Seabed static small target detection method suitable for underwater vehicle

Abstract

The invention relates to the field of underwater detection, and provides a seabed static small target detection method suitable for an underwater vehicle, and the method comprises the steps: obtaining a plurality of underwater sonar images through a two-dimensional imaging sonar of the underwater vehicle; inputting a to-be-detected underwater sonar image into the underwater target detection model to extract features and detect the features to obtain an object category, bounding box coordinates and a confidence score; according to the method, target features are automatically extracted through a deep neural network in an end-to-end mode, a CAA module is introduced to capture a context dependency relationship between long-distance pixels in an image, a feature map with higher expressive power is generated by integrating context information, a model is helped to better distinguish a target object from a background, and the accuracy of the target object is improved. Therefore, the detection accuracy is improved.

Description

Technical Field

[0001] The invention belongs to the field of underwater detection and proposes a method for detecting small static seabed targets suitable for underwater vehicles. Background Art

[0002] Underwater unmanned vehicles are flexible and are widely used in underwater target detection, underwater landform survey and other fields. Side scan sonar has clear imaging and is an important mission payload for underwater unmanned vehicles. However, conventional side scan sonar mainly relies on manual reading of maps. When unmanned systems are used, the equipment needs to be recovered and the map reading and judgment must be completed manually, which is inefficient.

[0003] At present, the academic community has conducted a lot of research in the field of autonomous detection and recognition, mainly focusing on feature-based recognition methods and deep learning-based recognition methods. Existing recognition methods are mainly borrowed from the field of visual processing. Compared with underwater acoustic images, image information in the visual field has richer details and more complete samples. However, the edge details of sonar images are fuzzy and the identifiable features are limited, making it difficult to autonomously detect underwater targets. Summary of the invention

[0004] In order to solve the problem of low accuracy of underwater sinking target detection in existing methods, the present invention discloses a one-stage underwater sinking target detection method integrating a context anchor point attention mechanism based on a new imaging sonar, the method comprising:

[0005] Acquiring multiple underwater sonar images by using a two-dimensional imaging sonar of an underwater vehicle;

[0006] Input the underwater sonar image to be detected into the underwater target detection model to extract features and detect, and obtain object category, bounding box coordinates and confidence score; the object category is the target object category in the underwater sonar image, the bounding box coordinates are the bounding box coordinates of the target location in the underwater sonar image, and the confidence score is the possibility of the target object being contained in the bounding box;

[0007] Wherein, the underwater target detection model includes a backbone network and a head network;

[0008] The backbone network includes a convolution module and a c2f module. The convolution module is used to extract local features of different scales and different spatial positions in underwater sonar images. The c2f module is used to fuse the local features of different scales to obtain a unified feature representation containing local details and global information. The local features include texture, edge, and shape features.

[0009] The c2f module includes a contextual attention mechanism, which captures the dependency between long-distance pixels, extracts the global context information of the underwater sonar image and maps it back to the original feature map, outputs an attention factor A reflecting the context information of the sonar image, adjusts the attention weight of the feature map according to the attention factor A, and enhances the features of the target area;

[0010] The head network generates multiple anchor boxes of different sizes and aspect ratios with each pixel as the center, predicts the bounding box regression value, confidence and target category of each anchor box, retains the anchor box with the highest confidence as the predicted box through non-maximum suppression operation, and outputs the target category, bounding box coordinates and confidence score corresponding to the predicted box.

[0011] Acquire a historical image dataset, and expand the historical image dataset by means of data enhancement; the historical image dataset includes historical sonar images and label results corresponding to the historical sonar images;

[0012] The historical image data set is input into the initial underwater target detection model to obtain the classification result of the historical image data set, and the initial underwater target detection model parameters are iterated based on the label difference between the classification result and the historical image data set to obtain the underwater target detection model.

[0013] The context anchor attention mechanism includes average pooling, convolution, depth-separable convolution, and Sigmoid activation function operations;

[0014] The contextual anchor attention mechanism first uses average pooling to extract global features from the entire feature map, then uses 1×1 convolution to obtain local area features with contextual information, and then uses two depth-separable strip convolutions to capture the dependencies between long-distance pixels. Finally, the attention factor of each anchor point is calculated, and after a 1×1 convolution layer and a Sigmoid activation function, the attention factor A is output.

