Underwater target recognition method, device and equipment

Abstract

The invention relates to the technical field of underwater target detection, and discloses an underwater target recognition method, device and equipment, comprising: collecting an acoustic target echo signal of a water area to be recognized; preprocessing the acoustic target echo signal to obtain a corresponding bright spot model map and a Mel spectrum map; inputting the bright spot model map into a first feature extractor, using a block segmentation module to segment the bright spot model map into a plurality of blocks, and then merging and position encoding to obtain sequence data; using a shift window transformer to extract features of the sequence data to obtain bright spot model map features; extracting features of the Mel spectrum map through a second feature extractor composed of a plurality of serially connected two-dimensional convolutional layers to obtain Mel spectrum map features; using a block embedding module to fuse the bright spot model map features with the Mel spectrum map features to obtain fused features; inputting the fused features into a classifier, outputting a classification result of whether there is an underwater target, and completing underwater target recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater target detection, and in particular to an underwater target recognition method, device and equipment. Background Art

[0002] The underwater target recognition technology identifies underwater targets by analyzing the target echo signal. Traditional underwater target recognition mainly relies on human observers to perform real-time analysis of the signals received by the array element to identify the type of target. Later, through the application of acoustic signal theory and modern spectrum theory, the recognition accuracy and efficiency have been improved to a certain extent. However, with the influence of various types of clutter and sea environment and the higher requirements for recognition accuracy, the traditional underwater target recognition method has become stretched at this stage. Therefore, the method of using artificial intelligence to identify underwater acoustic targets has emerged. The intelligent recognition method of underwater acoustic targets analyzes and extracts features from the original acoustic data, and then inputs these features into the intelligent algorithm after processing, trains a classifier, and realizes automatic target recognition. This process has two main parts: feature extraction and classifier design. For these two parts, two different intelligent development directions have been generated: feature-based methods and model-based methods.

[0003] Feature-based methods capture the characteristic differences of signals by analyzing the original signals; effective features can not only simplify the design of classifiers, but also have a direct impact on the accuracy of recognition. However, different sea areas and environmental information will have a great impact on the characteristic information of the target. Model-based methods focus more on how to apply complex intelligent classification algorithms to the field of underwater acoustic target recognition, such as convolutional neural networks (CNN), shallow neural networks (SNN) and deep Q networks (DQN). When there is sufficient data, they can automatically extract valuable features from training data without expert knowledge; however, in the field of underwater acoustic target recognition where data is scarce and data annotation is difficult, the amount of available data is usually insufficient to meet the requirements of training, which will lead to insufficient model training or overfitting problems. In addition, model-based methods lack physical interpretability, and robustness and transferability are difficult to guarantee, and cannot provide observers with useful auxiliary decision information. Therefore, model-based methods cannot be directly used for underwater acoustic target recognition.

[0004] Patent CN202110377531, titled "An underwater target recognition method based on active deep learning", pre-processes the raw data of sonar signals and divides them into candidate sets and test sets; selects a portion of data from the divided candidate sets, moves it from the candidate sets to the training set after annotation, and uses it as initial training data; uses the data in the training set to train the CNN model, and selects a certain amount of data from the remaining candidate sets according to the active learning strategy, moves it to the training set after annotation, and trains the final underwater target recognition model after repeated multiple times. However, the CNN model it uses is relatively complex. Although it significantly reduces the cost of data annotation, it still has high requirements for the amount of data for training samples, and still cannot achieve a high recognition accuracy in the case of small samples. Patent CN202010427316, entitled "A method for underwater acoustic target recognition based on weighted support vector machine", extracts frame features from the original underwater acoustic target signal, builds a weighted support vector machine recognition model, uses the grid search method to find the optimal kernel function parameters and penalty factors, and trains the weighted support vector machine recognition model to obtain a support vector machine model with underwater target recognition function, with a recognition accuracy of more than 80%. However, the extracted underwater acoustic target features have poor discrimination, low recognition accuracy, and require the collection of a large number of underwater acoustic signals emitted by underwater acoustic targets before marking the data, which has high manpower and material costs, and the recognition accuracy is significantly reduced under small sample data. Patent CN201910661350, entitled Intelligent Identification Method for Hydroacoustic Targets Based on Big Data, collects a large number of hydroacoustic signals emitted by hydroacoustic targets, extracts appropriate features, and then marks the data as having targets or not, and establishes a training sample set to train the artificial intelligence model; in actual use, the surrounding hydroacoustic signals are collected and features are extracted, and then input into the trained artificial intelligence model to determine the presence and type of hydroacoustic targets, and can identify hydroacoustic targets at a longer distance; this method also requires first collecting a large number of hydroacoustic signals emitted by hydroacoustic targets, and then marking the data, which has high manpower and material costs, and the recognition accuracy is significantly reduced under small sample data.

