Underwater sonar image recognition method with adaptive trigger de-noising and feature decoupling enhancement

Abstract

This invention discloses an adaptive triggered denoising and feature decoupling enhancement method for underwater sonar image recognition, belonging to the field of image recognition technology. The invention constructs an adaptive triggered decision function for multi-index quality assessment, initiating subspace denoising only when the image quality fails to meet detection requirements. A lightweight backbone network is constructed to extract multi-scale features, and a feature decoupling enhancement network is introduced, utilizing physical priors to output sparse components to suppress background clutter and preserve target structure. The enhanced features are fused with the backbone network features and fed into the neck and detection head to complete target detection. An improved ACmix attention module is deployed at the head of the highest-level features. This invention achieves end-to-end optimization of preprocessing and detection tasks, preserving target edges and texture details while suppressing noise interference, effectively improving the accuracy and robustness of target recognition in complex underwater environments, and achieving a balance between computational resources and detection performance through an adaptive mechanism.

Description

Technical Field

[0001] This invention relates to underwater sonar image recognition methods, specifically to an underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement, belonging to the field of image recognition technology. Background Technology

[0002] Underwater sonar imaging technology plays a vital role in marine resource exploration, underwater security, and submarine pipeline inspection. However, the complexity of the underwater environment leads to common problems in sonar images, such as low resolution, poor signal-to-noise ratio, and blurred edges, which seriously interfere with subsequent target identification tasks.

[0003] Numerous research methods have been developed for denoising and enhancing underwater sonar images. Traditional methods include spatial domain-based Gaussian filtering and median filtering, as well as transform domain-based wavelet thresholding. Furthermore, image decomposition-based methods have also attracted attention. These methods model the image as a superposition of different components and achieve noise separation and target enhancement through optimization. In recent years, data-driven methods based on deep learning have been introduced into the field of sonar image processing. By using convolutional neural networks to learn noise distribution characteristics, they have improved denoising performance to some extent.

[0004] However, existing methods still have some shortcomings. First, most methods have fixed processing parameters and cannot adaptively adjust the enhancement intensity according to the actual image quality. Second, existing preprocessing is usually independent of the detection task, and its optimization objective differs from that of subsequent target detection, which may result in the loss of weak target information that is crucial for detection. Finally, small target detection performs poorly in low-quality sonar images and is easily affected by background interference, leading to missed detections or false detections. Summary of the Invention

[0005] This invention provides an adaptive triggering denoising and feature decoupling enhancement method for underwater sonar image recognition, addressing the problems of complex noise in underwater sonar images and the fixed parameters and easy loss of weak target information in existing preprocessing methods. This method employs an adaptive triggering decision mechanism based on multi-index quality assessment, initiating subspace denoising preprocessing only when the image quality does not meet detection requirements. By introducing a physically prior-driven feature decoupling enhancer and an improved ACmix attention module, sparse target priors and high-level semantic features are deeply fused, suppressing background clutter while preserving target edges and texture details, significantly improving the accuracy and robustness of target recognition in complex underwater environments.

[0006] To achieve the above objectives, this invention provides an underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement, comprising the following steps:

[0007] Step 1: Acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment;

[0008] Step 2: Construct an adaptive triggering decision function to evaluate the sonar monitoring image and trigger subspace denoising preprocessing based on the decision result to obtain the sonar monitoring image after decision;

[0009] Step 3: Input the sonar monitoring image after decision-making into the pre-trained network model for processing to obtain the underwater target detection result.

[0010] Further, step 2 includes:

[0011] Calculate multiple quality indicators based on sonar monitoring images and reference images;

[0012] Based on the quality indicator threshold and tolerance threshold, an adaptive triggering decision function is constructed, and the quality indicator is input into the adaptive triggering decision function to obtain the decision result.

[0013] When the decision result is 1, subspace denoising preprocessing is performed on the sonar monitoring image; when the decision result is 0, the sonar monitoring image remains unchanged.

[0014] The final output is the sonar monitoring image after the decision is made.

[0015] Furthermore, the adaptive triggering decision function is expressed as:

[0016] Let the sonar detection image to be evaluated be... The reference image is ,but:

[0017] ,

[0018] in, This is an indicator function that outputs 1 if the condition is true, and 0 otherwise. for and The structural similarity index; for and Peak signal-to-noise ratio; for The Laplace variance; , , These are the threshold values ​​for the corresponding indicators; This is the tolerance parameter; Indicates "or", that is Less than With tolerance parameters The difference or Less than With tolerance parameters The difference or Less than With tolerance parameters The difference ensures that if the image has defects in any dimension of structure, noise, or sharpness, it can be processed in time to avoid missed detections. This is an adaptive triggering decision function used to determine whether the input sonar image needs to be preprocessed for denoising.

[0019] Furthermore, the subspace denoising preprocessing includes:

[0020] The sonar monitoring images are processed to obtain the sonar observation matrix;

[0021] Based on the sonar observation matrix and similarity matrix, an optimized target model is constructed. The image denoising process of the optimized target model is iteratively optimized using the alternating direction multiplier method until the termination condition is met. The final low-rank matrix is ​​then converted into an image format to obtain the denoised sonar monitoring image.

[0022] Furthermore, the expression for optimizing the target model is: ,

[0023] The constraints for optimizing the target model are: ,

[0024] Where D is the sonar observation matrix, C is the low-rank matrix, S is the sparse noise matrix, and A is the similarity matrix. , These are low-rank temporary variables, similarity matrix temporary variables, and sparse noise temporary variables, respectively. For nuclear norm, For weighted norm, It is an L1 norm; , and These are the regularization parameters for the sparse noise term, the low-rank constraint term, and the residual term, respectively.

[0025] Further, step 3 includes:

[0026] The sonar monitoring images after the decision are input into the backbone network of the pre-trained network model for processing, extracting multi-scale features and outputting the highest-level features.

[0027] The original input image is enhanced by using a feature decoupling enhancement network to decouple and enhance its features. The output is a sparse component. The sparse component is upsampled to the same spatial resolution as the highest level feature and then spliced ​​and fused to obtain the enhanced highest level feature.

