Sonar weak target detection method based on AA-CLSCM and SE-STAL

Abstract

The application relates to the technical field of underwater acoustic signal processing, in particular to a sonar weak target detection method based on AA-CLSCM and SE-STAL, which comprises the following steps: acquiring a plurality of side-scan sonar images as sample data, constructing a training sample set after double labeling of a highlight reflection area and an acoustic shadow area; constructing a sonar weak target detection model composed of a multi-scale feature extraction network, a multi-scale feature fusion network and a decoupling detection head; wherein the multi-scale feature fusion network is arranged at a cross-layer jump connection in a top-down path of a feature pyramid; the sonar weak target detection model is trained based on the training sample set; in a label allocation stage, target scale relaxation of a real target and allocation of positive and negative samples are performed based on an SE-STAL mechanism; classification loss and regression loss are calculated; and network parameters of the sonar weak target detection model are updated through back propagation.

Description

Technical Field

[0001] This application relates to the field of underwater acoustic signal processing technology, and proposes a sonar weak target detection method based on AA-CLSCM and SE-STAL. Background Technology

[0002] Side-scan sonar is an active acoustic imaging device widely used in underwater exploration, seabed mapping, shipwreck search, foreign object identification, and underwater engineering safety inspection. Compared with traditional optical imaging methods, sonar imaging does not depend on natural lighting conditions and can obtain relatively stable observation results in turbid, dark, and long-distance underwater environments, thus possessing significant application value. Side-scan sonar transmits sound waves towards the seabed or underwater targets and receives the echo intensity to form a two-dimensional acoustic image (side-scan sonar image). Target areas in side-scan sonar images generally exhibit strong bright reflections, and due to the target's blocking effect on sound waves, an acoustic shadow of a certain length and shape is often present behind it. The bright reflective area and the acoustic shadow together constitute important visual and physical characteristics that distinguish side-scan sonar targets from the ordinary seabed background.

[0003] However, the real underwater environment is very complex, and due to the limitations of acoustic imaging physics, target detection based on side-scan sonar images faces the following serious problems.

[0004] First, weak target features are easily lost in deep networks. Underwater weak targets often occupy only a few pixels in an image (i.e., the bright reflection area). After multiple downsampling operations, traditional deep learning networks easily smooth out or lose the spatial details of weak targets. Although some research teams have attempted to introduce high-level semantics into low-level layers through cross-layer fusion, conventional isotropic spatial attention mechanisms ignore the physical topological characteristics of "highlights first, shadows later" in sonar images, which can easily lead to semantic misalignment and fail to accurately compensate for the edge details of real targets.

[0005] Second, the positive sample allocation for weak targets fails (i.e., gradient truncation). Mainstream deep learning detection frameworks typically base their training on the area of ​​the target's physical bounding box (...). The model assigns positive and negative samples based on the intersection-over-union (IoU) ratio between the predicted and ground truth bounding boxes. Because the bright reflective area of ​​weak targets is extremely small, their IoU ratio with the predicted bounding box is often low, making them easily identified as background samples and discarded. This results in weak targets failing to generate sufficient supervisory signals during training, hindering gradient backpropagation and ultimately leading to a high false negative rate for weak targets. Summary of the Invention

[0006] To address the aforementioned technical issues, embodiments of this application propose a sonar weak target detection method based on AA-CLSCM (Acoustic Asymmetric Cross Layer Semantic Compensation Module) and SE-STAL (Shadow Enhanced Small Target Aware Label Assignment). By combining the paired topological physical characteristics of "highlight first, shadow later" in side-scan sonar images, cross-layer semantic compensation is performed on the spatial features of weak targets, and the positive sample assignment threshold for weak targets is scientifically and dynamically relaxed, thereby significantly improving the detection recall and localization accuracy of weak targets in complex underwater environments.

[0007] To achieve the above objectives, embodiments of this application propose a sonar weak target detection method based on AA-CLSCM and SE-STAL, applicable to complex underwater environments. Specifically, the method includes: acquiring several real side-scan sonar images as sample data; performing fine-grained double-polygon labeling on real targets, including highlight reflection areas and acoustic shadow areas; constructing training, validation, and test sample sets from the labeled sample data according to a preset ratio; and constructing a sonar weak target detection model composed of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head. The multi-scale feature extraction network extracts features from the input side-scan sonar images. In the top-down path of the feature pyramid, the multi-scale feature fusion network deploys the AA-CLSCM module at cross-layer skip connections to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise. The decoupled detection head... The model is used to output the category confidence score and bounding box regression results for weak targets. Based on the training sample set, the sonar weak target detection model is trained. In the label allocation stage, based on the SE-STAL mechanism, the normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. The dynamic matching threshold is determined based on the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, the positive and negative samples are allocated according to the intersection-union ratio between the candidate predicted boxes and the real labeled boxes output by the sonar weak target detection model and the dynamic matching threshold. Finally, the total loss function, including classification loss and regression loss, is calculated, and the network parameters of the sonar weak target detection model are updated through backpropagation until the preset convergence condition is met, resulting in the trained sonar weak target detection model. The side-scan sonar image to be detected is input into the trained sonar weak target detection model for detection, and its output weak target detection results are obtained.

