A dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba

Abstract

The application discloses a kind of dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba, belong to underwater image processing technical field.Method includes: obtaining the underwater image to be enhanced, and input to frequency domain-space domain dual branch enhancement network;The network includes color cast preprocessing module, dual-domain encoder and dual-domain decoder structure;Wherein, color cast preprocessing module is used to carry out adaptive white balance and color cast correction to input image;Encoder-decoder structure carries out multi-scale feature extraction and reconstruction by discrete wavelet transform and inverse transform, and utilizes high-frequency module, low-frequency module and spatial domain module to carry out feature interaction fusion and refinement, finally generates the underwater image after sharpening.The present application effectively solves the problems of color cast, low contrast, fuzzy details and other problems existing in underwater images by jointly using the frequency domain information and spatial domain information of the image, and has the advantages of significant enhancement effect and strong robustness.

Description

Technical Field

[0001] This invention relates to the fields of marine environmental perception, digital image processing and image enhancement technology, specifically to a dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba. Background Technology

[0002] With approximately 70% of the Earth's surface covered by oceans, underwater vision technology plays a crucial role in various fields such as ocean exploration, resource discovery, and underwater robot navigation. However, due to the complexity of the underwater environment, underwater imaging faces numerous challenges, primarily including light absorption and scattering, color distortion, low contrast, blurring, and noise. First, light absorption and scattering are the main causes of degraded underwater image quality. Water absorbs different wavelengths of light to varying degrees. Longer wavelengths such as red, orange, and yellow light attenuate the fastest in water, while blue and green light attenuate relatively slowly. Therefore, underwater images often exhibit a blue-green hue and lack red information, leading to color distortion. Furthermore, suspended particles in the water cause forward and backscattering, resulting in image blurring, reduced resolution, and a haze effect.

[0003] Currently, underwater image enhancement methods are mainly divided into three categories: physical model-based methods, traditional image processing methods, and deep learning methods. Physical model-based methods rely on idealized assumptions and struggle to accurately model complex underwater environments. Traditional image processing methods, while computationally simple, are difficult to adapt to different water conditions. Deep learning methods, although excelling in visual quality improvement, mostly only enhance images in the spatial or frequency domains, failing to fully integrate information from both domains. This results in limitations in processing texture details and structural information, insufficient color cast correction capabilities, and difficulty in achieving high-quality, comprehensive enhancement. Summary of the Invention

[0004] Due to the selective absorption of light by water and scattering caused by suspended particles, underwater imaging faces degradation problems such as color distortion, low contrast, blurring, and noise in existing technologies. Existing enhancement methods still have limitations in dual-domain information fusion and color cast correction. This invention discloses a dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba. By jointly utilizing the frequency and spatial domain information of the image and integrating a color cast preprocessing module and a multi-scale feature fusion unit, it effectively solves the problems of color cast, low contrast, and blurred details in underwater images. Specific steps include:

[0005] S1. Construct an underwater image enhancement network, the network including: a color shift preprocessing module, a dual-domain encoder and a dual-domain decoder; S2. Input the original underwater image into the color cast preprocessing module, perform color adaptive balancing, and obtain the color cast adjusted image. S3. The color-shifted image and the original image are stitched together and input into a dual-domain encoder for multi-scale feature extraction to obtain feature maps at different scales. The encoder includes discrete wavelet transform, a low-frequency enhancement module, a high-frequency enhancement module, and a spatial domain enhancement module. Wavelet decomposition is used to divide the frequency domain features into low-frequency and high-frequency components, which are then enhanced separately. S4. The feature map is input into the dual-domain decoder, which gradually restores image details and size, ultimately yielding the enhanced image. The decoder includes an inverse discrete wavelet transform module, a low-frequency enhancement module, a high-frequency enhancement module, and a spatial domain enhancement module, which gradually restore the features output by the encoder into the enhanced image.

[0006] S5. Calculate the difference between the enhanced image and the clear image using pixel-level loss, perceptual loss, SSIM loss, color loss, and edge loss. Perform error backpropagation based on the loss values ​​and update the model parameters.