[0028] Multi-scale features and enhanced highest-level features are input into the neck network of the pre-trained network model for multi-scale feature fusion to obtain neck output features at different levels.

[0029] The neck output features at different levels are input into the corresponding detection head to obtain underwater target detection results at different scales. Among them, the improved ACmix attention module is used to process the highest level neck output features, and the enhanced highest level neck output features are input into the detection head for processing.

[0030] Furthermore, the process of improving the ACmix attention module's feature processing includes:

[0031] The first step is to generate query, key, and value characteristics:

[0032] The input features are passed through three independent convolutional layers to generate query features, key features, and value features, which are denoted as query feature Q, key feature K, and value feature V, respectively.

[0033] The second step is to reshape the query features, key features, and value features:

[0034] The query features, key features, and value features are reshaped according to the number of attention heads and the head dimension, respectively, to obtain the head feature representation.

[0035] The third step involves parallel dual-branch processing, where feature processing is divided into two parallel paths: an attention branch and a convolution branch.

[0036] Attention Branch: The reshaped query features, key features, and value features are converted into a window sequence through a window expansion operation. Each window contains a feature vector containing the window size multiplied by the window size. The attention score is calculated using ScaledDot-Product Attention (SDPA): the query window sequence is multiplied by the transpose of the key window sequence and divided by a scaling factor to obtain the base attention score. If relative position encoding is enabled, positional bias entries are retrieved from the bias table using relative position indices and superimposed on the base attention score, enabling the model to perceive the relative spatial relationships between positions within the window. If relative position encoding is disabled, the base attention score output by the ScaledDot-Product Attention mechanism is directly used, achieving efficient computation through hardware optimization.

[0037] Then, a normalized exponential function is used to normalize the attention weights. The attention weights are multiplied by the value window sequence, and the window features are reconstructed into a spatial feature map with the same resolution as the input through feature folding operations. During the reconstruction process, a sliding strategy with half-window overlap is used to create overlapping regions between adjacent windows. The overlapping regions are then normalized and corrected by counting, i.e., calculating the number of times each pixel is covered by the window, dividing the feature value of the overlapping region by the number of coverages, eliminating the numerical inflation caused by overlap, and outputting the attention branch features.

[0038] Convolutional Branch: The reshaped query features, key features, and value features are concatenated along the channel dimension and then passed through a 1×1 convolutional layer to generate dynamic convolutional kernel parameters. After reshaping and permutation, the dynamic convolutional kernel parameters are passed to a depthwise separable convolutional layer. The depthwise separable convolutional layer adopts a grouped convolution strategy, with the number of groups equal to the head dimension. The dynamically generated parameters are used as the convolutional kernel to perform grouped convolution processing on the input features to obtain convolutional branch features.

[0039] Step 4, Integration and Output:

[0040] The attention branch features and convolution branch features are weighted and fused: the fused features are processed by a random deep regularization layer (DropPath, which randomly discards paths with a certain probability to enhance the model's generalization ability) to obtain the output features.

[0041] Furthermore, this invention constructs an underwater sonar target recognition system that features adaptive triggering denoising and feature decoupling enhancement, comprising:

[0042] The acquisition unit is used to acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment.

[0043] The monitoring unit is used to construct an adaptive triggering decision function, evaluate the sonar monitoring image, and trigger subspace denoising preprocessing based on the decision result to obtain the sonar monitoring image after the decision.

[0044] The detection unit is used to input the sonar monitoring image after decision-making into the pre-trained network model for processing to obtain the underwater target detection result.

[0045] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0046] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] This invention introduces a multi-index quality assessment mechanism based on structural similarity index, peak signal-to-noise ratio, and Laplacian variance, which adaptively judges sonar image quality by triggering a decision function. Subspace denoising preprocessing is only triggered when the image falls below the difference between a threshold and a tolerance in any dimension of structural fidelity, noise level, or edge sharpness. Other high-quality images are directly bypassed and enter the detection process, avoiding the waste of computational resources and excessive smoothing caused by a one-size-fits-all approach, and achieving dynamic matching between preprocessing intensity and actual image quality.

[0049] Unlike general low-rank denoising frameworks for natural images, this invention addresses the physical generation mechanism of marine sonar images. It unifies the modeling of water reverberation and the low-rank structure of the seabed background, as well as the sparsity characteristics of target and particle noise, and explicitly introduces spatially relevant priors about the marine background through a similarity matrix. Compared to the original sonar image, the denoised image shows significantly reduced background clutter, significantly improved structural similarity and peak signal-to-noise ratio, and overall improved visual quality. However, it should be noted that unified denoising may cause over-smoothing of the originally sharp image, leading to loss of fine target edges, manifested as a slight decrease in the strict localization metric (mAP@0.5-0.95). The adaptive triggering mechanism of this invention avoids this problem by selectively bypassing high-quality images, improving average detection accuracy.

[0050] This invention deploys an improved ACmix attention module at the highest level of network features, introduces learnable relative position encoding to dynamically adapt to the positional relationships of targets at different scales, employs hardware-optimized SDPA attention computation to reduce memory usage and accelerate inference, enhances information interaction between windows through half-window overlap and overlap region normalization, and combines DropPath regularization to suppress overfitting across datasets, significantly enhancing the discriminative ability of high-level semantic features and the model's generalization performance, thereby improving the target recognition accuracy in complex underwater environments. Attached Figure Description

[0051] Figure 1 is a flowchart illustrating the underwater sonar image recognition method of the present invention;

[0052] Figure 2 is a schematic diagram of the adaptive triggering decision-making process of the present invention;

[0053] Figure 3 is a schematic diagram of the network model structure of the present invention;

[0054] Figure 4 is a schematic diagram of the improved ACmix attention module of the present invention;

[0055] Figure 5 is a bar chart comparing the performance of the present invention. Detailed Implementation

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0057] Example 1:

[0058] As shown in Figure 1, this embodiment provides an underwater sonar image recognition method that combines adaptive triggering denoising and feature decoupling enhancement, including:

[0059] Step 1: Acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment, specifically including:

[0060] The underwater sonar device is configured as a synthetic aperture sonar detection device, and the synthetic aperture sonar detection device is controlled to continuously emit 300kHz sonar signals in a uniform motion along a straight track in the monitoring area; the 300kHz sonar signal is used for target detection in a first range and a second range, wherein the first range is a long-range detection mode and the second range is a short-range high-resolution imaging mode.

