A sonar image denoising method based on self-supervised learning

Abstract

The present application relates to the technical field of sonar image denoising processing, in particular to a sonar image denoising method based on self-supervised learning, which specifically comprises the following steps: converting the noise type of a noisy sonar image; using a nearest neighbor sampling strategy to generate a pair of noisy images from a single preprocessed noisy sonar image; constructing a denoising neural network model comprising a multi-branch fusion module and a high-frequency information enhancement module; training the denoising neural network model using a loss function comprising a reconstruction loss term and a regularization loss term; using the trained denoising neural network model to denoise a target noisy sonar image; and post-processing the denoised image to restore the original image domain. The present application can effectively suppress multiplicative speckle noise in sonar images without the need for clean noise-free images as labels, and can also enhance the high-frequency information of the image, thereby improving the clarity and detail performance of the denoised image.

Description

Technical Field

[0001] This invention relates to the field of sonar image denoising technology, specifically a sonar image denoising method based on self-supervised learning. Background Technology

[0002] Sonar images, due to their unique underwater imaging capabilities, have important applications in fields such as ocean exploration, resource surveying, and underwater security. However, the imaging mechanism of sonar images and the complex underwater environment often result in them containing severe noise, especially multiplicative speckle noise, and are often accompanied by the loss of high-frequency information (such as edges and texture details). These problems seriously affect image quality and the accuracy of subsequent target detection, recognition, and other tasks.

[0003] Existing sonar image denoising methods mainly include traditional methods (such as filtering and transform domain methods) and deep learning-based methods. Traditional methods have limited effectiveness in handling complex noise, especially speckle noise unique to sonar images, and are prone to losing image details. Deep learning-based methods are generally divided into supervised and unsupervised (including self-supervised) methods. Supervised methods require a large number of noisy images and corresponding clean, noise-free image pairs for training. However, in practical sonar applications, obtaining truly clean, noise-free images is extremely difficult, limiting their application. Self-supervised learning methods (such as Noise2Noise, Neighbor2Neighbor, etc.) construct training pairs from the noisy images themselves, avoiding the dependence on clean images.

[0004] However, existing self-supervised methods still face the following challenges when directly applied to sonar images:

[0005] 1. Existing self-supervised methods are theoretically designed mainly for additive noise. For multiplicative speckle noise, additional transformations (such as logarithmic transformations) are required, which may introduce other new problems and increase the denoising steps.

[0006] 2. The network structure usually draws on natural image processing, but fails to fully consider the characteristics of missing high-frequency information in sonar images, resulting in insufficient restoration of image details after denoising and loss of overall integrity;

[0007] 3. Existing networks (such as standard U-Net or networks based on self-attention mechanisms) are not efficient in capturing long-range dependencies and fusing multi-scale features, and have high computational costs. Summary of the Invention

[0008] The purpose of this invention is to provide a sonar image denoising method based on self-supervised learning to address the shortcomings of existing technologies. This method is specifically designed for the characteristics of sonar images. Without requiring clean, noise-free images as labels, it can effectively suppress multiplicative speckle noise in sonar images and specifically enhance the high-frequency information of the image, thereby improving the clarity and detail of the denoised image and solving the problems existing in the prior art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A sonar image denoising method based on self-supervised learning includes the following steps:

[0011] S1) Preprocess the acquired noisy natron image;

[0012] S2) A nearest neighbor sampling strategy is adopted to generate a pair of noisy images in the preprocessed noisy image, including a first noisy image and a second noisy image;

[0013] S3) Construct a denoising neural network model, which includes a multi-branch fusion module and a high-frequency information enhancement module;

[0014] S4) The first noisy image is used as the input of the denoising neural network model, and the second noisy image is used as the supervision target. The denoising neural network model is trained using a loss function that includes a reconstruction loss term and a regularization loss term.

