An underwater side-scan sonar image recognition method and system based on a double attention mechanism

Abstract

The application discloses a kind of based on dual attention mechanism's underwater side-scan sonar image recognition method and system, comprising: obtaining the monitoring image of underwater side-scan sonar equipment collection;Monitoring image is input to the EDA-Net model of pre-set and obtains underwater target detection result, wherein, EDA-Net model includes improved backbone network, improved neck network and head network;The improved backbone network is based on YOLOv11 framework construction, wherein double path heteromeric attention module is set;Improved neck network is provided with bottleneck convolution-parameterless attention module;The head network is used for target detection according to the feature map output by improved neck network.This application can more accurately identify target detail texture characteristics under the condition of small sample, multi-scale data set, improve the recognition accuracy and stability of model to multi-scale target in complex underwater environment.

Description

Technical Field

[0001] This invention relates to an underwater side-scan sonar image recognition method and system, and more particularly to an underwater side-scan sonar image recognition method and system based on a dual attention mechanism. Background Technology

[0002] Side-scan sonar (SSS) possesses wide coverage and high-resolution scanning capabilities. Leveraging this characteristic, SSS sonar has been widely applied in critical fields such as underwater target search and rescue, marine geological exploration, and underwater structural inspection. However, the unique acoustic physics of SSS images and the complex underwater scenes present extreme challenges to target identification, significantly increasing the difficulty, especially in tasks involving the localization and identification of targets at multiple scales with limited samples.

[0003] With the development of deep learning, detection frameworks based on convolutional neural networks have been widely applied in the field of object recognition, and can be divided into two categories: two-stage methods and one-stage methods. Two-stage methods first generate candidate regions, and then classify and regress their locations. Typical methods include Faster R-CNN and improved models based on feature pyramid networks. In contrast, one-stage object detection methods directly predict the object category and location from the input image, resulting in faster detection speeds. Typical methods include You Only Look Once (YOLO) and Detachment Transformer (DETR). In the field of sonar object detection, researchers have conducted numerous improvements on the one-stage detection framework.

[0004] However, existing models still have some shortcomings. First, targets in side-scan sonar images often exhibit scale variations and geometric distortions. Existing methods lack sufficient multi-scale feature representation and robust detection capabilities. Second, the quality of sonar imaging fluctuates significantly, which further weakens the feature learning ability of deep learning models and easily leads to overfitting. Finally, most existing models rely on large-scale datasets for training, making it difficult to fully learn the essential features and discriminative representations of targets in small-sample scenarios. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to propose an underwater side-scan sonar image recognition method and system based on a dual attention mechanism, which can more accurately identify the detailed texture features of targets under small sample and multi-scale dataset conditions, and improve the accuracy and stability of the model in recognizing multi-scale targets in complex underwater environments.

[0006] Technical solution: This invention includes the following steps:

[0007] Acquire monitoring images from underwater side-scan sonar equipment;

[0008] The monitoring images are input into a preset EDA-Net model to obtain underwater target detection results. The EDA-Net model includes an improved backbone network, an improved neck network, and a head network.

[0009] The improved backbone network is built on the YOLOv11 framework and includes a dual-path heterogeneous attention module. The dual-path heterogeneous attention module is used to divide the input feature map into a lightweight branch and an enhancement branch in the channel dimension according to a preset ratio. The lightweight branch extracts local information through depthwise convolution and pointwise convolution. The enhancement branch extracts deep information through stacked convolutional layers, and then performs joint enhancement through channel attention mechanism and spatial attention mechanism in sequence, and then merges with the lightweight branch for output.

[0010] The improved neck network includes a bottleneck convolution-parameterless attention module. This module allows the input features to be initially enhanced by a parameterless attention mechanism, progressively enhanced layer by layer by multi-level bottleneck convolutional blocks, and dynamically fused with learnable weights before output. The parameterless attention mechanism generates attention weights based on the channel statistical bias, spatial local contrast, and gradient information of the input feature map, and the calculation of these attention weights does not introduce learnable parameters, expressed as follows:

[0011]

[0012] in, Indicates the input feature map, This represents the output feature map; Channel deviation; Indicates local contrast; This represents the gradient response calculated using spatial difference. This represents element-wise multiplication; and These represent the mean and standard deviation, respectively. It is a constant;

[0013] The head network is used for target detection based on the feature map output by the improved neck network.