[0007] Furthermore, the underwater images to be enhanced are first... The input image is fed into the color cast preprocessing module. The processed image and the input image are then added together using learnable weights to obtain the final image. In the color cast preprocessing module, the determination of blue or green cast is based on whether the ratio of the channel means exceeds a learnable threshold, and the gain of each channel is adjusted through learnable parameters. Specifically, the module first calculates the mean of each channel of the input image, and determines whether the image has a blue or green cast based on the relationship between the means of the red, green, and blue channels. For example, if the ratio of the blue channel mean to the green channel mean exceeds the threshold, it is determined to be a blue cast; if the ratio of the green channel mean to the blue channel mean exceeds the threshold, it is determined to be a green cast. Different correction strategies are adopted for different types of color cast: (1) For blue biased images, adjust the intensity of the red and green channels by calculating the gain coefficients of red and green.

[0008] (2) For green-biased images, adjust the red and green channels similarly.

[0009] (3) Perform white balance for images with no color cast or slight color cast.

[0010] By introducing a learnable gain parameter, this module can adaptively adjust the correction intensity and ensure that the pixel values ​​of the output image are within a reasonable range.

[0011] Furthermore, the output of the color shift preprocessing module... High-dimensional feature maps are obtained by using convolution to increase the dimensionality. Where C is the number of feature channels, which is set to 64 in this method. Subsequently, F is divided into two parts along the channel dimension to obtain the frequency domain branch input. and spatial domain branch input The data is then input into a dual-domain encoder. Each stage of the encoder includes: The spatial domain enhancement module performs multi-scale convolution on the spatial domain branch input and concatenates them, then combines channel attention and spatial attention and downsamples them before passing them to the next level.

[0012]

[0013] in, Represents the spatial domain branch input, Represents convolution operations with different receptive fields, i=1,3,5,7; This represents the output of the spatial domain enhancement module.

[0014] The spatial attention calculation process is as follows:

[0015] in, 7 Convolution operation of 7, and These represent max pooling and average pooling operations, respectively. For the Sigmoid function; The channel attention calculation process is as follows:

[0016] MLP includes: 1 1 convolution, ReLU activation function and 1 1. Convolution.

[0017] The discrete wavelet transform unit decomposes the frequency domain branch input into one low-frequency subband and three high-frequency subbands;

[0018] in, Represents the low-frequency subband. Represents high-frequency sub-bands. This represents the frequency domain branch input.

[0019] The low-frequency enhancement module fuses the spatial features obtained from the low-frequency subband with those from the spatial domain module, and then enhances them through multi-scale processing and Mamba long sequence modeling. It includes multiple parallel branches, each performing downsampling at different scales using average pooling. After convolutional enhancement and restoration of the original dimensions, the samples are summed, further refined by spatial attention and the Mamba module, and then connected to the residuals of the original low-frequency features.

[0020]

[0021] in, It is the output from the encoder spatial domain enhancement module. This represents downsampling to 1 / i of the original size using pooling; Indicates 3 3 convolutions; This represents an upsampling operation that restores the original size; SA stands for Spatial Attention; and SS2D represents Mamba's scanning module, used for long sequence feature modeling. This represents the output of the low-frequency enhancement module.

[0022] The high-frequency enhancement module first connects the three high-frequency sub-bands. Merged into a single high-frequency feature The output of the low-frequency module Gradient information and high-frequency features obtained through the Sobel operator After fusion, the output is obtained through normalization, convolution, and activation functions. This process can be described as follows:

[0023] in Representative instance normalization, Represents the output of the low-frequency module Gradient information obtained through the Sobel operator Including 3 3 convolution, ReLU activation function and 1 1 convolution, This represents the output of the high-frequency enhancement module.

[0024] Furthermore, each stage of the decoder includes: The high-frequency enhancement module enhances the high-frequency features from the encoder, and the output of the high-frequency enhancement is transmitted to the inverse wavelet transform unit and the low-frequency enhancement module respectively. The low-frequency enhancement module utilizes the enhanced high-frequency features for guided enhancement and performs the same multi-scale processing and Mamba long sequence modeling as the encoder's low-frequency enhancement module. The output of the low-frequency enhancement is then fed to the inverse wavelet transform unit and the spatial domain enhancement module; this process can be described as follows:

[0025] in, Represents the output of the high-frequency enhancement module. This represents low-frequency characteristics from the lower layers.

[0026] The inverse wavelet transform unit merges and reconstructs the enhanced low-frequency and high-frequency features;

[0027] in, This represents the output of the low-frequency enhancement module. Represents the output of the high-frequency enhancement module. Represented by 3 3. Convolution is used to perform dimensionality increase, so that the channel dimensions of the high-frequency components meet the requirements of inverse wavelet transform. This represents the output of the frequency domain branch.