[0005] In summary, most of the existing underwater target recognition methods are model-based methods. When identifying underwater acoustic targets in different sea areas and with different emission signal forms, traditional CNN networks have difficulty achieving high accuracy in underwater acoustic target recognition under small sample conditions. In addition, model-based methods lack physical interpretability, and robustness and transferability cannot be guaranteed. Summary of the invention

[0006] To this end, the technical problem to be solved by the present invention is to overcome the problems in the prior art of low accuracy in underwater target recognition and lack of interpretability of the model when faced with small sample data.

[0007] In order to solve the above technical problems, the present invention provides an underwater target recognition method, comprising:

[0008] Collecting echo signals of hydroacoustic targets in the water area to be identified;

[0009] Preprocessing the underwater acoustic target echo signal to obtain a corresponding bright spot model map and a Mel-spectrogram;

[0010] Input the bright spot model image into the first feature extractor, use the block segmentation module to segment the bright spot model image into multiple blocks, merge and position encode them, and obtain sequence data; use the shift window transformer to extract features from the sequence data to obtain the bright spot model image features;

[0011] The mel-spectrogram is subjected to feature extraction by a second feature extractor composed of a plurality of two-dimensional convolutional layers connected in series to obtain mel-spectrogram features;

[0012] The block embedding module is used to fuse the bright spot model graph features with the Mel-spectrogram features to obtain the fused features;

[0013] The fused features are input into the classifier, and the classification result of whether there is an underwater target is output to complete the underwater target recognition.

[0014] Preferably, preprocessing the underwater acoustic target echo signal to obtain a corresponding bright spot model map includes: constructing a matched filter, performing matched filtering on the underwater acoustic target echo signal, and obtaining a corresponding bright spot model map.

[0015] Preferably, the underwater acoustic target echo signal is preprocessed to obtain the corresponding Mel spectrogram, including: framing, windowing, and Fourier transforming the underwater acoustic target echo signal, and then using a Mel filter bank to perform discrete cosine transform and size transformation to obtain the Mel spectrogram.

[0016] Preferably, the block segmentation module is used to segment the bright spot model image into multiple blocks, and then merge and position encode them to obtain sequence data, including:

[0017] According to the preset side length, the bright spot model graph is divided into multiple small blocks of the same size, and a position encoding tensor is preset for each small block;

[0018] After flattening multiple small blocks of the same size into one-dimensional vectors, they are projected to the preset embedding dimension through a linear layer to obtain the corresponding initial tensor;

[0019] Add the preset position encoding tensor corresponding to each small block to the initial tensor to obtain the embedded tensor;

[0020] The embedding tensors of all small blocks in the bright spot model graph are stacked to obtain the sequence data corresponding to the bright spot model graph.

