Underwater target detection method and system based on deep learning multi-beam sonar

Abstract

The application discloses a multi-beam sonar underwater target detection method and system based on deep learning, belongs to the technical field of underwater detection and artificial intelligence, and comprises the following steps: collecting original echo signals of an underwater detection area through a multi-beam sonar device, and completing signal denoising and distortion correction preprocessing.The application can accurately separate effective signals from interference signals in the original echo signals by adopting an adaptive Wiener filtering algorithm for denoising processing, and simultaneously updates a filtering gain coefficient through dynamic iteration, so that the denoising process can adaptively match the change of underwater environmental noise intensity.On the basis of denoising, signal correction is performed in combination with the underwater sound propagation time delay distortion law, further eliminating signal distortion caused by underwater medium non-uniformity and propagation distance difference in the sound signal propagation process, ensuring the purity and integrity of the echo signals after preprocessing, effectively improving the definition and contrast of the sonar image, and significantly reducing detection errors caused by signal interference and distortion.

Description

Technical Field

[0001] This invention belongs to the field of underwater detection and artificial intelligence technology, specifically referring to a method and system for underwater target detection using multi-beam sonar based on deep learning. Background Technology

[0002] Underwater target detection is one of the core technologies in marine engineering, underwater security, resource exploration and other fields. Its core requirement is to accurately identify the location, size and category of various underwater targets to provide reliable data support for subsequent operational decisions. Due to the complex and special underwater environment, severe light attenuation and high water turbidity, traditional optical detection methods are limited by underwater propagation conditions, resulting in short detection distances and low accuracy, making it difficult to meet the detection needs of large-scale and complex underwater environments.

[0003] However, existing multibeam sonar underwater target detection methods still have certain shortcomings. Current technologies generally suffer from weak anti-interference capabilities, insufficient detection accuracy, and inadequate environmental adaptability, making it difficult to meet the demands of complex underwater detection. The signal preprocessing stage lacks an effective combination of adaptive noise reduction and distortion correction, failing to adequately eliminate interference from underwater clutter and environmental noise, and also struggling to eliminate distortion caused by acoustic signal propagation delays, resulting in poor subsequent image quality. The construction of sonar image sample libraries suffers from problems such as scarce samples, uneven target distribution, and non-standard annotation, failing to provide high-quality support for model training. Deep learning detection networks have not been specifically improved for low-quality, low-signal-to-noise-ratio sonar images, lacking effective feature attention allocation and multi-scale feature fusion mechanisms, resulting in insufficient feature representation capabilities for small, blurred, and occluded targets, making them prone to errors. The current model suffers from insufficient feature extraction; the lack of a scientific iterative optimization and convergence verification mechanism during model training results in limited generalization ability; the initial detection results lack effective redundancy removal and false detection screening measures, resulting in a large number of overlapping redundant detection boxes and invalid false detection targets, leading to low accuracy of the final detection results and high false detection and false negative rates. Furthermore, the model lacks robustness, is easily affected by changes in the underwater environment and equipment differences, and is difficult to adapt to different underwater detection scenarios, failing to provide reliable technical support for underwater exploration, resource exploration, and other scenarios. Therefore, this paper proposes a deep learning-based multibeam sonar underwater target detection method and system. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for underwater target detection using multi-beam sonar based on deep learning, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-beam sonar underwater target detection method based on deep learning, comprising the following steps:

[0006] S1. Acquire the raw echo signal of the underwater detection area using a multibeam sonar device, and complete the signal noise reduction and distortion correction preprocessing.

[0007] S2. Perform beamforming processing on the preprocessed echo signal to generate a multi-beam sonar grayscale image dataset.

[0008] S3. Perform data augmentation, annotation, and normalization on the sonar image dataset to construct a standardized underwater sonar target sample library;

[0009] S4. Construct an improved deep learning object detection network that integrates channel attention and multi-scale feature fusion;

[0010] S5. Using the constructed sample library, iteratively train, optimize parameters, and verify convergence of the deep learning detection network to obtain the optimal object detection model;

[0011] S6. Input the sonar image to be detected into the optimal detection model to complete underwater target feature extraction, feature fusion and target prediction;

[0012] S7. Perform non-maximum suppression and confidence screening on the initial detection results output by the model, and output the final underwater target detection results.