[0008] To achieve the above objectives, embodiments of this application also propose a sonar weak target detection system based on AA-CLSCM and SE-STAL, suitable for complex underwater environments, used to implement the sonar weak target detection method based on AA-CLSCM and SE-STAL as described above. The system specifically includes: a sample set construction module, a model construction module, a model iterative training and update module, and a model usage module. The sample set construction module acquires several real side-scan sonar images as sample data, performs double polygonal fine-grained labeling of the real target's high-brightness reflection area and acoustic shadow area, and constructs the labeled sample data into a training sample set, a validation sample set, and a test sample set according to a preset ratio. The model construction module constructs a sonar weak target detection model composed of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head. The multi-scale feature extraction network extracts features from the input side-scan sonar images, and the multi-scale feature fusion network deploys the AA-CLSCM module across layers in a top-down path of the feature pyramid. At the connection point, to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise, the decoupled detection head is used to output the category confidence and bounding box regression results of weak targets. The model iteration training and update module is used to train the sonar weak target detection model based on the training sample set. In the label allocation stage, based on the SE-STAL mechanism, the normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. The dynamic matching threshold is determined according to the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, the positive and negative samples are assigned according to the intersection-union ratio between the candidate predicted box and the real labeled box output by the sonar weak target detection model and the dynamic matching threshold. Then, the total loss function including classification loss and regression loss is calculated, and the network parameters of the sonar weak target detection model are updated by backpropagation until the preset convergence condition is met, and the trained sonar weak target detection model is obtained. The model usage module is used to input the side-scan sonar image to be detected into the trained sonar weak target detection model for detection and obtain its output weak target detection results.

[0009] To achieve the above objectives, embodiments of this application also propose an electronic device, which specifically comprises a processor and a memory. The memory stores instructions that the processor can execute. When the processor is configured to execute the instructions, the electronic device enables the electronic device to implement the sonar weak target detection method based on AA-CLSCM and SE-STAL as described above.

[0010] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, enables a sonar weak target detection method based on AA-CLSCM and SE-STAL as described above.

[0011] In some optional embodiments, the step of acquiring several real side-scan sonar images as sample data, performing double-polygon fine-grained labeling of the real targets in the bright reflection area and acoustic shadow area, and constructing the sample data after labeling the real targets into a training sample set, a validation sample set, and a test sample set according to a preset ratio includes: acquiring several real side-scan sonar images as sample data; performing preprocessing on all sample data, including introducing underwater additive white Gaussian noise and simulating sonar blind zone enhancement; performing double-polygon fine-grained labeling of the real targets in the preprocessed sample data in the bright reflection area and acoustic shadow area to obtain sample data after labeling the real targets; and dividing the sample data after labeling the real targets into training sample data, validation sample data, and test sample data in a ratio of 7:2:1 to construct the training sample set, validation sample set, and test sample set.

[0012] In some optional embodiments, the multi-scale feature extraction network consists of five CNN layers, with the first three CNN layers being shallow networks and the last two CNN layers being deep networks, for the input side-scan sonar image. Multi-scale feature extraction networks utilize shallow networks to extract low-level spatial feature maps rich in texture and boundary details. Deep semantic feature maps containing strong class information are extracted using deep networks. .

[0013] In some optional embodiments, the multi-scale feature fusion network includes multiple AA-CLSCM modules, deployed at cross-layer skip connections in the top-down path of the feature pyramid, to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise; within the AA-CLSCM module, The semantic information is not broadcast globally. Instead, it is used to find highlight-shadow pairs through asymmetric convolution, generating a specific attention mask for the sonar target, and then accurately compensating across layers. .

[0014] In some optional embodiments, the AA-CLSCM module achieves cross-layer compensation through the following steps: The AA-CLSCM module received and Afterwards, Applying asymmetric orientation-aware convolution, asymmetric orientation-aware convolution utilizes... Horizontal asymmetric convolution kernel pairs Perform cross-course scanning along the sound wave to extract one-dimensional gradient features of pixel intensity transitioning from bright reflections to acoustic shadows. ; use Generate adaptive guided weight mask Only when the bright reflective area and the acoustic shadow area appear in pairs, High response is output in the corresponding area, and isolated bright reflection areas and isolated acoustic shadow areas are suppressed; Will Transform Alignment to After the scale, with As a gating signal and To integrate and achieve Cross-layer compensation yields fusion features. ; This can be expressed by the formula: ; in, This indicates element-wise multiplication. It is the Sigmoid activation function. This indicates feature transformation alignment.

[0015] In some optional embodiments, during the label allocation stage, based on the SE-STAL mechanism, a normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflective area of ​​the real target and the acoustic shadow area associated with the bright reflective area. A dynamic matching threshold is then determined based on the target scale relaxation coefficient and the normalized shadow evidence coefficient. Finally, the allocation of positive and negative samples is completed based on the intersection-union ratio (IU) between the candidate predicted bounding boxes output by the sonar weak target detection model and the real labeled bounding boxes, and the dynamic matching threshold. This includes: During the label assignment phase, based on the SE-STAL mechanism, the area of ​​the highlight reflective region is obtained for each real target. and the area of ​​the acoustic shadow region associated with the bright reflection area. ; Based on the following formula, and Calculate the total area of ​​shadow augmentation. : ; in, This is the preset shadow gain coefficient; Based on the following formula, and Calculate the normalized shadow evidence coefficient : ; in, It is a stability constant greater than zero, used to prevent division by zero; Based on the target scale relaxation factor, the following formula is used. and Calculate the effective threshold relaxation factor : ; in, This is the preset gain coefficient for shadow evidence; Based on the following formula, Calculate the dynamic matching threshold : ; in, Based on the matching threshold, The total area of ​​the input side-scan sonar image; Calculate the intersection-union ratio (IU) between the candidate predicted bounding boxes and the ground truth bounding boxes output by the sonar weak target detection model. If the IU is not less than the dynamic matching threshold, the candidate predicted bounding box is assigned as a positive sample; otherwise, it is assigned as a negative sample. Among them, the area of ​​the bright reflection zone, the area of ​​the acoustic shadow zone, and the shadow evidence coefficient are only used for the dynamic matching threshold calculation in the label assignment stage, and do not change the bounding box regression results of the real target.