[0061] When performing sonar imaging of targets at multiple synthetic aperture locations on a continuous flight path, an automatic target detection algorithm based on density clustering is used to replace manual annotation. Specifically, the complex image data after synthetic aperture processing is subjected to amplitude detection and local peak extraction. The DBSCAN clustering algorithm is used to spatially cluster the peak points to generate target connected regions, and the target bounding box is automatically drawn. Finally, a sonar monitoring image with annotation information is output.

[0062] Step 2: Construct an adaptive triggering decision function to evaluate the sonar monitoring image and trigger subspace denoising preprocessing based on the decision result to obtain the sonar monitoring image after decision;

[0063] Step 2 specifically includes:

[0064] Calculate the Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Laplacian Variance (LV) of the image.

[0065] Furthermore, the structural similarity index is used to measure the structural similarity between two images, and its formula is: ,

[0066] in, , are the local means of the image x and y, respectively. , For local variance, For covariance, , It is the stability constant.

[0067] In the specific implementation, an 11×11 Gaussian kernel is used to smooth the original image, and the smoothed image is used as the reference image. Then, the statistics within the local window are calculated pixel by pixel, and finally the average value of all pixel positions in the entire image is taken as the SSIM score of the image. The SSIM value ranges from [0,1]. The closer the value is to 1, the better the image structure is preserved and the higher the sharpness; the closer the value is to 0, the more severe the image degradation.

[0068] Furthermore, the peak signal-to-noise ratio (PSNR) is used to evaluate the noise level of an image, and its calculation formula is as follows:

[0069] ,

[0070] Where MaxValue is the maximum value that a pixel can take in the image, and is 255 for an 8-bit grayscale image. MSE is the mean square error between the original image and the reference image, calculated using the following formula:

[0071] ,

[0072] in, , These are two different images, where M and N are the height and width of the image, respectively, and i and j are pixel indexes. PSNR values ​​are measured in decibels (dB). A higher PSNR value indicates a lower level of noise in the image; a lower PSNR value usually indicates a higher level of noise in the image.

[0073] Furthermore, the Laplacian variance is used to detect the degree of edge blurring in sonar images, and its calculation formula is as follows:

[0074] ,

[0075] in, For the Laplacian operator response of the image, This represents the input sonar image, in coordinates Pixel value at that location, and These represent the second-order partial derivatives of the image in the horizontal and vertical directions, respectively. A larger Laplacian variance value indicates sharper image edges and richer details; a smaller value indicates a blurrier image.

[0076] Furthermore, based on the above three quality indicators, an adaptive triggering decision function is constructed as follows:

[0077] Let the sonar detection image to be evaluated be... The reference image is ,but:

[0078] ,

[0079] in, This is an indicator function that outputs 1 if the condition is true, and 0 otherwise. for and The structural similarity index; for and Peak signal-to-noise ratio; for The Laplace variance; , , These are the threshold values ​​for the structural similarity index, peak signal-to-noise ratio, and Laplace variance, respectively. This is the tolerance parameter; Indicates "or", that is Less than With tolerance parameters The difference or Less than With tolerance parameters The difference or Less than With tolerance parameters The difference ensures that any defects in the image, whether in structure, noise, or sharpness, are processed promptly to avoid missed detections. This is an adaptive triggering decision function used to determine whether the input sonar image needs to be preprocessed for denoising.

[0080] The adaptive triggering decision process described above is shown in Figure 2. The technical feature of this decision function is that it adopts an OR logic triggering mechanism. When any one of the SSIM value, PSNR value, or LV value is lower than its corresponding threshold minus the tolerance parameter δ (i.e., ... This means that the image quality is determined to be insufficient for detection and subspace denoising preprocessing is triggered.

[0081] This design ensures that if an image has quality defects in any dimension—structural fidelity, noise level, or edge sharpness—the adaptive trigger decision function outputs 1, triggering subspace denoising preprocessing. This avoids overlooking defects in other dimensions due to a normal performance in one metric. In this embodiment, the tolerance parameter δ is set to 0.05, forming a buffer with a width of 0.05 near the threshold boundary: when the metric is within a certain range... When the value is within the specified range, the original state remains unchanged, avoiding frequent switching caused by fluctuations of the indicator near the threshold, thus enhancing the stability of decision-making.

[0082] The reason for using a unified tolerance parameter is that although the physical meanings and dimensions of the three indicators SSIM, PSNR, and LV are different, they all pass through their respective thresholds. , , Independent quantitative assessment and uniform tolerance δ are applied to each threshold boundary in absolute numerical form, which simplifies parameter configuration, reduces the complexity of parameter tuning during actual deployment, and ensures that the buffer strength of each dimension is consistent at the threshold boundary, avoiding decision bias caused by different tolerances.

[0083] Considering the multipath scattering of sonar beams by ocean water and the non-uniform composite interference caused by suspended particles, this invention decomposes the sonar observation matrix D into a low-rank matrix C and a sparse matrix S. The low-rank matrix C corresponds to the clean sonar image components with low-rank characteristics of the ocean background, consisting of a large area of ​​uniform water and a flat seabed background; the sparse matrix S corresponds to the sparse anomaly components encompassing water reverberation, particle scattering, and equipment thermal noise. Simultaneously, a similarity matrix A constrained by underwater physical laws is embedded to explicitly encode the spatial structural correlation between the seabed sediment layer and the uniform water body. The optimization objective is as follows:

[0084] ,

[0085] The constraints for optimizing the target model are:

[0086] ,

[0087] Where D is the sonar observation matrix, C is the low-rank matrix, S is the sparse noise matrix, and A is the similarity matrix. , These are low-rank temporary variables, similarity matrix temporary variables, and sparse noise temporary variables, respectively. For nuclear norm, For weighted norm, It is an L1 norm; , and These are the regularization parameters for the sparse noise term, the low-rank constraint term, and the residual term, respectively.