[0013] Preferably, in step S1, a multi-beam sonar array is used to collect echo signals from the underwater detection area in all directions. The original echo signals contain interference signals such as water clutter, environmental noise, and equipment background noise. To eliminate signal distortion and noise interference caused by the complex underwater environment, an adaptive Wiener filtering algorithm is first used to perform signal denoising. The denoising is implemented as follows:

[0014] ,

[0015] In the formula, This represents the clean echo signal after noise reduction at time n. This represents the raw echo signal acquired by the multibeam sonar at time n. This represents the underwater environmental background noise signal at time n. Represents the adaptive filter gain coefficient; where the adaptive filter gain coefficient Through iterative solution, the gain is ensured to dynamically and adaptively adjust with the ambient noise intensity.

[0016] ,

[0017] In the formula, This represents the updated filter gain coefficient at time n+1. This represents the iteration step size constant, ranging from 0.01 to 0.1, used to control the gain update rate. After signal denoising, based on the underwater acoustic propagation time delay distortion law, combined with the pure echo signal... To complete the propagation distortion correction, the correction formula is:

[0018] ,

[0019] In the formula, This represents the final preprocessed echo signal after distortion correction, where f represents the center frequency of the multibeam sonar transmission signal. This represents the underwater propagation delay of the acoustic signal at time n; through layer-by-layer noise reduction, adaptive gain update, and distortion correction, signal interference caused by underwater turbulence, suspended matter, and seabed clutter is eliminated.

[0020] Preferably, in step S2, a digital beamforming algorithm is used to perform spatial filtering and beam pointing and focusing on the multi-channel echo signal to achieve two-dimensional imaging of the underwater detection area. The single-beam output signal strength is achieved as follows:

[0021] ,

[0022] In the formula, Indicates the beam pointing angle is The echo signal strength at time N represents the number of elements in the sonar array. Let represent the weighting coefficient of the i-th element. This represents the echo signal after preprocessing of the i-th array element. The distance between the i-th array element and the reference array element is represented by c, and the underwater acoustic propagation speed is represented by c; based on the beam signal strength at each angle. Complete full-domain beam scanning imaging and generate an initial sonar grayscale image. The image pixel grayscale values ​​are implemented as follows:

[0023] ,

[0024] In the formula, This represents the pixel grayscale value at image coordinates (x, y). , These represent the maximum and minimum values ​​of the beam signal intensity within a single frame of imaging; continuous imaging is performed at a fixed sampling frame rate to generate batches of original grayscale images of multi-beam sonar, which are then aggregated to form an initial sonar image dataset.

[0025] Preferably, in step S3, the generated initial sonar image dataset is augmented by random flipping, grayscale fine-tuning, Gaussian blurring, and local cropping to increase the number of samples. Simultaneously, the LabelImg tool is used to accurately annotate all images, with annotation categories including common underwater targets such as underwater reefs, shipwrecks, underwater robots, fish, and obstacles, generating annotation files containing target bounding box coordinates, categories, and confidence levels. After annotation, the images are uniformly normalized. Based on the normalized image data and corresponding annotation files, a training set, a validation set, and a test set are divided in a ratio of 7:2:1, ultimately constructing a standardized underwater sonar target sample library with balanced samples, unified format, and accurate annotation.

[0026] Preferably, in step S4, a single-stage YOLO basic detection network is used as the main framework. The network backbone, feature fusion layer, and detection head are improved to build a detection network adapted to low-quality, low-signal-to-noise ratio sonar images. The backbone network introduces an efficient layer aggregation module to complete the extraction of shallow texture features and deep semantic features. At the same time, a channel attention module is embedded to adaptively allocate the weights of different feature channels. The channel weights are implemented as follows:

[0027] ,

[0028] In the formula, This represents the global average pooling value of the c-th feature channel. This represents the feature pixel value at coordinates (h, w) in the c-th channel, where H and W represent the height and width of the feature map, respectively.

[0029] Generate channel weight coefficients based on the pooling results:

[0030] ,

[0031] In the formula, This represents the weight coefficient of the c-th feature channel. This represents the weight parameters of the fully connected layer in the attention network. Represents the ReLU activation function. This represents the Sigmoid activation function; ultimately, feature weighted optimization is completed.