[0028] The spatial domain enhancement module fuses the spatial domain features from the lower layer and the output from the low-frequency enhancement module, and then performs the same enhancement operation as the encoder spatial domain enhancement module. This process can be described as follows:

[0029] in, This represents the output of the low-frequency enhancement module. Represents spatial domain characteristics from the lower layer. The output represents the spatial domain branch.

[0030] Furthermore, the discrete wavelet transform employs Haar wavelets to achieve efficient downsampling of features.

[0031] Furthermore, the loss function specifically comprises: pixel-level L1 loss, perceptual loss based on pre-trained VGG, SSIM loss, LAB color space loss, and edge loss based on the Sobel operator. The calculation of the loss can be expressed as:

[0032] in, Indicates the total loss. This indicates the loss between the enhanced image and the sharpened image. This represents the loss between the output of the frequency domain branch and the sharp image. This represents the loss between the spatial domain branch and the sharp image.

[0033] The calculation can be described as follows:

[0034] in, For pixel-level L1 loss, For perceptual loss based on pre-trained VGG, For SSIM loss, For LAB color space loss, For edge loss based on Sobel operator The calculation can be described as follows:

[0035] The calculation can be described as follows:

[0036] Furthermore, the method employs a dual-domain fusion approach, specifically in the encoder stage, where the low-frequency component... Output characteristics of the receiving spatial domain branch Fusion enhancement, high-frequency features Gradient information from low-frequency components Guided enhancement; in the decoder stage, low-frequency components are passed through the output of the high-frequency module. Structural information guidance enhancement, input of spatial domain branches Receive the output of the low-frequency enhancement module The process involves fusion and refinement; in the final output stage, the outputs of the frequency domain and spatial domain branches are fused using learnable weights. Through this bidirectional, multi-layered fusion mechanism, joint enhancement of frequency domain and spatial domain information is achieved.

[0037] Compared with the prior art, the present invention has the following advantages: This invention combines frequency domain information and spatial domain information through a dual-branch structure, uses discrete wavelet transform to separate the low-frequency and high-frequency components of the image, and designs dedicated modules for enhancement of each component. This approach can more effectively restore the texture details and overall structure of the image, overcoming the limitations of single spatial domain enhancement methods.

[0038] The color cast preprocessing module proposed in this invention can detect the type of color cast in an image and employ a targeted adaptive processing strategy to narrow the range of color deviation to a certain extent, which helps the subsequent network to efficiently learn color features. This module introduces learnable parameters, enabling the correction process to be trained end-to-end, better adapting to complex and variable underwater lighting conditions. This invention integrates a high-frequency enhancement module, a low-frequency enhancement module, and a spatial domain enhancement module in the encoder-decoder. These modules, through gradient guidance, multi-scale pooling, attention mechanisms, and a state-space model, achieve efficient feature extraction, refined processing, and effective fusion, significantly improving the network's representational capabilities. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1This is a schematic diagram of the network method flow of a dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba in an embodiment of the present invention.

[0041] Figure 2 This is the main diagram of the network architecture of the dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba in this embodiment of the invention.

[0042] Figure 3 This is an encoder-decoder architecture diagram of a dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba in an embodiment of the present invention.

[0043] Figure 4 This refers to the underwater degraded image and the clear underwater image output by the dual-domain fusion underwater image sharpening network based on wavelet transform and Mamba in the embodiments of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0045] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0046] like Figure 1 The underwater image sharpening method based on wavelet transform and Mamba, as shown, mainly includes the following steps: S1. Construct an underwater image enhancement network, the network including: a color shift preprocessing module, a dual-domain encoder and a dual-domain decoder; In this embodiment, the encoder and decoder are configured with a total of 3 levels. Figure 2 This is a diagram of the overall network structure. Figure 3 This is a diagram of the encoder-decoder architecture.

[0047] S2. Input the original underwater image into the color cast preprocessing module, perform color adaptive balancing, and obtain the color cast adjusted image. Specifically, the network receives an underwater image that needs to be enhanced. First, the image is sent to the color cast preprocessing module. This module internally determines the type of color cast in the image. For example, it calculates the average value of each channel. ,like , If the threshold value is blue, then it is judged as blue-biased; if , If the threshold for green bias is reached, then it is judged as green biased.