[0088] To address the computational burden caused by the high-dimensional data of ocean sonar images, this invention employs the Alternating Direction Method of Multipliers (ADMM) to iteratively optimize the constrained problem, sequentially updating low-rank temporary variables. , sparse noise component S, sparse noise temporary variable Temporary variables of similarity matrix A clean image estimation C, a similarity matrix A, and Lagrange multipliers are used to achieve matrix factorization.

[0089] Unlike traditional ADMM frameworks that introduce additional reconstruction matrices, this invention decouples low-rank, sparsity, and structured constraints into independent subproblems through variable splitting. Each subproblem has a closed-form solution, eliminating the need for nested iterative solutions. This results in faster convergence and lower storage overhead in long-sequence, high-dimensional ocean sonar data. The specific iterative process is as follows:

[0090] The first step is to split the original problem into variables:

[0091] Introducing low-rank temporary variables sparse noise temporary variables Temporary variables of similarity matrix By decoupling low-rank constraints, sparse constraints, and structured constraints, the following constraint conditions are obtained: ,

[0092] The second step is to construct the augmented Lagrange function:

[0093] ,

[0094] Where Y1, Y2, Y3, and Y4 are Lagrange multiplier matrices. For penalty parameters, Represents the matrix inner product. It is the Frobenius norm. For nuclear norm, For weighted norm,

[0095] The third step involves iteratively updating each variable. In each iteration, the algorithm updates each variable sequentially, following the following update rules:

[0096] 1. Update the low-rank temporary variable C1. The corresponding subproblem is:

[0097] ,

[0098] make Singular value decomposition of M1 ,in It is a left singular vector. It is a singular value diagonal matrix. The transpose of the right singular vector, with superscript This indicates that the decomposition corresponds to The update steps involve soft-thresholding the singular values ​​to obtain the updated low-rank temporary variable. : ,

[0099] in, For a soft thresholding operator, for each diagonal element of a singular value diagonal matrix... implement , will be less than the threshold The singular values ​​are set to zero, thereby achieving low-rank constraints.

[0100] 2. Update the sparse noise component S, the corresponding subproblem is:

[0101] ,

[0102] make The solution is obtained directly using element-wise soft thresholding: ; The soft threshold operator is defined as follows: Its function is to shrink the input value x towards zero: when If the difference is retained and the sign is restored, otherwise the value is set to zero. For threshold parameter, here it refers to This operation will consider cases where the absolute value is less than the threshold. The elements are compressed to zero, and only significant outlier components are retained in the sparse noise matrix S.

[0103] This is a sign function that outputs +1 when x is greater than zero, -1 when x is less than zero, and 0 when x is equal to zero. It is used to restore the sign of the value after shrinking. This indicates taking the absolute value of x; Indicates will Subtract threshold If the larger value is taken after zero, Greater than If the difference is not found, then keep the difference; otherwise, output zero directly.

[0104] 3. Update the sparse noise temporary variable S1, the corresponding subproblem is:

[0105] ,

[0106] make Singular value decomposition of M3 ,in It is a left singular vector. It is a singular value diagonal matrix. The transpose of the right singular vector, with superscript This indicates that the decomposition corresponds to The update steps. Soft thresholding shrinkage of singular values: , For a soft thresholding operator, for each diagonal element of a singular value diagonal matrix... implement , will be less than the threshold The singular values ​​are set to zero, thereby constraining the sparse noise temporary variables. Rank.

[0107] 4. Update the temporary variable of the similarity matrix. The corresponding subproblem is:

[0108] ,

[0109] make Singular value decomposition of M1 ,in It is a left singular vector. It is a singular value diagonal matrix. The transpose of the right singular vector, with superscript This indicates that the decomposition corresponds to The update steps involve using weighted nuclear norm minimization to perform weighted soft thresholding shrinkage on singular values: , For weighted soft thresholding operators, singular values Apply weights : Through weight vector The shrinkage intensity is adaptively adjusted for different singular values, preserving important structured information in the similarity matrix A.

[0110] 5. Update the clean image estimate C, the corresponding subproblem is:

[0111] ,

[0112] Expanding both sides of the formula and differentiating with respect to C, we get:

[0113] ,

[0114] in, It is the identity matrix. It is the inverse of the matrix;

[0115] 6. Update the similarity matrix A. The corresponding subproblem is:

[0116] ,

[0117] Expanding and differentiating both sides of the formula, we get... ;

[0118] 7. Update the Lagrange multipliers:

[0119] ,

[0120] ,

[0121] ,

[0122] ,

[0123] The use of the same notation on both sides of the equation for updating Lagrange multipliers is due to the iterative update mechanism. For example, the right side of the equation This represents the old value of the Lagrange multipliers before this round of iteration update, on the left side of the equation. This represents the new value of the Lagrange multipliers after residual correction. For penalty parameters, Represents the corresponding constraint conditions The original residuals are used to measure the deviation between the current variable and the degree to which the constraint is satisfied.

[0124] In each iteration, the algorithm updates the low-rank temporary variable sequentially. , sparse noise component S, sparse noise temporary variable Temporary variables of similarity matrix The clean image is estimated as C, and the similarity matrix is ​​A, yielding the latest variable values ​​for this round. Subsequently, the residuals of each constraint are calculated based on these latest variable values, and the Lagrange multipliers are updated accordingly. , , , .

[0125] Step 3: Input the sonar monitoring image after decision into the pre-trained network model for processing to obtain the underwater target detection results, including target category, bounding box location and confidence score.

[0126] First, the sonar monitoring images after the decision are scaled and normalized: they are uniformly scaled and filled to a size of 640 pixels × 640 pixels, and the pixel values ​​are normalized to the range of [0,1].

[0127] Furthermore, the pre-trained network model includes the following components: a backbone network, a feature decoupling enhancement network, a neck network, an improved ACmix attention module, and a detection head.