[0015] S5) Use the trained denoising neural network model to denoise the target noisy image;

[0016] S6) Post-process the denoised image to restore the original image domain;

[0017] In step S4, the first noisy image Input denoising neural network model The second noisy image As a monitoring target, a method including reconstruction loss items is adopted. and regularization loss term Total loss function For denoising neural network models The training process is calculated using the following formula:

[0018]

[0019] in, These are the weighting parameters used to balance the two losses;

[0020] Reconstruction Loss Items The calculation formula is:

[0021]

[0022] in, To make the first noisy image Input Denoising Neural Network The actual output obtained later is the second noisy image. As a monitoring target, Indicates the expected value calculation, symbol Indicates mean square error;

[0023] Regularization loss term The calculation formula is:

[0024]

[0025] in, To make the first noisy image Input Denoising Neural Network The actual output obtained afterwards and These are respectively for the denoising neural network Applied to the original noisy image Ideal denoising result Process and Generation and Images obtained after the exact same nearest neighbor sampling operation. Indicates the expected value calculation, symbol Indicates mean square error;

[0026] Minimize the total loss function using the gradient descent optimizer To iteratively update network parameters The optimization process continues until convergence, and is as follows:

[0027] Gradient calculation: In each training batch, the total loss function is calculated using the backpropagation algorithm. Relative to the network trainable parameters gradient ;

[0028] Parameter update: The optimizer updates the parameters based on the calculated gradients. and the preset learning rate And, in conjunction with its internal state, calculate the parameter update amount. And update the parameters according to the following rules:

[0029]

[0030] This iterative update process will continue until the network performance converges or the preset number of training epochs is reached.

[0031] Furthermore, the specific preprocessing steps in step S1 are as follows:

[0032] For noisy images Perform a logarithmic transformation to obtain a noisy image in the logarithmic domain. The specific formula is as follows:

[0033]

[0034] in, For noisy images in the logarithmic domain, For noise-free images, It is additive Gaussian noise;

[0035] Among them, noisy images The model is:

[0036]

[0037] in, For the input noisy image, For an ideal noise-free image, It is multiplicative speckle noise.

[0038] Furthermore, in step S2, the specific process of the nearest neighbor sampling strategy is as follows:

[0039] S21) The logarithmic domain noisy image Divided into multiple non-overlapping Cell;

[0040] S22) For each Cell, randomly select two different pixel positions and ;

[0041] S23) Pixel position The pixel values ​​are used as the first noisy image. The pixel value at the corresponding position in this cell;

[0042] S24) Pixel position The pixel values ​​are used as the second noisy image. The pixel value at the corresponding position in this cell;

[0043] S25) Merge the results of all cell merges to form an image with the same dimensions as the original. Noisy images , serving as both input and target for the denoising neural network model.

[0044] Furthermore, in step S3, the specific process of constructing the denoising neural network model is as follows:

[0045] S31) A U-shaped network is used as the basic architecture to build a denoising neural network model. Denoising Neural Network Model It includes an encoder path and a decoder path, wherein, These are the trainable parameters of the network;

[0046] S32) In the encoder path, a multi-branch fusion module is used to replace the traditional convolutional block. The multi-branch fusion module includes multiple different convolutional paths and fuses the features extracted by the convolutional paths through a specific fusion strategy.

[0047] S33) A high-frequency information enhancement module is embedded at a specific location in the denoising neural network model to enhance the high-frequency information that is easily lost in the sonar image. The high-frequency information enhancement module includes a large kernel convolutional layer and an inverted bottleneck structure.

[0048] Further, in step S32, the multi-branch fusion module includes a first convolutional path and a second convolutional path, wherein the first convolutional path includes three convolutional layers, namely... and Let the input be the first The input feature maps of each encoder are Its dimensions are ,make Indicates the first The first encoder A multi-branch fusion module, the first convolutional layer The kernel size is ,make The output is represented as Then output feature map for:

[0049]

[0050] in, , and Indicates the spatial dimension of the output feature map;

[0051] Second convolutional layer The kernel size is The third convolutional layer The kernel size is ,but and The output feature maps are as follows:

[0052]

[0053] in, and They represent Output and The output;

[0054] The second convolution path consists of a single convolution kernel.

[0055] Furthermore, in step S33, the calculation process for the inverted bottleneck structure is as follows:

[0056] The inverted bottleneck structure passes through the first convolutional layer For input features Obtain intermediate features by performing channel dimensionality upgrade The calculation process is as follows:

[0057]

[0058] Through the second convolutional layer For intermediate features Channel dimensionality reduction is obtained :

[0059]

[0060] in, For input features, As an intermediate feature, The output characteristics after processing with the inverted bottleneck structure. It is a non-linear activation function.