[0015] The c2f module also includes convolution, split, and concat operations;

[0016] The sonar image is first passed through a 1x1 convolution kernel in the c2f module of the backbone network to reduce the number of input channels to 1 / 2 of the original number, and then a contextual attention mechanism module is used to extract a feature map containing contextual information. Then, a residual link composed of split and concat is used to first split the feature map into sub-feature maps of different scales, and then the features enhanced by the contextual anchor attention mechanism are concatenated with the feature map after convolution to generate the final feature map. Finally, a 1x1 convolution kernel is used to restore the number of channels of the feature map.

[0017] The head network includes a convolutional layer and a prediction layer;

[0018] The convolution layer is used to fuse features and enhance contextual information and semantic information;

[0019] The prediction layer includes an anchor box, a detection head and a non-maximum suppression operation. The anchor box is a plurality of square boxes of different sizes and aspect ratios generated with each pixel as the center, which are used to regress the bounding box of the object and classify the object. The non-maximum suppression operation is used to retain the anchor box with the highest confidence as the prediction box. The detection head is used to predict the bounding box regression value of each anchor box and the confidence of the target existence, and output the object category, bounding box coordinates and confidence score corresponding to the prediction box.

[0020] The data enhancement means include geometric transformation, adding noise, changing contrast, random cropping or filling operations.

[0021] 1. Since the present invention selects YOLOv8 as the benchmark model, an end-to-end approach is adopted to automatically extract target features through a deep neural network, which reduces human intervention and dependence on professional knowledge in target detection and obtains more implicit features.

[0022] 2. Aiming at the problem of small target detection of underwater sinking targets, this paper introduces the CAA module to capture the contextual dependencies between long-distance pixels in the image. By integrating contextual information to generate more expressive feature maps, it helps the model better distinguish between target objects and backgrounds, thereby improving detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a method for detecting a small static submarine target according to an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of a detection network of a method for detecting a small static underwater target according to an embodiment of the present invention;

[0025] Figure 3 The target acoustic image after detection obtained by one embodiment of the present invention; DETAILED DESCRIPTION

[0026] The best mode of implementation of the present invention is described below by way of example. It should be understood that the specific embodiments herein are used to explain the present invention in detail and should not be construed as limiting the present invention. It should be noted that various changes and modifications may be made under the premise of following the principles and core scope of the present invention, and these changes should all be deemed to be within the scope of protection of the present invention. In conjunction with the accompanying drawings, the specific implementation steps of the present invention are described in detail.

[0027] Step 1: Acquire multiple underwater sonar images through the two-dimensional imaging sonar of the underwater vehicle;

[0028] Step 2: Input the historical image dataset into the initial underwater target detection model to obtain the classification result of the historical image dataset, iterate the model parameters based on the label difference between the classification result and the historical image dataset, and obtain the underwater target detection model;

[0029] In a specific embodiment, step 2 includes the following steps:

[0030] Step 2.1: Obtain a historical image dataset, and expand the historical image dataset by data enhancement means; the historical image dataset includes historical sonar images and label results corresponding to the historical sonar images;

[0031] In a specific embodiment, the historical image dataset can be enhanced in a variety of ways for training, such as simulating target echoes of different perspectives and postures by geometric transformations such as rotation, scaling, translation or mirroring; simulating target echoes of different water quality conditions by adding noise or changing contrast; and increasing data diversity by random cropping or padding.

[0032] Step 2.2: Input the historical image dataset into the initial underwater target detection model to obtain the classification result of the historical image dataset, iterate the initial underwater target detection model parameters based on the label difference between the classification result and the historical image dataset, and obtain the underwater target detection model;

[0033] In one embodiment of the present invention, the initial underwater target detection model is improved based on YOLOv8, including a backbone network and a head network.

[0034] Among them, the backbone network is a lightweight deep convolutional neural network, which is used to extract features from the input image. It includes multiple convolutional layers, normalization layers and activation functions. It extracts multi-scale feature information through layer-by-layer operations, thereby providing high-quality feature representation for target detection in the head network.

[0035] like Figure 2 As shown in FIG. 1 , it is a schematic diagram of the structure of the backbone network, including a convolution module and a c2f module, wherein the convolution module is used to perform downsampling operations to reduce the size of the feature map and increase the number of channels, extract local features of different scales and different spatial positions in the underwater sonar image, and the c2f module is used to fuse the local features of different scales to capture the local details and global structure of the target; the local features include texture, edge, and shape features. The targets in the seabed environment may have fuzzy boundaries or incomplete shapes due to factors such as lighting changes, water flow, and sediments. Multi-scale feature fusion can extract detail information from multiple scales and enhance the edge, texture, and morphological features of the target. Especially in small target detection, the capture of details is crucial. Multi-scale feature fusion can provide richer contextual information and improve the accuracy of target recognition.

