Underwater sonar image buried pipe cable target detection method

Abstract

The present application belongs to the technical field of underwater sonar image detection, and specifically relates to a kind of underwater sonar image buried pipe cable target detection method.Firstly, the framework of the underwater buried pipe cable target detector is constructed, including three main networks: adaptive feature extractor, channel enhancement pyramid and adaptive cascade detection head;Then, the target detector is trained through classification loss and dynamic regression loss;Finally, the buried pipe cable target in the underwater sonar image to be detected is identified;The present application uses convolutional neural network to adaptively extract the effective feature map of the underwater buried pipe cable sonar image, then inputs the feature map into the pyramid feature fusion enhancement, and finally sends it into the adaptive cascade detection head to complete classification and positioning.The present application has universality, can accurately identify the buried target in the underwater sonar image, and obtain excellent results from multiple evaluation standards.

Description

Technical Field

[0001] This invention belongs to the field of underwater sonar image detection technology, specifically relating to a method for detecting buried pipe and cable targets in underwater sonar images. Background Technology

[0002] Underwater sonar imaging is achieved by an active sonar system transmitting sound waves into the underwater space and receiving echo information scattered or reflected by target objects, ultimately forming a sonar image representing the underwater scene. Underwater sonar image target detection refers to the automatic and accurate identification and location of targets from sonar images. In specific tasks such as submarine pipeline inspection, shipwreck salvage, mine detection, and marine archaeology, target detection technology is a crucial factor in ensuring operational safety. The detection of buried targets on the seabed is also vital for marine safety. For example, buried pipelines are the lifeline of energy supply at sea; regularly inspecting their condition to prevent damage and breakage is essential, playing an irreplaceable role in preventing leaks and ensuring the safety of energy transmission. Deep learning-based networks can effectively extract the texture and semantic information of targets for locating and identifying buried pipeline targets, making the research of more advanced deep learning detection networks crucial.

[0003] Existing underwater sonar image detection methods often focus on underwater targets above the seabed, where the target edges are relatively clear and the background noise is high but not similar in height to the target. However, buried pipe and cable targets are partially or completely covered by sediment, resulting in extremely weak acoustic echo signals that are highly mixed with background clutter. In addition, the targets are easily deformed by the pressure of sediment, resulting in imaging characteristics of low signal-to-noise ratio, low contrast, and high deformation. This poses a great challenge to the detection of buried pipe and cable targets in underwater sonar images. Summary of the Invention

[0004] Therefore, the purpose of this invention is to propose an underwater sonar image buried cable target detection method with high identification accuracy, so as to achieve effective identification of underwater buried cable objects.

[0005] The present invention provides a method for detecting buried pipe and cable targets in underwater sonar images, comprising constructing an underwater buried pipe and cable target detector, specifically including three main networks: an adaptive feature extractor, a channel enhancement pyramid, and an adaptive cascaded detection head; then, training the target detector using classification loss and dynamic weighted regression loss; finally, identifying buried pipe and cable targets in the underwater sonar image to be detected; wherein:

[0006] The adaptive feature extractor includes an inverse residual module and a deformable feature extraction module, and outputs a multi-scale basic feature map; the inverse residual module is used to extract basic features such as texture of the target in the sonar image, and the deformable feature extraction module is used to extract the complex deformation features of the target;

[0007] The channel enhancement pyramid includes a dual-branch channel attention mechanism module and a top-down feature pyramid structure, and outputs a fused and enhanced multi-level feature map.

[0008] The dual-branch channel attention mechanism module is used to modulate the channel weights of the high-level feature map to enhance key features and suppress background interference; the feature pyramid structure is used to achieve multi-scale feature fusion.

[0009] The adaptive cascaded detection head is used to predict the target category and location of the feature map generated by the feature pyramid, and output the confidence score of each target.