[0021] Preferably, the shift window converter comprises:

[0022] A first feature extraction layer, comprising a first normalization layer, a window multi-head attention mechanism, a second normalization layer and a first feedforward network connected in series in a forward propagation direction; the output of the window multi-head attention mechanism is residually connected to the input of the first normalization layer as the input of the second normalization layer; the input of the second normalization layer is residually connected to the output of the first feedforward network as the output of the first feature extraction layer;

[0023] The second feature extraction layer includes a third normalization layer, a sliding window multi-head attention mechanism, a fourth normalization layer and a second feedforward network which are sequentially connected in series along the forward propagation direction; the output of the first feature extraction layer is residually connected with the output of the sliding window multi-head attention mechanism as the input of the fourth normalization layer; the input of the fourth normalization layer is residually connected with the output of the second feedforward network as the output of the second feature extraction layer.

[0024] Preferably, the second feature extractor comprises an input layer, a first two-dimensional convolutional layer, an activation function layer, a second two-dimensional convolutional layer, an activation function layer, a maximum pooling layer and an output layer which are sequentially connected in series along a forward propagation direction.

[0025] Preferably, the block embedding module is used to fuse the bright spot model graph features with the mel spectrum graph features to obtain the fused features, including:

[0026] Process the bright spot model graph features into feature graph blocks of a preset size, and convert them into bright spot vectors of a preset dimension;

[0027] Process the Mel-spectrogram features into feature blocks of a preset size and convert them into Mel-spectrogram vectors of a preset dimension;

[0028] Concatenate the bright spot vector in the bright spot model graph feature with its corresponding mel spectrum vector in the mel spectrum graph feature to obtain multiple block vectors;

[0029] Concatenate all block vectors to obtain fused features.

[0030] Preferably, the classifier is a binary classifier.

[0031] This embodiment provides an underwater target recognition device, including:

[0032] A signal acquisition module, used for acquiring echo signals of hydroacoustic targets in the water area to be identified;

[0033] A signal preprocessing module, used to preprocess the underwater acoustic target echo signal to obtain a corresponding bright spot model map and a Mel spectrum map;

[0034] The bright spot model graph feature acquisition module is used to input the bright spot model graph into the first feature extractor, use the block segmentation module to segment the bright spot model graph into multiple blocks, merge and position encode them, and obtain sequence data; use the shift window transformer to extract features from the sequence data to obtain the bright spot model graph features;

[0035] A mel-spectrogram feature acquisition module is used to extract features from the mel-spectrogram through a second feature extractor composed of a plurality of two-dimensional convolutional layers connected in series, thereby acquiring mel-spectrogram features;

[0036] A feature fusion module is used to fuse the bright spot model graph features with the Mel spectrum graph features using the block embedding module to obtain fused features;

[0037] The classification module is used to input the fusion features into the classifier, output the classification result of whether there is an underwater target, and complete the underwater target recognition.

[0038] This embodiment provides an underwater target recognition device, including:

[0039] An underwater target recognition device is used to collect and execute the steps of the underwater target recognition method as described above based on the echo signal of the hydroacoustic target in the water area to be recognized, so as to obtain the classification result of the underwater target recognition;

[0040] The display is communicatively connected with the underwater target recognition device and is used to obtain and display the classification results of the underwater target recognition.

[0041] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0042] The underwater target recognition method of the present invention pre-processes the hydroacoustic target echo signal collected in the water area to be identified, obtains the bright spot model map and the mel spectrum map, and embeds the size information of the target represented by the hydroacoustic target echo signal into the bright spot model map feature as prior knowledge through the bright spot model map. By explicitly embedding the size information, a physical attribute of the target, into the feature, the model can more directly consider this important attribute when making classification decisions, thereby making the decision-making process of the model more intuitive and understandable, and improving the interpretability of the model classification. The mel spectrum map feature and the bright spot model map feature are fused by block embedding, and features are extracted in each small block, which is conducive to introducing local information of the image, helping the model to better understand the features of different areas in the image, and improving the classification performance. The fused features are used for classification, which reduces the difficulty of model training, improves the effectiveness and robustness of the features extracted from the hydroacoustic target echo signal, and improves the classification accuracy. At the same time, the present invention does not rely on a specific priori data set for training, and when facing small sample data sets of different sea areas, different environments, and different signal forms, it has good discrimination performance, high underwater target recognition accuracy, and strong transferability.