[0014] In the dual-path heterogeneous attention module, the channel dimension division ratio of the lightweight branch and the enhancement branch is 1:3; the enhancement branch sequentially passes through two stacked DPC layers to extract deep information.

[0015] The improved neck network includes an upsampling part and a downsampling part; the upsampling part includes an upsampling module, a splicing module and a bottleneck convolution-parameter-free attention module connected in sequence; the downsampling part includes a bottleneck convolution-parameter-free attention module, a splicing module and a CBS module connected in sequence.

[0016] The upsampling module and splicing module of the upsampling part are sequentially connected to the dual-path heterogeneous attention module of the second feature extraction layer, the third feature extraction layer and the fourth feature extraction layer of the backbone network.

[0017] The splicing module and CBS module of the downsampling part are connected to the bottleneck convolution-parameterless attention module corresponding to the upsampling part and the dual-path heterogeneous attention module of the fourth feature extraction layer of the backbone network.

[0018] The DPC layer consists of a 3×3 depthwise convolution, a 1×1 pointwise convolution, and a 1×1 projection layer.

[0019] The spatial attention mechanism extracts the average response map, the maximum response map, and the edge response map generated by learnable convolution from the input features. After concatenating the three, it performs multi-scale modeling using convolutional kernels with multiple receptive fields, adaptively fuses them through learnable scale weights, and outputs them after generating spatial weights through Sigmoid activation.

[0020] The present invention also provides a recognition system based on the aforementioned dual-attention mechanism for underwater side-scan sonar image recognition, comprising:

[0021] The detection result output unit receives the sonar image output by the underwater side-scan sonar device, sends the sonar image into the preset EDA-Net model, and outputs the corresponding underwater target detection result.

[0022] The model building unit obtains the EDA-Net model based on the improved backbone network, the improved neck network, and the head network of the YOLOv11 network model;

[0023] The sample acquisition unit collects multiple target side-scan sonar training images in a preset underwater environment, preprocesses the acquired training images, and adds real annotation information to construct training samples.

[0024] The model training unit trains the EDA-Net model based on training samples. It uses classification loss function and bounding box regression loss function to calculate the loss value between the training output and the true label. The model weights are iteratively updated according to the loss value until the preset termination condition is reached, and then the trained EDA-Net model is obtained.

[0025] The preprocessing in the sample acquisition unit includes: randomly rotating the sonar training image by ±15°, randomly flipping it horizontally and vertically, and randomly cropping and scaling the sonar training image to obtain training samples.

[0026] The preset termination condition in the model training unit is: the accuracy of the EDA-Net model exceeds 90% or the average accuracy exceeds 70%.

[0027] Beneficial effects: The present invention has the following advantages:

[0028] (1) In this invention, a dual-path heterogeneous attention mechanism is constructed in the backbone network. This mechanism uses a channel-splitting heterogeneous convolution strategy to extract local spatial features with lightweight convolution branches and enhances convolution branches to capture global dependencies using improved channel attention and spatial attention. This multi-dimensional feature extraction strategy ensures that the model can fully learn the essential attributes and discriminative representations of the target, and significantly alleviates the problem of feature loss and insufficient learning caused by information scarcity in small sample scenarios.

[0029] (2) This invention utilizes a strategy of combining bottleneck convolution and parameterless channel-space attention mechanism in the neck network. This strategy does not require the introduction of additional learnable parameters and can achieve adaptive weighting of channel dimension and spatial position based solely on the statistical information of feature maps. This design reduces model complexity and overfitting risk while improving the model's ability to model cross-scale robust features, thereby enhancing the generalization performance of feature representation. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention;

[0031] Figure 2 This is a structural diagram of the EDA-Net model of the present invention;

[0032] Figure 3 This is a structural diagram of the BC-PFAM module of the present invention. Detailed Implementation

[0033] The invention will now be further described with reference to the accompanying drawings.