[0128] The overall process of the network model is as follows: the sonar monitoring image after image preprocessing is sent to the backbone network to extract multi-scale features; the feature decoupling enhancement network decouples the features of the original input image, outputs sparse components and fuses them with the highest level features of the backbone network to generate enhanced features; the enhanced features and multi-scale features are sent to the neck network for cross-scale fusion; the highest level output of the neck network is enhanced by the improved ACmix attention module and sent to the detection head along with other level features to complete target classification and bounding box regression, and output underwater target detection results at different scales.

[0129] The specific calculation process is as follows:

[0130] The first step is backbone network feature extraction: The preprocessed image is input into the backbone network for processing. The backbone network consists of an initial layer, a first feature extraction layer, a second feature extraction layer, and a third feature extraction layer. A lightweight structure is constructed through convolutional layers and a cross-stage local fusion module to extract multi-scale features sequentially, which are denoted as third-level features, fourth-level features, and fifth-level features, and the highest-level feature, namely the fifth-level feature, is output.

[0131] The second step is feature decoupling enhancement: The feature decoupling enhancement network (DRPCANet) is used to perform feature decoupling enhancement on the original input image, deployed as an independent pre-feature decoupling enhancer outside the backbone network. DRPCANet consists of three parts: an encoder, a low-rank sparse decomposition layer, and a decoder. The encoder extracts multi-scale feature representations of the input image through multiple convolutional layers. The low-rank sparse decomposition layer applies low-rank and sparsity constraints to the encoder output, decomposing the features into background low-rank components and target sparse components. The decoder reconstructs the target sparse component and outputs a single-channel sparse component. This component effectively suppresses background clutter while preserving the target structure. To inject physical priors into the detection process, [the following is implemented / implemented]. The features are upsampled to the same spatial resolution as the fifth-level features through bilinear interpolation, then concatenated with the fifth feature layer, and fused through lightweight convolution to generate the enhanced fifth-level features.

[0132] The third step is multi-scale feature fusion of the neck network: The third-level features, fourth-level features, and enhanced fifth-level features are input into the neck network of the pre-trained network model. The neck network fuses deep semantic features and shallow position features across scales through a bidirectional feature fusion path from top to bottom and from bottom to top, to obtain neck output features of different levels, which are denoted as neck third-level features, neck fourth-level features, and neck fifth-level features in turn.

[0133] The fourth step is to improve the ACmix attention enhancement: the improved ACmix attention module is used to process the fifth-level features of the neck to obtain the enhanced fifth-level features of the neck.

[0134] Fifth step, detection head output: Input the third-level features of the neck into the first detection head to obtain the underwater target detection results at a small scale; input the fourth-level features of the neck into the second detection head to obtain the underwater target detection results at a medium scale; input the enhanced fifth-level features of the neck into the third detection head to obtain the underwater target detection results at a large scale.

[0135] Furthermore, as shown in Figure 4, the specific construction and operation process of the improved ACmix attention module is as follows:

[0136] The first step is to generate query, key, and value characteristics:

[0137] The input features are passed through three independent convolutional layers to generate query features, key features, and value features, which are denoted as query feature Q, key feature K, and value feature V, respectively.

[0138] The second step is to reshape the query feature Q, key feature K, and value feature V:

[0139] The query feature Q, key feature K, and value feature V are reshaped according to the number of attention heads and the head dimension, respectively, to obtain the head-specific feature representation.

[0140] The third step is parallel dual-branch processing:

[0141] Feature processing is divided into two parallel paths: the attention branch and the convolution branch.

[0142] Attention Branch: The split query feature Q, key feature K, and value feature V are transformed into a window sequence (each window contains a feature vector multiplied by the window size) through a window expansion operation. The attention score is calculated using Scaled Dot-Product Attention (SDPA): the query window sequence is multiplied by the transpose of the key window sequence and divided by a scaling factor, which is the head dimension. The square root of the value is used to obtain the basic attention score. Based on this, a learnable relative position bias is introduced: if enabled, the position bias item is retrieved from the bias table through the relative position index and superimposed on the basic attention score, enabling the model to perceive the relative spatial relationship between positions within the window; if disabled, the basic attention score output by the scaling dot product attention mechanism is directly used, and efficient computation is achieved through hardware optimization.

[0143] Then, the attention weights are normalized using the Softmax function, with the specific formula as follows:

[0144] ,

[0145] in, The learnable relative position offset is retrieved from the offset table using the relative position index.

[0146] Subsequently, the attention weights are multiplied by the value window sequence, and the window features are reconstructed into a spatial feature map with the same resolution as the input through a feature folding operation. During the reconstruction process, a sliding strategy with half-window overlap is adopted to create overlapping regions between adjacent windows. A normalized count correction is applied to the overlapping regions, i.e., the number of times each pixel is covered by the window is calculated, and the feature value of the overlapping region is divided by the number of coverages to eliminate numerical inflation caused by overlap, thus outputting the attention branch features. .

[0147] Convolutional Branch: The query feature Q, key feature K, and value feature V generated in the first step are concatenated along the channel dimension and then passed through a 1×1 convolutional layer to generate dynamic convolutional kernel parameters. The dynamic convolutional kernel parameters are reshaped and dimensionally permuted and used as input features, which are then fed into a depthwise separable convolutional layer. The depthwise separable convolutional layer adopts a grouped convolution strategy, with the number of groups equal to the head dimension, to perform grouped convolution processing on the input features and obtain convolutional branch features.

[0148] Step 4: Integration and Output;

[0149] Set learnable parameters rate1 and rate2, both with an initial value of 0.5, and output the weighted fusion of the attention branch output and the convolution branch output: These two parameters are scalar parameters, which are automatically updated during model training through the backpropagation algorithm: the gradient of the fusion weight parameters is calculated based on the detection loss, and the parameter values ​​are adjusted along the gradient descent direction so that the network adaptively learns the optimal fusion ratio of the attention branch and the convolution branch.

[0150] When the stride is greater than 1, the attention branch output is downsampled using average pooling to match the spatial resolution of the convolution branch. The fused features are processed through a random depth regularization (DropPath) layer to enhance the model's generalization ability.