[0016] In some optional embodiments, after each backpropagation update of the network parameters of the sonar weak target detection model, the updated sonar weak target detection model is validated based on the validation sample set to prevent model overfitting and to adjust hyperparameters and select the optimal version. After meeting the preset convergence conditions, the trained sonar weak target detection model is obtained. The performance of the trained sonar weak target detection model is evaluated based on the test sample set. The evaluation criteria are average intersection-union ratio, weak target recall, AP50 and AP75. If all performance evaluations are qualified, the trained sonar weak target detection model is allowed to be deployed; otherwise, it is retrained. Among them, the average intersection-union ratio is used to evaluate the degree of overlap between the predicted bounding box and the label, the weak target recall rate is used to evaluate the model's ability to detect weak targets, the AP50 is used to measure the model's detection performance under a loose IoU threshold, and the AP75 is used to measure the model's fine localization ability under a strict IoU threshold.

[0017] This application proposes a sonar weak target detection method based on AA-CLSCM and SE-STAL, which achieves the following improvements compared with traditional target detection based on side-scan sonar images.

[0018] First, it effectively preserves weak features and suppresses seabed noise interference. This application embeds the AA-CLSCM module into the multi-scale feature fusion network, abandoning the traditional isotropic convolution and using horizontal asymmetric convolution to extract the unique bright shadow pair topological gradient features of sonar images, generating a dedicated compensation attention mask. High-level semantic information is supplemented only for target regions that match the bright shadow pair features. Irrelevant interference such as isolated bright noise and seabed pit shadows are suppressed by the weight mask, thereby solving the problems of semantic misalignment in conventional cross-layer fusion and loss of shallow details of weak targets during downsampling, strengthening the edge and structural features of weak targets, and significantly reducing false detections caused by complex seabed reverberation.

[0019] Second, it significantly improves the recall rate of weak targets. This application employs the SE-STAL mechanism during the training phase. Building upon scale-adaptive label allocation using the area of ​​the bright reflective region of the real target, it converts the acoustic shadow region associated with the target into a normalized shadow evidence coefficient, and uses this coefficient to increase the effective threshold relaxation factor. For targets with small bright reflective regions and reliable acoustic shadows, the dynamic matching threshold can be further reduced, thereby increasing the probability that reasonable candidate prediction boxes are assigned as positive samples, enhancing gradient supervision of weak targets, and reducing the false negative rate.

[0020] Third, it significantly improves the positioning accuracy of weak targets. This application relies on the precise feature compensation capability of AA-CLSCM and the sufficient gradient supervision of SE-STAL to comprehensively optimize the model's fine positioning capability. In real underwater environments with low signal-to-noise ratio and multiple clutter interferences, it can achieve accurate matching and simultaneously identify the main weak target and its associated small debris, adapting to complex detection scenarios such as turbidity, multiple reefs, and strong reverberation.

[0021] Fourth, they are highly versatile and applicable to engineering. Both AA-CLSCM and SE-STAL are independent pluggable modules that can be directly embedded into end-to-end detection architectures. They are compatible with mainstream detection head structures such as decoupled detection heads, without requiring significant modifications to the original detection network framework. The dataset construction only requires highlighting and shadowing polygon annotations on side-scan sonar images. During the training phase, model convergence can be achieved by combining conventional underwater noise data augmentation. They have low deployment costs and wide applicability. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. The following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0023] Figure 1 This is a flowchart of a sonar weak target detection method based on AA-CLSCM and SE-STAL provided in one embodiment of this application; Figure 2 This is a structural diagram of a sonar weak target detection model provided in one embodiment of this application; Figure 3 This is a schematic diagram of the working principle of the AA-CLSCM module provided in one embodiment of this application; Figure 4 This is a schematic diagram of the SE-STAL mechanism provided in one embodiment of this application; Figure 5 This is a quantitative performance comparison chart of different methods provided in one embodiment of this application for weak target detection using side-scan sonar; Figure 6 This is a visual comparison of the side-scan sonar weak target detection results of different methods provided in one embodiment of this application in a low signal-to-noise ratio scenario; Figure 7 This is a structural diagram of a sonar weak target detection system based on AA-CLSCM and SE-STAL provided in another embodiment of this application; Figure 8 This is a structural diagram of an electronic device provided in another embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0025] One embodiment of this application proposes a sonar weak target detection method based on AA-CLSCM and SE-STAL, which is suitable for complex underwater environments. The implementation details of the sonar weak target detection method based on AA-CLSCM and SE-STAL proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding only and are not necessary for implementing this solution.

