A seabed static small target detection method suitable for underwater vehicle

Abstract

The present disclosure relates to the field of underwater detection, and proposes a seabed static small target detection method suitable for underwater vehicles, comprising: obtaining multiple underwater sonar images through a two-dimensional imaging sonar of an underwater vehicle; inputting the underwater sonar image to be detected into an underwater target detection model to extract features and detect, obtaining object categories, boundary box coordinates and confidence scores; the present application adopts an end-to-end mode to automatically extract target features through a deep neural network, introduces a CAA module to capture the context dependency relationship between distant pixels in the image, generates a more expressive feature map by integrating context information, helps the model to better distinguish target objects and backgrounds, thereby improving the accuracy of detection.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of underwater detection, and proposes a seabed static small target detection method suitable for underwater vehicles. BACKGROUND

[0002] Underwater unmanned vehicles are flexible and are widely used in underwater target search and exploration, underwater topography survey and other fields. Side-scan sonar imaging is clear, and is an important task load of underwater unmanned vehicles. However, conventional side-scan sonar mainly relies on manual reading, and when unmanned systems are used, reading and mapping need to be completed by manual reading after the equipment is recovered, which is low in operation efficiency.

[0003] At present, a large amount of research has been carried out in the academic field in the aspect of autonomous detection and identification, mainly focusing on feature-based identification methods and deep learning-based identification methods. Existing identification methods are mainly borrowed from the field of visual processing. Compared with underwater acoustic images, image information in the visual field is more detailed and more complete. However, the edge details of sonar images are fuzzy, and the recognizable features are limited, so there is great difficulty in autonomous detection of underwater bottom targets. SUMMARY

[0004] To solve the problem of low detection precision of existing methods for underwater bottom targets, the application discloses a one-stage underwater bottom target detection method based on a new system imaging sonar and fusing context anchor point attention mechanism, and the method comprises the following steps:

[0005] Obtaining a plurality of underwater sonar images through a two-dimensional imaging sonar of an underwater vehicle;

[0006] Inputting the underwater sonar image to be detected into an underwater target detection model to extract features and detect, to obtain an object category, a bounding box coordinate and a confidence score; the object category is a target object category in the underwater sonar image, the bounding box coordinate is a bounding box coordinate of a position of a target in the underwater sonar image, and the confidence score is a possibility that the target object is contained in the bounding box.

[0007] The underwater target detection model comprises a backbone network and a head network.

[0008] The backbone network comprises a convolution module and a c2f module, the convolution module is used for extracting local features of different scales and different spatial positions in the underwater sonar image, and the c2f module is used for fusing the local features of different scales to obtain a unified feature representation containing local details and global information; the local features comprise texture, edge and shape features.

[0009] The c2f module comprises a context attention mechanism, which captures the long-distance inter-pixel dependency, 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 region;

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

[0011] A historical image dataset is obtained, and the historical image dataset is expanded through data enhancement means; the historical image dataset comprises historical sonar images and label results corresponding to the historical sonar images;

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

[0013] The context anchor attention mechanism comprises an average pooling, a convolution, a depth separable convolution and a Sigmoid activation function operation;

[0014] The context anchor attention mechanism first extracts global features from the entire feature map through average pooling, then obtains local region features with context information through 1x1 convolution, then captures the long-distance inter-pixel dependency through two depth separable strip convolutions, and finally calculates the attention factor of each anchor point, and outputs the attention factor A through a 1x1 convolution layer and a Sigmoid activation function.

[0015] The c2f module further comprises a convolution, a split and a concat operation;

[0016] The sonar image is first subjected to 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, then a context attention mechanism module is used to extract a feature map containing context information, then a residual link composed of split and concat is used, the feature map is first split into sub-feature maps of different scales, the features enhanced by the context anchor attention mechanism are spliced with the feature map after convolution to generate a final feature map, and finally a 1x1 convolution kernel is used to restore the number of channels of the feature map.

[0017] The head network comprises a convolution layer and a prediction layer;

[0018] The convolutional layer is used for fusing features, enhancing context information and semantic information;

[0019] The prediction layer includes an anchor frame, a detection head and a non-maximum suppression operation, the anchor frame is a plurality of square frames with different sizes and aspect ratios generated with each pixel as the center, used for regressing the bounding box of the object and classifying the object, the non-maximum suppression operation is used for retaining the anchor frame with the highest confidence as the prediction frame, the detection head is used for predicting the bounding box regression value and the confidence of the existence of the target of each anchor frame, and outputs the object category corresponding to the prediction frame, the bounding box coordinates and the confidence score.

[0020] The data enhancement means includes geometric transformation, adding noise, changing contrast, random cropping or padding operation.

[0021] 1. Since the YOLOv8 is selected as the benchmark model, the target features are automatically extracted through the deep neural network in an end-to-end manner, the human intervention and the dependence on professional knowledge in target detection are reduced, and more implicit features are obtained.