[0036] In the prior art, the C2f module includes convolution, split, bottleneck, and concat operations. The composition of split and concat is residual module processing, that is, the split module splits the input data into two parts, one part is processed by n Bottleneck modules to extract features of different levels, and the other part is directly concat-ed with the result processed by the Bottleneck module, and finally processed by a convolution module to output the features of different scales and levels.

[0037] Among them, the Bottleneck module mainly includes two convolutional layers, which use multiple 3x3 convolution kernels to perform convolution operations to extract feature information. At the same time, it has the option of whether to perform residual links.

[0038] Since sonar images usually have unique texture, noise and contrast characteristics, which are different from ordinary visible light images, in order to more effectively identify targets in sonar images, the present invention replaces the Bottleneck in all C2f modules in the backbone network with the Context Anchor Attention (CAA) module.

[0039] The Context Anchor Attention (CAA) mechanism uses global average pooling and 1D strip convolution to capture the relationship between distant pixels, extract features with local and global context information, and automatically adjust the attention weight according to the importance of each position or region (i.e., the attention factor A) and the characteristics of the anchor. When the target is in a complex background or occluded in the image, CAA can dynamically adjust the weight according to the relationship between the target and the surrounding environment, so as to better locate the target and avoid excessive attention to irrelevant areas.

[0040] In a specific embodiment, the context anchor attention mechanism mainly includes average pooling (avgpool), convolution (conv), depthwise separable convolution (DWconv), and Sigmoid activation function operations, and its main calculation steps are as follows:

[0041] (a) Average pooling is used to extract relatively macro context information from the entire feature map. After extracting the global features, 1×1 convolution is used to obtain local area features with context information:

[0042] F conv =Conv 1×1 (P avg (X)

[0043] Among them, avgrepresents the average pooling operation, X is the input feature map, Conv 1×1 is a 1×1 convolution operation, F conv is the extracted local area features.

[0044] (b) Then, two 1D depthwise separable strip convolutions are applied to extract the features of the target:

[0045] F w =DWConv(F conv )

[0046] F h =DWConv(F w )

[0047] Among them, DWConv is a deep strip convolution operation, F h is a deep feature.

[0048] The receptive field is expanded by using two depth-wise separable strip convolutions. Compared with the standard large-kernel depth-wise separable convolution, the effect is similar while the parameters are reduced. It can better capture the dependencies between long-distance pixels and contain more contextual information.

[0049] (c) Calculate the attention factor of each anchor point, pass it through a 1×1 convolution layer and a Sigmoid activation function, and output the attention factor A:

[0050] A=Sigmoid(Conv 1×1 (F h ))

[0051] The calculated attention factor is usually transformed nonlinearly through the Sigmoid activation function to ensure that its value is between 0 and 1, indicating the degree of attention of the anchor point.

[0052] In a specific embodiment, the head network of the initial underwater target detection model includes a convolutional layer and a prediction layer.

[0053] The convolution layer includes convolution processing and activation functions, which are used to fuse features so that the feature map contains richer contextual information and semantic information, enhances the model's ability to detect objects of different scales, and introduces nonlinearity to improve feature expression capabilities. The convolution layer outputs feature maps of three scales so that objects of different sizes can be detected.

[0054] Among them, the prediction layer includes anchor boxes, detection heads, and non-maximum suppression operations:

[0055] The anchor boxes are used as prior boxes to provide an initial box shape and adjust it to accommodate target objects of different sizes and aspect ratios. Each pixel position in the input feature map is associated with a different anchor box to regress the bounding box of the object and classify the object.

[0056] Among them, the detection head is used to learn the offset between the anchor box and the actual object bounding box, adjust the position and size of the anchor box, generate detection results, and predict the bounding box regression value of each anchor box and the confidence of the target existence.

[0057] Specifically, the detection head consists of the following parts:

[0058] (a) Bounding Box Regression: Regress the four bounding box parameters of the object (usually the x, y coordinates of the center point, as well as the width and height).

[0059] (b) Classification Branch: Predict the category of the object in the box. Each candidate box will output a probability distribution of a category.