[0026] The specific process of the sonar weak target detection method based on AA-CLSCM and SE-STAL proposed in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Obtain several real side-scan sonar images as sample data, and perform double polygon fine labeling on the real targets, including the bright reflection area and the acoustic shadow area. Construct the sample data after labeling the real annotation boxes into training sample set, validation sample set and test sample set according to a preset ratio.

[0027] In the specific implementation, this embodiment is divided into two parts: model training and model use. In model training, several real side-scan sonar images are first obtained as sample data. After the real targets are labeled with double polygonal fine labels for the bright reflection area and the acoustic shadow area, the sample data after the labeled real annotation box is constructed into training sample set, validation sample set and test sample set according to a preset ratio, so as to lay a solid data foundation for model training.

[0028] In one example, several real side-scan sonar images are first acquired as sample data. All sample data undergo preprocessing, including the introduction of underwater additive white Gaussian noise and simulation enhancement of sonar blind zones. Then, real targets in the preprocessed sample data are annotated with double polygonal labels for both bright reflective areas and acoustic shadow areas, resulting in sample data with labeled ground truth bounding boxes. Finally, the sample data with labeled ground truth bounding boxes is divided into training, validation, and test sets in a 7:2:1 ratio, constructing the training, validation, and test sets.

[0029] Step 102: Construct a sonar weak target detection model consisting of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head.

[0030] In the specific implementation, after constructing the training sample set, validation sample set, and test sample set, it is also necessary to construct a sonar weak target detection model consisting of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head. The multi-scale feature extraction network is used to extract features from the input side-scan sonar images. The multi-scale feature fusion network deploys the AA-CLSCM module at the cross-layer skip connections in the top-down path of the feature pyramid to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise. The decoupled detection head is used to output the category confidence and bounding box regression results of the weak targets.

[0031] In one example, the specific structure of the sonar weak target detection model is as follows: Figure 2 As shown.

[0032] The multi-scale feature extraction network (Backbone) consists of five CNN layers (such as CSPNet), with the first three CNN layers being shallow and the last two being deep. For the input side-scan sonar image... Multi-scale feature extraction networks utilize shallow networks to extract low-level spatial feature maps rich in texture and boundary details. Deep semantic feature maps containing strong class information are extracted using deep networks. .

[0033] The multi-scale feature fusion network (Neck) incorporates multiple AA-CLSCM modules, deployed at cross-layer skip connections along the top-down path of the feature pyramid. This enhances the edge and structural representation of weak targets while filtering complex underwater reverberation noise. Within the AA-CLSCM modules, The semantic information is not broadcast globally. Instead, it is used to find highlight-shadow pairs through asymmetric convolution, generating a specific attention mask for the sonar target, and then accurately compensating across layers. . Figure 2 The multi-scale feature fusion network in the model incorporates three AA-CLSCM modules.

[0034] The decoupled detection head, based on the fused multi-scale feature map, outputs the class confidence of weak targets and the bounding box regression results, which can reduce redundant candidate box screening steps according to deployment requirements.

[0035] In one example, the AA-CLSCM module works as follows: Figure 3 As shown.

[0036] Side-scan sonar images are highly directional; a true protruding target will inevitably exhibit a one-dimensional spatial sequence distribution of "high brightness (extremely high pixel value) + accompanying shadow (extremely low pixel value)". The AA-CLSCM module receives... and Afterwards, Applying asymmetric orientation-aware convolution, asymmetric orientation-aware convolution utilizes... ( or ) horizontal asymmetric convolution kernel pairs Perform cross-course scanning along the sound wave to extract one-dimensional gradient features of pixel intensity transitioning from bright reflections to acoustic shadows. .

[0037] Next, dynamic weight calculation will be performed, utilizing... Generate adaptive guided weight mask If and only if the bright reflective area and the acoustic shadow area appear in pairs, High response is output in the corresponding area, and isolated bright reflection areas (suspended object noise) and isolated acoustic shadow areas (seabed pits) are suppressed.

[0038] Finally, there is cross-layer compensation injection, which will... Transform Alignment to After the scale, with As a gating signal and To integrate and achieve Cross-layer compensation yields fusion features. .

[0039] This can be expressed by the formula: ; in, This indicates element-wise multiplication. It is the Sigmoid activation function. This indicates feature transformation alignment.

[0040] Step 103: Train the sonar weak target detection model based on the training sample set. In the label allocation stage, based on the SE-STAL mechanism, calculate the normalized shadow evidence coefficient according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. Determine the dynamic matching threshold according to the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, complete the allocation of positive and negative samples according to the intersection-union ratio between the candidate predicted box and the real labeled box output by the sonar weak target detection model and the dynamic matching threshold. Then, calculate the total loss function including classification loss and regression loss. Update the network parameters of the sonar weak target detection model through backpropagation until the preset convergence condition is met, and obtain the trained sonar weak target detection model.

[0041] In the specific implementation, after constructing the sonar weak target detection model, it can be trained based on the training sample set. In the label allocation stage, based on the SE-STAL mechanism, the normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. The dynamic matching threshold is determined according to the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, the positive and negative samples are allocated according to the intersection-union ratio between the candidate predicted boxes and the real labeled boxes output by the sonar weak target detection model and the dynamic matching threshold. Then, the total loss function including classification loss and regression loss is calculated, and the network parameters of the sonar weak target detection model are updated by backpropagation until the preset convergence condition is met, and the trained sonar weak target detection model is obtained.

