Underwater image enhancement method and system

Abstract

The invention belongs to the technical field of image processing, and discloses an underwater image enhancement method and system, and the method comprises the following steps: S1, enabling an inputted underwater image to pass through a wide area sensing module, and obtaining a feature map; s2, inputting the feature map obtained in the step S1 into an attention down-sampling module to obtain a down-sampled feature map; s3, inputting the feature map obtained in the S2 into a feature purification module, and purifying the feature map to obtain an optimized feature map; s4, inputting the feature map obtained in the S3 into an attention up-sampling module to obtain a feature map after up-sampling; s5, inputting the feature map obtained in the step S4 into a feature map smoothing module, and outputting an enhanced underwater image; and S6, training the underwater image enhancement model constructed in S1 to S5, and enhancing the input underwater image by using the underwater image enhancement model after the training is completed. The subjective quality of underwater images can be remarkably improved, the operation speed is high, and the requirements of underwater operation can be met.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular to an underwater image enhancement method and system. This patent is supported by the National Natural Science Foundation of China (62201538), the Natural Science Foundation of Shandong Province (ZR2022QF006), and the Natural Science Foundation of Shandong Province (ZR2024MF116). Background Art

[0002] Underwater imaging is widely used in underwater operations and is an important part of underwater operations. Due to the refraction and scattering of light by water, the quality of underwater imaging is low, often showing problems such as blue-green bias, blur, and low contrast, which cannot meet the needs of underwater operations. Therefore, the collected underwater images often need to be appropriately enhanced to improve their quality, so as to provide accurate information for underwater operations.

[0003] Existing underwater image enhancement methods can be divided into spatial domain methods, transform domain methods and deep learning methods according to the processing method. The spatial domain method performs enhancement processing on the spatial domain of the image, mainly including grayscale transformation and spatial domain filtering. Grayscale transformation (or contrast stretching) changes the grayscale value of the image pixel by pixel according to the correspondence between a certain pixel point and the target task (grayscale transformation function). Histogram equalization (HE) is often used for contrast enhancement of underwater images. It enhances the contrast of the image by globally uniformly stretching the histogram of the image. However, the HE method also increases the sparsity of the histogram, resulting in the loss of some details of the underwater image. The transform domain method transforms the underwater image into other domains for enhancement processing through image transformation.

[0004] In recent years, deep learning technology has achieved great success in the field of computer vision. With the development of CNN methods in the field of underwater image enhancement, the problem of data scarcity has received more and more attention. The research on Generative Adversarial Networks (GANs) in the field of image domain transformation has inspired scholars to transform the underwater image enhancement task into the transformation of the image underwater domain to the enhancement domain.

[0005] However, the performance of current underwater image enhancement methods is still unsatisfactory, the scope of application is limited, and the robustness of the method is not high. Therefore, the present invention proposes an underwater image enhancement method to effectively improve the subjective quality of underwater images, thereby improving the efficiency of underwater operations. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes an underwater image enhancement method and system, which constructs an enhancement model through wide-area perception of underwater images, feature map downsampling, feature map purification, feature map upsampling and feature map smoothing, and trains the image enhancement model in an end-to-end manner to achieve high-quality enhancement of underwater images.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0008] An underwater image enhancement method, firstly constructing an underwater image enhancement model architecture, including a wide-area perception module, an attention downsampling module, a feature purification module, an attention upsampling module, and a feature smoothing module, and then training the model; finally, using the trained model to enhance a new underwater image; specifically comprising the following steps:

[0009] S1, the input underwater image passes through the wide area perception module, undergoes convolution operation and activation function, and obtains the feature map F1;

[0010] S2, input the feature map F1 obtained in S1 into the attention downsampling module to obtain the downsampled feature map F2;

[0011] S3, inputting the feature map F2 obtained in S2 into a feature purification module, purifying the feature map to obtain a purified feature map F3;

[0012] S4, input the feature map F3 obtained in S3 into the attention upsampling module to obtain the upsampled feature map F4; S5, input the feature map F4 obtained in S4 into the feature smoothing module to output the enhanced underwater image;

[0013] S6. Train the underwater image enhancement model constructed by S1-S5, and use it to enhance the input underwater image after the training is completed.