[0061] Furthermore, in step S33, the output characteristics of the inverted bottleneck structure are connected via residual connections. Its original input features The final output of the high-frequency information enhancement module is obtained by adding elements one by one. The calculation formula is as follows:

[0062]

[0063] Furthermore, in step S5, the sonar image to be denoised is input into the trained denoising neural network model. In the process, a forward propagation calculation is performed to obtain the image after preliminary denoising. .

[0064] Furthermore, in step S6, the denoised image... An exponential transform is performed to obtain the final denoised image. The calculation formula is as follows:

[0065]

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] The sonar image denoising method based on self-supervised learning provided by this invention can not only adapt to the self-supervised learning paradigm to get rid of the dependence on clean labels, but also use a specially designed network structure to specifically process multiplicative speckle noise and actively recover the missing high-frequency information of the sonar image, thereby improving the clarity and detail of the denoised image to ensure the integrity of the image. At the same time, it can also efficiently perform multi-scale feature fusion, simplifying the denoising steps to the greatest extent and reducing the computational cost. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0069] Figure 2 This is a schematic diagram illustrating the generation of training image pairs through nearest neighbor sampling in this invention;

[0070] Figure 3 This is a schematic diagram of the U-shaped architecture used in the denoising neural network of this invention;

[0071] Figure 4 This is a schematic diagram of the multi-branch fusion module in this invention;

[0072] Figure 5 This is a schematic diagram of the high-frequency information enhancement module in this invention. Detailed Implementation

[0073] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0074] For easier understanding, please refer to Figures 1 to 5 This invention provides a sonar image denoising method based on self-supervised learning, comprising the following steps:

[0075] S1) Preprocess the acquired noisy na image, that is, perform a logarithmic transformation on the input noisy na image to approximately convert the multiplicative speckle noise into additive noise.

[0076] Considering that the main noise in noisy nag images is multiplicative speckle noise, the model for noisy nag images can be simplified as follows:

[0077]

[0078] in, For the input noisy image, For an ideal noise-free image, It is multiplicative speckle noise.

[0079] To apply self-supervised learning frameworks and their core modules (such as MBF and HFIE) that are more direct and efficient at handling additive noise, for input noisy images... Perform a logarithmic transformation to obtain a noisy image in the logarithmic domain. The specific formula is as follows:

[0080]

[0081] in, For noisy images in the logarithmic domain, For noise-free images, the original multiplicative speckle noise is removed. Approximates as additive Gaussian noise in the logarithmic domain. This is for subsequent network processing.

[0082] The subsequent denoising neural network will be used on this logarithmic domain noisy image. Perform the operation on it.

[0083] S2) A nearest neighbor sampling strategy is employed to generate a pair of noisy images from a preprocessed single noisy natron image, comprising a first noisy image and a second noisy image. The nearest neighbor sampling strategy includes the following steps:

[0084] S21) The processed logarithmic domain noisy image Divided into multiple non-overlapping Cells, for example, taking .

[0085] S22) For each Cell, randomly select two different pixel positions and .

[0086] S23) Position The pixel values ​​are used as the first noisy image. The pixel value at the corresponding position in this cell.

[0087] S24) Position The pixel values ​​are used as the second noisy image. The pixel value at the corresponding position in this cell.

[0088] S25) Merge the results of all cell merges to form an image with the same dimensions as the original. Noisy images , serving as both the input and target of the denoising neural network.

[0089] For specific details and to facilitate understanding, please refer to... Figure 2 Nearest neighbor sampler Input a single logarithmic domain noisy image (For ease of subsequent description, let's call it) ), and divide it into For each cell For each cell, randomly select two different pixel positions (e.g., top left and bottom right, or two randomly adjacent pixels), and assign these two pixels to the sub-image respectively. and At the corresponding positions, the processing results of all cells are concatenated to form two images of the same size as the original image. (Right now ,in For width, Noisy image (for height) and These two images are highly similar in content, but there are pixel-level differences, which meets the requirements of self-supervised training.