[0032] ,

[0033] In the formula, This represents the weighted and optimized feature map; at the same time, a bidirectional feature pyramid fusion module is constructed to complete the splicing and fusion of multi-scale feature maps, thereby enhancing the feature representation capabilities of small and fuzzy targets.

[0034] Preferably, in step S5, the constructed standardized sample library is input into the improved deep learning detection network, and gradient descent is used to complete iterative training. During the training process, the model detection error is evaluated in real time through the loss function. The overall loss function consists of coordinate loss, confidence loss and class loss. After each iteration, the model precision, recall and mAP value are calculated based on the validation set data, and the network weight parameters are continuously updated through backpropagation.

[0035] Preferably, in step S6, a multibeam sonar image of the area to be detected is acquired, and after preprocessing and imaging processing, it is input into the trained optimal detection model. The backbone network of the model completes the basic feature extraction of the image to be detected and outputs a multi-scale original feature map. The channel attention module optimizes the weights of the original feature map to enhance effective target features and suppress ineffective features of water clutter. The bidirectional feature pyramid module completes the cross-layer fusion of shallow detail features and deep semantic features to obtain a highly recognizable fused feature map. The detection head performs traversal prediction on the fused feature map and outputs the class probability of potential targets in the image, the predicted bounding box coordinates, and the initial confidence score, thus completing the preliminary identification and localization of underwater multi-scale targets.

[0036] Preferably, in step S7, for the multiple sets of output initial target prediction boxes, non-maximum suppression is used to remove overlapping redundant detection boxes, and the intersection-union ratio (CIU) of any two prediction boxes is calculated. The CIU is implemented as follows:

[0037] ,

[0038] In the formula, This represents the intersection-union ratio of prediction boxes B1 and B2. This represents the area of ​​the overlapping region between the two prediction boxes. This represents the total area of ​​the merged region of the two prediction boxes.

[0039] Preferably, in step S7, an intersection-union ratio (IU) threshold is set. If the IU of two predicted boxes is greater than the threshold, redundant detection boxes with low confidence are removed. At the same time, a target confidence screening threshold is set to remove invalid false detection targets with confidence below the threshold. Finally, the target detection box with the highest accuracy and the target classification result are retained, and the final underwater target detection result containing the target location, size, and category is output.

[0040] Preferably, a deep learning-based multibeam sonar underwater target detection system includes a sonar signal acquisition module, a signal preprocessing module, a sonar imaging module, a dataset construction module, a deep learning model training module, an underwater target detection module, and a result optimization output module.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. To address the technical challenges of existing multibeam sonar underwater target detection, such as strong signal interference, poor imaging quality, unbalanced sonar image samples, weakened target features, low accuracy and high false positive / false negative rates in multi-scale and weak target identification, and to improve the accuracy, stability, and environmental adaptability of underwater target detection, this invention combines adaptive signal processing and improved deep learning technology to achieve efficient and accurate detection of multiple types and scales of targets in complex underwater environments. This effectively overcomes the shortcomings of weak anti-interference capability and poor adaptability, providing reliable technical support for underwater exploration, marine resource exploration, underwater security, and other scenarios, significantly improving the efficiency and practicality of underwater target detection.

[0043] 2. This invention employs an adaptive Wiener filtering algorithm for noise reduction, which can accurately separate the effective signal from the interference signal in the original echo signal. Simultaneously, by dynamically iteratively updating the filter gain coefficient, the noise reduction process can adaptively match changes in underwater environmental noise intensity, avoiding the problems of incomplete noise reduction or loss of effective signal caused by fixed-gain filtering. Based on noise reduction, signal correction is performed by combining the underwater sound propagation time delay distortion law, further eliminating signal distortion caused by the inhomogeneity of the underwater medium and differences in propagation distance during sound signal propagation. This ensures the purity and integrity of the pre-processed echo signal, effectively improving the clarity and contrast of the sonar image and significantly reducing detection errors caused by signal interference and distortion.