[0014] Furthermore, the step S1 specifically includes: performing a convolution operation on the input underwater image using a 7×7 convolution kernel, and obtaining a feature map after a ReLU activation function, which is expressed as:

[0015] F1 = ReLU (Conv 7×7 (I))

[0016] Where I represents the input underwater image, Conv 7×7 (·) indicates the convolution operation using a 7×7 convolution kernel, and ReLU(·) indicates the rectified linear unit activation function.

[0017] Furthermore, step S2 is to obtain a feature map by subjecting the feature map obtained in S1 to convolution, normalization, ReLU, and channel space attention modules, and repeat this process 5 times to output the feature map:

[0018] The step S2 specifically includes: defining an attention downsampling module including convolution, normalization, ReLU, and channel space attention modules, which is expressed as:

[0019] AD(·)=CBAM(ReLU(Norm(Conv 4×4 (·))))

[0020] Among them, AD(·) represents the attention downsampling operation, CBAM(·) represents the channel space attention module, and the calculation method is:

[0021]

[0022] Among them, F represents the input feature map, F' represents the feature map adjusted by the channel attention module, and F" represents the feature map adjusted by the spatial attention module. Represents element-wise multiplication operation, M c represents the channel attention part of the channel spatial attention module, M S represents the spatial attention part of the channel spatial attention module, M c (F) and M s (F) is calculated as follows:

[0023] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0024] M s (F) = σ(conv 7×7 [AvgPool(·); MaxPool(·)])

[0025] Among them, MLP represents multi-layer perceptron, AvgPool(·) and MaxPool(·) represent average pooling and maximum pooling, conv 7×7 (·) represents the convolution operation, the convolution kernel size is 7×7, σ(·) represents the sigmoid function, and Norm(·) represents normalization;

[0026] After F1 is subjected to 5 attention downsampling operations, the feature map F2 is obtained, which is expressed as:

[0027] F2=AD(AD(AD(AD(AD(F1)))))).

[0028] Furthermore, in step S3, the feature map obtained in S2 is input into the feature purification module to purify the feature map to obtain an optimized feature map. The specific method is as follows: the feature map obtained in S2 is input into the residual block for feature purification. The residual block includes convolution, normalization, ReLU, convolution and normalization operations. This process is repeated three times to obtain a purified feature map.

[0029] Furthermore, the step S3 specifically includes: defining the residual block operation as follows:

[0030]

[0031] Among them, F r represents the feature map obtained by the residual block, F i represents the input feature map, represents the element-wise addition operation, Conv3×3 represents the convolution operation, the convolution kernel size is 3×3, and Norm(·) represents normalization. The feature map obtained by S2 is subjected to three residual block operations to obtain the purified feature map F3, which is expressed as:

[0032] F3=ResBlock(ResBlock(ResBlock(F2)))

[0033] Among them, ResBlock(·) represents the residual operation.

[0034] Furthermore, in step S4, the feature map obtained in S3 is input into the attention upsampling module to obtain the upsampled feature map. The specific method is as follows: the feature map obtained in S3 is input into the attention upsampling module, the attention upsampling module includes deconvolution, normalization, ReLU, and channel space attention module, and this process is repeated 5 times to obtain the upsampled feature map.

[0035] Furthermore, step S4 specifically includes: defining an attention upsampling module, including deconvolution, normalization, ReLU, and channel space attention modules, expressed as:

[0036]

[0037] Among them, AU(·) represents the attention upsampling operation, CBAM(·) represents the channel space attention module, Represents the deconvolution operation; the feature map obtained by S3 is subjected to 5 attention upsampling operations to obtain the upsampled feature map F4, which is expressed as:

[0038] F4=AU(AU(AU(AU(AU(F3))))))).

[0039] Further, in step S5, the feature map obtained in S4 is input into the feature map smoothing module to output the enhanced underwater image. The specific method is as follows: the feature map obtained in S4 is input into the feature map smoothing module, including convolution and sigmoid activation operations, to obtain the enhanced underwater image, which is expressed as:

[0040]

[0041] in, Represents the enhanced underwater image.