[0090] S3) Construct a denoising neural network model The model includes a Multi-Branch Fusion (MBF) module and a High-Frequency Information Enhancement (HFIE) module. The specific model construction steps are as follows:

[0091] S31) Please refer to Figure 3 The denoising neural network model constructed in this embodiment It adopts the classic U-Net as its basic architecture, which consists of an encoder path ( Figure 3 On the left, there is a series of downsampling modules and a decoder path. Figure 3 The right side consists of a series of upsampling modules, among which... These are the trainable parameters of the network (the set of network parameters). The encoder path progressively reduces the spatial resolution of the feature maps while increasing the number of channels to extract abstract features from different levels; the decoder path progressively restores the spatial resolution while reducing the number of channels, and combines feature information from the corresponding level of the encoder (through skip connections). Figure 3 (As shown by the horizontal arrow in the middle) to accurately reconstruct image details.

[0092] S32) Please refer to Figure 4The embedded multi-branch fusion module (MBF module) replaces traditional convolutional blocks in at least one downsampling module (preferably multiple or all downsampling modules) of the encoder path. The core design idea of ​​the MBF module is to process input features in parallel to capture information at different scales or frequencies and to perform effective fusion without relying on computationally intensive self-attention mechanisms. That is, it effectively captures and fuses feature information from different receptive fields or different frequency channels without relying on computationally expensive self-attention mechanisms. The MBF module contains at least two parallel convolutional processing paths. These paths use convolutional layers with different receptive fields (e.g., different kernel sizes) to extract high-frequency and low-frequency information from the input feature map, and fuse the features extracted from multiple branches without relying on self-attention mechanisms to simultaneously capture global context and local details.

[0093] Specifically, in this embodiment, the MBF module consists of two different convolutional paths, one of which uses a smaller convolutional kernel (e.g., ...). One approach focuses on capturing and extracting high-frequency details and texture information from local image regions, while another approach uses larger convolutional kernels (such as...). Alternatively, dilated convolutions can be used to expand the receptive field, or deeper network structures can be constructed to capture broader low-frequency structures, low-frequency information, contour features, and contextual information. This design allows the network to process information at different scales simultaneously, enhancing its feature representation capabilities. The features extracted by the parallel paths described above are integrated at the end of the module using specific fusion strategies (such as element-wise addition, channel concatenation followed by convolution for feature selection and dimensionality reduction). This means that the input features from different parallel paths are fused into a single output feature tensor. This output feature tensor is a comprehensive representation of the input features across multiple scales and frequencies, and it is passed as the final output of the MBF module to subsequent network layers.

[0094] More specifically, the first convolutional path branch consists of three convolutional layers, each using... , and This indicates that for each convolutional layer within an MBF module, the stride will be... Set to equal to the kernel size per layer (Right now ), and without using fill (Right now This generates non-overlapping spatial blocks, thereby reducing spatial dimensionality. The output spatial resolution of any convolutional layer can be derived from the following general formula:

[0095]

[0096] in, The size (i.e., height or width) of the input feature map in the corresponding spatial dimension. To determine the size of the output feature map in the same dimension, Indicates filling. Indicates stride length. Indicates the kernel size.

[0097] In the MBF module, set , And rewrite the above formula as follows:

[0098]

[0099] Let the input be the first The input feature maps of each encoder are Its dimensions are ,make Indicates the first The first encoder One MBF module, the first convolutional layer The kernel size is It will be derived from the input feature map. Features of the aggregation space in the neighborhood. Let The output is represented as Then the output feature map can be represented as:

[0100]

[0101] in , and This indicates the spatial dimension of the output feature map.

[0102] Similarly, the second convolutional layer The kernel size is It will further from the upper layer Information is aggregated within these regions. These regions correspond to the original input feature maps. The region, therefore the output of the second convolution can be described as the region from the input feature map. Aggregated information of the neighborhood.

[0103] Accordingly, the last convolutional layer The kernel size is Continuing to aggregate the information from the previous layer, then... and The output feature map can be represented as:

[0104]

[0105] in and They represent Output and The output of .

