Underwater side-scan sonar image target identification method based on shadow features

Abstract

The invention discloses an underwater side-scan sonar image target identification method based on shadow features. The method comprises the following steps: 1) acquiring a plurality of underwater side-scan sonar images and marking the images to obtain a plurality of training samples; 2) training an underwater sonar image target identification model by using the training sample; the underwater sonar image target recognition model comprises a feature extraction module, a feature fusion module and a target recognition module; the feature fusion module comprises a first self-attention unit, a second self-attention unit, a first cross attention unit, a second cross attention unit, a third self-attention unit and a fourth self-attention unit; and 3) inputting a to-be-identified underwater side-scan sonar image into the trained model to obtain a target identification result of the to-be-identified underwater side-scan sonar image. According to the invention, by applying the self-attention layer and the bidirectional cross-attention layer, important association inside each feature and between different features is revealed, and the performance of the system is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of underwater sonar side-scan image target recognition technology, specifically relating to an underwater sonar side-scan image target recognition method based on image shadow features. Background Technology

[0002] Underwater sonar imagery plays a vital role in numerous fields. In the military, it is used for mine countermeasures, anti-submarine warfare, and naval tactical analysis. In marine engineering, sonar imagery aids in seabed topography exploration, seabed modeling, and the laying and maintenance of subsea pipelines and cables. In marine resource development, sonar technology is used to detect and assess oil and gas reservoirs, mineral resources, and fishery resources. In environmental protection, sonar imagery is used for seabed ecological research, coral reef monitoring, and marine pollution assessment. In scientific research, sonar imagery is used for marine geological surveys, biodiversity studies, and marine process analysis. Furthermore, in underwater archaeology and salvage operations, sonar imagery helps locate and identify shipwrecks and other sunken objects. These applications demonstrate the practical value of underwater sonar imagery in various fields, and its significance for the development and protection of marine resources, marine scientific research, and national security.

[0003] Currently, underwater sonar target recognition technology is experiencing rapid development. With the increasing development and utilization of marine resources, underwater sonar image target recognition has become a research hotspot, involving multiple aspects such as underwater acoustic signal processing, sonar image target recognition, and intelligent information processing. Imaging sonar technologies such as forward-looking sonar, side-scan sonar, and synthetic aperture sonar play a crucial role in underwater target recognition. Underwater target detection methods include traditional methods and deep learning-based methods, among which the combination of deep learning and transfer learning shows particular promise. Furthermore, research on sonar image denoising, segmentation, and target recognition is of great significance for understanding the underwater target processing process, identifying key scientific questions, and exploring possible solutions.

[0004] However, the challenges facing underwater sonar target recognition technology are also driving continuous innovation and development. The complexity of the underwater environment and the scarcity of samples increase the difficulty of the task, prompting researchers to constantly explore new technologies and methods to improve the accuracy and efficiency of underwater target detection. Deep learning-based methods have shown particular promise in sonar image target recognition, but problems such as signal noise and complex underwater environments still need to be addressed. In summary, underwater sonar target recognition technology is constantly progressing, but the challenges it faces are also driving continuous innovation and development. Summary of the Invention

[0005] The purpose of this invention is to provide a method for underwater sonar image target recognition based on shadow features. This method consists of three core modules: a feature extraction module, a feature fusion module, and a target recognition module. The feature extraction module uses digital image processing technology to extract corresponding features from the image, providing strong support for subsequent target recognition. The feature fusion module uses a multi-head attention mechanism to fuse multiple features, enabling the model to capture more comprehensive information. The target recognition module classifies the target based on the extracted features.

[0006] This invention adds shadow features to traditional underwater sonar image recognition methods, enabling the network to understand partial semantic information of the image and thus achieve better classification results. This improvement automatically processes shadow areas based on image features and the needs of the actual task, increasing the model's information extraction capabilities and thereby improving the accuracy, efficiency, and robustness of underwater sonar image target recognition. The application of this invention can help experts better understand the marine environment and perceive underwater conditions; it also improves the model's performance and generalization ability, providing strong technical support for target recognition tasks in underwater operations and military applications.

