A complex underwater side-scan sonar exploration detection method and device based on a multi-dimensional attention collaborative lightweight anti-noise detection framework

Abstract

The application discloses a complex underwater side-scan sonar exploration detection method and device based on a multi-dimensional attention collaborative lightweight anti-noise detection framework. The device comprises: an underwater side-scan sonar imaging device for acquiring a sonar image of a detected target; and a computer connected to the underwater side-scan sonar imaging device, comprising: a backbone feature extraction network module for processing an input sonar image through a multi-scale edge refining module to output feature maps of three scales; a neck feature fusion network module for realizing information fusion across channels and spaces through a focusing space adaptive local attention module, and for adaptively modulating multi-scale features and aggregating depth information through channel frequency aggregation and an attention mechanism to output three enhanced feature maps; and a YOLOhead detection head module for generating a target detection frame according to the enhanced feature maps and outputting a detection result after non-maximum suppression processing.

Description

Technical Field

[0001] This invention belongs to the field of underwater side-scan sonar exploration and detection technology, and particularly relates to a complex underwater side-scan sonar exploration and detection method and device based on a multi-dimensional attention collaborative lightweight anti-noise detection framework. Background Technology

[0002] Side-scan sonar (SSS) provides high-resolution images of the seabed even in zero-visibility waters, making it extremely useful in various military and civilian applications such as mine countermeasures, marine mapping, offshore oil exploration, and underwater search and rescue. Particularly in underwater search and rescue missions, SSS is widely used to detect sunken bodies, shipwrecks, and crashed aircraft. However, prolonged search operations can easily lead to operator fatigue, potentially causing them to miss potential targets. Therefore, utilizing intelligent image processing technology to assist in the discovery of potential targets and accurate identification of desired objects is particularly important.

[0003] Despite advancements in SSS image processing technology in recent years, multi-class underwater target detection tasks still face significant challenges. The complexity of the underwater acoustic environment (signal attenuation, multipath effects, scattering) exacerbates the blurring of target edges and the diffusion of contours. Many underwater targets (such as partially buried shipwrecks, objects with acoustic properties similar to sand waves or rocks) have extremely low contrast, making them difficult to distinguish from complex noisy backgrounds. The significant variations in target scale and shape, coupled with potential overlap or partial occlusion, further increase the difficulty. While deep learning, particularly convolutional neural networks (CNNs), has demonstrated powerful capabilities in target detection tasks and has been applied to mine or seabed detection in SSS images, directly applying general CNN detectors to multi-class SSS environments has limitations. Standard CNNs lack explicit mechanisms to effectively handle severe edge blurring and low contrast, and general feature fusion strategies struggle to cope with the heterogeneous multi-scale features of SSS. More importantly, existing detectors face a significant trade-off between accuracy and efficiency: lightweight models like the YOLO series often sacrifice accuracy in pursuit of real-time performance, performing poorly especially for small or blurred targets with low signal-to-noise ratios; while high-performance models like the DETR series incur enormous computational overhead, making it difficult to meet the real-time deployment requirements of resource-constrained underwater platforms. Therefore, developing an efficient and robust deep learning framework that can effectively address the unique challenges of SSS images, achieving both high detection accuracy and computational efficiency, and enabling accurate multi-class target detection in underwater SSS images, has become a more challenging and practically significant key issue. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes a method and device for detecting complex underwater side-scan sonar targets based on a multi-dimensional attention-based lightweight noise-resistant detection framework. It introduces a deep refinement module for multi-scale analysis and edge information, a spatially adaptive local attention module, and a C3-based channel frequency aggregation and attention mechanism. This effectively utilizes feature information from side-scan sonar images, achieving high-precision, lightweight, and robust target detection in complex underwater side-scan sonar images. Simultaneously, it lightweights the original YOLOv11 network, reducing network complexity and improving computational efficiency. With the continuous development and widespread application of side-scan sonar imaging technology, underwater detection algorithms combined with deep learning are extremely useful in various military and civilian applications such as mine countermeasures, marine mapping, offshore oil exploration, and underwater search and rescue. Utilizing intelligent image processing technology to assist in discovering potential targets and accurately identifying desired objects is particularly important.

