation module and the global channel attention module on the model performance. On the U45 and T90 test sets, after adding the MFA and GCAM modules, the model has significantly improved the PSNR and UCIQE indicators, respectively, and the increase in calculation time is within a reasonable range. This shows that these modules have little impact on computational efficiency while enhancing image quality.

[0128] After adding the multi-scale feature aggregation module alone, the PSNR is significantly improved despite the increase in computation time, showing the key role of MFA in improving image quality and processing complex images. However, UIQM and UCIQE have not changed much, probably because MFA focuses more on enhancing local details and specific frequency information. In addition, after adding the global channel attention module alone, the PSNR, UIQM and UCIQE indicators are significantly improved, indicating that the global and local features are effectively fused, especially in the fusion of high-frequency and low-frequency information, and more important channel information is selected, which significantly helps the overall image quality.

[0129] 2 Qualitative comparison

[0130] This paper gives an intuitive comparison of the visual effects with previous SOTA methods. Figure 5 As shown in Figure 2, PUIE-Net exhibits poor perception of details, resulting in significant erosion of texture details in the enhanced photo. On the other hand, NU2Net lacks precise color control, thus resulting in visible color differences that can be noticed by the human eye. The method of the present invention exhibits fairly compatible color and detail restoration, enhancing the entire degraded image and making its contrast and texture details meet the sensory requirements of the human eye.

[0131] 3 Quantitative comparison

[0132] Table 2 Quantitative comparison

[0133]

[0134] LAFF-Net shows significant advantages in PSNR, achieving a 0.62dB improvement on the T90 dataset compared to NU2Net, which shows that the architecture performs well in detail texture, contrast, and color restoration. In addition, the data in Table 2 show that LAFF-Net achieves good results on the C60 and U45 datasets, highlighting its strong generalization ability. The number of parameters of the LAFF-Net algorithm is only 1 / 16 of that of the Shallow-UWnet algorithm, which is about 14k. However, LAFF-Net is significantly better than the Shallow-UWnet algorithm in various performance indicators. This shows that LAFF-Net can still achieve good performance with fewer parameters, showing its high parameter efficiency. Such a low number of parameters also improves the applicability of the algorithm in resource-constrained environments.

[0135] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0136] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0137] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0139] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive various other forms of lightweight underwater image enhancement methods based on adaptive feature fusion under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention should be covered by this patent.

Claims

1. A lightweight underwater image enhancement method based on adaptive feature fusion, characterized by: A lightweight underwater image enhancement network based on adaptive feature fusion is adopted to decompose the underwater image enhancement task into a strong prior enhancement task and a fine-grained enhancement task, so as to make full use of the multi-scale information in the image; the strong prior enhancement task is used to extract the global and low-level features of the image; The fine-grained enhancement task is used to enhance the local details and color restoration of the image, thereby improving the visual quality; the lightweight underwater image enhancement network includes a multi-scale feature aggregation module and a global channel attention module composed of an adaptive feature fusion module and a segmented convolutional feedforward network, which are used to combine the global and local information of the features and adaptively adjust the importance of each channel through information association between channels.

2. The lightweight underwater image enhancement method based on adaptive feature fusion according to claim 1, characterized in that: In the lightweight underwater image enhancement network, the input image is first subjected to the strong prior enhancement task after point-by-point convolution. In the strong prior enhancement task, one branch is subjected to a multi-branch color enhancement module, a multi-scale feature aggregation module and a global channel attention module, and the other branch is subjected to a multi-scale enhancement network and then jointly input into a mixed domain feature interaction module for fusion and then transferred to the fine-grained enhancement task; in the fine-grained enhancement task, it is first subjected to point-by-point convolution, and then sequentially input into a multi-scale feature aggregation module, a global channel attention module, a multi-branch color enhancement module and a channel attention layer.

3. The lightweight underwater image enhancement method based on adaptive feature fusion according to claim 1, characterized in that: The multi-scale feature aggregation module combines an adaptive feature fusion module and a piecewise convolutional feedforward network; firstly, the adaptive feature fusion module is used to perform spatial enhancement of features, extract global features and enhance local details; then, the piecewise convolutional feedforward network is used to retain some original features while enhancing some features; and feature normalization is applied to the feature input of each step to ensure that the input features remain numerically stable.