[0010] Furthermore, in the adaptive feature extractor of the present invention:

[0011] The inverse residual module consists of an up-dimensional convolutional layer, a depthwise separable convolutional layer, a down-dimensional convolutional layer, and an adaptive residual connection; the deformable feature extraction module consists of an up-dimensional convolutional layer, a deformable convolutional layer, a down-dimensional convolutional layer, and an adaptive scaling connection. (See [link to relevant documentation]). Figure 2 ;

[0012] Furthermore, in the channel enhancement pyramid of the present invention:

[0013] The dual-branch channel attention module structure includes: a global average pooling layer followed by two parallel branches. Branch one is a one-dimensional convolutional layer, and branch two consists of channel rearrangement, one-dimensional convolution, and inverse channel rearrangement. It also includes: concat connections, fused convolutional layers, a sigmoid activation function, and weighted multiplication. (See [link]). Figure 3 .

[0014] Furthermore, the adaptive cascaded detection head of the present invention consists of three cascaded detection stages, each stage including an RoI pooling layer, an RCNN detection layer, and a fully connected head. This cascaded structure improves performance stage by stage. Thresholds are used to optimize the regression boxes layer by layer, thereby achieving adaptive improvement in the accuracy of object detection.

[0015] The dynamic weighted regression loss is defined as:

[0016] ,

[0017] in, , , This regression loss function enables the detector head to adaptively adjust the regression strategy, thereby locating the target more accurately.

[0018] The underwater sonar image buried cable target detection method provided by this invention includes the following specific steps:

[0019] Step 1: Construct an underwater buried pipe and cable target detector.

[0020] The adaptive feature extractor takes underwater sonar images as input. In the first three stages, it uses an inverse residual module to extract shallow features. In the fourth stage, it uses a deformable feature extraction module to extract deformation features of the buried target. These four stages contain 3, 3, 9, and 3 feature extraction modules, respectively. The inverse residual module processes the image sequentially using increasing-dimensional convolutional layers, depthwise separable convolutional layers, decreasing-dimensional convolutional layers, and adaptive residual connections. The deformable feature extraction module processes the image matrix sequentially using increasing-dimensional convolutional layers, deformable convolutional layers, decreasing-dimensional convolutional layers, and adaptive scaling. After processing by the four stages of the network, each stage outputs a set of basic feature maps. .

[0021] The aforementioned channel-enhanced pyramid to As input to the feature pyramid, the feature map is obtained through a four-layer convolutional network structure within the feature pyramid. ;

[0022] Channel enhancement is performed using a dual-branch channel attention mechanism. The specific steps are as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] A global average pooling layer is used to obtain channel feature vectors. These channel vectors are then input into two branches simultaneously. Branch one extracts local channel relationships through one-dimensional convolution, while branch two performs channel rearrangement, one-dimensional convolution, and inverse rearrangement operations to obtain cross-channel interaction information. The outputs of the two branches are concatenated, then fused through a convolutional layer, and channel attention weights are generated using the Sigmad function. These attention weights are then compared with the feature maps. Multiplication yields the enhanced feature map ; feature map Upsampling is performed through a convolutional network layer, and then combined with... Added output feature map ; feature map Upsampling is performed through a convolutional network layer, and then combined with... Added output feature map ; feature map Upsampling is performed through a convolutional network layer, and then combined with... Added output feature map Finally, the pyramid-level feature map was obtained. ;

[0023] The adaptive cascaded detection head is based on As input, the target is adaptively optimized step-by-step through three cascaded detection stages. Specifically, the feature map... , , and Perform the same operation as below, each stage performs RoI pooling and RCNN detection based on the bounding boxes output by the previous stage, and gradually increases the intersection-union threshold as a localization requirement. Finally, output the category, location and confidence prediction value of the target in each image through two convolutions.

[0024] Step 2: Train the object detector using classification loss and dynamic weighted regression loss. Calculate the classification loss based on the predicted and ground truth classes, and the dynamic regression loss based on the matching degree between the predicted and ground truth boxes. Use the sum of these two losses to guide gradient updates, thus progressively training the object detector.

[0025] Step 3: Identify buried cable targets in the underwater sonar image to be detected. This involves inputting the sonar image to be detected into the buried cable target detector, and the output is a classification confidence score and a predicted bounding box.