[0043] The present invention utilizes a block segmentation module to segment the image and extract many small blocks, and merges these small blocks to reduce the dimension of the data; then the merged small blocks are position-encoded so that the subsequent model can process the sequence data more orderly; a shifting window transformer is used to extract features from the sequence data, and based on the window multi-head attention mechanism, the sliding window multi-head attention mechanism and the residual connection, the accuracy of the features extracted based on the bright spot model graph is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0045] Figure 1 It is a flow chart of the steps of the underwater target identification method provided by the present invention;

[0046] Figure 2 It is a structural schematic diagram of the shift window converter provided by the present invention;

[0047] Figure 3 It is a feature extraction and feature fusion strategy diagram provided by the present invention;

[0048] Figure 4 This is the underwater acoustic target recognition strategy diagram provided by the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0050] Reference Figure 1 As shown, the step flow chart of the underwater target recognition method of the present invention includes the following steps:

[0051] S101: collecting echo signals of hydroacoustic targets in the water area to be identified;

[0052] S102: Preprocessing the underwater acoustic target echo signal to obtain a corresponding bright spot model map and a Mel-spectrogram;

[0053] S103: inputting the bright spot model image into the first feature extractor, using the block segmentation module to segment the bright spot model image into multiple blocks, then merging and position encoding, to obtain sequence data; using the shift window transformer to extract features from the sequence data, to obtain bright spot model image features;

[0054] S104: extracting features from the mel-spectrogram by a second feature extractor composed of a plurality of two-dimensional convolutional layers connected in series, and obtaining mel-spectrogram features;

[0055] S105: using a block embedding module to fuse the bright spot model graph features with the mel spectrum graph features to obtain fused features;

[0056] S106: Input the fused features into the classifier, output the classification result of whether there is an underwater target, and complete the underwater target recognition.

[0057] Specifically, in step S102, preprocessing is performed to obtain a bright spot model image and a mel-spectrogram, including:

[0058] S102-1: construct a matched filter, perform matched filtering on the underwater acoustic target echo signal, and obtain a corresponding bright spot model map;

[0059] S102-2: After framing, windowing, and Fourier transforming the underwater acoustic target echo signal, a Mel filter bank is used to perform discrete cosine transform and size transformation to obtain a Mel spectrogram.

[0060] Since the echo signal of the underwater target contains size information, this embodiment uses the bright spot model to characterize the relevant physical information and embeds it into the original features as a priori knowledge. By explicitly embedding the size information, a physical attribute of the target, into the features, the model can more directly consider this important attribute when making classification decisions, thereby making the decision-making process of the model more intuitive and understandable, and improving the interpretability of the model classification.

[0061] Specifically, in step S103, the acquisition of the bright spot model graph features includes:

[0062] S103-1: Use the block segmentation module to segment the bright spot model image into multiple blocks, then merge and position encode them to obtain sequence data, including:

[0063] According to the preset side length, the bright spot model graph is divided into multiple small blocks of the same size, and a position encoding tensor is preset for each small block;

[0064] After flattening multiple small blocks of the same size into one-dimensional vectors, they are projected to the preset embedding dimension through a linear layer to obtain the corresponding initial tensor;

[0065] Add the preset position encoding tensor corresponding to each small block to the initial tensor to obtain the embedded tensor;

[0066] Stack the embedding tensors of all small blocks in the bright spot model graph to obtain the sequence data corresponding to the bright spot model graph;

[0067] S103-2: extracting features from the sequence data using a shift window transformer to obtain bright spot model graph features;

[0068] The present invention utilizes a block segmentation module to segment the image and extract many small blocks, and merges these small blocks to reduce the dimension of the data; then the merged small blocks are position-encoded so that the subsequent model can process the sequence data more orderly; a shifting window transformer is used to extract features from the sequence data, and based on the window multi-head attention mechanism, the sliding window multi-head attention mechanism and the residual connection, the accuracy of the features extracted based on the bright spot model graph is significantly improved.