[0034] Example 1

[0035] like Figure 1 As shown, the underwater side-scan sonar image recognition method based on the dual attention mechanism in this embodiment includes the following steps:

[0036] S1. Acquire sonar monitoring images of the monitoring area collected by the underwater sonar equipment, and annotate the acquired side-scan sonar images. The annotation content includes:

[0037] The targets in the image are classified and labeled as mine-like objects and non-mine seabed objects.

[0038] The target is labeled with a bounding box, and the target area is identified by a rectangle. The bounding box is represented in a normalized way, including: the center coordinates (x, y) of the bounding box, the width (w) of the bounding box, and the height (h) of the bounding box.

[0039] The underwater sonar equipment is a side-scan sonar detection device. The operating frequency range of the side-scan sonar detection device is set from 900 kHz to 1800 kHz. It obtains high-quality side-scan sonar image data by continuously scanning the target area to obtain high-resolution underwater images.

[0040] S2. Input the sonar monitoring images into the preset EDA-Net model to obtain the underwater target detection results. For example... Figure 2 The construction and training process of the EDA-Net model shown includes:

[0041] S21. Construction of the EDA-Net model: The EDA-Net model is constructed based on the YOLOv11 framework, a dual-path heterogeneous attention module, and a bottleneck convolution-parameterless attention module. Specifically, this includes:

[0042] S211. An improved backbone network is obtained by combining a YOLOv11-based backbone network with a Two-Path Heterogeneous Attention Module (DHAM). The improved backbone network includes an initial layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer. The initial layer is set as a CBS module. The DHAM module is constructed based on the two-path heterogeneous attention mechanism, with the first, second, third, and fourth feature extraction layers sequentially connected to the CBS module and the DHAM module. After processing by the CBS module, low-level features, such as edges and textures, are extracted from the sonar image. Then, the DHAM module enhances the feature extraction of the sonar image.

[0043] The specific process of feature extraction in the DHAM module includes: dividing the input feature map into two sub-feature branches in a 1:3 ratio along the channel dimension; the lightweight branch passes through a DPC layer consisting of a 3×3 depthwise convolution, a 1×1 pointwise convolution, and a 1×1 projection layer to extract local information; the enhancement branch first passes through two stacked DPC layers to extract deep information, then passes through an enhanced channel attention mechanism and an edge-aware pyramid spatial attention mechanism to jointly enhance the features, then passes through two sets of depthwise separable convolutions for feature fusion, and finally adds the feature to the lightweight branch to obtain the final output.

[0044] The enhanced channel attention mechanism first extracts channel statistics using global average pooling and global max pooling, and then captures local cross-channel dependencies through one-dimensional convolution. Subsequently, the two statistical responses are fused in a weighted manner, and channel weights are generated through sigmoid activation. Finally, residual scaling is used, i.e. This yields the channel attention output.

[0045] The edge-aware pyramid spatial attention mechanism first extracts the average response map, maximum response map, and edge response map generated through learnable convolution from the input features, and then concatenates these three maps using Concat. Subsequently, it performs multi-scale modeling using convolutional kernels with multiple receptive fields, and adaptively fuses them using learnable scale weights. Finally, it generates spatial weights through sigmoid activation. Ultimately, spatial attention output is generated. .

[0046] By introducing the Dual-Path Heterogeneous Attention Module (DHAM), the problem of feature loss caused by data scarcity in small sample scenarios is significantly alleviated, ensuring the comprehensiveness and effectiveness of discriminative feature information learning. The Dual-Path Heterogeneous Attention Module (DHAM) can accurately capture the key discriminative features of small sample targets from multiple dimensions, thereby better improving the model's recognition accuracy.