[0007] The technical solution adopted in this invention is as follows:

[0008] The technical solution of this invention for target identification using shadow features of underwater side-scan sonar images mainly relies on the shadow effects produced by target objects in the side-scan sonar images. Due to the limited light in the underwater environment, sonar equipment generates images by emitting sound waves and recording the echoes. When sound waves encounter underwater objects, they form acoustic shadows behind the objects. These shadows contain information about the object's shape, size, and location. By analyzing these shadow features, underwater objects such as shipwrecks, reefs, and schools of fish can be identified. In practical applications, this technical solution typically requires combining image processing algorithms and machine learning techniques to improve the accuracy and efficiency of identification.

[0009] To achieve target recognition using shadow features from underwater side-scan sonar images, a comprehensive technical workflow was developed. First, the side-scan sonar system collects a large amount of underwater side-scan sonar image data. This data can vary significantly under different environmental conditions, including factors such as water turbidity, temperature, and depth. The collected side-scan sonar image data needs to be labeled, specifying the object and its corresponding shadow area; these labels will serve as training labels for subsequent models. Next, the collected side-scan sonar images are preprocessed to reduce noise and enhance shadow features. Preprocessing steps may include filtering, gain adjustment, and contrast enhancement. Subsequently, the image histogram is analyzed, and a suitable threshold is selected based on the bimodal characteristics of the histogram. This threshold should be able to distinguish shadow areas from other areas. Using this threshold, the image is segmented, with pixels having gray values ​​below the threshold considered shadow areas, and pixels with gray values ​​above or equal to the threshold considered non-shadow areas. Dilation and erosion operations from computer graphics are used to remove holes in the shadow areas. Once the shadow region is identified, a deep learning model can be trained. This model requires training on a labeled dataset. A VGG network is used to extract semantic features from the image, obtaining semantic feature vectors. The shadow region's height and width are calculated using methods described later to obtain shadow features. An encoder structure transforms these shadow features to obtain shadow feature vectors. The semantic and shadow feature vectors are then fused using a hybrid structure combining self-attention and cross-attention mechanisms. Finally, a linear layer outputs labels to identify different types of underwater targets. In practical applications, the recognition system processes side-scan sonar images in real time and uses the trained model to identify targets within the images. To improve accuracy and robustness, it may be necessary to combine other side-scan sonar image features, such as reflection intensity and echo structure, as well as multi-sensor data fusion techniques. Furthermore, to adapt to different water areas and mission requirements, the recognition system may need to be regularly updated and optimized to adapt to new environments and target types. These technologies can effectively utilize the shadow features of underwater sonar images for target identification, providing support for underwater search and rescue, resource exploration, environmental monitoring, and other fields.

[0010] The technical solution of this invention is as follows:

[0011] A method for target recognition in underwater side-scan sonar images based on shadow features, comprising the following steps:

[0012] 1) Acquire multiple underwater side-scan sonar images and annotate them to obtain multiple training samples; the annotation content is the object and the corresponding shadow area;

[0013] 2) Train an underwater sonar image target recognition model using the training samples; the underwater sonar image target recognition model includes a feature extraction module, a feature fusion module, and a target recognition module; the feature fusion module includes a first self-attention unit, a second self-attention unit, a first cross-attention unit, a second cross-attention unit, a third self-attention unit, and a fourth self-attention unit; wherein,

[0014] The feature extraction module extracts semantic feature vectors and shadow feature vectors from the training samples and then inputs them into the feature fusion module.

[0015] The feature fusion module first processes the semantic feature vector through a first self-attention unit and then inputs it into a first cross-attention unit and a second cross-attention unit, respectively. It then processes the shadow feature vector through a second self-attention unit and inputs it into the first and second cross-attention units, respectively. Next, the first cross-attention unit fuses the input features and inputs them into a third self-attention unit, and the second cross-attention unit fuses the input features and inputs them into a fourth self-attention unit. Then, the features processed by the third self-attention unit are fused with the semantic feature vector to obtain a first fused feature vector, and the features processed by the fourth self-attention unit are fused with the shadow feature vector to obtain a second fused feature vector. Finally, the first fused feature vector and the second fused feature vector are fused to obtain the final fused feature of the training sample and input into the target recognition module.

[0016] The target recognition module predicts the category of the training sample based on the final fusion features of the training sample, and then optimizes the underwater sonar image target recognition model based on the prediction results and the annotation of the training sample.