[0042] Understandably, during the label assignment phase of training, the SE-STAL mechanism is invoked. Even if the bounding box output by the decoupled detection head only covers a very small bright reflection area, the SE-STAL mechanism will still retrieve and use the associated shadow annotation area to reconstruct the area, lower the assignment threshold, and then calculate the classification loss and regression loss. The above losses are backpropagated to update the network weights, thereby improving the detection capability of weak targets.

[0043] In one example, the principle of the SE-STAL mechanism is as follows: Figure 4 As shown.

[0044] During the label assignment phase, based on the SE-STAL mechanism, the area of ​​the highlight reflective region is obtained for each real target. and the area of ​​the acoustic shadow region associated with the bright reflection area. .

[0045] Based on the following formula, and Calculate the total area of ​​shadow augmentation. : ; in, This is the preset shadow gain coefficient.

[0046] Based on the following formula, and Calculate the normalized shadow evidence coefficient : ; in, It is a stability constant greater than zero, used to prevent division by zero.

[0047] Based on the target scale relaxation factor, the following formula is used. and Calculate the effective threshold relaxation factor : ; in, This is the preset shadow evidence gain coefficient.

[0048] Based on the following formula, Calculate the dynamic matching threshold : ; in, Based on the matching threshold, The total area of ​​the input side-scan sonar image.

[0049] Calculate the intersection-union ratio (IUR) between the candidate predicted bounding boxes and the ground truth bounding boxes output by the sonar weak target detection model. If the IUR is not less than the dynamic matching threshold, the candidate predicted bounding box is assigned as a positive sample; otherwise, it is assigned as a negative sample.

[0050] It should be noted that the area of ​​the bright reflection zone, the area of ​​the acoustic shadow zone, and the shadow evidence coefficient are only used for the dynamic matching threshold calculation in the label assignment stage and do not change the bounding box regression results of the real target.

[0051] In one example, after each backpropagation update of the network parameters of the sonar weak target detection model, the updated sonar weak target detection model is validated based on the validation sample set to prevent overfitting and to adjust hyperparameters and select the optimal version.

[0052] After meeting the preset convergence conditions, the trained sonar weak target detection model is obtained. The performance of the trained sonar weak target detection model is evaluated based on the test sample set. The evaluation criteria include mean intersection-over-union ratio (Intersection over Union), weak target recall, AP50, and AP75. If all performance evaluations are satisfactory, the trained sonar weak target detection model is allowed to be deployed; otherwise, it is retrained. Specifically, the mean intersection-over-union ratio evaluates the overlap between the predicted bounding box and the label; the weak target recall evaluates the model's ability to detect weak targets; AP50 measures the model's detection performance under a relaxed IoU threshold; and AP75 measures the model's fine-grained localization ability under a strict IoU threshold.

[0053] Step 104: Input the side-scan sonar image to be detected into the trained sonar weak target detection model for detection, and obtain its output weak target detection result.

[0054] In practice, after the deployment of the trained sonar weak target detection model is completed, the side-scan sonar image to be detected can be input into the trained sonar weak target detection model for detection, thereby obtaining its output weak target detection result.

[0055] This embodiment proposes a sonar weak target detection method based on AA-CLSCM and SE-STAL, which achieves the following improvements compared with traditional target detection based on side-scan sonar images.

[0056] First, it effectively preserves weak features and suppresses seabed noise interference. This embodiment embeds the AA-CLSCM module into the multi-scale feature fusion network, abandoning the traditional isotropic convolution and using horizontal asymmetric convolution to extract the unique bright shadow pair topological gradient features of sonar images, generating a dedicated compensation attention mask. High-level semantic information is supplemented only for target regions that match the bright shadow pair features. Irrelevant interference such as isolated bright noise and seabed pit shadows are suppressed by the weight mask, solving the problems of semantic misalignment in conventional cross-layer fusion and loss of shallow details of weak targets during downsampling. It strengthens the edge and structural features of weak targets and significantly reduces false detections caused by complex seabed reverberation.

[0057] Second, it significantly improves the recall rate of weak targets. This embodiment employs the SE-STAL mechanism during the training phase. Building upon scale-adaptive label allocation using the area of ​​the bright reflective region of the real target, it converts the acoustic shadow region associated with the target into a normalized shadow evidence coefficient, and uses this coefficient to increase the effective threshold relaxation factor. For targets with small bright reflective regions and reliable acoustic shadows, the dynamic matching threshold can be further reduced, thereby increasing the probability that reasonable candidate prediction boxes are assigned as positive samples, enhancing gradient supervision of weak targets, and reducing the false negative rate.

[0058] Third, it significantly improves the positioning accuracy of weak targets. This embodiment relies on the precise feature compensation capability of AA-CLSCM and the sufficient gradient supervision of SE-STAL to comprehensively optimize the model's fine positioning capability. In real underwater environments with low signal-to-noise ratio and multiple clutter interferences, it can achieve accurate matching and simultaneously identify the main weak target and its associated small debris, adapting to complex detection scenarios such as turbidity, multiple reefs, and strong reverberation.

[0059] Fourth, they are highly versatile and applicable to engineering. Both AA-CLSCM and SE-STAL are independent pluggable modules that can be directly embedded into end-to-end detection architectures. They are compatible with mainstream detection head structures such as decoupled detection heads, without requiring significant modifications to the original detection network framework. The dataset construction only requires highlighting and shadowing polygon annotations on side-scan sonar images. During the training phase, model convergence can be achieved by combining conventional underwater noise data augmentation. They have low deployment costs and wide applicability.