[0026] Detecting buried pipes and cables in underwater sonar images remains challenging in real-world seabed applications. For example, the texture features of seabed impurities and buried pipes and cables are extremely similar, easily causing severe background interference; targets are prone to deformation due to long-term burial, making key feature extraction difficult; and the multi-scale nature of target dimensions makes it difficult for the detection head to maintain accuracy during regression localization. Traditional detection methods are often complex and fail to fully consider these challenges, resulting in low detection accuracy and poor recall and precision. Furthermore, when pipes and cables are buried at greater depths, the reflected acoustic signals attenuate significantly, further weakening the identifiable target features in the sonar image and increasing the detection difficulty.

[0027] This invention leverages the powerful image feature learning capabilities of convolutional neural networks to automatically extract effective features from underwater buried pipe and cable sonar images. The resulting feature maps are then input into an adaptive cascaded detection head for classification and localization. The proposed underwater sonar image buried pipe and cable target detection framework can adaptively extract target deformation features, enhance key discriminative features, and suppress background interference. Considering the characteristics of underwater buried object sonar images, this invention can effectively extract multi-scale features, thereby achieving reliable detection of underwater buried pipe and cable targets. This method has good universality, accurately identifies buried targets in sonar images, and achieves good performance on multiple evaluation metrics.

[0028] Compared with existing methods, the present invention has the following characteristics and advantages:

[0029] (1) The deformable feature extraction module is designed as a key component of the feature extraction network. The deformable feature extraction module can adaptively capture complex deformation features, such as when the target is compressed or deformed. This extraction module makes the feature extraction of the target not limited to fixed convolution kernels, but adaptive.

[0030] (2) The designed dual-channel attention mechanism module performs channel enhancement processing on the extracted high-level features. This module can enhance key features by suppressing background interference. This channel enhancement strategy enables the pyramid to capture key feature information during feature fusion.

[0031] (3) Design a dynamic regression loss function so that the detection head can... The value dynamically adjusts the regression strategy, enabling the adaptive cascaded detection head to adaptively regress multi-size targets during model training, compared to... Regression loss function, dynamic regression can locate target boxes more flexibly. Attached Figure Description

[0032] Figure 1 This is a schematic diagram illustrating the framework of the underwater sonar image buried cable target detection method of the present invention.

[0033] Figure 2 This is a block diagram of the deformable feature extraction module.

[0034] Figure 3 This is a structural diagram of a dual-branch channel attention mechanism.

[0035] Figure 4 This is a diagram illustrating the effect of the underwater sonar image buried cable target detection method of the present invention. (a)-(d) represent the detection results for targets of different sizes. Detailed Implementation

[0036] The present invention will be further described below with reference to specific examples and accompanying drawings to facilitate a clearer understanding of the essence of the invention, as well as its effectiveness and advantages. However, it should be understood that the specific examples described herein are not intended to limit the invention.

[0037] like Figure 2 As shown, the deformable feature extraction module consists of an up-dimensional convolutional layer, a deformable convolutional layer, an down-dimensional convolutional layer, and an adaptive scaling connection. Figure 3As shown, the dual-branch channel attention mechanism structure consists of a global average pooling layer, followed by two parallel branches. Branch one is a one-dimensional convolutional layer, and branch two involves channel rearrangement, one-dimensional convolution, and inverse channel rearrangement, followed by concat connections, fused convolutional layers, a sigmoid activation function, and weighted multiplication. Each stage of the detection head comprises an RoI pooling layer, an RCNN detection layer, and two convolutional layers, forming three cascaded stages. Each stage progressively optimizes the regression boxes to achieve more accurate target detection results. The adaptive feature extractor is the fundamental component of the entire detection framework, used to extract key features of underwater buried cables from sonar images. The deformable feature extraction module extracts the deformation features of underwater sonar images. The channel enhancement pyramid network aims to enhance the key channel features of the target and fuse multi-scale feature maps, outputting multi-level feature maps for multi-scale detection. The dual-branch channel attention mechanism enhances channel contrast, highlights key discriminative features, and suppresses background interference. The adaptive cascaded detection head uses a cascaded design for the classification and regression networks, adaptively adjusting the target localization accuracy through progressive optimization. Among them, the dynamic regression loss function enables the detection head to... The value dynamically adjusts the regression strategy, enabling adaptive and accurate regression prediction for multi-scale targets. The underwater buried pipe and cable target detector, through the synergistic effect of three key networks—an adaptive feature extractor, a channel enhancement pyramid, and an adaptive cascaded detection head—is able to extract rich feature information of buried targets from multiple scales.