[0044] 3. This invention, based on a single-stage detection network, introduces an efficient layer aggregation module to fully extract shallow texture features and deep semantic features from sonar images, taking into account both target detail information and global features. The embedded channel attention module adaptively allocates weights to different feature channels, strengthening the representation of effective target features and suppressing interference from invalid features such as water clutter, thus improving feature recognition. Simultaneously, the construction of a bidirectional feature pyramid fusion module enables cross-layer splicing and fusion of multi-scale feature maps, effectively solving the problem of large differences in underwater target scale, enhancing the feature extraction capabilities for small and blurred targets, and enabling the network to accurately capture underwater targets of different sizes and resolutions. The dedicated detection network built in this step significantly improves adaptability to complex underwater environments and substantially enhances the accuracy and generalization ability of underwater target detection.

[0045] 4. This invention employs a non-maximum suppression algorithm to filter predicted bounding boxes with high overlap, eliminating redundant detection boxes and retaining the most accurate target predicted boxes. This avoids the problems of repeated target counting and inaccurate positioning caused by multiple overlapping boxes. Simultaneously, by setting a confidence threshold, it effectively eliminates invalid false detections with low confidence, reducing false detections caused by water clutter, image blur, and other factors, further improving the accuracy of the detection results. Through dual-filter optimization, the final output detection results are more accurate and clear, accurately presenting the location, size, and category information of underwater targets, effectively reducing the model's false detection rate and false negative rate. Attached Figure Description

[0046] Figure 1 The following is the operational flow of the deep learning-based multibeam sonar underwater target detection method of this invention. Figure 1 ;

[0047] Figure 2 The following is the operational flow of the deep learning-based multibeam sonar underwater target detection method of this invention. Figure 2 ;

[0048] Figure 3 The following is the operational flow of the deep learning-based multibeam sonar underwater target detection method of this invention. Figure 3 ;

[0049] Figure 4 The following is the operational flow of the deep learning-based multibeam sonar underwater target detection method of this invention. Figure 4 ;

[0050] Figure 5 This is a schematic diagram of the structure of the deep learning-based multibeam sonar underwater target detection system of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example

[0053] Please see Figures 1-5 As shown, the present invention provides a technical solution comprising the following steps:

[0054] S1. Acquire the raw echo signal of the underwater detection area using a multibeam sonar device, and complete the signal noise reduction and distortion correction preprocessing.

[0055] S2. Perform beamforming processing on the preprocessed echo signal to generate a multi-beam sonar grayscale image dataset.

[0056] S3. Perform data augmentation, annotation, and normalization on the sonar image dataset to construct a standardized underwater sonar target sample library;

[0057] S4. Construct an improved deep learning object detection network that integrates channel attention and multi-scale feature fusion;

[0058] S5. Using the constructed sample library, iteratively train, optimize parameters, and verify convergence of the deep learning detection network to obtain the optimal object detection model;

[0059] S6. Input the sonar image to be detected into the optimal detection model to complete underwater target feature extraction, feature fusion and target prediction;

[0060] S7. Perform non-maximum suppression and confidence screening on the initial detection results output by the model, and output the final underwater target detection results.

[0061] In this embodiment, in step S1, a multi-beam sonar array is used to collect echo signals from the underwater detection area in all directions. The original echo signals contain interference signals such as water clutter, environmental noise, and equipment background noise. To eliminate signal distortion and noise interference caused by the complex underwater environment, an adaptive Wiener filtering algorithm is first used to complete signal denoising. The denoising is implemented as follows:

[0062] ,

[0063] In the formula, This represents the clean echo signal after noise reduction at time n. This represents the raw echo signal acquired by the multibeam sonar at time n. This represents the underwater environmental background noise signal at time n. Represents the adaptive filter gain coefficient; where the adaptive filter gain coefficient Through iterative solution, the gain is ensured to dynamically and adaptively adjust with the ambient noise intensity.

[0064] ,

[0065] In the formula, This represents the updated filter gain coefficient at time n+1. This represents the iteration step size constant, ranging from 0.01 to 0.1, used to control the gain update rate. After signal denoising, based on the underwater acoustic propagation time delay distortion law, combined with the pure echo signal... To complete the propagation distortion correction, the correction formula is:

[0066] ,

[0067] In the formula, This represents the final preprocessed echo signal after distortion correction, where f represents the center frequency of the multibeam sonar transmission signal. This represents the underwater propagation delay of the acoustic signal at time n. Through layer-by-layer noise reduction, adaptive gain update, and distortion correction, signal interference caused by underwater turbulence, suspended objects, and seabed clutter is eliminated, ensuring the accuracy and clarity of subsequent beam imaging.