[0048] For blue bias, the module adjusts the color by calculating the red and green gains: the gain coefficients are calculated based on the average of each channel and a learnable parameter is introduced for fine-tuning.

[0049] A similar strategy is adopted for green bias.

[0050] If there is no obvious color cast, perform white balance.

[0051] The processed image and the input image are added together using learnable weights to obtain the final image. .

[0052] S3. The color-shifted image and the original image are stitched together and input into a dual-domain encoder for multi-scale feature extraction to obtain feature maps at different scales. The encoder includes a discrete wavelet transform module, a low-frequency enhancement module, a high-frequency enhancement module, and a spatial domain enhancement module. The features are divided into low-frequency and high-frequency components through wavelet decomposition and enhanced separately. Specifically, the output of the color shift preprocessing module High-dimensional feature maps are obtained by using convolution to increase the dimensionality. Where C is the number of feature channels, which is set to 64 in this method. Subsequently, F is divided into two parts along the channel dimension to obtain the frequency domain branch input. and spatial domain branch input The inputs are respectively fed into the frequency domain module and the spatial domain module of the encoder.

[0053] S301, Encoder Spatial Domain Enhancement Module, performs multi-scale convolution on the spatial domain branch input and concatenates them, combines channel attention and spatial attention, downsamples and then passes them to the next level.

[0054]

[0055] in, The output of the spatial domain enhancement module is passed to the next level encoder as the input of the spatial domain branch.

[0056] S302, Discrete Wavelet Transform Unit, decomposes the frequency domain branch input into one low-frequency sub-band and three high-frequency sub-bands;

[0057] in, Represents the low-frequency subband. Represents high-frequency sub-bands. This represents the frequency domain input.

[0058] S303, the encoder low-frequency enhancement module, includes multiple parallel branches. Each branch is downsampled at different scales through average pooling, enhanced by convolution and restored to its original size, then added together, further refined by spatial attention and the Mamba module, and connected to the original low-frequency feature residual.

[0059]

[0060] in, The output of the low-frequency enhancement module is passed to the next level encoder as the input of the frequency domain branch.

[0061] S304, encoder high-frequency enhancement module, which enhances the three high-frequency sub-bands Merged into a single high-frequency feature The output of the low-frequency module Gradient information and high-frequency features obtained through the Sobel operator After fusion, the output is obtained through normalization, convolution, and activation functions.

[0062]

[0063] in, The output of the high-frequency enhancement module is passed to the decoder as the input of the decoder's high-frequency module. Including 3 3 convolution, ReLU activation function and 1 1. Convolution.

[0064] S305. The structure of the subsequent encoder is the same as described above, but the number of feature channels processed increases step by step, for example, from 64 to 128, and then to 256, to achieve multi-scale feature extraction of the image.

[0065] S4. The feature map is input into the dual-domain decoder, which gradually restores image details and size, ultimately yielding the enhanced image. The decoder includes an inverse discrete wavelet transform module, a low-frequency enhancement module, a high-frequency enhancement module, and a spatial feature upsampling module, which gradually restores the features output by the encoder into the enhanced image.

[0066] Specifically, the lower-level decoder receives input from the lower-level encoder.

[0067] S401, the high-frequency enhancement module enhances the high-frequency features from the encoder, and the output of the high-frequency enhancement is transmitted to the inverse wavelet transform unit and the low-frequency enhancement module respectively. S402, the low-frequency enhancement module, utilizes the enhanced high-frequency features for guided enhancement and performs the same multi-scale processing and Mamba long sequence modeling as the encoder's low-frequency enhancement module. The output of the low-frequency enhancement is then passed to the inverse wavelet transform unit and the spatial domain enhancement module; this process can be described as follows:

[0068] in, Represents the output of the high-frequency enhancement module. This represents low-frequency characteristics from the encoder or lower-level decoder.

[0069] S403, the inverse wavelet transform unit, merges the enhanced low-frequency and high-frequency features and reconstructs them into spatial domain features;

[0070] in, This represents the output of the low-frequency enhancement module. Represents the output of the high-frequency enhancement module. Represented by 3 3. Convolution is used to perform dimensionality increase, so that the channel dimensions of the high-frequency components meet the requirements of inverse wavelet transform. This represents the output of the frequency domain branch.