[0038] Basic model parameter settings: 640×640 resolution is used as the baseline input size for the detection network. A multi-scale training strategy is introduced during the training phase to enhance the model's generalization ability to targets of different scales. The training batch size is set to 8. The adaptive feature extractor in the detection network consists of four layers, with the number of channels in each layer configured as 96, 192, 384, and 768 respectively. This batch and channel setting balances model performance with existing hardware resources, ensuring efficient and stable training.

[0039] For network weight optimization, the SGD optimizer was selected, with an initial learning rate of 0.1. Linear warm-up and cosine annealing were used to gradually adjust the learning rate, with a weight decay coefficient of 0.005 to improve training stability and convergence efficiency. The model adopted a de novo training initialization strategy, with a total training cycle of 30 epochs to ensure the network fully learns the feature representations of sonar images. All modules were developed based on the PyTorch framework and integrated into the mmdetection2 object detection platform to ensure code maintainability and scalability. Model training and testing were both performed on the same computer equipped with an NVIDIA RTX3090 GPU.

[0040] Effect Analysis: Figure 4The diagram illustrates the detection results of the proposed AEBPNet algorithm for buried targets. The green boxes represent ground truth annotations, and the red boxes represent the predicted output of AEBPNet. It is evident that AEBPNet exhibits high localization and recognition accuracy in underwater buried pipe and cable detection tasks, primarily due to the deformable feature extraction module and channel enhancement pyramid structure designed within the network. This method effectively extracts the complex deformation features of targets, generates multi-scale feature maps, and suppresses interference from sonar images, highlighting key information of buried targets, thereby enhancing the network's ability to extract and learn features from buried pipes and cables.

[0041] Table 1 shows the mean precision (AP) and recall (Recall) evaluation results of each model for detecting buried pipe and cable targets in underwater sonar images when the IoU threshold is 0.5. Experiments show that the AEBPNet detection model proposed in this invention significantly outperforms existing mainstream detection models in both of these key metrics, achieving 91.1% and 94.0% respectively, demonstrating superior overall detection performance. This performance improvement is attributed to the collaborative design of the three core modules introduced in the model: an adaptive feature extractor, a channel enhancement pyramid, and an adaptive cascaded detection head. This enables it to more accurately and comprehensively identify buried pipe and cable targets in complex underwater scenarios.

[0042] Table 1

[0043] .

Claims

1. A method for detecting buried pipe and cable targets in underwater sonar images, characterized in that, First, an underwater buried pipeline target detector is constructed, which specifically includes three main networks: An adaptive feature extractor, a channel enhancement pyramid, and an adaptive cascaded detection head are used; then, a target detector is trained using classification loss and dynamically weighted regression loss; finally, buried cable targets in the underwater sonar image to be detected are identified; where: The adaptive feature extractor includes an inverse residual module and a deformable feature extraction module, and outputs a multi-scale basic feature map; the inverse residual module is used to extract basic features such as texture of the target in the sonar image, and the deformable feature extraction module is used to extract the complex deformation features of the target; The channel enhancement pyramid includes a dual-branch channel attention mechanism module and a top-down feature pyramid structure, and outputs a fused and enhanced multi-level feature map. The dual-branch channel attention mechanism module is used to modulate the channel weights of the high-level feature map to enhance key features and suppress background interference; the feature pyramid structure is used to achieve multi-scale feature fusion. The adaptive cascaded detection head is used to predict the target category and location of the feature map generated by the feature pyramid, and output the confidence score of each target.