[0060] The steps described above are merely for clarity in describing the technical solution. In actual implementation, they can be combined into one step, or certain steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Any insignificant modifications or designs added to the algorithm or process, as long as they do not change the core of the algorithm or process, are also within the scope of protection of this application.

[0061] In one embodiment, to verify the superiority of the sonar weak target detection method based on AA-CLSCM and SE-STAL proposed in this application (hereinafter referred to as Ours), we conducted relevant simulation experiments. The evaluation criteria selected were mean intersection-union ratio, weak target recall, AP50, and AP75. The comparison models included the general visual models YOLOv8 and RT-DETR, as well as the original BGSC-Net and the original YOLO26.

[0062] Figure 5 , Figure 6 For comparison data from simulation experiments, Figure 5 This is a comparison chart of the quantitative performance of different methods in the task of weak target detection in side-scan sonar images. Figure 6 A visualization comparison of the results of weak target detection in side-scan sonar images under low signal-to-noise ratio scenarios using different methods is presented, showing the original image, the detection results of the original YOLO26 baseline model, the detection results of this application, and manually annotated ground truth values.

[0063] Figure 5 YOLOv8, RT-DETR, BGSC-Net, and YOLO26 were selected as comparison models. YOLOv8 is the latest generation of real-time object detection model. Built upon previous versions of YOLO, it introduces improvements such as anchor-free box design and decoupled head structure. YOLO stands for "You Only Look Once," referring to a one-stage real-time object detection algorithm that treats object detection as a single regression problem, completing detection with a single forward propagation. RT-DETR, or Real-Time Detection Transformer, is a real-time end-to-end object detection model that directly outputs detection results, completely eliminating the non-maximum suppression post-processing step. BGSC-Net is a hybrid model, short for Boundary Guided Semantic Compensation Network, which addresses issues such as poor segmentation of small objects, blurred object boundaries, and misalignment of cross-layer feature semantics. YOLO26 is the latest generation object detection model, released in September 2025 and officially open-sourced in January 2026. It is designed for edge devices and low-power hardware. Figure 5 Four evaluation criteria were selected: mean intersection-union ratio (mIoU), weak target recall (Recall_tiny), AP50, and AP75. AP50 refers to the average precision when the intersection-union ratio (IoU) threshold is set to 0.5, and AP75 refers to the average precision when the intersection-union ratio (IoU) threshold is set to 0.75.

[0064] The advantages of this application are mainly reflected in the following two aspects.

[0065] The recall rate for weak targets has been significantly improved. For pixels with an area smaller than [missing information], the recall rate is significantly improved. For faint targets, general visual detection models are prone to missing detection in side-scan sonar images. Figure 5 As can be seen, the performance of YOLOv8, RT-DETR, the original BGSC-Net, and the original YOLO26 in terms of weak target recall is limited by the scale of the weak target and acoustic noise. The weak target recall of the original YOLO26 is [missing data]. This application utilizes the SE-STAL mechanism to incorporate acoustic shadow regions as supplementary evidence for bright reflection regions in label allocation. This allows weak targets to obtain more positive samples and more sufficient gradient supervision during the training phase, thereby improving the recall rate of weak targets. This effectively reduced the false negative rate for small targets.

[0066] Noise resistance and positioning accuracy are improved simultaneously. Regarding positioning-related metrics, the average intersection-union ratio (IUU) of this application is significantly higher than that of the original YOLO26. Upgraded to AP75 is Upgraded to respectively improve and The results demonstrate that AA-CLSCM effectively utilizes the paired topological relationship of "high-brightness reflection areas - acoustic shadow areas" in side-scan sonar images to enhance the semantic response of real target regions during cross-layer feature fusion and suppress false detections caused by isolated bright noise and complex seabed reverberation. Figure 6 The visualization results show that the original YOLO26 baseline model has areas of missed detection of small targets and suspected false detections in low signal-to-noise ratio scenarios. However, this application can detect both bright targets and their weak accompanying targets at the same time. The detection results are more consistent with the manually labeled ground truth. This indicates that this application has better detection robustness and fine positioning capabilities in complex seabed environments.

[0067] Another embodiment of this application proposes a sonar weak target detection system based on AA-CLSCM and SE-STAL, suitable for complex underwater environments, to implement the sonar weak target detection method based on AA-CLSCM and SE-STAL as described in the above method embodiments. The implementation details of the sonar weak target detection system based on AA-CLSCM and SE-STAL proposed in this embodiment are described below. The following implementation details are provided for ease of understanding only and are not essential for implementing this solution. Figure 7This is a schematic diagram of the structure of a sonar weak target detection system based on AA-CLSCM and SE-STAL proposed in this embodiment, including: a sample set construction module 201, a model construction module 202, a model iterative training and update module 203, and a model usage module 204.

[0068] The sample set construction module 201 is used to acquire several real side-scan sonar images as sample data, perform double polygon fine labeling on the real targets for the bright reflection area and the acoustic shadow area, and construct the sample data after labeling the real annotation boxes into training sample set, validation sample set and test sample set according to a preset ratio.

[0069] The model building module 202 is used to build a sonar weak target detection model consisting of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head. The multi-scale feature extraction network is used to extract features from the input side-scan sonar images. In the top-down path of the feature pyramid, the multi-scale feature fusion network deploys the AA-CLSCM module at the cross-layer skip connection to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise. The decoupled detection head is used to output the category confidence and bounding box regression results of weak targets.