[0071] S404, the spatial domain enhancement module, fuses the spatial domain features from the lower layer and the output from the low-frequency enhancement module, and then performs the same enhancement operation as the encoder spatial domain enhancement module. This process can be described as follows:

[0072] in, This represents the output of the low-frequency enhancement module. Represents spatial domain features from the encoder or lower-level decoder. The output represents the spatial domain branch.

[0073] S405. The subsequent decoder structure is similar to the decoder described above, gradually restoring the resolution of the feature map and reducing the number of channels (from 256 to 128, and then to 64), ultimately obtaining the frequency domain branch output. and spatial domain branch output .

[0074] S406, Frequency Domain Branch Output and spatial domain branch output First, through a learnable weight The weighted fusion process can be described as follows:

[0075] right Channel dimensionality reduction is performed using convolution, mapping the number of feature channels from C back to a 3-channel RGB image, compared with the input image. Perform residual connections to obtain the final enhanced image.

[0076] S5. Calculate the difference between the enhanced image and the clear image using pixel-level loss, perceptual loss, SSIM loss, color loss, and edge loss. Perform error backpropagation based on the loss values ​​and update the model parameters.

[0077] Specifically, a dataset containing a large number of paired underwater images is used. The enhanced images output by the network are used as supervisory signals to evaluate the quality of the output results, with the corresponding sharpened images serving as the supervisory signals. The loss functions used include: pixel-level L1 loss, perceptual loss based on pre-trained VGG, SSIM loss, LAB color space loss, and edge loss.

[0078] The calculation of Loss can be expressed as:

[0079] in, Indicates the total loss. This indicates the loss between the enhanced image and the sharpened image. This represents the loss between the output of the frequency domain branch and the sharp image. This represents the loss between the spatial domain branch and the sharp image.

[0080] The calculation can be described as follows:

[0081] in, For pixel-level L1 loss, For perceptual loss based on pre-trained VGG, For SSIM loss, For LAB color space loss, This is the edge loss based on the Sobel operator.

[0082] The calculation can be described as follows:

[0083] The calculation can be described as follows:

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dual-domain fusion underwater image sharpening method based on wavelet transform and Mamba, characterized in that, include: An underwater image enhancement network is constructed, which includes a color shift preprocessing module, a dual-domain encoder, and a dual-domain decoder. The original underwater image is input into the color shift preprocessing module for color adaptive balancing to obtain the color shift adjusted image. The color-shifted image and the original image are stitched together and input into a dual-domain encoder for multi-scale feature extraction to obtain feature maps at different scales. The dual-domain encoder includes a discrete wavelet transform unit, a low-frequency enhancement module, a high-frequency enhancement module, and a spatial domain enhancement module. The frequency domain features are divided into low-frequency and high-frequency components through wavelet decomposition and enhanced separately. The feature map is input into the dual-domain decoder to gradually restore the image details and size, and finally obtain the enhanced image. The dual-domain decoder includes an inverse discrete wavelet transform module, a low-frequency enhancement module, a high-frequency enhancement module, and a spatial domain enhancement module, which gradually restores the features output by the dual-domain encoder into the enhanced image. The difference between the enhanced and sharp images is calculated using pixel-level loss, perceptual loss, SSIM loss, color loss, and Sobel-based edge loss. Error backpropagation is then performed based on the loss values ​​to update the model parameters.

2. The underwater image sharpening method based on wavelet transform and Mamba according to claim 1, characterized in that: underwater images to be enhanced The input image is fed into the color cast preprocessing module, which adds the processed image and the input image using learnable weights to obtain the final image. In the color cast preprocessing module, the determination of blue or green cast is based on whether the ratio of channel mean values ​​exceeds a learnable threshold, and the gain of each channel is adjusted through learnable parameters. For different types of color cast, the color cast preprocessing module adopts the following correction strategy: For blue-biased images, the intensity of the red and green channels is adjusted by calculating the gain coefficients for red and green. For images with a green tint, adjust the red and green channels; For images with no color cast or only a slight color cast, perform white balance.