[0047] S212. An improved neck network is obtained by combining a neck network based on the YOLOv11 network model with a bottleneck convolutional-parameterless attention module (BC-PFAM). The improved neck network consists of an upsampling part and a downsampling part. The upsampling part includes an upsampling module (Upsample), a concatenation module (Concat), and a BC-PFAM module. The upsampled features of each layer are fused with the corresponding low-level features to construct a top-down feature pyramid, achieving initial fusion of multi-scale features. The Concat module and the Upsample module are sequentially connected to the DHAM modules of the second, third, and fourth feature extraction layers of the backbone network. The downsampling part includes a BC-PFAM module, a Concat module, and a CBS module. Downsampling starts from the bottom-level features, passing low-level features layer by layer into high-level features to further enrich multi-scale features. The CBS module and the Concat module are sequentially connected to the BC-PFAM module of the upsampling part and the DHAM module of the fourth feature extraction layer of the backbone network.

[0048] like Figure 3As shown, the specific processing steps of the BC-PFAM module include: First, the input features undergo preliminary enhancement through a parameterless channel-spatial joint attention mechanism to obtain shallow enhanced features. Then, layer-by-layer feature enhancement is performed through two consecutive bottleneck convolutional blocks of "pointwise convolutional dimensionality reduction—depth convolution—pointwise convolutional dimensionality increase," with an attention mechanism applied again after each level to improve the response of key regions and effective channels, resulting in mid-layer and deep enhanced features, respectively. To fully utilize the complementarity of features at different stages, the module uses learnable parameters to generate normalized fusion weights. It uses three weight coefficients, w0, w1, and w2, to dynamically fuse the three layers of features, where w0 is the weight coefficient for the shallow enhanced features; w1 is the weight coefficient for the second enhanced features; and w2 is the weight coefficient for the third enhanced features. Finally, a 1×1 projection layer is used for channel adjustment and output mapping.

[0049] The parameter-free attention mechanism achieves adaptive feature recalibration by fusing channel statistical bias with spatial local contrast and gradient information. Specifically, it first obtains channel descriptors through global average pooling and calculates their deviation from the overall mean to construct channel importance. Simultaneously, it constructs contrast responses using local average pooling in the spatial dimension and extracts gradient information using a fixed-difference operator to characterize structural changes. Finally, it standardizes and fuses the responses from each branch, generates the final attention weights using the sigmoid function, reweights the input features element-wise, and outputs the residuals. Its overall expression is as follows:

[0050]

[0051] in, Indicates the input feature map, This represents the output feature map, where C, H, and W represent the number of channels, height, and width, respectively. Channel deviation, This represents the channel descriptor obtained through global average pooling. For channel descriptors The mean; Indicates local contrast. Indicates to Perform local average pooling; This represents the gradient response calculated by spatial difference, i.e. , and These are the features after being translated one position horizontally and vertically, respectively. This represents element-wise multiplication; and These represent the mean and standard deviation, respectively. To prevent numerically unstable constants.

[0052] By introducing a bottleneck convolution-parameterless attention module, it can perform adaptive weighting processing based on the channel dimension and spatial location of the feature map, thereby significantly enhancing the model's feature representation ability and generalization performance. While reducing model complexity and the risk of overfitting, the EDA-Net model can still accurately identify targets, achieving a balance between model size and accuracy.

[0053] S213. The EDA-Net model is constructed by building a head network based on an improved backbone network, an improved neck network, and a YOLOv11 network model. The head network receives feature maps output from the BC-PFAM module in the neck network. First, the head network predicts the class of each grid cell in the feature map, outputting the possible target classes and their probability distributions for each grid cell. Next, for each grid cell, the head network performs bounding box regression, calculating the position and size parameters of the box. Furthermore, the head network calculates a confidence score for each predicted box to indicate whether the box contains a target. A high confidence score indicates that the box contains a target, while a low confidence score indicates that the box does not contain a target. Through this series of steps, the head network can transform the fused feature maps into complete target detection results, providing accurate information for the final target recognition task.