[0070] The model iterative training and update module 203 is used to train the sonar weak target detection model based on the training sample set. In the label allocation stage, based on the SE-STAL mechanism, the normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. The dynamic matching threshold is determined according to the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, the positive and negative samples are allocated according to the intersection-union ratio between the candidate predicted box and the real labeled box output by the sonar weak target detection model and the dynamic matching threshold. Then, the total loss function including classification loss and regression loss is calculated, and the network parameters of the sonar weak target detection model are updated by backpropagation until the preset convergence condition is met, and the trained sonar weak target detection model is obtained.

[0071] The model uses module 204 to input the side-scan sonar image to be detected into the trained sonar weak target detection model for detection, and obtain the weak target detection results output by it.

[0072] It is worth noting that all modules involved in this embodiment are logical modules. In practical applications, a logical module can be a physical module, a part of a physical module, or an organic combination of multiple physical modules. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce modules that are not closely related to solving the technical problems proposed in this application. However, this does not mean that other modules are absent from this embodiment.

[0073] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0074] Another embodiment of this application proposes an electronic device, the specific structure of which is as follows: Figure 8 As shown, the electronic device consists of a processor 301 and a memory 302. The memory 302 stores instructions that the processor 301 can execute. When the processor 301 is configured to execute the instructions, the electronic device can realize a sonar weak target detection method based on AA-CLSCM and SE-STAL as described in the above method embodiment.

[0075] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors and the memory. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which also receives and transmits data back to the processor. The processor manages the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. The memory can be used to store data used by the processor during operation.

[0076] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a sonar weak target detection method based on AA-CLSCM and SE-STAL as described in the above method embodiments.

[0077] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0078] It will be understood by those skilled in the art that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form, detail, and description without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A sonar-based weak target detection method based on AA-CLSCM and SE-STAL, suitable for complex underwater environments, characterized in that... The method includes: Several real side-scan sonar images were acquired as sample data. The real targets were labeled with double polygonal fine labels for the bright reflection area and the acoustic shadow area. The sample data after the real labeling boxes were constructed into training sample set, validation sample set and test sample set according to a preset ratio. A sonar weak target detection model is constructed, consisting of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head. The multi-scale feature extraction network is used to extract features from the input side-scan sonar images. In the top-down path of the feature pyramid, the multi-scale feature fusion network deploys the AA-CLSCM module at the cross-layer skip connection to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise. The decoupled detection head is used to output the category confidence and bounding box regression results of weak targets. The sonar weak target detection model is trained based on the training sample set. In the label allocation stage, based on the SE-STAL mechanism, the normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. The dynamic matching threshold is determined according to the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, the positive and negative samples are allocated according to the intersection-union ratio between the candidate predicted box and the real labeled box output by the sonar weak target detection model and the dynamic matching threshold. The total loss function including classification loss and regression loss is calculated, and the network parameters of the sonar weak target detection model are updated by backpropagation until the preset convergence condition is met, and the trained sonar weak target detection model is obtained. The side-scan sonar image to be detected is input into the trained sonar weak target detection model for detection, and the weak target detection result output by the model is obtained. The multi-scale feature fusion network has multiple AA-CLSCM modules, which are deployed at the cross-layer jump connections in the top-down path of the feature pyramid, in order to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise. In the AA-CLSCM module, deep semantic feature maps The semantic information is not broadcast globally. Instead, it is used to find highlight-shadow pairs through asymmetric convolution to generate a specific attention mask for the sonar target, and then accurately compensated across layers to the lower-level spatial feature map. ; The AA-CLSCM module achieves cross-layer compensation through the following steps: The AA-CLSCM module received and Afterwards, Applying asymmetric orientation-aware convolution, asymmetric orientation-aware convolution utilizes... Horizontal asymmetric convolution kernel pairs Perform cross-course scanning along the sound wave to extract one-dimensional gradient features of pixel intensity transitioning from bright reflections to acoustic shadows. ; use Generate adaptive guided weight mask Only when the bright reflective area and the acoustic shadow area appear in pairs, High response is output in the corresponding area, and isolated bright reflection areas and isolated acoustic shadow areas are suppressed; Will Transform Alignment to After the scale, with As a gating signal and To integrate and achieve Cross-layer compensation yields fusion features. ; This can be expressed by the formula: ； in, This indicates element-wise multiplication. It is the Sigmoid activation function. Indicates feature transformation alignment; In the label allocation stage, based on the SE-STAL mechanism, a normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflective area of ​​the real target and the acoustic shadow area associated with the bright reflective area. A dynamic matching threshold is then determined based on the target scale relaxation coefficient and the normalized shadow evidence coefficient. Finally, the allocation of positive and negative samples is completed based on the intersection-union ratio (IU) between the candidate predicted bounding boxes output by the sonar weak target detection model and the real labeled bounding boxes, and the dynamic matching threshold. This includes: During the label assignment phase, based on the SE-STAL mechanism, the area of ​​the highlight reflective region is obtained for each real target. and the area of ​​the acoustic shadow region associated with the bright reflection area. ; Based on the following formula, and Calculate the total area of ​​shadow augmentation. : ； in, This is the preset shadow gain coefficient; Based on the following formula, and Calculate the normalized shadow evidence coefficient : ； in, It is a stability constant greater than zero, used to prevent division by zero; Based on the target scale relaxation factor, the following formula is used. and Calculate the effective threshold relaxation factor : ； in, This is the preset gain coefficient for shadow evidence; Based on the following formula, Calculate the dynamic matching threshold : ； in, Based on the matching threshold, The total area of ​​the input side-scan sonar image; Calculate the intersection-union ratio (IU) between the candidate predicted bounding boxes and the ground truth bounding boxes output by the sonar weak target detection model. If the IU is not less than the dynamic matching threshold, the candidate predicted bounding box is assigned as a positive sample; otherwise, it is assigned as a negative sample. Among them, the area of ​​the bright reflection zone, the area of ​​the acoustic shadow zone, and the shadow evidence coefficient are only used for the dynamic matching threshold calculation in the label assignment stage, and do not change the bounding box regression results of the real target.