[0068] In this embodiment, in step S2, a digital beamforming algorithm is used to perform spatial filtering and beam pointing and focusing on the multi-channel echo signal to achieve two-dimensional imaging of the underwater detection area. The single-beam output signal strength is achieved as follows:

[0069] ,

[0070] In the formula, Indicates the beam pointing angle is The echo signal strength at time N represents the number of elements in the sonar array. Let represent the weighting coefficient of the i-th element. This represents the echo signal after preprocessing of the i-th array element. The distance between the i-th array element and the reference array element is represented by c, and the underwater acoustic propagation speed is represented by c; based on the beam signal strength at each angle. Complete full-domain beam scanning imaging and generate an initial sonar grayscale image. The image pixel grayscale values ​​are implemented as follows:

[0071] ,

[0072] In the formula, This represents the pixel grayscale value at image coordinates (x, y). , These represent the maximum and minimum values ​​of the beam signal intensity within a single frame of imaging. Continuous imaging is performed at a fixed sampling frame rate to generate batches of original grayscale images of multi-beam sonar. These images are then aggregated to form an initial sonar image dataset, ensuring complete coverage of the underwater detection area and preserving key features such as target contours, textures, and sizes.

[0073] In this embodiment, in step S3, to address the issues of scarce underwater sonar image samples, imbalanced target samples, and low image contrast, the generated initial sonar image dataset is augmented by random flipping, grayscale fine-tuning, Gaussian blurring, and local cropping to increase the number of samples. Simultaneously, the LabelImg tool is used to accurately annotate all images, with annotation categories including common underwater targets such as underwater reefs, shipwrecks, underwater robots, fish, and obstacles, generating an annotation file containing target bounding box coordinates, categories, and confidence scores. After annotation, the images undergo unified normalization, which is implemented as follows:

[0074] ,

[0075] In the formula, This represents the normalized pixel value. Represents the grayscale value of the original image pixels. This represents the average grayscale value of all pixels in a single image. This represents the standard deviation of pixel grayscale in a single image. Based on the normalized image data and corresponding annotation files, the training set, validation set, and test set are divided in a ratio of 7:2:1. Finally, a standardized underwater sonar target sample library with balanced samples, unified format, and accurate annotation is constructed to provide data support for subsequent network training.

[0076] In this embodiment, in step S4, a single-stage YOLO basic detection network is used as the main framework. The network backbone, feature fusion layer, and detection head are improved to build a detection network adapted to low-quality, low-signal-to-noise ratio sonar images. The backbone network introduces an efficient layer aggregation module to complete the extraction of shallow texture features and deep semantic features. At the same time, a channel attention module is embedded to adaptively allocate the weights of different feature channels. The channel weights are implemented as follows:

[0077] ,

[0078] In the formula, This represents the global average pooling value of the c-th feature channel. This represents the feature pixel value at coordinates (h, w) in the c-th channel, where H and W represent the height and width of the feature map, respectively.

[0079] Generate channel weight coefficients based on the pooling results:

[0080] ,

[0081] In the formula, This represents the weight coefficient of the c-th feature channel. This represents the weight parameters of the fully connected layer in the attention network. Represents the ReLU activation function. This represents the Sigmoid activation function; ultimately, feature weighted optimization is completed.

[0082] ,

[0083] In the formula, This represents the weighted and optimized feature map; at the same time, a bidirectional feature pyramid fusion module is constructed to complete the splicing and fusion of multi-scale feature maps, enhance the feature representation capability of small and fuzzy targets, and solve the detection problem of underwater sonar targets with large scale differences and weakened features.

[0084] In this embodiment, in step S5, the constructed standardized sample library is input into the improved deep learning detection network, and gradient descent is used to complete iterative training. During the training process, the model detection error is evaluated in real time through a loss function. The overall loss function consists of coordinate loss, confidence loss, and class loss, and is implemented as follows:

[0085] ,

[0086] In the formula, This represents the overall training loss value of the model. This represents the loss of the target bounding box coordinates. Indicates the target confidence loss. This represents the target classification loss; the bounding box coordinate loss is calculated using GIoU loss.