[0042] Further, the underwater image enhancement model constructed by S1-S5 described in step S6 is trained, and after the training is completed, the input underwater image is enhanced by using the model, and the specific method is as follows: N underwater images and their enhanced reference enhanced images are selected, and the N original underwater images are enhanced by using the proposed model to obtain enhanced underwater images, and the mean square error between the enhanced underwater image and the reference enhanced image is calculated, and the model is end-to-end trained using the mean square error as a loss function, and after the training is completed, it is used for underwater image enhancement;

[0043] The mean square error is defined as:

[0044]

[0045] Among them, I i represents the reference enhanced image corresponding to the i-th underwater image, represents the i-th underwater image enhanced by the proposed model;

[0046] After training, the model is used to enhance the input underwater image:

[0047]

[0048] in, represents the enhanced underwater image, I is the input underwater image, and M(·) is the model after training.

[0049] The present invention also provides an underwater image enhancement system for implementing an underwater image enhancement method as described above, the system comprising an underwater image enhancement model and a training module, the training module being used to train the underwater image enhancement model, the underwater image enhancement model comprising a wide area perception module, an attention downsampling module, a feature purification module, an attention upsampling module, and a feature smoothing module; the input underwater image passes through the wide area perception module, the attention downsampling module, the feature purification module, the attention upsampling module, and the feature smoothing module in sequence, and outputs an enhanced underwater image.

[0050] Compared with the prior art, the present invention has the advantages of:

[0051] The present invention realizes high-quality enhancement of underwater images and achieves performance superior to that of the prior art; the present invention introduces a wide-area perception module to expand the receptive field, capture the global features of the image, and facilitate overall image enhancement; an attention mechanism is introduced in the upsampling and downsampling processes to better enhance important information such as edge texture in the image; the degradation of underwater images is relatively complex, therefore, a feature purification module is introduced between the upsampling and downsampling modules to more fully learn the deep-level features of the image and effectively improve the image enhancement performance; finally, a feature smoothing module is introduced to effectively avoid the checkerboard effect in the enhanced image and improve the image enhancement performance.

[0052] In addition, the method proposed by the present invention runs fast and meets the application requirements of underwater operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0054] Figure 1 is a flow chart of the method of the present invention;

[0055] Figure 2 is a system structure diagram of the present invention;

[0056] Figure 3 It is a structural diagram of each module of the underwater image enhancement model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0058] The embodiment of the present invention inputs the underwater image into a wide-area perception module, an attention downsampling module, a feature purification module, an attention upsampling module and a feature map smoothing module to achieve underwater image enhancement, and trains the underwater image enhancement model using an end-to-end training method. After the training is completed, the underwater image is enhanced.

[0059] like Figure 1 , Figure 2 , Figure 3As shown, an embodiment of the present invention provides an underwater image enhancement method, which first constructs an underwater image enhancement model architecture, including a wide-area perception module, an attention downsampling module, a feature purification module, an attention upsampling module, and a feature smoothing module, and then trains the model; finally, the trained model is used to enhance the new underwater image. Specifically, the following steps are included:

[0060] S1, the input underwater image passes through the wide-area perception module, undergoes convolution operation and activation function, and obtains the global feature map F1;

[0061] S2, input the global feature map F1 obtained in S1 into the attention downsampling module to obtain the downsampled feature map F2;

[0062] S3, inputting the feature map F2 obtained in S2 into a feature purification module, purifying the feature map to obtain a purified feature map F3;

[0063] S4, input the feature map F3 obtained in S3 into the attention upsampling module to obtain the upsampled feature map F4; S5, input the feature map F4 obtained in S4 into the feature smoothing module to output the enhanced underwater image;

[0064] S6. Train the underwater image enhancement model constructed by S1-S5, and use it to enhance the input underwater image after the training is completed.

[0065] This invention can significantly improve the quality of underwater images and at the same time has a fast running speed, meeting the application requirements of underwater operations.