[0005] To achieve the above objectives, this invention provides a complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework, comprising:

[0006] Sonar images of the target under inspection are acquired using underwater side-scan sonar imaging equipment;

[0007] The sonar image is input into a deep learning network model, which includes a backbone feature extraction network, a neck feature fusion network, and a YOLO head detection head connected in sequence.

[0008] In the backbone feature extraction network, the input features are processed by a multi-scale edge refinement module, and feature maps of three scales are output.

[0009] In the aforementioned neck feature fusion network, cross-channel and cross-space information fusion is achieved through a spatial adaptive local attention module, and multi-scale features are adaptively modulated and deep information aggregated through channel frequency aggregation and attention mechanism, outputting three enhanced feature maps;

[0010] The YOLOhead detection head generates a target detection box, and the detection result is output after non-maximum suppression processing.

[0011] Optionally, the processing steps of the multi-scale edge refinement module include: performing local convolution on the input feature map to extract preliminary features;

[0012] Multi-scale pooling is performed through parallel branches, and each branch uses an adaptive average pooling layer with different bin parameters to obtain multi-resolution features.

[0013] For each branch's pooled features, convolution is performed sequentially to adjust the number of channels, and grouped convolution is used to extract spatial features.

[0014] The extracted spatial features are input into the edge refinement unit submodule. Low-frequency background information is obtained through average pooling. The difference between the original features and the low-frequency information is calculated to obtain high-frequency edge features. After refining the edges through convolution and Sigmoid activation, the refined edge features are superimposed on the original features through residual connections.

[0015] The features processed from each branch are upsampled and aligned, then concatenated with the initial features for output.

[0016] Optionally, the processing procedure of the focusing space adaptive local attention module includes:

[0017] The input feature map is compressed into two one-dimensional feature vectors by adaptive average pooling in both the height and width dimensions.

[0018] The two feature vectors are reshaped and then input into a shared one-dimensional convolutional module, which includes a convolutional layer, a group normalization layer and a sigmoid activation function to generate attention weights in the height and width directions.

[0019] When the module is configured in feature enhancement mode, the original feature map is multiplied element-wise with the attention weights in both directions;

[0020] When the module is configured in fusion-guided mode, it directly outputs the product of the attention weights in the two directions as the spatial focus map.

[0021] Optionally, the channel frequency aggregation and attention mechanism processing includes:

[0022] In the channel-frequency aggregation attention unit, the input features are processed by an interactive channel transpose attention module. The module generates a Query, Key, and Value matrix through linear projection, captures global contextual dependencies through multi-head scaling dot product attention, and introduces a learnable temperature parameter to adjust the sharpness of the attention distribution.

[0023] Attention features are input into a dual-frequency aggregation feedforward network. After the features are processed by linear transformation and activation function, the frequency gate unit performs a transformation from the spatial domain to the frequency domain. Attention weights are assigned in the frequency domain and then the network is inversely transformed back to the spatial domain.

[0024] Attention features and frequency domain processing results are integrated through residual connections.

[0025] Optionally, the processing procedure of the frequency gating unit includes:

[0026] Transform spatial domain features to the frequency domain;

[0027] Filtering or attention weighting is performed in the frequency domain based on differences in acoustic texture;

[0028] The weighted frequency domain features are inversely transformed back to the spatial domain.

[0029] Optionally, the training process of the deep learning network model includes: freezing the weights of the backbone feature extraction network and training for 100 epochs;

[0030] Unfreeze all network weights and train for 500 epochs;

[0031] The SGD optimizer was used with a learning rate of 0.01 and a batch size of 64.

[0032] Optionally, the nonmaximum suppression process includes:

[0033] The prediction boxes are stitched together and their dimensions are transformed on the three scale feature maps output by the detection head.