[0054] S22. Training the EDA-Net model: Acquire sonar training images of targets under various underwater scenarios. Preprocess the sonar training images and add real labels to obtain training samples. Use the training samples to train the EDA-Net model to obtain training recognition results. Calculate the training loss value between the training recognition results and the real labels using the YOLOv11 loss function. The calculation formula is as follows:

[0055]

[0056] in, The training loss value is the difference between the training recognition result and the real label; For classification loss value, For bounding box regression loss, For confidence loss; , and These represent the weight coefficients of classification loss, bounding box regression loss, and confidence loss, respectively.

[0057] The weight parameters of the EDA-Net model are optimized based on the training loss value. The training process of the EDA-Net model is repeated iteratively until the iteration termination condition is met. Then, the final weights of the trained EDA-Net model are output for inference on the test set. In this embodiment, the iteration termination condition is: the accuracy of the EDA-Net network model exceeds 90% or the mean average precision (mAP) exceeds 70% when tested on the validation set.

[0058] The preprocessing of sonar training images includes: denoising the sonar training images using a median filtering algorithm, where the sliding window size is set to 3×3 or 5×5. A 5×5 median filter is used when the image noise is strong, while a 3×3 median filter is used when it is necessary to preserve target edge details; the sonar training images are randomly rotated ±15°; the sonar training images are randomly flipped horizontally and vertically; and the sonar training images are randomly cropped and scaled to obtain training samples.

[0059] This embodiment is trained on the publicly available MILCO dataset. A Stochastic Degree Descent (SGD) optimizer with specific settings is employed. The SGD optimizer's momentum is set to 0.949, weight decay to 0.0005, and batch size to 64. The number of iterations is set to 800 epochs, and the initial learning rate is set to 0.01.

[0060] The EDA-Net, YOLOv10n, and YOLOv11n models were compared with the RT-DETR-L model, and the mean average precision (mAP) was used as the evaluation metric. mAP represents the average accuracy of all target classes under a fixed cross-merge ratio. mAP@0.5 represents the average accuracy of all target classes at a fixed cross-merge ratio threshold of 0.5. mAP@0.5:0.95 represents the average mAP calculated at multiple IoU thresholds (from 0.5 to 0.95, with a step size of 0.05).

[0061] AP is the area under the precision-recall curve, reflecting the average precision of a single class. The formula is:

[0062]

[0063] in, Indicates recall rate, It is the precision and recall curve.

[0064] The average AP value across all categories is used to obtain mAP, calculated using the following formula:

[0065]

[0066] in, Indicates the number of target categories; This represents the category index, used to represent the first... One category; Indicates the first Average precision for each category.

[0067] In addition, the number of parameters, model size, and GFLOPs are other metrics for evaluating a model. Specifically, model size refers to the space the model occupies on storage devices, parameters refer to the total number of weights and biases in the model, and GFLOPs are the total number of floating-point operations performed by the model during a single forward propagation. These metrics are commonly used to assess model complexity.

[0068] Table 1 Performance comparison of this embodiment with other target detection algorithms

[0069]

[0070] As shown in Table 1, EDA-Net performs excellently across all metrics. Firstly, in terms of detection accuracy, EDA-Net's mAP@0.5 is 72.3%, significantly higher than YOLOv10n's 63.0%, YOLOv11n's 58.6%, and RT-DETR-L's 52.8%. This indicates that EDA-Net possesses stronger target localization and recognition capabilities in target detection tasks. Furthermore, in the more stringent evaluation metric of mAP@0.5:0.95, EDA-Net achieves 46.3%, again significantly outperforming YOLOv10n's 39.5%, YOLOv11n's 35.9%, and RT-DETR-L's 21.9%. This demonstrates that EDA-Net maintains strong robustness and generalization ability even under multiple threshold averaging conditions.