[0106] The second convolution path branch consists of a single convolution kernel and focuses more on the low-frequency information of the image.

[0107] By combining small kernel convolutional layers and dual-path design, global context information is effectively captured, while achieving deep fusion of high-frequency and low-frequency information.

[0108] (S33) Please refer to Figure 5 Embedded high-frequency information enhancement module (HFIE module): in the denoising neural network model At least one HFIE module is embedded at a specific location (e.g., after the MBF module of each encoder, or integrated into the decoding path). HFIE modules are specifically designed to enhance high-frequency information such as edges and textures that are easily lost in sonar images. An HFIE module includes a large-kernel convolutional layer (e.g., It uses depthwise separable convolutions and an inverted bottleneck structure (ConvFFN), where large-kernel convolutional layers provide better performance than standard convolutional layers. A larger receptive field in convolution helps capture a wider range of structural information, especially edge contours. To improve efficiency, this large-kernel convolution is equivalently implemented post-training using structural reparameterization techniques, while parallel small kernels (such as a single kernel) are used during training. Convolutional (combination). The inverted bottleneck structure (ConvFFN) borrows from the feedforward network design in the Transformer, it first passes through a... Convolution inputs it into the feature map (denoted here as ). Increase the dimensionality to a higher dimension (e.g., increase the dimensionality by a factor of 4, i.e., 4C), then apply a non-linear activation function (such as GELU) for non-linear activation, and finally pass it through another... Convolution restores the number of channels to the original number (i.e., reduces the dimensionality to C). This dimensionality increase-decrease process facilitates the full interaction and refinement of features. The inverted bottleneck structure is beneficial for input features. Processing yields output features The calculation process includes:

[0109] Through the first convolutional layer right Obtain intermediate features by performing channel dimensionality upgrade :

[0110]

[0111] Through the second convolutional layer right Channel dimensionality reduction is obtained :

[0112]

[0113] in, For input features (feature maps received by the HFIE module). As an intermediate feature, This represents the output characteristics after processing with the inverted bottleneck structure.

[0114] Furthermore, to stabilize the training process and preserve the original input information, the final output of the HFIE module... The output features of the inverted bottleneck structure are obtained through a residual connection. Its original input features The formula for calculating it, obtained by adding each element one by one, is:

[0115]

[0116] This operation ensures that the HFIE module primarily learns incremental or residual information from the input features, thereby improving the network's performance and training stability.

[0117] S4) Pair the noisy image Used for training denoising neural network models This step aims to optimize the denoising neural network model. (i.e., a denoiser) that enables it to effectively recover clean images from noisy sonar images, and is specifically optimized for the characteristics of sonar images.

[0118] Specifically, the first noisy image Input denoising neural network model The second noisy image As a monitoring target, a method including reconstruction loss items is adopted. and regularization loss term Total loss function For denoising neural network models The training process is calculated using the following formula:

[0119]

[0120] in, These are the weighting parameters used to balance the two losses.

[0121] Reconstruction Loss Items The L2 norm (mean squared error) is typically used to measure the difference between the network output and the supervised target. Its formula is as follows:

[0122]

[0123] in, To make the first noisy image Input Denoising Neural Network The actual output obtained later is the second noisy image. As a monitoring target, This indicates the expected calculation. The mean squared error (MSE) is the square of the L2 norm of the difference between the network output and the second noisy image.

[0124] Regularization loss term Used to compensate for the introduction of the nearest neighbor sampling strategy and The tiny differences between the corresponding clean underlying images This ensures that the model learns a universal denoising capability, rather than a specific one. arrive The specific mapping of is in the following form:

[0125]

[0126] in, To make the first noisy image Input Denoising Neural Network The actual output obtained afterwards and These are respectively for the denoising neural network Applied to the original noisy image Ideal denoising result Process and Generation and Images obtained after the exact same nearest neighbor sampling operation. This indicates the expected calculation. This represents the mean square error.