[0034] The predicted boxes are sorted in descending order based on a confidence threshold of 0.001;

[0035] Calculate the intersection-union ratio (IU) of the predicted boxes and remove redundant boxes with an IU greater than 0.6;

[0036] The filtered detection boxes are restored to the original image scale, with a maximum output of 300.

[0037] On the other hand, to achieve the above objectives, the present invention also provides a complex underwater side-scan sonar exploration and detection device based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework, comprising:

[0038] Underwater side-scan sonar imaging equipment is used to acquire sonar images of the target being inspected;

[0039] A computer, connected to the underwater side-scan sonar imaging device, includes:

[0040] The backbone feature extraction network module is used to process the input sonar image through the multi-scale edge refinement module and output feature maps at three scales.

[0041] The neck feature fusion network module is used to achieve cross-channel and cross-space information fusion through the focus space adaptive local attention module, and to adaptively modulate and aggregate multi-scale features and deep information through channel frequency aggregation and attention mechanism, and output three enhanced feature maps.

[0042] The YOLOhead detection module is used to generate target detection boxes based on the enhanced feature map, and output the detection results after non-maximum suppression processing.

[0043] Technical Effects of this Invention: This invention discloses a method and apparatus for complex underwater side-scan sonar exploration and detection based on a multi-dimensional attention-based lightweight anti-noise detection framework. It aims to address the inherent limitations of underwater side-scan sonar (SSS) images, such as edge blurring, low signal-to-noise ratio, multi-scale variations, and complex background interference, and the resulting difficulty in balancing detection accuracy and computational efficiency. To this end, this invention introduces a multi-scale edge refinement module, a spatially adaptive local attention module, and the synergistic application of channel frequency aggregation and attention mechanisms. This effectively utilizes feature information in side-scan sonar images to sharpen blurred target contours and enhances the extraction and recognition capabilities of weak target signals under low signal-to-noise ratio conditions. This significantly improves detection accuracy and robustness to complex underwater scenes, particularly suitable for situations with blurred target edges, extremely low contrast, or similar acoustic characteristics. Simultaneously, this invention features a lightweight design of the original YOLOv11 network, significantly reducing network complexity and parameter count, and substantially improving computational efficiency, meeting the real-time deployment requirements of resource-constrained underwater platforms. This method achieves high-precision, lightweight, and robust detection of targets in complex underwater SSS images. It is applicable to military and civilian scenarios such as underwater search and rescue, mine countermeasures, marine surveying, and offshore oil exploration, and has significant practical value. Attached Figure Description

[0044] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0045] Figure 1 This is a schematic diagram of the complex underwater side-scan sonar exploration and detection device according to an embodiment of the present invention, wherein 101-computer, 102-underwater side-scan sonar imaging equipment, and 103-target to be inspected;

[0046] Figure 2 This is a schematic diagram of the overall architecture of the deep learning detection network according to an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the network architecture of the multi-scale edge refining module in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the spatially adaptive local attention module network in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the network architecture of the channel frequency aggregation and attention mechanism module in an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of the network architecture of the interactive channel transpose attention module according to an embodiment of the present invention;

[0051] Figure 7 This is a schematic diagram of the dual-frequency aggregation feedforward network architecture according to an embodiment of the present invention. Detailed Implementation

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0054] like Figure 1 As shown, this embodiment provides a complex underwater side-scan sonar exploration and detection device based on a multi-dimensional attention-based lightweight noise-resistant detection framework, including: a computer 101, an underwater side-scan sonar imaging device 102, and a target 103; the computer 101 is connected to the underwater side-scan sonar imaging device 102. The computer acquires side-scan sonar images of the target 103 through a deep learning image detection algorithm and performs reconnaissance, exploration, and salvage.