[0017] 3) Input an underwater side-scan sonar image to be identified into the trained underwater sonar image target recognition model to obtain the target recognition result of the underwater side-scan sonar image to be identified.

[0018] Furthermore, the feature extraction module uses the VGG16 neural network model to extract semantic feature vectors from the training samples.

[0019] Furthermore, the method by which the feature extraction module extracts shadow feature vectors from the training samples is as follows:

[0020] 21) After converting the input underwater side-scan sonar image into a grayscale image, a filter is used for noise reduction. Then, a histogram of the underwater side-scan sonar image is generated. A threshold is selected based on the histogram to distinguish between shadow areas and other areas. The underwater side-scan sonar image is segmented using the threshold. Pixels with grayscale values ​​below the threshold are considered shadow areas, while pixels with grayscale values ​​above or equal to the threshold are considered non-shadow areas. Then, morphological operations are used to process the segmentation results.

[0021] 22) Calculate the height and width of the object corresponding to the shaded area;

[0022] 23) Perform embedding processing on the height and width of the object to obtain the corresponding height embedding vector and width embedding vector;

[0023] The height embedding vector is input into the encoder to obtain the first vector sequence, and the width embedding vector is input into the encoder to obtain the second vector sequence; then the encoder generates a shadow feature vector containing the object's height and width information based on the first vector sequence and the second vector sequence.

[0024] Furthermore, the encoder uses the average of the first vector sequence and the second vector sequence as the shadow feature vector; or the encoder merges the first vector sequence and the second vector sequence as the shadow feature vector.

[0025] Furthermore, according to the formula Calculate the height H of the object corresponding to the shaded area. t Wherein, Hs is the vertical distance of the towed sonar device from the water surface, Rs is the straight-line distance from the sonar transmitter to the object corresponding to the shadow area, Ls is the length of the shadow area, Rh is the horizontal distance from the towed sonar device to the object corresponding to the shadow area, Ht is the height of the object corresponding to the shadow area from the seabed, and Lt is the horizontal length of the object corresponding to the shadow area.

[0026] Furthermore, according to the formula Calculate the width W of the object corresponding to the shaded area. t Where Ws is the width of the shadow area and Wt is the actual width of the object corresponding to the shadow area.

[0027] Furthermore, the underwater side-scan sonar images are preprocessed and then annotated; the preprocessing steps include filtering, gain adjustment, and contrast enhancement.

[0028] A server is characterized by comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the methods described above.

[0029] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program implements the above-described method when executed by a processor.

[0030] The beneficial effects of this invention are as follows:

[0031] This invention provides a method for constructing an underwater side-scan sonar image target recognition model based on shadow features. This method achieves efficient and accurate underwater target recognition by combining shadow feature extraction and feature fusion. Its advantages are mainly reflected in the following aspects: First, this invention increases the information available to the model through shadow features, enabling better recognition of objects that are difficult to identify due to angle limitations. This helps the model better handle noise, occlusion, or other image problems that may affect performance, improving the accuracy and robustness of target recognition. Second, this invention utilizes a feature fusion module to further fuse multiple feature maps, reducing information loss during processing and analysis, providing more interpretability for decision-making, and making the results easier to understand. By combining these features, the target recognition module can more accurately identify underwater targets, improving the precision and efficiency of target recognition. Attached Figure Description

[0032] Figure 1 This is an overall framework diagram of the present invention.

[0033] Figure 2 This is a side view illustrating the principle of side-scan sonar imaging.

[0034] Figure 3 This is a top view illustrating the principle of side-scan sonar imaging. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0036] This invention provides a method for constructing an underwater sonar image target recognition model based on shadow features, such as... Figure 1 As shown, it mainly includes three modules: feature extraction, feature fusion, and target recognition. The feature extraction module includes shadow features and deep semantic information extraction from the image; the feature fusion module fuses shadow feature vectors and semantic feature vectors; the target recognition module, based on the feature representation, further applies a classifier to classify the target, converts the model's output into a probability distribution, thereby determining the probability that the input image belongs to each category, and finally outputs the recognition result.

[0037] 1. Feature Extraction Module

[0038] The feature extraction module refers to the use of two methods to extract semantic features and shadow features from the underwater side-scan sonar image to be detected. The specific steps of the feature extraction module are described below.