[0022] 2. For the small target detection problem of underwater bottom targets, the CAA module is introduced to capture the context dependency relationship between distant pixels in the image, more expressive feature maps are generated by integrating context information, the model is helped to better distinguish target objects and backgrounds, and the detection accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The flow chart of the seabed static small target detection method of an embodiment of the present application;

[0024] Figure 2 The detection network schematic diagram of the seabed static small target detection method of an embodiment of the present application;

[0025] Figure 3 The detected target acoustic image obtained by an embodiment of the present application; DETAILED DESCRIPTION

[0026] The best mode of the present application will be described below through embodiments, and it should be understood that the specific embodiments herein are used to explain the present application in detail, and should not be understood as a limitation of the present application. It should be noted that various changes and modifications can be made within the scope of the principles and cores of the present application, and these changes should be regarded as belonging to the protection scope of the present application. The specific implementation steps of the present application are described in detail in combination with the drawings.

[0027] Step one: obtain a plurality of 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 results of the historical image dataset, iterate the model parameters based on the label difference between the classification results 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 through data enhancement; the historical image dataset includes historical sonar images and label results corresponding to the historical sonar images;

[0031] In a specific embodiment, historical image datasets can be enhanced for training in a variety of ways. For example, geometric transformations such as rotation, scaling, translation, or mirroring can be used to simulate target echoes from different perspectives and postures; noise can be added or contrast can be changed to simulate target echoes under different water quality conditions; and random cropping or padding can be used to increase data diversity.

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

[0034] Among them, the backbone network is a lightweight deep convolutional neural network 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 Figure 1 shows the structure of the backbone network, which includes a convolutional module and a C2F module. The convolutional module is used to perform downsampling operations to reduce the size of the feature map and increase the number of channels, extracting local features at different scales and spatial locations in underwater sonar images. The C2F module is used to fuse these local features at different scales to capture the local details and global structure of the target. These local features include texture, edge, and shape features. Targets in the seabed environment may have relatively blurred boundaries or incomplete shapes due to factors such as lighting changes, water flow, and sediment. Multi-scale feature fusion can extract detailed information from multiple scales, enhancing the edge, texture, and morphological features of the target. This is especially important in small target detection, where capturing details is crucial. Multi-scale feature fusion can provide richer contextual information and improve the accuracy of target recognition.

[0036] In the existing technology, 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 at different levels, and the other part is directly concat-ed with the result processed by the Bottleneck module. Finally, it is processed by a convolution module to output features that fuse 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] Because sonar images typically have unique texture, noise, and contrast characteristics that differ 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 a Context Anchor Attention (CAA) module.

[0039] The Context Anchor Attention (CAA) mechanism uses global average pooling and one-dimensional strip convolution to capture relationships between distant pixels, extracting features with both local and global contextual information, and automatically adjusts attention weights based on the importance of each location or region (i.e., the attention factor A) and the characteristics of the anchor. When an object is located in a complex background or is obscured in an image, CAA dynamically adjusts the weights based on the relationship between the object and its surroundings, thereby better locating the object and avoiding 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 macroscopic contextual information from the entire feature map. After extracting global features, 1×1 convolution is used to obtain local area features with contextual information:

[0042] F conv =Conv 1×1 (P avg (X)

[0043] Among them, avgdenotes the average pooling operation, X is an input feature map, Conv 1×1 is a 1x1 convolution operation, F conv is the extracted local region feature.

[0044] (b) Then, two one-dimensional depth separable bar convolutions are applied to extract the features of the target:

[0045] F w = DWConv(F conv )

[0046] F h = DWConv(F w )

[0047] wherein DWConv is a depth bar convolution operation, F h is a depth feature.

[0048] The receptive field is expanded by two depth separable bar convolutions, which can better capture the dependence between long-distance pixels and contain more context information compared with the standard large kernel depth separable convolution.

[0049] (c) The attention factor of each anchor point is calculated, which is output after a 1x1 convolution layer and a Sigmoid activation function:

[0050] A = Sigmoid(Conv 1×1 (F h ))

[0051] The calculated attention factor is usually nonlinearly transformed by a Sigmoid activation function, ensuring that its value is between 0 and 1, representing the degree of attention of the anchor point.

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

[0053] The convolution layer includes convolution processing and an activation function, which is used to fuse features so that the feature map contains richer context information and semantic information, enhances the detection ability of the model for targets of different scales, and introduces nonlinearity to improve the feature expression ability. The convolution layer outputs three scale feature maps to enable detection of targets of different sizes.

[0054] The prediction layer includes an anchor box, a detection head, and a non-maximum suppression operation:

[0055] Wherein the anchor boxes are used as prior boxes to provide an initial box shape, and the box shape is adjusted to adapt to objects of different sizes and aspect ratios. Each pixel position in the input feature map is associated with a different anchor box, which is used to regress the object bounding box and classify the object.