[0055] This embodiment also provides a complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-based lightweight noise-resistant detection framework, including: constructing a network using a YOLO-based target detector and attention mechanism, capturing sonar images of the target object using an underwater side-scan sonar imaging device, and performing underwater target detection on the target images. The implementation steps are as follows:

[0056] Step 1: The underwater side-scan sonar imaging device 102 captures images of the target 103 under inspection, constructing a sonar image dataset I. h =[I h1 I h2 ,...I hK ], where dataset I h The total number of elements is K = 2000, and the image size is 3×640×640. Image annotation was performed using the open-source tool labelImg, with annotations including the underwater target category and the coordinates of the top-left and bottom-right corners of defective targets. The annotated information file is in XML format.

[0057] Step 2.1: Set up as follows Figure 2 The network model shown includes a backbone feature extraction network, a neck feature extraction network, a spatial adaptive local attention module, a channel frequency aggregation and attention mechanism, and a Yolohead detection head structure.

[0058] An RGB image of size 3×640×640 is used as input to the backbone of the network model. It passes through convolutional layers, multi-scale edge refinement module layers, SPPF layers, and C2PSA layers, outputting three feature maps of different scales. b1 I b2 and I b3 The scales are 128×80×80, 256×40×40, and 512×20×20, respectively. The original C3K2 module is replaced with a multi-scale edge refinement module to improve accuracy and reduce the number of parameters. Feature maps P3, P4, and P5 output from the backbone are connected to the multi-scale edge refinement module layer and the neck layer via a focus space adaptive local attention module for feature transfer. This module enables cross-channel and cross-space information fusion between feature maps. The output and input feature scales of the focus space adaptive local attention module are kept consistent. In the neck of the network model, fine adjustments are made to achieve efficient feature map fusion and further reduce the number of parameters. Blurred target contours are sharpened through refined edge information enhancement techniques, ultimately effectively integrating these heterogeneous but complementary information to output a more discriminative enhanced feature map.

[0059] In the Neck network model, refined adaptive modulation and deep information aggregation of multi-scale features across spatial, channel, and latent frequency dimensions fundamentally alter the flow and interaction of features within the pyramid. This addresses the challenges of accurate extraction and effective enhancement of weak target signals in traditional feature pyramid networks. Within the overall structure of Neck, its core lies in the strategic embedding of two deeply customized attention and aggregation units: a spatially adaptive local attention module and a channel-frequency aggregation and attention mechanism. Neck outputs feature maps at three different scales. n1 I n2 and I n3 The scales are 128×80×80, 256×40×40, and 512×20×20, respectively. A channel frequency aggregation and attention mechanism was designed to fuse the feature maps output from the neck to the detection head pairwise. This mechanism provides powerful feature extraction capabilities while maintaining relatively low computational complexity and parameter count, which is crucial for efficient deployment on resource-constrained underwater platforms. The collaborative work of the channel frequency aggregation and attention mechanism with the spatially adaptive local attention module enables comprehensive processing of SSS images, from coarse spatial localization to fine multidimensional feature extraction. This significantly improves detection accuracy and adaptability to complex underwater scenes while ensuring detection efficiency.

[0060] Step 2.2: Multi-scale edge refining module, such as Figure 3As shown, the overall architecture design aims to achieve multi-scale analysis of input features and in-depth refinement of edge information. Specifically, for the input feature map F... in (Its dimensions are W×H×C). The module first extracts preliminary local contextual features from a standard 3×3 convolution using a local convolution, providing a foundation for subsequent multi-scale analysis. Subsequently, the feature map enters the core processing flow for multi-scale edge information enhancement. In this flow, we designed multiple parallel feature extraction branches, each dedicated to capturing information at a specific scale to address variations in target size and situations where the target is partially covered by sediment.