[0039] 1.1 Semantic Feature Extraction

[0040] Step 1: The semantic extraction part of the side-scan image uses a VGG16 network, requiring a pre-trained VGG16 model. A pre-trained VGG16 model refers to a VGG16 neural network model already trained on a large image database (such as ImageNet). The input image size is adjusted to a size acceptable to the VGG16 model, typically 224x224 pixels. Simultaneously, the image needs to be preprocessed according to the data normalization method used during VGG16 training.

[0041] Step Two: The input image first enters the first layer of the VGG16 network, the convolutional layer. The convolutional layer extracts local features of the image through a series of convolutional kernels (also called filters or feature detectors). Each kernel slides across the image, calculating its dot product with a local region, generating a feature map. These feature maps represent the feature responses of the image at different locations. Pooling layers, such as max pooling layers, are typically inserted between consecutive convolutional layers. Pooling layers reduce the spatial size of the feature map by sliding a small window (e.g., 2x2) across the feature map and taking the maximum value within the window as the pooled feature value, while retaining the most important feature information. This process is repeated multiple times, increasing the number of convolutional kernels each time to extract more complex features. As the network depth increases, the number of feature maps also increases, while the spatial size gradually decreases.

[0042] Step 3: After a series of convolutional and pooling layers, the image is transformed into a series of feature maps, each representing a specific feature of the image. At the end of the network, the fully connected layers of VGG16 are removed and replaced with a global average pooling layer after the last convolutional layer. This pooling layer averages the values ​​of each feature map, transforming each feature map into a single numerical value. The result of global average pooling is a one-dimensional vector, denoted as vector j, which is the semantic feature vector of the image and can be used as part of the input to the subsequent feature fusion module. The length of this vector is equal to the number of feature maps output by the last convolutional layer.

[0043] 1.2 Shadow Feature Extraction

[0044] Step 1: Calculate the shadow region of the object in the side-scan image. First, convert the original side-scan sonar image to grayscale. Use filters such as Gaussian filtering or median filtering to denoise the grayscale sonar image, reducing the impact of noise on the threshold selection. Then, analyze the histogram of the grayscale sonar image and select an appropriate threshold based on the bimodal characteristics of the histogram. This threshold should be able to distinguish shadow areas from other areas. Use this threshold to segment the image, considering pixels with grayscale values ​​below the threshold as shadow areas and pixels with grayscale values ​​above or equal to the threshold as non-shadow areas. To improve the segmentation results, morphological operations such as erosion and dilation can be used to remove noise and fill small holes. Finally, mark the detected shadow areas. The corresponding shadow range can be obtained through contour calculation methods. This process can be accomplished using the findContours() function provided by OpenCV, which calculates the minimum bounding rectangle of a closed region. The length and width of the rectangle can be used for subsequent calculations of the actual height and width of the object.

[0045] Step 2: Model the object based on the imaging characteristics of side-scan sonar. Calculate various parameters of the object using the model. The model should be feasible and simple. Feasible means that the data can be obtained from side-scan sonar data, and simple means that a simple model should be obtained to represent the actual object.

[0046] To simplify the process, the object imaging process can be modeled using side and top views. The following details how to model side-scan sonar images and the calculation methods for specific parameters.

[0047] 1) How to calculate the height of an object:

[0048] This section corresponds to Figure 1 The height calculation part in the mid-shading feature extraction. Figure 2 In this context, "Hs" refers to the vertical distance between the sonar towed fish (i.e., the towed sonar device) and the water surface; "Rs" refers to the straight-line distance from the sonar towed fish to the reflecting target; "Ls" refers to the length of the shadow area produced by the target in the sonar scan image, which can be used to estimate the size of the target; "Rh" refers to the horizontal distance from the sonar towed fish to the target; "Ht" refers to the vertical height of the target from the seabed; and "Lt" refers to the horizontal length of the target. After obtaining these parameters through side-scan sonar, Ht, the height of the object, can be calculated using the following formula.

[0049]

[0050] 2) How to calculate the width of an object:

[0051] This section corresponds to Figure 1 The width calculation part in shadow feature extraction. Ws is the width of the shadow area, and Wt is the actual width of the object. It's difficult to directly calculate the metrics when processing top views. Obtaining the corresponding metrics from the side view can be considered. Figure 3 By drawing a horizontal auxiliary line at the contact point between the object and the sonar, and calculating the width of the object using the corresponding proportions, the width of the object can be obtained.