[0066] In one embodiment, the specific implementation process and details of the above-mentioned underwater enhanced image quality assessment method are as follows: a 7×7 convolution kernel is used to perform a convolution operation on the input underwater image, and a ReLU activation function is performed to obtain a feature map, which is expressed as:

[0067] F1 = ReLU (Conv 7×7 (I))

[0068] Among them, I represents the input underwater image, F1 is the feature map obtained by S1, Conv 7×7 (·) indicates the convolution operation using a 7×7 convolution kernel, and ReLU(·) indicates the rectified linear unit activation function.

[0069] Define the attention downsampling module, including convolution, normalization, ReLU, and channel space attention modules, expressed as:

[0070] AD(·)=CBAM(ReLU(Norm(Conv 4×4 (·))))

[0071] Among them, AD(·) represents the attention downsampling operation, CBAM(·) represents the channel space attention module, and the calculation method is:

[0072]

[0073] Among them, F represents the input feature map, F′ represents the feature map adjusted by the channel attention module, and F” represents the feature map adjusted by the spatial attention module. Represents element-wise multiplication operation, M c represents the channel attention part of the channel spatial attention module, M S represents the spatial attention part of the channel spatial attention module, M c (F) and M s (F) is calculated as follows:

[0074] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0075] M s (F) = σ(conv 7×7 [AvgPool(·); MaxPool(·)])

[0076] Among them, MLP represents multi-layer perceptron, AvgPool(·) and MaxPool(·) represent average pooling and maximum pooling, conv 7×7 (·) represents the convolution operation, the convolution kernel size is 7×7, and σ(·) represents the sigmoid function, which is defined as:

[0077]

[0078] Norm(·) represents normalization and is calculated as:

[0079]

[0080] Among them, ε is 0.0001, x is the input variable, Var[x] is the variance of variable x, E[x] is the mean of variable x, and y is the normalized output value.

[0081] After F1 undergoes 5 AD operations, the feature map F2 is obtained, which is expressed as:

[0082] F2=AD(AD(AD(AD(AD(F1)))))

[0083] The residual block operation (ResBlock) is defined as follows:

[0084]

[0085] Among them, F r represents the feature map obtained by the residual block, F i represents the input feature map, represents the element-wise addition operation, Conv3×3 represents the convolution operation, the convolution kernel size is 3×3, and Norm(·) represents normalization. F2 is subjected to three residual block operations to obtain the purified feature map F3, which is expressed as:

[0086] F3=ResBlock(ResBlock(ResBlock(F2)))

[0087] Among them, ResBlock(·) represents the residual operation.

[0088] Define the attention upsampling module, including deconvolution, normalization, ReLU, and channel space attention modules, expressed as:

[0089]

[0090] Among them, AU(·) represents the attention upsampling operation, CBAM(·) represents the channel space attention module, Denotes the deconvolution operation. After F3 is subjected to 5 attention upsampling operations, the upsampled feature map F4 is obtained, which is expressed as:

[0091] F4=AU(AU(AU(AU(AU(F3))))).

[0092] F4 is input into the smoothing module, including convolution and sigmoid activation operations, to obtain the enhanced underwater image, which is expressed as:

[0093]

[0094] in, Represents the enhanced underwater image.

[0095] The underwater image enhancement model constructed by S1-S5 described in step S6 is trained, and after the training is completed, the input underwater image is enhanced by using it. The specific method is as follows: N underwater images and their enhanced reference enhanced images are selected, the N original underwater images are enhanced by using the proposed model to obtain the enhanced underwater image, the mean square error between the enhanced underwater image and the reference enhanced image is calculated, the mean square error is used as the loss function to perform end-to-end training on the model, and after the training is completed, it is used for underwater image enhancement.

[0096] In this embodiment, 1000 underwater images are selected and enhanced using the proposed model to obtain 1000 enhanced images. The model is trained end-to-end using the mean square error, and the mean square error is defined as:

[0097]

[0098] in, represents the reference enhanced image corresponding to the i-th underwater image, It represents the i-th underwater image enhanced by the proposed model. The model is optimized by Adam algorithm, the learning rate is set to 0.0001, the number of samples is set to 8, and the training iteration is 300 times.