[0127] The total loss function is minimized by using a gradient descent optimizer (i.e., an optimization algorithm such as Adam, where the learning rate decays according to the policy). To iteratively update network parameters The optimization process continues until convergence, and is as follows:

[0128] Gradient calculation: In each training batch, the total loss function is calculated using the backpropagation algorithm. Compared to all trainable parameters of the network gradient Specifically, the backpropagation algorithm is based on the chain rule, calculating the gradient layer by layer from the output layer to the input layer. For any layer in the network... parameters The gradient calculation can be schematically represented as follows:

[0129]

[0130] in, For the first The output of the layer.

[0131] Parameter update: The optimizer updates the parameters based on the calculated gradients. and the preset learning rate Update network parameters. Specifically, the amount of parameter updates. It is estimated by maintaining the first moment of the gradient (momentum). and second-order moment estimation The core update rule used for calculation can be simplified as follows:

[0132]

[0133] in, and The first and second moments are estimated after bias correction. It is a tiny constant used to maintain numerical stability.

[0134] Update parameters according to the following rules:

[0135]

[0136] in, The parameter values ​​at the start of the current iteration step. This is the updated parameter value.

[0137] This iterative update process will continue until the network performance converges or the preset number of training epochs is reached.

[0138] S5) Denoising application, which uses a pre-trained denoising neural network model. Denoising is performed on the noisy image of the target.

[0139] Specifically, the sonar image to be denoised (if it has undergone step S1, it is a logarithmic domain noisy image after logarithmic transformation) is... The input is fed into the trained denoising neural network model. In the process, a forward propagation calculation is performed to obtain the image after preliminary denoising. Denoising image This can be understood as an estimate of the ideal image in the logarithmic domain.

[0140] More specifically, the forward propagation calculation process includes: using the image to be denoised as a denoiser. First layer input And according to the network's preset topology, the output of each layer is calculated recursively layer by layer. For any layer in the network... Its output The calculation can be schematically represented as:

[0141]

[0142] in, and For the first The pre-trained weights and bias parameters of the layer, This is the activation function for this layer.

[0143] Image after initial denoising This refers to the output of the last layer (output layer) of the network. This process does not involve loss calculation or parameter updates.

[0144] S6) Denoising the image Post-processing is performed to recover the original image domain. Specifically, for the denoised image... Perform an exponential (inverse logarithmic) transform to obtain the final denoised image. The calculation formula is as follows:

[0145]

[0146] If step S1 is not executed, then .

[0147] Based on the above denoising methods, this embodiment also provides a sonar image denoising system based on self-supervised learning, which includes:

[0148] Input interface: Used to receive noisy natron images to be processed.

[0149] Preprocessing module: Used to perform logarithmic transformation.

[0150] Training pair generation module: Noisy image pairs are generated using a nearest neighbor sampling strategy.

[0151] Denoising network module: Stores the trained denoiser, which includes the MBF module and HFIE module as described above.

[0152] Processing unit: Calls the denoising network module to denoise the input image (or the pre-processed image).

[0153] Post-processing module: Used to perform exponential transformation.

[0154] Output interface: Outputs the final denoised image.

[0155] Furthermore, the system can be implemented by a processor executing instructions stored on a non-transitory computer-readable medium.

[0156] Although the present invention has been described using the above preferred embodiments, it is not intended to limit the scope of protection of the present invention. Any changes and modifications made by those skilled in the art to the above embodiments without departing from the spirit and scope of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. A sonar image denoising method based on self-supervised learning, characterized in that, Includes the following steps: S1) Preprocess the acquired noisy natron image; S2) A nearest neighbor sampling strategy is adopted to generate a pair of noisy images in the preprocessed noisy image, including a first noisy image and a second noisy image; S3) Construct a denoising neural network model, which includes a multi-branch fusion module and a high-frequency information enhancement module; S4) The first noisy image is used as the input of the denoising neural network model, and the second noisy image is used as the supervision target. The denoising neural network model is trained using a loss function that includes a reconstruction loss term and a regularization loss term. S5) Use the trained denoising neural network model to denoise the target noisy image; S6) Post-process the denoised image to restore the original image domain; In step S3, the specific process of constructing the denoising neural network model is as follows: S31) A U-shaped network is used as the basic architecture to build a denoising neural network model. Denoising Neural Network Model It includes an encoder path and a decoder path, wherein, These are the trainable parameters of the network; S32) In the encoder path, a multi-branch fusion module is used to replace the traditional convolutional block. The multi-branch fusion module includes multiple different convolutional paths and fuses the features extracted by the convolutional paths through a specific fusion strategy. S33) A high-frequency information enhancement module is embedded at a specific location in the denoising neural network model to enhance the high-frequency information that is easily lost in the sonar image. The high-frequency information enhancement module includes a large kernel convolutional layer and an inverted bottleneck structure. In step S4, the first noisy image is input into the denoising neural network model. The second noisy image is used as the supervision target, and a reconstruction loss term is employed. and regularization loss term Total loss function For denoising neural network models Conduct training; Minimize the total loss function using the gradient descent optimizer To iteratively update network parameters The optimization process continues until convergence, and is as follows: Gradient calculation: In each training batch, the total loss function is calculated using the backpropagation algorithm. Relative to the network trainable parameters gradient ; Parameter update: The optimizer updates the parameters based on the calculated gradient. and the preset learning rate And, in conjunction with its internal state, calculate the parameter update amount. And update the parameters; This iterative update process will continue until the network performance converges or the preset number of training epochs is reached.