[0052]

[0053] Step 3: Use the Encoder structure to obtain the corresponding shadow feature vector. First, the object's height and width inputs need to be treated as a sequence: [H t W t The input vectors are then embedded to transform them into vectors with higher dimensions. This is typically done through an embedding layer, which maps the input to a predefined vector space as a sequence of vectors. In other common feature vector encoding processes using encoder structures, positional encoding information—the relationship between actual index positions in the vector sequence—is required. However, in this process, since there is no contextual relationship between the sequences, this can be handled directly by the encoder.

[0054] Next, these vector sequences are fed into the Encoder, which contains multiple layers, each including a multi-head self-attention mechanism and a feedforward network. The self-attention mechanism allows the model to consider information from other inputs when processing each input, while the feedforward network performs further non-linear transformations on the output of the attention mechanism. After processing through all layers, the Encoder's output is a series of vectors, each corresponding to an element from the input sequence. Since there are only two inputs, the output will be a sequence containing two vectors. These two vectors can be averaged or otherwise combined to obtain a shadow feature vector, denoted as vector k, which will contain combined information from both the height and width inputs.

[0055] 2. Feature Fusion Module

[0056] The specific steps of the feature fusion module process are explained below.

[0057] Step 1: Process the input using a self-attention mechanism. During feature fusion, self-attention layers are first applied independently to the semantic and shading feature vectors. This is done to allow each feature to focus on its own characteristics, thereby strengthening the connections between internal elements.

[0058] Step Two: Building upon this, a cross-attention mechanism is used. The purpose of this layer is to establish a connection between the two types of features, promoting information interaction and fusion between them. The semantic feature vector and the shadow feature vector, processed by the self-attention mechanism, are input into the cross-attention layer. The cross-attention mechanism can learn the correlation between different feature vectors, enabling the model to focus on the most relevant parts between different feature vectors. In this way, the two feature vectors can complement each other, improving the model's ability to understand and process data.

[0059] Step 3: Connecting different feature vectors. Let 'a' be the result of the semantic feature vector after passing through the cross-attention mechanism, and let 'b' be the result of the shadow feature vector after passing through the cross-attention mechanism. Then, pass 'a' and 'b' through a self-attention mechanism module respectively, and denote the results as 'a'. ′ With b ′ Then, add it to the two original feature vectors output by the feature extraction layer, i.e., a ′ +j and b ′ +k, yielding two outputs. This feature fusion method is more efficient because the fusion process considers the internal connections between identical features and the interconnections between different features, thus providing a more powerful representation of both features. Finally, the Concentrate function is used to concatenate the two outputs as the final output of the feature fusion module.

[0060] 3. Target Recognition Module

[0061] The specific process of the target recognition module is described below.

[0062] Step 1: A classifier network typically consists of an input layer, one or more hidden layers, and an output layer. The number of neurons in the input layer matches the dimension of the final feature vector obtained after feature fusion. The number and size of the hidden layers depend on the complexity of the problem and the size of the dataset; each layer uses an activation function (set to ReLU) to introduce non-linearity. The number of neurons in the output layer matches the number of classes; for multi-class problems, the Softmax activation function is used.

[0063] Step 2: Input the feature vector output by the feature fusion module into the classifier network model, and calculate the probability of each category through the internal computing unit of the model.

[0064] Step 3: Output the name of the recognition result based on the category with the highest probability in Step 3.

[0065] The key points of this invention are mainly reflected in the following aspects:

[0066] 1. Diverse Feature Extraction Module: Utilizing digital image processing techniques, shadow features are extracted from underwater sonar images; deep learning techniques are employed to obtain high-order semantic features from side-scan sonar images. These features comprehensively reflect the attributes of targets in the images, enabling an understanding of the inherent relationships within the data and providing strong support for subsequent target recognition. Diverse feature extraction methods improve the accuracy and robustness of target recognition.

[0067] 2. High Efficiency of the Feature Fusion Module. During cross-modal fusion, attention maps for different features were successfully obtained by applying self-attention layers and bidirectional cross-attention layers. These attention maps reveal important relationships within each feature and between different features. This fusion method, which comprehensively considers both internal and inter-feature connections, significantly improves system performance because it can more effectively utilize and integrate information from multiple sources.