[0069] Reference Figure 2 As shown, it is a structural schematic diagram of a shift window converter; the shift window converter comprises:

[0070] A first feature extraction layer, comprising a first normalization layer, a window multi-head attention mechanism, a second normalization layer and a first feedforward network connected in series in a forward propagation direction; the output of the window multi-head attention mechanism is residually connected to the input of the first normalization layer as the input of the second normalization layer; the input of the second normalization layer is residually connected to the output of the first feedforward network as the output of the first feature extraction layer;

[0071] The second feature extraction layer includes a third normalization layer, a sliding window multi-head attention mechanism, a fourth normalization layer and a second feedforward network which are sequentially connected in series along the forward propagation direction; the output of the first feature extraction layer is residually connected with the output of the sliding window multi-head attention mechanism as the input of the fourth normalization layer; the input of the fourth normalization layer is residually connected with the output of the second feedforward network as the output of the second feature extraction layer.

[0072] Specifically, in step S104, the second feature extractor for extracting mel-spectrogram features includes an input layer, a first two-dimensional convolutional layer, an activation function layer, a second two-dimensional convolutional layer, an activation function layer, a maximum pooling layer and an output layer which are sequentially connected in series along the forward propagation direction.

[0073] Specifically, in step S105, the acquisition of fusion features includes:

[0074] S105-1: Processing the bright spot model image features into feature image blocks of a preset size, and converting them into bright spot vectors of a preset dimension;

[0075] S105-2: Processing the Mel-spectrogram features into feature blocks of a preset size, and converting them into Mel-spectrogram vectors of a preset dimension;

[0076] S105-3: concatenating the bright spot vector in the bright spot model graph feature with its corresponding mel spectrum vector in the mel spectrum graph feature to obtain multiple block vectors;

[0077] S105-4: Concatenate all block vectors to obtain fusion features.

[0078] Reference Figure 3 As shown, it is a feature extraction and feature fusion strategy diagram provided in this embodiment; the feature fusion method proposed in the present invention fuses the Mel-spectrogram features and the bright spot model features by block embedding, which reduces the difficulty of model training and improves the effectiveness and robustness of the features.

[0079] In the embodiment of the present invention, the classifier is a binary classifier, which is used to output two classification results: the existence of an underwater target or the absence of an underwater target.

[0080] The underwater target recognition method of the present invention pre-processes the hydroacoustic target echo signal collected in the water area to be identified, obtains the bright spot model map and the mel spectrum map, and embeds the size information of the target represented by the hydroacoustic target echo signal into the bright spot model map feature as prior knowledge through the bright spot model map. By explicitly embedding the size information, a physical attribute of the target, into the feature, the model can more directly consider this important attribute when making classification decisions, thereby making the decision-making process of the model more intuitive and understandable, and improving the interpretability of the model classification. The mel spectrum map feature and the bright spot model map feature are fused by block embedding, and features are extracted in each small block, which is conducive to introducing local information of the image, helping the model to better understand the features of different areas in the image, and improving the classification performance. The fused features are used for classification, which reduces the difficulty of model training, improves the effectiveness and robustness of the features extracted from the hydroacoustic target echo signal, and improves the classification accuracy. At the same time, the present invention does not rely on a specific priori data set for training, and when facing small sample data sets of different sea areas, different environments, and different signal forms, it has good discrimination performance, high underwater target recognition accuracy, and strong transferability.

[0081] Reference Figure 4 As shown, it is a diagram of the underwater acoustic target recognition strategy of this embodiment. The specific strategy includes: first, by performing data preprocessing on the underwater acoustic target echo signal, including framing, windowing, Fourier transform, beamforming and matched filtering, etc., the corresponding bright spot model map and Mel spectrum map are obtained, and corresponding feature extraction modules are designed for the bright spot model map and Mel spectrum map respectively. The extracted image features are subjected to feature fusion by a suitable method, and after the feature fusion, they are input into the classifier for training, and the trained model can be used for the recognition of underwater acoustic targets.