[0061] In each parallel branch i, an adaptive average pooling layer (AdaptiveAvgPoo12d(·)) with different bin parameters is used to extract features from the original input F. in To obtain feature representations F at different spatial resolutions pool_i :

[0062] F pool_i =AdaptiveAvgPoo12d bin_i (F in );

[0063] This multi-scale pooling helps the model capture multi-level target information, from coarse contours to fine textures. Then, each pooled feature F... pool_i It involves two convolutional operations: the first layer is a convolution (Conv(·)) used to adjust the number of channels and promote cross-channel information interaction, reducing subsequent computation; the second layer is a 3×3 grouped convolution (Conv... g (·) is used to further extract spatial features at this scale, resulting in F scale_feat_i :

[0064] F scale_feat_i =Conv g，i (Conv 1×1，i (F pool_i ));

[0065] Crucially, each multi-scale feature F processed by convolution... scale_feat_i They will all be sent to a specially designed edge refining unit submodule (its detailed structure is shown in...). Figure 2 (As shown). This edge refinement unit submodule addresses the issues of blurred edges and low contrast in SSS images by enhancing the network's sensitivity to target contours through the extraction and enhancement of high-frequency edge signals. This allows the network to better distinguish boundaries even when the target and background acoustic properties are similar.

[0066] First, it uses an average pooling layer AvgPool2d(·) to smooth x and extract the low-frequency component x. lowThis information approximates a relatively uniform seabed background or a smooth region within an object:

[0067] x low =AvgPool2d(x);

[0068] Where, x low The overall trend of the feature map is captured. Then, the original input x is compared with the extracted low-frequency information x... low Subtracting the two yields the high-frequency information x. edge This difference operation effectively enhances rapidly changing details in the image, such as edges and textures, especially potentially weak but crucial target-background boundary signals.

[0069] x edge =xx low ;

[0070] Where, x edge To initially extract and enhance edge features. For further processing and optimization, and to introduce non-linear expressive power, x... edge The input is a convolutional layer followed by a Sigmoid (·) activation function. The convolution learns local edge patterns, and the Sigmoid (·) normalizes the responses to (0,1), assigning higher weights to salient edges that remain discernible in noisy backgrounds.

[0071] x refined_edge =Sigmoid(Conv(x) edge ));

[0072] Where, x refined_edge The refined edge features are obtained after convolution and Sigmoid (·) activation. Finally, the refined edge features x are connected via residual connections. refined_edge Added back to the original input feature map x, forming the final output x of the edge refinement unit submodule. REU_out This additive blending mechanism enhances edges that are difficult to distinguish at low contrast levels while avoiding excessive destruction of the original information and preserving useful context.

[0073] x ERU_out =x+x refined_edge ;

[0074] Where, x ERU_out This is the final output feature map of the edge refinement unit submodule. Through this series of operations, the edge refinement unit submodule significantly improves the network's ability to perceive target edges in SSS images, effectively solving the problems of blurred underwater target contours and difficulty in distinguishing them from the background. It is a key link in improving the overall detection accuracy.

[0075] After processing by the edge refinement unit submodule, the edge-refined multi-scale features FERU_processed_i are obtained. All FERU_processed_i from different branches are upsampled to the same spatial size as the initial local convolutional output features through bilinear interpolation to ensure scale alignment for subsequent fusion. Finally, these scale-aligned and edge-refined multi-scale feature maps are concatenated with the initial local convolutional feature maps along the channel dimension, and then deep fusion and information integration are performed through a final convolutional layer to output the final enhanced feature map F of the multi-scale edge refinement module. out This strategy of first extracting multi-scale and edge details separately and then fusing them together enables the model to simultaneously focus on the global structure and local fine features of the target, effectively addressing the challenges of underwater targets with varying sizes, shapes, and high similarity to the background, thereby significantly improving detection accuracy.

[0076] Step 2.3: Focusing on the spatially adaptive local attention module, such as... Figure 4 As shown, first, the input feature map... Spatial dimension compression is performed to extract contextual information. To achieve this, the spatial attention unit utilizes two parallel adaptive average pooling layers, performing global average pooling (AdaptiveAvgPool2d(·)) on the height (H) and width (W) dimensions respectively, to obtain x. h and x w :

[0077]

[0078] In this way, two-dimensional spatial information is efficiently compressed into two one-dimensional feature vectors, which aggregate information along the width and height dimensions, respectively. This compression operation significantly reduces the computational cost of subsequent processing and is the foundation for achieving lightweight modules.