[0060] (c) Objectness Prediction Branch: Predicts whether the current box contains an object and generates a confidence score, which indicates the possibility that the box contains the object. The confidence score is generally in the range of 0-1. The closer the score is to 1, the greater the possibility that the box contains the object.

[0061] The detection head can convert the feature map extracted from the aforementioned network into a vector containing multiple feature values, including object category, bounding box coordinates, and confidence score.

[0062] Among them, non-maximum suppression processing is used to remove duplicate detection boxes, retain the bounding box with the highest confidence, and remove other bounding boxes with high overlap, ensuring that each target is detected only once.

[0063] Inputting the enhanced historical image dataset into the initial underwater target detection model to obtain the classification result of the historical image dataset, iterating the model parameters based on the label difference between the classification result and the historical image dataset, and obtaining the underwater target detection model with optimal parameters;

[0064] Step 3: Input the underwater sonar image data to be detected into the underwater target detection model to obtain the detection result.

[0065] In a specific embodiment, the sonar image to be detected is first downsampled in the backbone network of the underwater target detection model to reduce the size of the feature map and increase the number of channels, and then batch normalization and ReLU activation function are performed to enhance the nonlinear representation capability of the model.

[0066] In the improved c2f module, the number of input channels is reduced to 1 / 2 of the original through a 1x1 convolution kernel to reduce the amount of computation and memory consumption. Then, multiple contextual attention mechanism modules are used to capture the dependencies between long-distance pixels, extract the global context information of the underwater sonar image and map it back to the original feature map, output the attention factor A that reflects the context information of the sonar image, adjust the attention weight of the feature map according to the attention factor A, enhance the features of the target area, and then use the residual link composed of split and concat to add the convolution input directly to the output of the contextual attention mechanism module, fuse the features and avoid the problem of gradient disappearance in the deep network. Finally, the 1x1 convolution kernel is used again to restore the number of channels of the feature map.

[0067] Then it enters the head network of the underwater target detection model, first through the convolutional layer to fuse features, so that the feature map contains richer contextual information and semantic information, enhance the model's detection ability for targets of different scales, output feature maps of three scales, and then generate anchor boxes through the prediction layer, learn the offset between the anchor box and the actual object bounding box, adjust the position and size of the anchor box, generate detection results, and then remove duplicate detection boxes through non-maximum suppression processing, retain the bounding box with the highest confidence, and remove other bounding boxes with high overlap. Finally, the feature map extracted from the above network is converted into a vector containing multiple eigenvalues, and the object category, bounding box coordinates and confidence score corresponding to the underwater sonar image are output.

[0068] In a specific embodiment, the present invention is used to process and analyze the sonar data in the lake, and the conclusions are as follows:

[0069] mAP50 mAP75 GFLOPs YOLOv8 86.80％ 42.70％ 28.4 After Improvement 92.20％ 45.10％ 23.5 change 5.40％ 2.40％ 17.30％

[0070] This table shows the performance comparison of two different algorithms on different indicators. Among them, "YOLOv8" is the original algorithm, and "Improved" is the optimized method disclosed in the present invention.

[0071] Among them, mAP is an indicator for evaluating the performance of the target detection model, which indicates the degree of match between the detected objects and the real objects in different pictures. Specifically, mAP50 calculates the average accuracy of the model when the overlap between the predicted box and the real box reaches 50%, while mAP75 calculates the average accuracy of the model when the overlap between the predicted box and the real box reaches 75%. The higher the mAP50 and mAP75 values, the better the detection effect.

[0072] GFLOPS is an indicator to measure the amount of data processed by a computer, which means "billion floating-point operations" that can be performed per second. Floating-point operations are the basic operations for processing complex mathematical calculations. The larger the GFLOPS value, the more mathematical calculations the computer needs to process per second, and the greater the amount of calculations.

[0073] From the table data we can see that:

[0074] In terms of the "mAP50" indicator (i.e., the average precision when the IoU threshold is 0.5), the improved algorithm achieved an accuracy of 92.20%, an increase of 5.40% compared to the original algorithm.

[0075] In terms of the "mAP75" indicator (i.e., the average precision when the IoU threshold is 0.75), the improved algorithm achieved an accuracy of 45.10%, an increase of 2.40% compared to the original algorithm.