4. The lightweight underwater image enhancement method based on adaptive feature fusion according to claim 3, characterized in that: The adaptive feature fusion module combines two branches: self-attention approximation and local detail estimation. The self-attention approximation is used to capture global features through an adaptive pooling method to reduce computational complexity. The local detail estimation is used to enhance local detail processing capabilities and achieve detail recovery through 1×1 convolution: The input features are channel-added through the convolution layer, and then split into two parts in the channel dimension to obtain feature maps X1 and X2 respectively; For feature map X1, adaptive maximum pooling is first performed to generate a downsampled feature map to capture spatial context information; then, the pooling result is further processed by deep convolution to enhance local detail information; and the variance of feature map X1 in the spatial dimension is calculated to extract global statistical information and obtain the global context of the feature map; the local pooled features are multiplied by a learnable weight parameter alpha to amplify local feature information; the global variance features are multiplied by another learnable parameter belt to enhance global feature information; the local and global features are weighted summed to fuse information of different scales; the fused features are nonlinearly processed by the GELU activation function, and the fused features are further compressed and adjusted by 1x1 convolution; finally, the nearest neighbor interpolation is used to restore the fused features from the downsampled size to the original size; The feature map X2 is locally enhanced through a depthwise separable convolution layer, a pointwise convolution layer, and a GELU activation function; the depthwise separable convolution layer is composed of a combination of depthwise convolution and pointwise convolution.

5. The lightweight underwater image enhancement method based on adaptive feature fusion according to claim 3, characterized in that: The segmented convolutional feedforward network first performs dimension transformation through 1x1 convolution, and then uses the activation function GELU to enhance nonlinearity and expand the feature dimension of the input; Then, in the training phase, a split operation is used to divide the hidden features into two parts in the channel dimension: one part learns local features through 3x3 convolution, and uses the activation function GELU to enhance the nonlinear feature expression; the other part keeps the original features unchanged; after processing through different convolutional layers, these features are reconnected and dimensionality is transformed using 1x1 convolution; in the inference phase, the local features are directly convolved without using a split operation to simplify the inference process.

6. The lightweight underwater image enhancement method based on adaptive feature fusion according to claim 3, characterized in that: The global channel attention module performs weighted processing on the input feature map through a channel-level attention mechanism to enhance the representation capability of important features; The activation function Softsign is used to balance performance and computational complexity to obtain the weight coefficient of each channel; The global channel attention module first performs global average pooling on each channel and calculates the weight, expressed as: With f =Flatten(from m ), y1=S(W1z f ), y2=S(W2y1), Where X m represents the mth feature map that has been extracted; z m represents the average global pooling of the mth channel; z f It means that the tensor z m The process of flattening to a one-dimensional vector, W1 and W2 are weight matrices used to calculate z f Perform a linear transformation and use W1 to transform the one-dimensional vector z f Channel attenuation is performed to reduce the number of channels from the original to 1 / 4, and then W2 is used to increase the dimension of the one-dimensional vector y1 to obtain the weight of each channel; the symbol S represents the Softsign activation function, and y1 is activated by using the weight matrix z f After dimensionality reduction, y2 is obtained as the initial result of the channel attention weight after dimension recovery, which is a one-dimensional vector; Then resize and reweight the channels: A e =expand(y2), X′=X×A e , Among them A e Indicates adjusting the dimension of y2 to match the dimension of the input feature map X; Finally, the weight A e Multiply it with the input feature map X to get a new feature map X'.

7. The lightweight underwater image enhancement method based on adaptive feature fusion according to claim 1, characterized in that: The lightweight underwater image enhancement network is optimized by combining loss terms; The training objective is divided into perceptual loss L vgg , SSIM loss, Charbonnier Loss, combined in a weighted manner: L total =λ1L vgg1 +λ2L ssim +λ3L c +λ4L vgg2 Among them, λ1,λ2,λ3,λ4 are all weight parameters; The formula for Charbonnier Loss is as follows: Where X is the predicted value and Y is the target value; ∈ is used to prevent numerical instability when the sum of squares approaches zero.

8. A lightweight underwater image enhancement system based on adaptive feature fusion, characterized in that: include: A lightweight underwater image enhancement network based on adaptive feature fusion, including a strong prior enhancement task module and a fine-grained enhancement task module, to fully utilize the multi-scale information in the image; The strong prior enhancement task module is used to extract global and low-level features of the image; The fine-grained enhancement task module is used to enhance the local details and color restoration of the image, thereby improving the visual quality; the lightweight underwater image enhancement network includes a multi-scale feature aggregation module composed of an adaptive feature fusion module and a segmented convolutional feedforward network, and a global channel attention module, which is used to combine the global and local information of the features and adaptively adjust the importance of each channel through information association between channels.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the lightweight underwater image enhancement method based on adaptive feature fusion as described in any one of claims 1 to 7 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the lightweight underwater image enhancement method based on adaptive feature fusion as described in any one of claims 1 to 7 are implemented.

Images