[0079] Subsequently, in order to model local dependencies and generate attention weights on the compressed dimension, x h and x w Reshape each tensor into a shape suitable for one-dimensional convolution processing. and These two reshaped tensors share a one-dimensional convolutional module f. 1D (·). This module consists of a one-dimensional convolutional layer (kernel size K). c =7, fill with P c =3 (to maintain the same length), a group normalization layer, and a Sigmoid (·) activation function σ (·). One-dimensional convolutions capture local spatial patterns along their respective dimensions; group normalization helps stabilize the training process; the Sigmoid function maps the output to the (0, 1) range, forming the attention weights.

[0080]

[0081] These weights att h and att w These represent the importance score of each channel in its corresponding spatial dimension, calculated based on its local context.

[0082] Finally, the learned spatial attention weights will be applied to the original input feature map x according to the configuration parameter flag of the spatially adaptive local attention module. For this purpose, att needs to be... h and att w Reconstructed separately and The dimension. When the module is configured in feature enhancement mode (flag=True), the spatially adaptive local attention module achieves adaptive focusing and enhancement of the original feature map x by performing element-wise multiplication (Hadamard product ⊙) between the original feature map x and the attention weights learned in the height and width dimensions:

[0083] x out =x⊙att h .reshape(B, C, H, 1) ⊙att w .reshape(B, C, 1, W);

[0084] This approach effectively weights features based on their spatial importance, highlighting regions containing key information while suppressing background and noise, achieving a focused spatial adaptive effect. Conversely, when the module is configured in fusion-guided mode (flag=False), the focused spatial adaptive local attention module directly outputs the product of the attention weights in both directions:

[0085] x out =att h .reshape(B, C, H, 1) ⊙att w .reshape(B, C, 1, W);

[0086] at this time, This can be viewed as a purely spatial focus map or saliency map, where high-value regions indicate spatial locations considered important in both the horizontal and vertical directions. This focus map can be fused with other feature maps to guide the fusion process to focus more on these spatially salient regions, thereby optimizing information flow.

[0087] Step 2.4: Channel frequency aggregation and attention mechanism as follows Figure 5As shown, firstly, the interactive channel transpose attention module's global-local dual-path processing and cross-dimensional interaction mechanism enable it to effectively extract and enhance target features from cluttered SSS signals while suppressing background noise and artifacts. This is crucial for identifying targets with low signal-to-noise ratios, small sizes, or partial burial. We designed the interactive channel transpose attention module, where each channel-frequency aggregation attention unit (its internal structure is shown in...) Figure 5 (As shown) all employ residual design, connecting an interactive channel transpose attention module and a dual-frequency aggregation feedforward network, aiming to comprehensively optimize features from the channel, spatial, and even frequency dimensions. For features input to the channel-frequency aggregation attention unit... Its processing flow can be summarized as sequence transformation and residual learning after normalization layer stabilization:

[0088] x′ cfa =iCTA(LN(x cfa_in ));

[0089] x cfa_out =x′ cfa +DropPath(DFFN(LN(x′ cfa ))).

[0090] Interactive channel transpose attention module (detailed structure as follows) Figure 6 (As shown) is the perceptual core of the channel-frequency convergence attention unit, designed to overcome the limitations of traditional self-attention to address the subtle differences in target features and the complexity of the background in SSS images. For input features that have undergone dimensionality transformation... (where N = HW), the interactive channel transpose attention module first passes through three independent linear projections. Generate a matrix of Query(Q), Key(K), Value(V): Q = x ln W Q K = x ln W K V = x ln W V Subsequently, scaled dot product attention is computed under a multi-head mechanism to capture global context dependencies:

[0091]