[0056] Wherein, 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, and generate a detection result, predict the bounding box regression value of each anchor box and the confidence of the target existence.

[0057] Specifically, the detection head includes 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, and the width and height).

[0059] (b) Class Prediction Branch: Predict the class of the object in the box. Each candidate box outputs a probability distribution of a class.

[0060] (c) Objectness Prediction Branch: Predict whether the current box contains an object, generate a confidence score indicating the likelihood of the box containing an object. The confidence score is generally in the range of 0-1, and the closer the score is to 1, the greater the likelihood of the box containing an object.

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

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

[0063] Input the enhanced historical image dataset into the initial underwater target detection model to obtain the classification result of the historical image dataset, and based on the difference between the classification result and the label of the historical image dataset, iterate the model parameters to obtain an underwater target detection model with optimal parameters.

[0064] Step three: 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 subjected to a downsampling operation in the backbone network of the underwater target detection model, reducing the size of the feature map and increasing the number of channels, and then subjected to batch normalization and a ReLU activation function, thereby enhancing the nonlinear representation capability of the model.

[0066] In the improved c2f module, the input channel number is first reduced to 1 / 2 of the original number by using a 1x1 convolution kernel to reduce the amount of calculation and memory consumption. Then, a plurality of context attention mechanism modules are used to capture the dependency between long-distance pixels, extract the global context information of the underwater sonar image, and map it back to the original feature map. The attention factor A reflecting the context information of the sonar image is outputted. The attention weight of the feature map is adjusted according to the attention factor A, the features of the target region are enhanced, and then the residual link composed of split and concat is used to directly add the input after convolution to the output of the context attention mechanism module, fuse the features and avoid the problem of gradient disappearance in the deep network. Finally, the channel number of the feature map is restored by using a 1x1 convolution kernel again.

[0067] Subsequently, the head network of the underwater target detection model is entered. The features are first fused by a convolution layer, so that the feature map contains more rich context information and semantic information, and the detection capability of the model for targets of different scales is enhanced. Three scale feature maps are outputted, and then an anchor box is generated by a prediction layer. The offset between the anchor box and the actual object bounding box is learned, the position and size of the anchor box are adjusted, the detection result is generated, and then the repeated detection boxes are removed by non-maximum suppression processing. The bounding box with the highest confidence is retained, and other bounding boxes with high overlap are removed. Finally, the extracted feature mapping from the foregoing network is converted into a vector containing a plurality of feature values, and the object class corresponding to the underwater sonar image, the bounding box coordinates and the confidence score are outputted.

[0068] In a specific embodiment, the sonar image data in the lake is processed and analyzed using the present application, and the following table is obtained:

[0069] mAP50 mAP75 GFLOPs YOLOv8 86.80％ 42.70％ 28.4 After improvements 92.20％ 45.10％ 23.5 Changes 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 application.

[0071] Among them, mAP is an indicator for evaluating the performance of a target detection model, which represents the matching degree of 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%, and 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 a measure of the amount of data a computer can process, representing the number of "gigafloating-point operations" it can perform per second. Floating-point operations are fundamental to complex mathematical calculations. A higher GFLOPS value indicates a greater number of mathematical calculations a computer must perform per second, representing a greater computational workload.

[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 it 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 The figure shows the detected target acoustic image obtained by one embodiment of the present invention. The red square box is the bounding box output by the head network, "cylinder" is the predicted classification, that is, the object is predicted to be 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 well detect and classify objects.

Claims

1. A method for detecting small static seabed targets suitable for underwater vehicles, comprising: Acquiring multiple underwater sonar images through a two-dimensional imaging sonar of an underwater vehicle; Inputting the underwater sonar image to be detected into the underwater target detection model to extract features and detect the object, and obtaining an object category, bounding box coordinates, and a confidence score; the object category is the category of the target object 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 probability that the bounding box contains the target object; 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 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 that contains local details and global information. The local features include texture, edge, and shape features. The c2f module includes a contextual attention mechanism that captures dependencies between long-distance pixels, extracts global context information of underwater sonar images, 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 centered on each pixel, 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 through data enhancement; the historical image dataset includes historical sonar images and label results corresponding to the historical sonar images; The historical image dataset is input into the initial underwater target detection model to obtain a classification result of the historical image dataset. The initial underwater target detection model parameters are iterated based on the label difference between the classification result and the historical image dataset 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 Sigmoid activation function, the attention factor A is output.

4. The method according to claim 1, wherein The c2f module also includes convolution, split, and concat operations; The sonar image is first processed 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. The context attention mechanism module is then used to extract a feature map containing context information. A residual link consisting of split and concat is then used to split the feature map into sub-feature maps of different scales. The features enhanced by the context anchor attention mechanism are then concatenated with the convolved feature map 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, wherein The head network includes a convolutional layer and a prediction layer; The convolutional 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