[0099] After training, the model is used to enhance the input underwater image:

[0100]

[0101] in, represents the enhanced underwater image, I is the input underwater image, and M(·) is the model after training.

[0102] The present invention also provides an underwater image enhancement system for implementing the underwater image enhancement method as described above. Figure 2 As shown, the system includes an underwater image enhancement model and a training module (not shown in the figure). The training module is used to train the underwater image enhancement model. The underwater image enhancement model includes a wide area perception module, an attention downsampling module, a feature purification module, an attention upsampling module, and a feature smoothing module. The input underwater image passes through the wide area perception module, the attention downsampling module, the feature purification module, the attention upsampling module, and the feature smoothing module in sequence to output an enhanced underwater image.

[0103] In a specific embodiment of the present invention, the performance of the method of the present invention is tested on the LSUI library (the paper "U-Shape Transformer for Underwater Image Enhancement" published by Peng et al. in "IEEE Transactions on Image Processing", Vol. 32, pp. 3066 to 3079), and the performance of the algorithm is measured using two indicators, PSNR and SSIM. The experimental results are shown in Table 1.

[0104] Table 1 Performance comparison on LSUI database

[0105] Performance Indicators UT-UIE SyreaNet WaterNet UGAN UWCNN The present invention PSNR 24.331 17.799 21.962 21.643 17.080 27.177 SSIM 0.817 0.755 0.852 0.867 0.731 0.873

[0106] Table 1 compares the latest underwater image quality evaluation methods, including UT-UIE, SyreaNet, WaterNet, UGAN, and UWCNN. It can be seen that the present invention achieves higher prediction performance than similar methods; the present invention achieves better subjective enhancement effect than similar methods, which fully proves the effectiveness and superiority of the present invention.

[0107] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. An underwater image enhancement method, characterized in that, firstly, an underwater image enhancement model architecture is constructed, including a wide-area perception module, an attention downsampling module, a feature purification module, an attention upsampling module, and a feature smoothing module, and then the model is trained; finally, the trained model is used to enhance a new underwater image; specifically, the following steps are included: S1, the input underwater image passes through the wide area perception module, undergoes convolution operation and activation function, and obtains the feature map F1; S2, input the feature map F1 obtained in S1 into the attention downsampling module to obtain the downsampled feature map F2; S3, inputting the feature map F2 obtained in S2 into a feature purification module, purifying the feature map to obtain a purified feature map F3; S4, input the feature map F3 obtained in S3 into the attention upsampling module to obtain the upsampled feature map F4; S5, inputting the feature map F4 obtained in S4 into a feature smoothing module, and outputting an enhanced underwater image; S6. Train the underwater image enhancement model constructed by S1-S5, and use it to enhance the input underwater image after the training is completed.

2. The underwater image enhancement method according to claim 1, characterized in that: The step S1 specifically includes: performing a convolution operation on the input underwater image using a 7×7 convolution kernel, and obtaining a feature map after a ReLU activation function, which is expressed as: F1=ReLU(Conv 7×7 (I)) Where I represents the input underwater image, Conv 7×7 (·) indicates the convolution operation using a 7×7 convolution kernel, and ReLU(·) indicates the rectified linear unit activation function.

3. The underwater image method according to claim 1, characterized in that: Step S2 is to obtain the feature map obtained by S1 through convolution, normalization, ReLU, and channel space attention modules. This process is repeated 5 times to output the feature map: The step S2 specifically includes: defining an attention downsampling module including convolution, normalization, ReLU, and channel space attention modules, which is expressed as: AD(·)=CBAM(ReLU(Norm(Conv 4×4 (·)))) Among them, AD(·) represents the attention downsampling operation, CBAM(·) represents the channel space attention module, and the calculation method is: Among them, F represents the input feature map, F' represents the feature map adjusted by the channel attention module, and F" represents the feature map adjusted by the spatial attention module. Represents element-wise multiplication operation, M c represents the channel attention part of the channel spatial attention module, M S represents the spatial attention part of the channel spatial attention module, M c (F) and M s (F) is calculated as follows: M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) M s (F)=σ(conv 7×7 [AvgPool(·);MaxPool(·)]) Among them, MLP represents multi-layer perceptron, AvgPool(·) and MaxPool(·) represent average pooling and maximum pooling, conv 7×7 (·) represents the convolution operation, the convolution kernel size is 7×7, σ(·) represents the sigmoid function, and Norm(·) represents normalization; After F1 is subjected to 5 attention downsampling operations, the feature map F2 is obtained, which is expressed as: F2=AD(AD(AD(AD(AD(F1)))))).