[0082] Specifically, for the bright spot model graph, the Patch Partition module is first used to segment the image and extract many small blocks, and these small blocks are merged to reduce the dimension of the data; then the merged small blocks are position encoded so that the subsequent model can better process the sequence data; then the Swin Transformer layer is used to extract features from the sequence data. In this embodiment, the shift window transformer Swin Transformer module consists of two feature extraction layers. The first feature extraction is divided into layer normalization, window multi-head attention mechanism, layer normalization and feedforward network, and two residual connections are introduced; the second feature extraction changes the window multi-head attention mechanism into a sliding window multi-head attention mechanism, and the rest remains unchanged. A Swin Transformer module composed of this is an important part of the bright spot model graph feature extractor.

[0083] The underwater acoustic target recognition method based on domain knowledge embedding proposed in this invention embeds the size information of the target as prior knowledge into the original features through the bright spot model map. The bright spot model map is extracted using the Swin Transformer layer, which is significantly improved compared with the traditional two-dimensional convolution extraction method.

[0084] Specifically, for the Mel-spectrogram, two layers of two-dimensional convolutional layers are selected to extract features from the input image through convolution operations, and ReLU is selected as the activation function. Finally, a maximum pooling layer is added to reduce the dimension and extract significant features. After feature extraction, the block embedding module is used to fuse the two types of features. By dividing the image into small blocks and extracting features in each small block, it is beneficial to introduce local information of the image, and help the model better understand the features of different areas in the image, thereby improving classification performance.

[0085] Based on the above embodiment, in an embodiment of the present invention, an underwater target recognition method provided by the present invention is used to conduct an experiment based on the underwater acoustic data audio data including the target echo signal and the background underwater acoustic signal, wherein there are 332 target echo signal data and 332 background underwater acoustic signal data; specifically, the experimental process includes:

[0086] S201: construct image dataset;

[0087] (1) Construct a matched filter and perform matched filtering on the audio signal to obtain a bright spot model graph;

[0088] (2) The audio signal is framed, windowed, Fourier transformed, a Mel filter bank is applied, a discrete cosine transform is performed, and the size is transformed to obtain a Mel spectrogram with a size of 224*224*3.

[0089] S202: Divide the data into a training set and a test set with a ratio of 8:2, and set the random seed to 42.

[0090] S203: Input the image data set to two feature extractors respectively to perform feature extraction.

[0091] S204: Input the extracted features to the Patch Embedding layer for feature fusion.

[0092] S205: Build a model, using 3 fully connected layers as classifiers. The parameters are set as follows: 19,600 input neurons, 2 hidden layers, 256 and 32 hidden neurons respectively, 2 output neurons, cross entropy loss function, stochastic gradient descent optimizer, learning rate 0.001, and maximum number of training times 100.

[0093] S206: Input the test sample set for testing, and the test accuracy is 90.98%.

[0094] The method for underwater acoustic target recognition based on domain knowledge embedding proposed in the present invention embeds the radial size information of the target into the model through the bright spot model graph, thereby improving the interpretability of the model. The bright spot model graph features and the mel-spectrogram features are integrated based on the block embedding module, which helps the model learn the sequential features of the input sequence, reduces the difficulty of model training, and improves the effectiveness and robustness of the features. In addition, when facing small sample data sets in different sea areas, different environments, and different signal forms, the model recognition accuracy is relatively high.