[0151] The improved ACmix attention module incorporates the following key improvements based on the original ACmix module:

[0152] First, a learnable relative position encoding mechanism is introduced. The original ACmix uses fixed absolute position encoding, generating position information through linear coordinates and then mapping it via convolution. This approach cannot adapt to changes in the positional relationships of targets at different scales. This invention introduces a learnable relative position encoding module. By constructing a relative position index matrix and a learnable bias parameter table, the model can dynamically learn the relative positional relationships between targets. This module supports interpolation expansion with arbitrary window sizes. Through bilinear interpolation, the pre-trained large window position bias is transferred to the actual window size used, enhancing the model's adaptability to sonar images at different resolutions.

[0153] Second, the attention calculation method is optimized. The original ACmix manually implemented local window self-attention, calculating attention weights pixel-by-pixel after an expansion operation and using Softmax normalization. This invention introduces a hardware-optimized scaling dot product attention mechanism, utilizing flash memory attention technology to accelerate attention calculation while reducing video memory usage. When relative position encoding is enabled, attention scores are calculated separately and then superimposed with learnable positional biases to maintain fine-grained positional awareness within the local window.

[0154] Third, we add window overlap and overlap region normalization. The original ACmix slides the window with a fixed step size, which limits the information interaction between windows. This invention adopts a half-window overlap sliding strategy to increase the information interaction between adjacent windows. At the same time, after feature folding reconstruction, we add an overlap region normalization operation. By calculating the number of times each pixel is covered, we perform division normalization to eliminate the numerical inflation caused by overlap and improve the accuracy of feature reconstruction.

[0155] Fourth, a random depth regularization mechanism is introduced. This invention adds random depth regularization after the fusion output of the attention branch and the convolution branch, randomly discarding the branch output with a certain probability, forcing the network to learn more robust feature representations, which can effectively alleviate overfitting and improve the model's generalization ability in complex underwater environments.

[0156] In this embodiment, the model is trained and validated on the publicly available underwater sonar image dataset SCTD1.0.

[0157] The model used in this invention is YOLOv8 as the detection framework, combined with the Feature Decoupling Enhancement Network (DRPCANet), to construct an end-to-end underwater sonar target detection model, YOLOv8-DRPOCANet, with the structure as follows: Figure 3 As shown. YOLOv11 is also introduced as a baseline model for comparison and verification of the detection performance of the network in this invention.

[0158] The AdamW optimizer was used for parameter updates during training, with a momentum parameter of 0.937, a weight decay coefficient of 0.0005, and a batch size of 8. The total number of training epochs was set to 300. All input images were uniformly scaled and padded to a size of 640×640 pixels. During training, the total loss function consisted of a weighted average of the bounding box regression loss, classification loss, and distribution focus loss, with corresponding weight coefficients of 7.5, 0.5, and 1.5, respectively.

[0159] To evaluate the impact of preprocessing strategies on detection performance and computational efficiency, this embodiment designs three preprocessing strategies for comparison during the model training and validation phases:

[0160] (1) No noise reduction preprocessing group: No noise reduction processing is performed on the sonar monitoring images. The original images are directly sent to the backbone network and DRPCANet for training and testing.

[0161] (2) All Denoising Group: Subspace denoising is uniformly performed on all sonar monitoring images in the dataset, and then the denoised images are sent to the backbone network and DRPCANet for training and testing.

[0162] (3) Adaptive Denoising Group: Based on the adaptive triggering decision mechanism of the present invention, subspace denoising is performed on the corresponding image only when any one of the SSIM value, PSNR value or LV value is lower than the corresponding threshold minus the tolerance parameter δ; the image that does not trigger the threshold remains unchanged and is directly sent to the subsequent detection process, so as to achieve an adaptive balance between computing resources and detection performance.

[0163] To quantitatively evaluate the detection capabilities of each model on underwater sonar images, this embodiment selects mean precision (mAP), precision (P), recall (R), and F1 score as core evaluation metrics.

[0164] mAP is used to comprehensively evaluate the detection performance of the model. First, the average accuracy (AP) for each class is calculated using the following formula:

[0165] ,

[0166] In the formula, R is the integration variable. This is the precision and recall curve, and AP is the area under the precision and recall curve. Then, the arithmetic mean of the AP values ​​for all classes is taken to obtain mAP, as shown in the following formula: ,

[0167] Where C represents the total number of categories. In this embodiment, mAP@0.5 is used as the core evaluation metric, which means that a correct detection is considered to occur when the intersection-over-union (IoU) between the predicted bounding box and the ground truth bounding box is not less than 0.5.

[0168] This example uses two evaluation metrics:

[0169] (1) mAP@0.5: Fix the IoU threshold to 0.5, that is, when the overlap between the predicted box and the ground truth box is not less than 50%, it is considered a correct detection. This is used to measure the overall performance of the model under relaxed localization requirements.

[0170] (2) mAP@0.5-0.95: Take 10 IoU thresholds in the range of 0.5 to 0.95 with a step size of 0.05, calculate the mAP under each threshold and take the average value to more rigorously evaluate the positioning accuracy of the model.

[0171] Precision (P) and recall (R) characterize model behavior from the dimensions of false positive control and false negative control, respectively. Precision is defined as the proportion of true positive samples out of all positive predictions output by the model, and its mathematical expression is:

[0172] ,

[0173] TP represents the number of correctly identified target instances, reflecting the number of accurately identified targets; FP represents the number of false alarm instances where the background area is incorrectly identified as a target.

[0174] Recall measures the proportion of all true targets that are successfully detected by the model, and is calculated as follows: ,

[0175] In the formula, FN represents the number of missed detection instances that actually exist but were not detected by the model.

[0176] The F1 score balances the accuracy and recall of a classification model, evaluating the model's overall performance in terms of both. The calculation formula is as follows: ,

[0177] Table 1 Performance Comparison

[0178]

[0179] As shown in Table 1, YOLOv8-DRPCANet exhibits significant differences under different preprocessing strategies. The introduction of subspace denoising results in a substantial leap in the detection capability of YOLOv8-DRPCANet on sonar images. When the raw, noisy image is directly input into the network without preprocessing, the ocean background clutter and sensor noise severely interfere with the feature decoupling process, resulting in a mAP@0.5 of only 81.8%, and low precision, recall, and F1 scores.