[0071] In terms of model complexity and deployment cost, EDA-Net also demonstrates a strong advantage. Its parameter count is only 1.95M, and its GFLOPs are 6.0, lower than other comparative methods, indicating that EDA-Net maintains high accuracy without incurring additional computational burden. Meanwhile, EDA-Net's model size is only 4.2M, the smallest among the four models, suggesting better deployment potential in underwater scenarios with limited storage resources.

[0072] Example 2

[0073] The underwater side-scan sonar image recognition system based on the dual attention mechanism in this embodiment includes:

[0074] The detection result output unit is used to acquire side-scan sonar monitoring images of the monitoring area collected by the underwater sonar equipment, and input the sonar monitoring images into the preset EDA-Net model to obtain the underwater target detection results;

[0075] The model building unit constructs the EDA-Net model based on the YOLOv11 framework, a dual-path heterogeneous attention module, and a bottleneck convolution-parameterless attention module; specifically, it includes:

[0076] An improved backbone network is obtained by combining a YOLOv11 network model backbone network with a dual-path heterogeneous attention module (DHAM). The improved backbone network includes an initial layer, a first feature extraction layer, a second feature extraction layer, a third feature extraction layer, and a fourth feature extraction layer; the initial layer is set as a CBS module; the DHAM module is constructed based on the dual-path heterogeneous attention mechanism, and the first, second, third, and fourth feature extraction layers are sequentially connected to the CBS module and the DHAM module.

[0077] An improved neck network is obtained based on the YOLOv11 network model and the bottleneck convolution-parameterless attention module (BC-PFAM). The improved neck network consists of upsampling and downsampling parts. The bottleneck convolution and parameterless attention mechanism constitute the BC-PFAM module. The upsampling part includes an upsampling module, a concatenation module, and the BC-PFAM module. The upsampled features of each layer are fused with the corresponding low-level features to construct a top-down feature pyramid, achieving initial multi-scale feature fusion. The concatenation module and the upsampling module are sequentially connected to the DHAM modules of the second, third, and fourth feature extraction layers of the backbone network. The downsampling part includes the BC-PFAM module, the concatenation module, and the CBS module. Downsampling starts from the bottom-level features, progressively passing low-level features to higher-level features to further enrich multi-scale features. The CBS module and the concatenation module connect the BC-PFAM module of the upsampling part to the DHAM module of the fourth feature extraction layer of the backbone network. The neck network is responsible for fusing and enhancing features at different levels.

[0078] The EDA-Net model is constructed based on the improved backbone network, the improved neck network, and the head network of the YOLOv11 network model.

[0079] The sample acquisition unit is used to acquire side-scan sonar training images of monitoring targets under various set underwater scenarios, preprocess the sonar training images and add real labels to obtain training samples.

[0080] The model training unit is used to train the EDA-Net model using training samples and obtain training recognition results. It uses the YOLOv11 loss function to calculate the training loss value between the training recognition result and the real label. Based on the training loss value, it optimizes the weight parameters of the EDA-Net model and repeats the training process of the EDA-Net model until the iteration termination condition is met, and then outputs the trained EDA-Net model.

Claims

1. A method for underwater side-scan sonar image recognition based on a dual attention mechanism, characterized in that, Includes the following steps: Acquire monitoring images from underwater side-scan sonar equipment; The monitoring images are input into a preset EDA-Net model to obtain underwater target detection results. The EDA-Net model includes an improved backbone network, an improved neck network, and a head network. The improved backbone network is built on the YOLOv11 framework and includes a dual-path heterogeneous attention module. The dual-path heterogeneous attention module is used to divide the input feature map into a lightweight branch and an enhancement branch in the channel dimension according to a preset ratio. The lightweight branch extracts local information through depthwise convolution and pointwise convolution. The enhancement branch extracts deep information through stacked convolutional layers, and then performs joint enhancement through channel attention mechanism and spatial attention mechanism in sequence, and then merges with the lightweight branch for output. The improved neck network includes a bottleneck convolution-parameterless attention module. This module enhances the input features sequentially through a parameterless attention mechanism, multi-level bottleneck convolutional blocks, and dynamic fusion with learnable weights before outputting the feature. The parameterless attention mechanism generates attention weights based on the channel statistical bias, spatial local contrast, and gradient information of the input feature map. The calculation of these attention weights does not introduce learnable parameters, and their expression is as follows: in, Indicates the input feature map, This represents the output feature map; Channel deviation; Indicates local contrast; This represents the gradient response calculated using spatial difference. This represents element-wise multiplication; and These represent the mean and standard deviation, respectively. It is a constant; The head network is used for target detection based on the feature map output by the improved neck network.