4. The underwater image enhancement method according to claim 1, characterized in that: Step S3 inputs the feature map obtained in S2 into the feature purification module, purifies the feature map, and obtains the optimized feature map. The specific method is as follows: input the feature map obtained in S2 into the residual block for feature purification. The residual block includes convolution, normalization, ReLU, convolution and normalization operations. Repeat this process three times to obtain the purified feature map.

5. The underwater image enhancement method according to claim 4, characterized in that: The step S3 specifically includes: defining the residual block operation as follows: Among them, F r represents the feature map obtained by the residual block, F i represents the input feature map, Represents element-wise addition operation, Conv 3×3 represents the convolution operation, the convolution kernel size is 3×3, and Norm(·) represents normalization. The feature map obtained by S2 is subjected to three residual block operations to obtain the purified feature map F3, which is expressed as: F3=ResBlock(ResBlock(ResBlock(F2))) Among them, ResBlock(·) represents the residual operation.

6. The underwater image enhancement method according to claim 1, characterized in that: The feature map obtained in S3 as described in step S4 is input into the attention upsampling module to obtain the upsampled feature map. The specific method is as follows: the feature map obtained in S3 is input into the attention upsampling module. The attention upsampling module includes deconvolution, normalization, ReLU, and channel space attention module. This process is repeated 5 times to obtain the upsampled feature map.

7. The underwater image enhancement method according to claim 6, characterized in that: Step S4 specifically includes: defining an attention upsampling module, including deconvolution, normalization, ReLU, and channel space attention modules, expressed as: Among them, AU(·) represents the attention upsampling operation, CBAM(·) represents the channel space attention module, Represents the deconvolution operation; the feature map obtained by S3 is subjected to 5 attention upsampling operations to obtain the upsampled feature map F4, which is expressed as: F4=AU(AU(AU(AU(AU(F3))))).

8. The underwater image enhancement method according to claim 7, characterized in that: The feature map obtained in step S4 is input into the feature map smoothing module in step S5 to output the enhanced underwater image. The specific method is as follows: the feature map obtained in step S4 is input into the feature map smoothing module, including convolution and sigmoid activation operations, to obtain the enhanced underwater image, which is expressed as: in, Represents the enhanced underwater image.

9. The underwater image enhancement method according to claim 1, characterized in that: The underwater image enhancement model constructed by S1-S5 described in step S6 is trained, and after the training is completed, the input underwater image is enhanced by using the model, and the specific method is as follows: N underwater images and reference enhanced images after enhancement are selected, and the N original underwater images are enhanced by using the proposed model to obtain enhanced underwater images, and the mean square error between the enhanced underwater images and the reference enhanced images is calculated, and the mean square error is used as the loss function to perform end-to-end training on the model, and after the training is completed, it is used for underwater image enhancement; The mean square error is defined as: Among them, I i represents the reference enhanced image corresponding to the i-th underwater image, represents the i-th underwater image enhanced by the proposed model; After training, the model is used to enhance the input underwater image: in, represents the enhanced underwater image, I is the input underwater image, and M(·) is the model after training.

10. An underwater image enhancement system, characterized in that: Used to implement an underwater image enhancement method as described in any one of claims 1 to 9, the system includes an underwater image enhancement model and a training module, the training module is used to train the underwater image enhancement model, the underwater image enhancement model includes a wide area perception module, an attention downsampling module, a feature purification module, an attention upsampling module, and a feature smoothing module; the input underwater image passes through the wide area perception module, the attention downsampling module, the feature purification module, the attention upsampling module, and the feature smoothing module in sequence, and the enhanced underwater image is output.

Images