[0180] When subspace denoising was applied to all samples, the detection quality improved: mAP@0.5 jumped from 81.8% to 91.9%, an increase of 10.1 percentage points, while precision and F1 improved to 89.6% and 84.4%, respectively. This validates the ability of low-rank sparse decomposition to suppress non-uniform noise in oceans. However, the mAP@0.5-0.95 of the entire denoising strategy decreased to 57.7%, a reduction of 0.3 percentage points compared to the no-denoising group, revealing the drawbacks of uniform preprocessing: excessive smoothing of clear images and loss of weak target localization information.

[0181] The adaptive triggering strategy achieves the optimal balance between denoising gains and preservation of original information. This mechanism, based on a decision function, only denoises low-quality sonar images below a threshold boundary, while the remaining images are passed directly to the detection network in their original state. Under this strategy, mAP@0.5 reaches 91.3%, essentially on par with all denoising strategies; while mAP@0.5-0.95 rebounds to 62.8%, an improvement of 5.1 percentage points compared to all denoising strategies and 4.8 percentage points compared to no denoising strategy. More importantly, the recall (85.0%) and F1 score (88.3%) are both at their highest levels, with recall improving by 8.0 percentage points compared to YOLOv11n, F1 score improving by 4.4 percentage points compared to YOLOv11n, and improving by 3.9 percentage points compared to all denoising strategies. This result demonstrates that adaptive denoising leverages the advantage of subspace denoising in suppressing complex ocean noise while avoiding excessive damage to high-quality sonar images caused by a one-size-fits-all approach, achieving the optimal trade-off between denoised image quality and preservation of original information.

[0182] Figure 5 is a bar chart comparing the performance of each model provided in this embodiment on the test set. From left to right, the figures show four configurations: YOLOv11n baseline model, YOLOv8-DRPCANet (without denoising), YOLOv8-DRPCANet (fully denoised), and YOLOv8-DRPCANet (adaptive denoising). The vertical axis represents the percentage value (%) of each evaluation metric, and the horizontal axis represents different model strategies. Specifically, the diagonally filled bars represent precision, the back-diagonally filled bars represent recall, the grid-filled bars represent F1 scores, gray bars represent mAP@0.5, and dotted filled bars represent mAP@0.5-0.95. Figure 5 As can be seen, the adaptive denoising strategy achieves the best results in terms of precision, recall, and F1 score, verifying the effectiveness and robustness of the adaptive preprocessing mechanism on the test set.

[0183] In addition, to verify the impact of the proposed improved ACmix module on sonar target detection performance, this embodiment conducts comparative experiments on the SCTD1.0 dataset (denoted as test1) and the LASA dataset (denoted as test2), respectively, comparing the baseline model (YOLOv8-DRPCANet) and the model with ACmix attention enhancement (YOLOv8-DRPCANet-ACmix). The experiments maintain the same training settings as in Example 1.

[0184] Table 2 Performance comparison of Baseline and ACmix augmented models on different datasets

[0185]

[0186] As shown in Table 2, on the SCTD1.0 test set, the overall detection accuracy improved after introducing the ACmix module: mAP@0.5 increased from 81.8% to 83.9%, mAP@0.5-0.95 increased from 58.0% to 60.4%, and recall increased from 74.8% to 79.6%. Meanwhile, precision decreased from 83.4% to 76.5%, and the F1 score slightly decreased from 78.9% to 78.0%. These results indicate that the ACmix module effectively reduced the false negative rate by enhancing feature discrimination capabilities, but introduced a small number of false positives in some high-confidence background regions, leading to a slight decrease in precision.

[0187] On the LASA test set, the performance gains from the ACmix module were even more significant. Overall mAP@0.5 improved from 61.5% to 66.5%, mAP@0.5:0.95 improved from 41.8% to 48.1%, recall increased dramatically from 47.5% to 77.6%, and F1 score improved from 54.9% to 59.5%. Similar to test1, precision decreased from 64.9% to 48.2%, a noticeable drop.

[0188] Combining the experimental results of test1 and test2, the ACmix module achieved positive growth in the mAP metric on the test sets of both datasets, verifying its effectiveness on sonar images with different distributions. In particular, the LASA dataset, due to its larger target scale variations and more complex background clutter, showed a significantly higher performance improvement than SCTD1.0, indicating that the module has stronger feature decoupling and enhancement capabilities for highly complex sonar images.

[0189] A trade-off between significantly improved recall and decreased precision was observed on both datasets. This stems from the fact that while the ACmix attention mechanism enhances feature discriminativeness, making the model more sensitive to low-contrast, weak targets, it also amplifies some seabed clutter and sidelobe interference similar to target features, leading to an increase in low-confidence false alarms. Furthermore, the learnable relative position encoding is insufficiently adapted to the target scale distribution across different datasets during cross-dataset transfer, introducing additional localization bias in the large-scale variation scenarios of the LASA dataset, further reducing precision. Future solutions could address this issue by enhancing the fidelity constraint for low-contrast targets in the adaptive triggering decision function, or by introducing a dynamic threshold mechanism to adjust the fusion weights rate1 and rate2 of the attention branch based on dataset characteristics.

[0190] This embodiment also provides an underwater sonar image recognition system based on adaptive processing and feature decoupling enhancement, including: an acquisition unit, a monitoring unit, and a detection unit;

[0191] In this embodiment, the acquisition unit is used to acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment;

[0192] In this embodiment, the monitoring unit is used to construct an adaptive triggering decision function, perform decision evaluation on the sonar monitoring image, and trigger subspace denoising preprocessing based on the decision result to obtain the sonar monitoring image after decision.

[0193] In this embodiment, the detection unit is used to input the sonar monitoring image after decision-making into the pre-trained network model for processing to obtain the underwater target detection result.

[0194] The aforementioned system can implement the underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement described in this invention. However, the implementation device of the underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement described in this invention includes, but is not limited to, the underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement described in this invention.