[0092] Where d k τ is the dimension of the key vector C′ / num_heads, and τ is a learnable temperature parameter used to adjust the sharpness of the attention distribution. This represents attention features with global context awareness, which is crucial for understanding large-scale scene structure and the relative relationships between objects and the environment in SSS images. Secondly, the frequency awareness capability of the dual-frequency aggregation feedforward network provides the model with a novel dimension for understanding and distinguishing content in SSS images. Its core frequency-gated unit f... FG (·, H, W) performs frequency-domain-aware processing on the features. In SSS images, different types of targets and the natural seabed often have different acoustic textures, and these texture differences manifest as different energy distributions or spectral characteristics in the frequency domain. Frequency-gated unit f FG It is designed to capture and utilize these frequency domain cues. Formally, for input... Dual-frequency aggregation feedforward network (detailed structure as follows) Figure 7 As shown, through linear transformation W1, activation function act(·) and frequency gated unit f FG Processing:

[0093]

[0094] Frequency gating unit f FG The specific implementation may involve transforming spatial domain features to the frequency domain, performing filtering or attention allocation in the frequency domain, and then transforming back to the spatial domain. The final output is obtained through a linear transformation W2:

[0095] DFFN(x dffn_in )=x gated W2+Dropout(x gated W2).

[0096] By leveraging the acoustic texture and spectral characteristics of targets, the system enhances adaptability to complex seabed environments and robustness in target recognition, particularly excelling at distinguishing samples with similar acoustic features but different textures. Finally, the integration of the C3 structure ensures that the frequency aggregation and attention mechanisms provide powerful feature extraction capabilities while maintaining relatively low computational complexity and parameter count, crucial for efficient deployment on resource-constrained underwater platforms. The collaborative work of the frequency aggregation and attention mechanisms with the spatially adaptive local attention module enables comprehensive processing of SSS images, from coarse spatial localization to fine multidimensional feature extraction, thereby significantly improving detection accuracy and adaptability to complex underwater scenes while maintaining detection efficiency.

[0097] Step 3: The training method includes first freezing the weights of the skeleton feature extraction network, training for 100 epochs, then unfreezing all weights and training for 500 epochs. The loss function used is the same as the loss function in the original YOLOv11. Network training parameters are set as follows: learning rate lr = 0.01, batch size = 64, mosaic = 0, training and validation set split at 0.9:0.1, SGD optimizer, and 100 training epochs.

[0098] Step 4: When using the trained network for prediction, a test image of size 3×640×640 is first input into the network. After inference, the YOLO head outputs feature maps at three different scales: 80×80, 40×40, and 20×20. Next, classification and regression prediction results are extracted from these feature maps and then concatenated and transformed in dimensions. The channel dimension is moved to the end, making the shapes of the class prediction branch and the bounding box (bbox) prediction branch (1, 8400, 80 and 1, 8400, 4, respectively). Then, all predicted bounding boxes are sorted in descending order based on their confidence level (set to 0.001), and non-maximum suppression (NMS) is performed. Starting with the box with the highest confidence level, the intersection-over-union (IOU) with the remaining predicted boxes is calculated, and all redundant boxes with an IOU greater than 0.6 are removed. Finally, the filtered detection boxes are restored to the original image scale, resulting in a final output of no more than 300 detection boxes. Finally, the coordinates (x, y, w, h) of these detection boxes and their corresponding normalized position coordinates (X, Y, W, H) are obtained and plotted on the test image. If a detection box is successfully output in the test image, it indicates the presence of a reconnaissance, exploration, or salvage target; otherwise, no target appears.