2. The sonar weak target detection method based on AA-CLSCM and SE-STAL according to claim 1, characterized in that, The process involves acquiring several real side-scan sonar images as sample data, performing fine-grained double-polygon labeling on the real targets in both bright reflective areas and acoustic shadow areas, and constructing the labeled sample data into training, validation, and test sample sets according to a preset ratio, including: Several real side-scan sonar images were acquired as sample data, and all sample data were preprocessed, including the introduction of underwater additive white Gaussian noise and sonar blind zone simulation enhancement. The real targets in the preprocessed sample data are labeled with double polygonal fine labels for the bright reflection area and the acoustic shadow area to obtain the sample data with the real label boxes labeled; The sample data after the ground truth bounding boxes are labeled are divided into training sample data, validation sample data and test sample data in a ratio of 7:2:1 to construct the training sample set, validation sample set and test sample set.

3. The sonar weak target detection method based on AA-CLSCM and SE-STAL according to claim 1, characterized in that, The multi-scale feature extraction network consists of five CNN layers. The first three CNN layers are shallow networks, and the last two CNN layers are deep networks. This is applied to input side-scan sonar images. Multi-scale feature extraction networks utilize shallow networks to extract low-level spatial feature maps rich in texture and boundary details. Deep semantic feature maps containing strong class information are extracted using deep networks. .

4. The sonar weak target detection method based on AA-CLSCM and SE-STAL according to claim 1, characterized in that, After each backpropagation update of the network parameters of the sonar weak target detection model, the updated sonar weak target detection model is validated based on the validation sample set to prevent overfitting and to adjust the hyperparameters and select the optimal version. After meeting the preset convergence conditions, the trained sonar weak target detection model is obtained. The performance of the trained sonar weak target detection model is evaluated based on the test sample set. The evaluation criteria are average intersection-union ratio, weak target recall, AP50 and AP75. If all performance evaluations are qualified, the trained sonar weak target detection model is allowed to be deployed; otherwise, it is retrained. Among them, the average intersection-union ratio is used to evaluate the degree of overlap between the predicted bounding box and the label, the weak target recall rate is used to evaluate the model's ability to detect weak targets, the AP50 is used to measure the model's detection performance under a loose IoU threshold, and the AP75 is used to measure the model's fine localization ability under a strict IoU threshold.

5. A sonar weak target detection system based on AA-CLSCM and SE-STAL, suitable for complex underwater environments, used to implement the sonar weak target detection method based on AA-CLSCM and SE-STAL as described in any one of claims 1 to 4, characterized in that, The system includes: The sample set construction module is used to acquire several real side-scan sonar images as sample data, perform double polygon fine labeling on the real targets for the bright reflection area and acoustic shadow area, and construct the sample data after labeling the real annotation boxes into training sample set, validation sample set and test sample set according to preset ratio. The model building module is used to construct a sonar weak target detection model consisting of a multi-scale feature extraction network, a multi-scale feature fusion network, and a decoupled detection head. The multi-scale feature extraction network is used to extract features from the input side-scan sonar images. In the top-down path of the feature pyramid, the multi-scale feature fusion network deploys the AA-CLSCM module at the cross-layer skip connections to enhance the edge and structural representation of weak targets while filtering complex underwater reverberation noise. The decoupled detection head is used to output the category confidence and bounding box regression results of weak targets. The model iterative training and update module is used to train the sonar weak target detection model based on the training sample set. In the label allocation stage, based on the SE-STAL mechanism, the normalized shadow evidence coefficient is calculated according to the area of ​​the bright reflection area of ​​the real target and the acoustic shadow area associated with the bright reflection area. The dynamic matching threshold is determined according to the target scale relaxation coefficient and the normalized shadow evidence coefficient. Then, the positive and negative samples are allocated according to the intersection-union ratio between the candidate predicted box and the real labeled box output by the sonar weak target detection model and the dynamic matching threshold. Then, the total loss function including classification loss and regression loss is calculated, and the network parameters of the sonar weak target detection model are updated by backpropagation until the preset convergence condition is met, and the trained sonar weak target detection model is obtained. The model uses a module to input the side-scan sonar image to be detected into the trained sonar weak target detection model for detection, and obtain the weak target detection results output by the model.

6. An electronic device comprising a processor and a memory, characterized in that, The memory stores instructions that the processor can execute, and when the processor is configured to execute the instructions, the electronic device is able to implement a sonar weak target detection method based on AA-CLSCM and SE-STAL as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement a sonar weak target detection method based on AA-CLSCM and SE-STAL as described in any one of claims 1 to 4.

Images