2. The underwater side-scan sonar image recognition method based on a dual attention mechanism according to claim 1, characterized in that, In the dual-path heterogeneous attention module, the channel dimension division ratio of the lightweight branch and the enhancement branch is 1:3; The enhanced branch sequentially extracts deep information through two stacked DPC layers.

3. The underwater side-scan sonar image recognition method based on a dual attention mechanism according to claim 1, characterized in that, The improved neck network includes an upsampling part and a downsampling part; the upsampling part includes an upsampling module, a splicing module and a bottleneck convolution-parameter-free attention module connected in sequence; the downsampling part includes a bottleneck convolution-parameter-free attention module, a splicing module and a CBS module connected in sequence.

4. The underwater side-scan sonar image recognition method based on a dual attention mechanism according to claim 3, characterized in that, The upsampling module and splicing module of the upsampling part are sequentially connected to the dual-path heterogeneous attention module of the second feature extraction layer, the third feature extraction layer and the fourth feature extraction layer of the backbone network.

5. The underwater side-scan sonar image recognition method based on a dual attention mechanism according to claim 3, characterized in that, The splicing module and CBS module of the downsampling part are connected to the bottleneck convolution-parameterless attention module corresponding to the upsampling part and the dual-path heterogeneous attention module of the fourth feature extraction layer of the backbone network.

6. The underwater side-scan sonar image recognition method based on a dual attention mechanism according to claim 1, characterized in that, The DPC layer consists of a 3×3 depthwise convolution, a 1×1 pointwise convolution, and a 1×1 projection layer.

7. The underwater side-scan sonar image recognition method based on a dual attention mechanism according to claim 1, characterized in that, The spatial attention mechanism extracts the average response map, the maximum response map, and the edge response map generated by learnable convolution from the input features. After concatenating the three, it performs multi-scale modeling using convolutional kernels with multiple receptive fields, adaptively fuses them through learnable scale weights, and outputs them after generating spatial weights through Sigmoid activation.

8. An underwater side-scan sonar image recognition system based on a dual-attention mechanism, used to implement the underwater side-scan sonar image recognition method based on a dual-attention mechanism as described in any one of claims 1 to 7, characterized in that, include: The detection result output unit receives the sonar image output by the underwater side-scan sonar device, sends the sonar image into the preset EDA-Net model, and outputs the corresponding underwater target detection result. The model building unit obtains the EDA-Net model based on the improved backbone network, the improved neck network, and the head network of the YOLOv11 network model; The sample acquisition unit collects multiple target side-scan sonar training images in a preset underwater environment, preprocesses the acquired training images, and adds real annotation information to construct training samples. The model training unit trains the EDA-Net model based on training samples. It uses classification loss function and bounding box regression loss function to calculate the loss value between the training output and the true label. The model weights are iteratively updated according to the loss value until the preset termination condition is reached, and then the trained EDA-Net model is obtained.

9. The underwater side-scan sonar image recognition system based on a dual attention mechanism according to claim 8, characterized in that, The preprocessing in the sample acquisition unit includes: randomly rotating the sonar training image, randomly flipping it horizontally and vertically, and randomly cropping and scaling the sonar training image to obtain training samples.

10. The underwater side-scan sonar image recognition system based on a dual attention mechanism according to claim 8, characterized in that, The preset termination condition in the model training unit is: the accuracy of the EDA-Net model exceeds 90% or the average accuracy exceeds 70%.

Images