[0195] Example 2:

[0196] This invention provides an electronic device, including a processor and a memory. The memory stores program instructions for implementing the adaptive triggering denoising and feature decoupling enhancement underwater sonar image recognition method of Embodiment 1. The processor executes the program instructions stored in the memory to perform underwater sonar image recognition.

[0197] The processor can also be called a Central Processing Unit (CPU). A processor may be an integrated circuit chip with signal processing capabilities. A processor can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0198] The methods described in Embodiment 1 of this invention are applicable to this electronic device embodiment. The specific functions implemented by this electronic device embodiment are the same as those in the above method embodiment, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0199] Example 3:

[0200] This embodiment provides a storage medium storing a computer program or instructions capable of implementing the above-described adaptive triggering denoising and feature decoupling enhancement underwater sonar image recognition method. When the computer program or instructions are run on a computer, the steps of the above-described adaptive triggering denoising and feature decoupling enhancement underwater sonar image recognition method are executed.

[0201] The computer program or instructions can be stored in the aforementioned storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0202] The methods described in Embodiment 1 of this invention are all applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0203] This invention provides an underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. An underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement, characterized in that, Includes the following steps: Step 1: Acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment; Step 2: Construct an adaptive triggering decision function to evaluate the sonar monitoring image and trigger subspace denoising preprocessing based on the decision result to obtain the sonar monitoring image after decision; Step 3: Input the sonar monitoring image after decision-making into the pre-trained network model for processing to obtain the underwater target detection result.

2. The underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement according to claim 1, characterized in that, Step 2 includes: Calculate multiple quality indicators based on sonar monitoring images and reference images; Based on the quality indicator threshold and tolerance threshold, an adaptive triggering decision function is constructed, and the quality indicator is input into the adaptive triggering decision function to obtain the decision result. When the decision result is 1, subspace denoising preprocessing is performed on the sonar monitoring image; when the decision result is 0, the sonar monitoring image remains unchanged. The final output is the sonar monitoring image after the decision is made.

3. The underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement according to claim 2, characterized in that, The adaptive triggering decision function is expressed as follows: Let the sonar detection image to be evaluated be The reference image is ,but: ， in, This is an indicator function that outputs 1 if the condition is true, and 0 otherwise. for and The structural similarity index; for and Peak signal-to-noise ratio; for The Laplace variance; , , These are the thresholds for structural similarity index, peak signal-to-noise ratio, and Laplace variance, respectively. For tolerance parameters, Indicates or; This is an adaptive triggering decision function.

4. The underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement according to claim 2, characterized in that, The subspace denoising preprocessing includes: The sonar monitoring images are processed to obtain the sonar observation matrix; An optimized target model is constructed based on the sonar observation matrix and the similarity matrix; The image denoising process of the target model is iteratively optimized using the alternating direction multiplier method until the termination condition is met. The resulting low-rank matrix is ​​then converted into an image format to obtain the denoised sonar monitoring image.

5. The underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement according to claim 4, characterized in that, The expression for optimizing the target model is: ， The constraints for optimizing the target model are: ， Where D is the sonar observation matrix, C is the low-rank matrix, S is the sparse noise matrix, and A is the similarity matrix. , , These are low-rank temporary variables, similarity matrix temporary variables, and sparse noise temporary variables, respectively. For nuclear norm, For weighted norm, It is an L1 norm; , and These are the regularization parameters for the sparse noise term, the low-rank constraint term, and the residual term, respectively.

6. The underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement according to claim 1, characterized in that, Step 3 includes: The sonar monitoring images after the decision are input into the backbone network of the pre-trained network model for processing, extracting multi-scale features and outputting the highest-level features. The original input image is enhanced by using a feature decoupling enhancement network to decouple and enhance its features. The output is a sparse component. The sparse component is upsampled to the same spatial resolution as the highest level feature and then spliced ​​and fused to obtain the enhanced highest level feature. Multi-scale features and enhanced highest-level features are input into the neck network of the pre-trained network model for multi-scale feature fusion to obtain neck output features at different levels. The neck output features at different levels are input into the corresponding detection head to obtain underwater target detection results at different scales. Among them, the improved ACmix attention module is used to process the highest level neck output features, and the enhanced highest level neck output features are input into the detection head for processing.

7. The underwater sonar image recognition method with adaptive triggering denoising and feature decoupling enhancement according to claim 6, characterized in that, The process of improving the ACmix attention module's feature processing includes: The input features are passed through three independent convolutional layers to generate query features, key features, and value features respectively; the query features, key features, and value features are then reshaped to obtain the reshaped query features, key features, and value features. The reshaped query features, key features, and value features are expanded into a window sequence. Position bias items are retrieved from the learnable bias table using relative position indexes and superimposed on the scaled dot product attention score. After Softmax normalization, the attention weights are calculated. The attention weights are multiplied by the value window sequence, and the overlapping regions are normalized and counted to correct for the output attention branch features. The reshaped query features, key features, and value features are concatenated along the channel dimension and then used to generate dynamic convolution kernel parameters through a 1×1 convolutional layer. After being reshaped and dimensionally replaced again, they are input into a depth-separable convolutional layer for grouped convolution processing to obtain convolutional branch features. The attention branch features and convolution branch features are weighted and fused, and the fused features are processed through a random depth regularization layer to obtain the output features.

8. An underwater sonar image recognition system with adaptive triggering denoising and feature decoupling enhancement, characterized in that, include: The acquisition unit is used to acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment. The monitoring unit is used to construct an adaptive triggering decision function, evaluate the sonar monitoring image, and trigger subspace denoising preprocessing based on the decision result to obtain the sonar monitoring image after the decision. The detection unit is used to input the sonar monitoring image after decision-making into the pre-trained network model for processing to obtain the underwater target detection result.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the adaptive triggering denoising and feature decoupling enhancement underwater sonar image recognition method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The device stores a computer program or instructions that, when run on a computer, execute the steps of the underwater sonar image recognition method as described in any one of claims 1 to 7, characterized in that the adaptive triggering denoising and feature decoupling enhancement method is used.

Images