[0099] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework, characterized in that, include: Sonar images of the target under inspection are acquired using underwater side-scan sonar imaging equipment; The sonar image is input into a deep learning network model, which includes a backbone feature extraction network, a neck feature fusion network, and a YOLO head detection head connected in sequence. In the backbone feature extraction network, the input features are processed by a multi-scale edge refinement module, and feature maps of three scales are output. In the aforementioned neck feature fusion network, cross-channel and cross-space information fusion is achieved through a spatial adaptive local attention module, and multi-scale features are adaptively modulated and deep information aggregated through channel frequency aggregation and attention mechanism, outputting three enhanced feature maps; The target detection box is generated by the YOLOhead detection head, and the detection result is output after non-maximum suppression processing. The processing steps of the multi-scale edge refinement module include: performing local convolution on the input feature map to extract preliminary features; Multi-scale pooling is performed through parallel branches, and each branch uses an adaptive average pooling layer with different bin parameters to obtain multi-resolution features. For each branch's pooled features, convolution is performed sequentially to adjust the number of channels, and grouped convolution is used to extract spatial features. The extracted spatial features are input into the edge refinement unit submodule. Low-frequency background information is obtained through average pooling. The difference between the original features and the low-frequency information is calculated to obtain high-frequency edge features. After refining the edges through convolution and Sigmoid activation, the refined edge features are superimposed on the original features through residual connections. The features processed from each branch are upsampled and aligned, then concatenated with the initial features for output.

2. The complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework as described in claim 1, characterized in that, The processing procedure of the focused spatial adaptive local attention module includes: The input feature map is compressed into two one-dimensional feature vectors by adaptive average pooling in both the height and width dimensions. The two feature vectors are reshaped and then input into a shared one-dimensional convolutional module, which includes a convolutional layer, a group normalization layer and a sigmoid activation function to generate attention weights in the height and width directions. When the module is configured in feature enhancement mode, the original feature map is multiplied element-wise with the attention weights in both directions; When the module is configured in fusion-guided mode, it directly outputs the product of the attention weights in the two directions as the spatial focus map.

3. The complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework as described in claim 1, characterized in that, The channel frequency aggregation and attention mechanism processing includes: In the channel-frequency aggregation attention unit, the input features are processed by an interactive channel transpose attention module. The module generates a Query, Key, and Value matrix through linear projection, captures global contextual dependencies through multi-head scaling dot product attention, and introduces a learnable temperature parameter to adjust the sharpness of the attention distribution. Attention features are input into a dual-frequency aggregation feedforward network. After the features are processed by linear transformation and activation function, the frequency gate unit performs a transformation from the spatial domain to the frequency domain. Attention weights are assigned in the frequency domain and then the network is inversely transformed back to the spatial domain. Attention features and frequency domain processing results are integrated through residual connections.

4. The complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework as described in claim 3, characterized in that, The processing procedure of the frequency gating unit includes: Transform spatial domain features to the frequency domain; Filtering or attention weighting is performed in the frequency domain based on differences in acoustic texture; The weighted frequency domain features are inversely transformed back to the spatial domain.

5. The complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework as described in claim 1, characterized in that, The training process of the deep learning network model includes: freezing the weights of the backbone feature extraction network and training for 100 epochs. Unfreeze all network weights and train for 500 epochs; The SGD optimizer was used with a learning rate of 0.01 and a batch size of 64.

6. The complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework as described in claim 1, characterized in that, The nonmaximum suppression process includes: The prediction boxes are stitched together and their dimensions are transformed on the three scale feature maps output by the detection head. The predicted boxes are sorted in descending order based on a confidence threshold of 0.001; Calculate the intersection-union ratio (IU) of the predicted boxes and remove redundant boxes with an IU greater than 0.6; The filtered detection boxes are restored to the original image scale, with a maximum output of 300.

7. An apparatus for a complex underwater side-scan sonar exploration and detection method based on a multi-dimensional attention-coordinated lightweight noise-resistant detection framework according to any one of claims 1-6, characterized in that, include: Underwater side-scan sonar imaging equipment is used to acquire sonar images of the target being inspected; A computer, connected to the underwater side-scan sonar imaging device, includes: The backbone feature extraction network module is used to process the input sonar image through the multi-scale edge refinement module and output feature maps at three scales. The neck feature fusion network module is used to achieve cross-channel and cross-space information fusion through the focus space adaptive local attention module, and to adaptively modulate and aggregate multi-scale features and deep information through channel frequency aggregation and attention mechanism, and output three enhanced feature maps. The YOLOhead detection module is used to generate target detection boxes based on the enhanced feature map, and output the detection results after non-maximum suppression processing.

Images