[0087] ,

[0088] In the formula, This indicates that the model predicts the target bounding box. This indicates that the bounding boxes of real targets are manually labeled. During training, the initial learning rate is set to 0.001, the batch size is 16, and the maximum number of iterations is 300. After each iteration, the model's precision, recall, and mAP value are calculated based on the validation set data, and the network weight parameters are continuously updated through backpropagation. When the loss value no longer decreases and the mAP value tends to stabilize after 20 consecutive iterations, the model is considered to have converged, and the deep learning underwater target detection model with the optimal weight parameters is saved.

[0089] In this embodiment, in step S6, multibeam sonar images of the area to be detected are acquired. After preprocessing and imaging, these images are input into the trained optimal detection model. The backbone network of the model extracts the basic features of the image to be detected and outputs a multi-scale original feature map. The channel attention module optimizes the weights of the original feature map to enhance effective target features and suppress ineffective features from water clutter. The bidirectional feature pyramid module performs cross-layer fusion of shallow detail features and deep semantic features to obtain a highly recognizable fused feature map. The detection head performs traversal prediction on the fused feature map and outputs the class probability of potential targets in the image, the predicted bounding box coordinates, and the initial confidence level. This completes the preliminary identification and localization of underwater multi-scale targets, enabling the effective capture of weak and occluded targets in complex underwater environments.

[0090] In this embodiment, in step S7, non-maximum suppression is used to remove overlapping redundant detection boxes from the multiple sets of output initial target prediction boxes, and the intersection-union ratio (CIU) of any two prediction boxes is calculated. The CIU is implemented as follows:

[0091] ,

[0092] In the formula, This represents the intersection-union ratio of prediction boxes B1 and B2. This represents the area of ​​the overlapping region between the two prediction boxes. This represents the total area of ​​the merged region of the two prediction boxes.

[0093] In this embodiment, in step S7, an intersection-union ratio (IU) threshold is set. If the IU of two predicted boxes is greater than the threshold, redundant detection boxes with low confidence are removed. At the same time, a target confidence screening threshold is set to remove invalid false detection targets with confidence below the threshold. Finally, the target detection box with the highest accuracy and the target classification result are retained, and the final underwater target detection result containing the target location, size, and category is output, effectively reducing the model's false detection rate and false negative rate.

[0094] In this embodiment, the deep learning-based multibeam sonar underwater target detection system includes a sonar signal acquisition module, a signal preprocessing module, a sonar imaging module, a dataset construction module, a deep learning model training module, an underwater target detection module, and a result optimization output module.

[0095] Working Principle: A multi-beam sonar array performs a 360-degree scan of the underwater detection area, capturing the raw echo signals reflected from underwater targets and the environment. This signal is contaminated with water clutter, environmental noise, and the equipment's own background noise, and is also distorted due to time delay differences during underwater propagation. To eliminate these adverse effects, an adaptive Wiener filtering algorithm is first used to denoise the raw echo signal. By dynamically adjusting the filter gain coefficient, the denoising process adapts to real-time changes in underwater environmental noise, accurately separating the effective signal from the interference signal. Then, considering the time delay distortion characteristics of underwater sound propagation, distortion correction is applied to the denoised pure echo signal, eliminating signal distortion caused by inhomogeneity of the medium and different propagation distances during sound propagation. Finally, a pure, distortion-free pre-processed echo signal is obtained. A digital beamforming algorithm is used to spatially filter the multi-channel pre-processed echo signal. By weighting the signals of each array element and focusing them at a specific beam pointing angle, directional scanning and signal enhancement of the underwater detection area are achieved. The signal intensity of the beam pointing angle is collected and integrated to complete the full-domain beam scanning. The intensity information of the echo signal is converted into image pixel grayscale values ​​to generate the initial multi-beam sonar grayscale image. Imaging operations are continuously performed at a fixed sampling frame rate to generate original sonar grayscale images in batches and summarize them into a dataset. To address the insufficiency of the initial sonar image dataset, various data augmentation methods are used to expand the number of samples. At the same time, image grayscale is adjusted, slight blurring or local cropping is added to improve the diversity of samples and the generalization ability of the model. Professional annotation tools are used to accurately annotate all sonar images, clarifying the specific information of various underwater targets in the images and generating corresponding annotation files to ensure the accuracy and completeness of the annotation information. The annotated images are uniformly normalized to eliminate grayscale differences between different images, so that the image pixel values ​​are within a uniform range. The processed image data and annotation files are then divided into training set, validation set and test set according to the ratio, finally constructing a standardized underwater sonar target sample library with balanced samples, uniform format and accurate annotation.