2. The sonar image denoising method based on self-supervised learning according to claim 1, characterized in that, In step S1, the specific preprocessing process is as follows: For noisy images Perform a logarithmic transformation to obtain a noisy image in the logarithmic domain. The specific formula is as follows: ； in, For noisy images in the logarithmic domain, For a noise-free image in the logarithmic domain, It is additive Gaussian noise; Among them, noisy images The model is: ； in, For the input noisy image, For an ideal noise-free image, It is multiplicative speckle noise.

3. The sonar image denoising method based on self-supervised learning according to claim 2, characterized in that, In step S2, the specific process of the nearest neighbor sampling strategy is as follows: S21) The logarithmic domain noisy image Divided into multiple non-overlapping Cell; S22) For each Cell, randomly select two different pixel positions and ; S23) Pixel position The pixel value is used as the pixel value of the first noisy image at the corresponding position in that cell; S24) Pixel position The pixel value is used as the pixel value of the second noisy image at the corresponding position in that cell; S25) Merge the results of all cell merges to form an image with the same dimensions as the original. The noisy image pairs are used as the input and target of the denoising neural network model.

4. The sonar image denoising method based on self-supervised learning according to claim 1, characterized in that, In step S32, the multi-branch fusion module includes a first convolutional path and a second convolutional path, wherein the first convolutional path includes three convolutional layers, namely... Let the input be the first The input feature maps of the encoders are Its dimensions are ,make Indicates the first The first encoder A multi-branch fusion module, the first convolutional layer The kernel size is ,make The output is represented as Then output feature map for: ； in, Indicates the spatial dimension of the output feature map; Second convolutional layer The kernel size is The third convolutional layer The kernel size is ,but and The output feature maps are as follows: ； in, They represent Output and The output; The second convolution path consists of a single convolution kernel.

5. The sonar image denoising method based on self-supervised learning according to claim 1, characterized in that, In step S33, the calculation process for the inverted bottleneck structure is as follows: The inverted bottleneck structure passes through the first convolutional layer For input features Obtain intermediate features by performing channel dimensionality upgrade The calculation process is as follows: ； Through the second convolutional layer For intermediate features Channel dimensionality reduction is obtained : ； in, As input features, As an intermediate feature, The output characteristics after processing with the inverted bottleneck structure. It is a non-linear activation function.

6. The sonar image denoising method based on self-supervised learning according to claim 5, characterized in that, In step S33, the output characteristics of the inverted bottleneck structure are connected via residual connections. Its original input features The final output of the high-frequency information enhancement module is obtained by adding elements one by one. The calculation formula is as follows: 。 7. The sonar image denoising method based on self-supervised learning according to claim 1, characterized in that, In step S5, the sonar image to be denoised is input into the trained denoising neural network model. In the process, a forward propagation calculation is performed to obtain the image after preliminary denoising. .

8. The sonar image denoising method based on self-supervised learning according to claim 1, characterized in that, In step S6, the denoised image An exponential transform is performed to obtain the final denoised image. The calculation formula is as follows: 。

Images