[0068] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A method for target recognition in underwater side-scan sonar images based on shadow features, comprising the following steps: 1) Acquire multiple underwater side-scan sonar images and annotate them to obtain multiple training samples; the annotation content is the object and the corresponding shadow area; 2) Train an underwater sonar image target recognition model using the training samples; the underwater sonar image target recognition model includes a feature extraction module, a feature fusion module, and a target recognition module; the feature fusion module includes a first self-attention unit, a second self-attention unit, a first cross-attention unit, a second cross-attention unit, a third self-attention unit, and a fourth self-attention unit; wherein, The feature extraction module extracts semantic feature vectors and shadow feature vectors from the training samples and then inputs them into the feature fusion module. The feature fusion module first processes the semantic feature vector through a first self-attention unit and then inputs it into a first cross-attention unit and a second cross-attention unit, respectively. It then processes the shadow feature vector through a second self-attention unit and inputs it into the first and second cross-attention units, respectively. Next, the first cross-attention unit fuses the input features and inputs them into a third self-attention unit, and the second cross-attention unit fuses the input features and inputs them into a fourth self-attention unit. Then, the features processed by the third self-attention unit are fused with the semantic feature vector to obtain a first fused feature vector, and the features processed by the fourth self-attention unit are fused with the shadow feature vector to obtain a second fused feature vector. Finally, the first fused feature vector and the second fused feature vector are fused to obtain the final fused feature of the training sample and input into the target recognition module. The target recognition module predicts the category of the training samples based on the final fusion features of the training samples. Then, based on the prediction results and the annotations of the training samples, the underwater sonar image target recognition model is optimized; 3) Input an underwater side-scan sonar image to be identified into the trained underwater sonar image target recognition model to obtain the target recognition result of the underwater side-scan sonar image to be identified.

2. The method according to claim 1, characterized in that, The feature extraction module uses the VGG16 neural network model to extract semantic feature vectors from the training samples.

3. The method according to claim 1, characterized in that, The method by which the feature extraction module extracts shadow feature vectors from the training samples is as follows: 21) After converting the input underwater side-scan sonar image into a grayscale image, a filter is used for noise reduction. Then, a histogram of the underwater side-scan sonar image is generated. A threshold is selected based on the histogram to distinguish between shadow areas and other areas. The underwater side-scan sonar image is segmented using the threshold. Pixels with grayscale values ​​below the threshold are considered shadow areas, while pixels with grayscale values ​​above or equal to the threshold are considered non-shadow areas. Then, morphological operations are used to process the segmentation results. 22) Calculate the height and width of the object corresponding to the shaded area; 23) The height and width of the object are embedded to obtain the corresponding height embedding vector and width embedding vector respectively; the height embedding vector is input into the encoder to obtain the first vector sequence, and the width embedding vector is input into the encoder to obtain the second vector sequence; then the encoder generates a shadow feature vector containing the height and width information of the object based on the first vector sequence and the second vector sequence.

4. The method according to claim 3, characterized in that, The encoder uses the average of the first vector sequence and the second vector sequence as the shadow feature vector; or the encoder merges the first vector sequence and the second vector sequence as the shadow feature vector.

5. The method according to claim 3, characterized in that, According to the formula Calculate the height H of the object corresponding to the shaded area. t Wherein, Hs is the vertical distance of the towed sonar device from the water surface, Rs is the straight-line distance from the sonar transmitter to the object corresponding to the shadow area, Ls is the length of the shadow area, Rh is the horizontal distance from the towed sonar device to the object corresponding to the shadow area, Ht is the height of the object corresponding to the shadow area from the seabed, and Lt is the horizontal length of the object corresponding to the shadow area.

6. The method according to claim 5, characterized in that, According to the formula Calculate the width W of the object corresponding to the shaded area. t Where Ws is the width of the shadow area and Wt is the actual width of the object corresponding to the shadow area.

7. The method according to claim 1, 2, or 3, characterized in that, The underwater side-scan sonar images are preprocessed and then annotated; the preprocessing steps include filtering, gain adjustment, and contrast enhancement.

8. A server, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images