3. The underwater image sharpening method based on wavelet transform and Mamba according to claim 2, characterized in that: Output of the color shift preprocessing module High-dimensional feature maps are obtained by using convolution to increase the dimensionality. Where C is the number of feature channels, and F is divided into two parts in the channel dimension to obtain the frequency domain branch input. and spatial domain branch input And input to a dual-domain encoder, wherein each stage of the dual-domain encoder includes: The spatial domain enhancement module performs multi-scale convolutions on the spatial domain branch inputs and concatenates them, then combines channel attention and spatial attention, downsamples the data, and passes it to the next stage. in, Represents the spatial domain branch input, Represents convolution operations with different receptive fields, i=1,3,5,7; Represents the output of the spatial domain enhancement module; The spatial attention calculation process is as follows: in, 7 Convolution operation of 7, and These represent max pooling and average pooling operations, respectively. For the Sigmoid function; The channel attention calculation process is as follows: MLP includes: 1 1 convolution, ReLU activation function and 1 1. Convolution; The discrete wavelet transform unit decomposes the frequency domain branch input into a low-frequency subband and three high-frequency subbands; in, Represents the low-frequency subband. Represents high-frequency sub-bands. Represents frequency domain branch input The low-frequency enhancement module fuses the low-frequency subband with the spatial features obtained from the spatial domain module, and then enhances them through multi-scale processing and Mamba long sequence modeling. The multi-scale processing includes multiple parallel branches, each of which performs downsampling at different scales through average pooling. After convolutional enhancement and restoration of the original size, the samples are summed, then refined by spatial attention and the Mamba module, and finally connected to the residuals of the original low-frequency features. in, This is output information from the spatial domain enhancement module. This represents downsampling to 1 / i of the original size using pooling; Indicates 3 3 convolutions; This represents an upsampling operation that restores the original size; SA stands for Spatial Attention; and SS2D represents Mamba's scanning module, used for long sequence feature modeling. This represents the output of the low-frequency enhancement module; The high-frequency enhancement module will integrate the three high-frequency subbands. Merged into a single high-frequency feature The output of the low-frequency module Gradient information and high-frequency features obtained through the Sobel operator After fusion, the output is obtained through normalization, convolution, and activation functions. This process can be represented as follows: in Representative instance normalization, Represents the output of the low-frequency module Gradient information obtained through the Sobel operator Including 3 3 convolution, ReLU activation function and 1 1 convolution, This represents the output of the high-frequency enhancement module.

4. The underwater image sharpening method based on wavelet transform and Mamba according to claim 1, characterized in that: Each stage of the dual-domain decoder includes: The high-frequency enhancement module enhances the high-frequency features from the encoder, and the output of the high-frequency enhancement is sent to the inverse wavelet transform unit and the low-frequency enhancement module respectively. The low-frequency enhancement module utilizes the enhanced high-frequency features for guided enhancement and performs the same multi-scale processing and Mamba long sequence modeling as the encoder's low-frequency enhancement module. The output information of the low-frequency enhancement is then transmitted to the inverse wavelet transform unit and the spatial domain enhancement module, respectively. This process is represented as follows: in, Represents the output of the high-frequency enhancement module. This represents low-frequency characteristics from the lower layers; The inverse wavelet transform unit merges and reconstructs the enhanced low-frequency and high-frequency features; in, This represents the output of the low-frequency enhancement module. Represents the output of the high-frequency enhancement module. Represented by 3 3. Convolution is used to perform dimensionality increase, so that the channel dimensions of the high-frequency components meet the requirements of inverse wavelet transform. The output represents the frequency domain branch; The spatial domain enhancement module fuses the spatial domain features from the lower layer and the output from the low-frequency enhancement module, and then performs the same enhancement operation as the spatial domain enhancement module. This process is represented as follows: in, This represents the output of the low-frequency enhancement module. Represents spatial domain characteristics from the lower layer. The output represents the spatial domain branch.

5. The underwater image sharpening method based on wavelet transform and Mamba according to claim 1, characterized in that: The pixel-level loss, perceptual loss, SSIM loss, color loss, and Sobel-based edge loss are expressed as follows: The calculation of Loss is expressed as follows: in, Indicates the total loss. This indicates the loss between the enhanced image and the sharpened image. This represents the loss between the output of the frequency domain branch and the sharp image. This represents the loss between the spatial domain branch and the sharp image. The calculation is described as follows: in, For pixel-level L1 loss, For perceptual loss based on pre-trained VGG, For SSIM loss, For LAB color space loss, For edge loss based on Sobel operator The calculation is described as follows: The calculation is described as follows: 。

Images