2. The underwater sonar image buried cable target detection method according to claim 1, characterized in that, In the adaptive feature extractor: the inverse residual module consists of an up-dimensional convolutional layer, a depthwise separable convolutional layer, a down-dimensional convolutional layer, and an adaptive residual connection; the deformable feature extraction module consists of an up-dimensional convolutional layer, a deformable convolutional layer, a down-dimensional convolutional layer, and an adaptive scaling connection.

3. The underwater sonar image buried cable target detection method according to claim 3, characterized in that, In the channel enhancement pyramid, the dual-branch channel attention module structure includes: a global average pooling layer followed by two parallel branches; branch one is a one-dimensional convolutional layer, and branch two is channel rearrangement, one-dimensional convolution, and channel inverse rearrangement; it also includes: concat connection, fused convolutional layer, sigmoid activation function, and weighted multiplication.

4. The underwater sonar image buried cable target detection method according to claim 4, characterized in that, The adaptive cascaded detection head consists of three cascaded detection stages, each containing an RoI pooling layer, RCNN detection, and a fully connected head. This cascaded structure improves performance step by step. Thresholds are used to optimize the regression boxes layer by layer, thereby achieving adaptive improvement in the accuracy of object detection.

5. The underwater sonar image buried cable target detection method according to claim 1, characterized in that, The dynamic weighted regression loss of the training target detector is defined as: , in, , , This regression loss function enables the detector head to adaptively adjust the regression strategy, thereby locating the target more accurately.

6. The underwater sonar image buried cable target detection method according to claim 5, characterized in that, The specific steps are as follows: Step 1: The underwater buried pipe and cable target detector detects the underwater sonar image; The underwater sonar image is input into the adaptive feature extractor; the inverse residual module is used to extract shallow features of the image in the first three stages, and the deformable feature extraction module is used to extract the deformation features of the buried target in the fourth stage. These four stages contain 3, 3, 9 and 3 feature extraction modules, respectively. The inverse residual module processes the image sequentially using an up-dimensional convolutional layer, a depthwise separable convolutional layer, a down-dimensional convolutional layer, and an adaptive residual connection; the deformable feature extraction module processes the image matrix sequentially using an up-dimensional convolutional layer, a deformable convolutional layer, a down-dimensional convolutional layer, and an adaptive scaling. After processing through four stages of the network, each stage outputs a set of basic feature maps. ;Will The input channel is enhanced with a feature pyramid, which generates feature maps through a four-layer convolutional network structure. ; Channel enhancement is performed using a dual-branch channel attention mechanism. The specific steps are as follows: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] A global average pooling layer is used to obtain channel feature vectors. These channel vectors are then input into two branches simultaneously. Branch one extracts local channel relationships through one-dimensional convolution, while branch two performs channel rearrangement, one-dimensional convolution, and inverse rearrangement operations to obtain cross-channel interaction information. The outputs of the two branches are concatenated, then fused through a convolutional layer, and channel attention weights are generated using the Sigmad function. These attention weights are then compared with the feature maps. Multiplication yields the enhanced feature map ; feature map Upsampling is performed through a convolutional network layer, and then combined with... Added output feature map ; feature map Upsampling is performed through a convolutional network layer, and then combined with... Added output feature map ; feature map Upsampling is performed through a convolutional network layer, and then combined with... Added output feature map ; Finally, the pyramid-level feature map was obtained. ; The adaptive cascaded detection head is based on As input, the target is adaptively optimized step-by-step through three cascaded detection stages; specifically, the feature map... , , and Perform the same operation as below, each stage performs RoI pooling and RCNN detection based on the bounding boxes output by the previous stage, and gradually increases the intersection-union threshold as a localization requirement. Finally, through two convolutions, output the category, location and confidence prediction value of the target in each image. Step 2: Train the target detector using classification loss and dynamic weighted regression loss; The classification loss is calculated based on the predicted category and the true category, and the dynamic regression loss is calculated based on the matching degree between the predicted box and the true box. The total loss after adding the two losses is used to guide the gradient update, and the target detector is trained step by step. Step 3: Identify buried cable targets in the underwater sonar image to be detected; that is, input the sonar image to be detected into the buried cable target detector, and output the classification confidence and prediction box.

Images