[0096] Based on a single-stage detection network as the main framework, and considering the characteristics of sonar images and the needs of underwater target detection, targeted improvements are made to the network backbone, feature fusion layer, and detection head. An efficient layer aggregation module is introduced into the network backbone to fully extract shallow texture features and deep semantic features from sonar images, taking into account both detailed information and global features of the target, avoiding the problem of insufficient target representation caused by single feature extraction. A channel attention module is embedded to adaptively allocate weights to different feature channels, strengthening the representation of effective target features and suppressing interference from invalid features such as water clutter, thereby improving feature recognition accuracy. Simultaneously, a bidirectional feature pyramid fusion module is constructed to achieve cross-layer stitching and fusion of multi-scale feature maps, enhancing the feature extraction capabilities for small and blurred targets. The researchers ultimately built an improved deep learning object detection network adapted to underwater sonar images. A standardized sample library was input into the improved deep learning detection network, and gradient descent was used for iterative training. During training, the model's detection error was evaluated in real time using a loss function that comprehensively considered target bounding box localization error, target confidence error, and target classification error, thus fully reflecting the model's detection performance. After each iteration, the model's detection performance was evaluated using validation set data, and relevant performance indicators were calculated. Based on the evaluation results, the network's weight parameters were updated using the backpropagation algorithm to continuously optimize the model's performance. Through continuous iterative training, when the model's loss value no longer decreased and the performance indicators tended to stabilize, the model was considered to have reached convergence. The model with the optimal weight parameters is saved at this point. Multibeam sonar images of the area to be detected are acquired, and preprocessed and beamforming images are processed to ensure the quality of the images is consistent with the training sample images. The processed images are then input into the optimal detection model. The model's backbone network first extracts basic features from the images, outputting multi-scale raw feature maps. Subsequently, the channel attention module optimizes the weights of the raw feature maps, strengthening effective target features and suppressing interference from ineffective features such as water clutter. Then, the bidirectional feature pyramid fusion module completes the cross-layer fusion of shallow detail features and deep semantic features, obtaining a more recognizable fused feature map. Finally, the detection head performs traversal prediction on the fused feature map, outputting the potential underwater targets in the image. The system uses category probability, predicted bounding box coordinates, and initial confidence scores to perform preliminary identification and localization of underwater multi-scale targets. For the multiple sets of initial target prediction boxes, a non-maximum suppression algorithm is used to filter out prediction boxes with high overlap, eliminating redundant prediction boxes and retaining those that most accurately reflect the target location, thus avoiding issues such as duplicate target counting and inaccurate localization. Simultaneously, a target confidence screening threshold is set to judge the target confidence in the initial detection results, eliminating invalid false detections with low confidence scores and reducing false detections caused by water clutter, image blur, and other factors. Through this dual screening optimization, the system ultimately retains the target detection boxes and target classification results with the highest accuracy, outputting the final detection results containing the underwater target's location, size, and category.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0098] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-based multibeam sonar underwater target detection method, characterized in that, Includes the following steps: S1. Acquire the raw echo signal of the underwater detection area using a multibeam sonar device, and complete the signal noise reduction and distortion correction preprocessing. S2. Perform beamforming processing on the preprocessed echo signal to generate a multi-beam sonar grayscale image dataset. S3. Perform data augmentation, annotation, and normalization on the sonar image dataset to construct a standardized underwater sonar target sample library; S4. Construct an improved deep learning object detection network that integrates channel attention and multi-scale feature fusion; S5. Using the constructed sample library, iteratively train, optimize parameters, and verify convergence of the deep learning detection network to obtain the optimal object detection model; S6. Input the sonar image to be detected into the optimal detection model to complete underwater target feature extraction, feature fusion and target prediction; S7. Perform non-maximum suppression and confidence screening on the initial detection results output by the model, and output the final underwater target detection results.

2. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In S1, a multi-beam sonar array is used to collect echo signals from the underwater detection area in all directions. An adaptive Wiener filtering algorithm is used to complete the signal noise reduction process. Through layer-by-layer noise reduction, adaptive gain update, and distortion correction, signal interference is eliminated.

3. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In step S2, a digital beamforming algorithm is used to perform spatial filtering and beam pointing and focusing on the multi-channel echo signal to achieve two-dimensional imaging of the underwater detection area. The single-beam output signal strength is achieved as follows: ， In the formula, Indicates the beam pointing angle is The echo signal strength at time N represents the number of elements in the sonar array. Let represent the weighting coefficient of the i-th element. This represents the echo signal after preprocessing of the i-th array element. The distance between the i-th array element and the reference array element is represented by c, and the underwater acoustic propagation speed is represented by c; based on the beam signal strength at each angle. Complete full-domain beam scanning imaging and generate an initial sonar grayscale image. By continuously imaging at a fixed sampling frame rate, batches of original grayscale images of multibeam sonar are generated and summarized to form an initial sonar image dataset.

4. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In step S3, the generated initial sonar image dataset is sample augmented by randomly flipping, fine-tuning grayscale, Gaussian blurring, and local cropping to increase the number of samples, label common underwater targets, and generate a labeling file containing target bounding box coordinates, category, and confidence level. After the annotation is completed, the images are uniformly normalized. Based on the normalized image data and the corresponding annotation files, the training set, validation set and test set are divided, and finally a standardized underwater sonar target sample library with balanced samples, uniform format and accurate annotation is constructed.

5. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In S4, a single-stage YOLO basic detection network is used as the main framework. The network backbone, feature fusion layer, and detection head are improved to build a detection network adapted to low-quality and low signal-to-noise ratio sonar images. The backbone network introduces an efficient layer aggregation module to complete the extraction of shallow texture features and deep semantic features. At the same time, a channel attention module is embedded to adaptively allocate the weights of different feature channels. Based on the pooling results, channel weight coefficients are generated to finally complete feature weighting optimization. Meanwhile, a bidirectional feature pyramid fusion module is constructed to complete the splicing and fusion of multi-scale feature maps and enhance the feature representation capabilities of small and blurred targets.

6. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In step S5, the standardized sample library is input into the improved deep learning detection network, and gradient descent is used to complete iterative training. During the training process, the model detection error is evaluated in real time through the loss function. The overall loss function consists of coordinate loss, confidence loss and class loss. After each iteration, the model precision, recall and mAP value are calculated based on the validation set data, and the network weight parameters are continuously updated through backpropagation.

7. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In step S6, multibeam sonar images of the area to be detected are acquired. After preprocessing and imaging, these images are input into the trained optimal detection model. The backbone network of the model extracts the basic features of the image to be detected and outputs a multi-scale original feature map. The channel attention module optimizes the weights of the original feature map to enhance effective target features and suppress ineffective features from water clutter. The bidirectional feature pyramid module performs cross-layer fusion of shallow detail features and deep semantic features to obtain a highly recognizable fused feature map. The detection head performs traversal prediction on the fused feature map and outputs the class probability of potential targets in the image, the predicted bounding box coordinates, and the initial confidence level, thus completing the preliminary identification and localization of underwater multi-scale targets.

8. The deep learning-based multibeam sonar underwater target detection method according to claim 1, characterized in that: In step S7, non-maximum suppression is used to remove overlapping redundant detection boxes from the multiple sets of output initial target prediction boxes, and the intersection-union ratio of any two prediction boxes is calculated.

9. The deep learning-based multibeam sonar underwater target detection method according to claim 8, characterized in that: In step S7, an intersection-union ratio (IU) threshold is set. If the IU of two predicted boxes is greater than the threshold, redundant detection boxes with low confidence are removed. At the same time, a target confidence screening threshold is set to remove invalid false detection targets with confidence below the threshold. Finally, the target detection box with the highest accuracy and the target classification result are retained, and the final underwater target detection result containing the target location, size, and category is output.

10. The deep learning-based multibeam sonar underwater target detection system implemented according to the deep learning-based multibeam sonar underwater target detection method of claim 1, characterized in that: It includes a sonar signal acquisition module, a signal preprocessing module, a sonar imaging module, a dataset construction module, a deep learning model training module, an underwater target detection module, and a result optimization and output module.

Images