[0076] In terms of the "GFLOPs" indicator (i.e., billion floating-point operations per second), the improved algorithm achieved a 17.30% performance improvement over the original algorithm, reducing from 28.4 GFLOPs to 23.5 GFLOPs.

[0077] These results demonstrate that the improved algorithm successfully improves computational efficiency while maintaining high accuracy, demonstrating good performance optimization capabilities.

[0078] like Figure 3 As shown, it is a target sound image after detection obtained by an embodiment of the present invention, wherein the red square box is the bounding box output by the head network, "cylinder" is the predicted classification, that is, it is predicted that the object is a cylinder, and 0.8 is the confidence score that there is an object in the box. It can be seen that the method used in the present invention can detect and classify objects well.

Claims

1. A method for detecting small static targets on the seabed applicable to underwater vehicles, comprising: Acquiring multiple underwater sonar images through a two-dimensional imaging sonar of an underwater vehicle; Input the underwater sonar image to be detected into the underwater target detection model to extract features and detect, and obtain object category, bounding box coordinates and confidence score; the object category is the target object category in the underwater sonar image, the bounding box coordinates are the bounding box coordinates of the target location in the underwater sonar image, and the confidence score is the possibility of the target object being contained in the bounding box; Wherein, the underwater target detection model includes a backbone network and a head network; The backbone network includes a convolution module and a c2f module. The convolution module is used to extract local features of different scales and different spatial positions in underwater sonar images. The c2f module is used to fuse the local features of different scales to obtain a unified feature representation containing local details and global information. The local features include texture, edge, and shape features. The c2f module includes a contextual attention mechanism, which captures the dependency between long-distance pixels, extracts the global context information of the underwater sonar image and maps it back to the original feature map, outputs an attention factor A reflecting the context information of the sonar image, adjusts the attention weight of the feature map according to the attention factor A, and enhances the features of the target area; The head network generates multiple anchor boxes of different sizes and aspect ratios with each pixel as the center, predicts the bounding box regression value, confidence and target category of each anchor box, retains the anchor box with the highest confidence as the predicted box through non-maximum suppression operation, and outputs the target category, bounding box coordinates and confidence score corresponding to the predicted box.

2. The method according to claim 1, characterized in that Also includes: Acquire a historical image dataset, and expand the historical image dataset by means of data enhancement; the historical image dataset includes historical sonar images and label results corresponding to the historical sonar images; The historical image data set is input into the initial underwater target detection model to obtain the classification result of the historical image data set, and the initial underwater target detection model parameters are iterated based on the label difference between the classification result and the historical image data set to obtain the underwater target detection model.

3. The method according to claim 1, characterized in that The context anchor attention mechanism includes average pooling, convolution, depth-separable convolution, and Sigmoid activation function operations; The contextual anchor attention mechanism first uses average pooling to extract global features from the entire feature map, then uses 1×1 convolution to obtain local area features with contextual information, and then uses two depth-separable strip convolutions to capture the dependencies between long-distance pixels. Finally, the attention factor of each anchor point is calculated, and after a 1×1 convolution layer and a Sigmoid activation function, the attention factor A is output.

4. The method according to claim 1, characterized in that: The c2f module also includes convolution, split, and concat operations; The sonar image is first passed through a 1x1 convolution kernel in the c2f module of the backbone network to reduce the number of input channels to 1 / 2 of the original number, and then a contextual attention mechanism module is used to extract a feature map containing contextual information. Then, a residual link composed of split and concat is used to first split the feature map into sub-feature maps of different scales, and then the features enhanced by the contextual anchor attention mechanism are concatenated with the feature map after convolution to generate the final feature map. Finally, a 1x1 convolution kernel is used to restore the number of channels of the feature map.

5. The method according to claim 1, characterized in that The head network includes a convolutional layer and a prediction layer; The convolution layer is used to fuse features and enhance contextual information and semantic information; The prediction layer includes an anchor box, a detection head and a non-maximum suppression operation. The anchor box is a plurality of square boxes of different sizes and aspect ratios generated with each pixel as the center, which are used to regress the bounding box of the object and classify the object. The non-maximum suppression operation is used to retain the anchor box with the highest confidence as the prediction box. The detection head is used to predict the bounding box regression value of each anchor box and the confidence of the target existence, and output the object category, bounding box coordinates and confidence score corresponding to the prediction box.

6. The method according to claim 2, characterized in that The data enhancement means include geometric transformation, adding noise, changing contrast, random cropping or filling operations.

Images