[0095] Based on the above embodiments, an underwater target recognition device provided by an embodiment of the present invention includes:

[0096] A signal acquisition module, used for acquiring echo signals of hydroacoustic targets in the water area to be identified;

[0097] A signal preprocessing module, used to preprocess the underwater acoustic target echo signal to obtain a corresponding bright spot model map and a Mel spectrum map;

[0098] The bright spot model graph feature acquisition module is used to input the bright spot model graph into the first feature extractor, use the block segmentation module to segment the bright spot model graph into multiple blocks, merge and position encode them, and obtain sequence data; use the shift window transformer to extract features from the sequence data to obtain the bright spot model graph features;

[0099] A mel-spectrogram feature acquisition module is used to extract features from the mel-spectrogram through a second feature extractor composed of a plurality of two-dimensional convolutional layers connected in series, thereby acquiring mel-spectrogram features;

[0100] A feature fusion module is used to fuse the bright spot model graph features with the Mel spectrum graph features using the block embedding module to obtain fused features;

[0101] The classification module is used to input the fusion features into the classifier, output the classification result of whether there is an underwater target, and complete the underwater target recognition.

[0102] The underwater target recognition device of the present embodiment is used to implement the aforementioned underwater target recognition method. Therefore, the specific implementation method of the underwater target recognition device can be seen in the embodiment part of the underwater target recognition method in the previous text. For example, the signal acquisition module, the signal preprocessing module, the bright spot model map feature acquisition module, the Mel spectrum map feature acquisition module, the feature fusion module and the classification module are respectively used to implement steps S101, S102, S103, S104, S105 and S106 in the aforementioned underwater target recognition method. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part, which will not be repeated here.

[0103] Based on the above embodiment, an embodiment of the present invention further provides an underwater target recognition device, including:

[0104] An underwater target recognition device is used to collect and execute the steps of the underwater target recognition method as described above based on the echo signal of the hydroacoustic target in the water area to be recognized, so as to obtain the classification result of the underwater target recognition;

[0105] The display is communicatively connected with the underwater target recognition device and is used to obtain and display the classification results of the underwater target recognition.

[0106] The underwater target recognition method of the present invention pre-processes the hydroacoustic target echo signal collected in the water area to be identified, obtains the bright spot model map and the mel spectrum map, and embeds the size information of the target represented by the hydroacoustic target echo signal into the bright spot model map feature as prior knowledge through the bright spot model map. By explicitly embedding the size information, a physical attribute of the target, into the feature, the model can more directly consider this important attribute when making classification decisions, thereby making the decision-making process of the model more intuitive and understandable, and improving the interpretability of the model classification. The mel spectrum map feature and the bright spot model map feature are fused by block embedding, and features are extracted in each small block, which is conducive to introducing local information of the image, helping the model to better understand the features of different areas in the image, and improving the classification performance. The fused features are used for classification, which reduces the difficulty of model training, improves the effectiveness and robustness of the features extracted from the hydroacoustic target echo signal, and improves the classification accuracy. At the same time, the present invention does not rely on a specific priori data set for training, and when facing small sample data sets of different sea areas, different environments, and different signal forms, it has good discrimination performance, high underwater target recognition accuracy, and strong transferability. The present invention utilizes a block segmentation module to segment the image and extract many small blocks, and merges these small blocks to reduce the dimension of the data; then the merged small blocks are position-encoded so that the subsequent model can process the sequence data more orderly; a shifting window transformer is used to extract features from the sequence data, and based on the window multi-head attention mechanism, the sliding window multi-head attention mechanism and the residual connection, the accuracy of the features extracted based on the bright spot model graph is significantly improved.

[0107] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0108] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0109] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0110] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0111] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for underwater target recognition, characterized in that: include: Collecting echo signals of hydroacoustic targets in the water area to be identified; Preprocessing the underwater acoustic target echo signal to obtain a corresponding bright spot model map and a Mel-spectrogram; Input the bright spot model image into the first feature extractor, use the block segmentation module to segment the bright spot model image into multiple blocks, merge and position encode them, and obtain sequence data; use the shift window transformer to extract features from the sequence data to obtain the bright spot model image features; The mel-spectrogram is subjected to feature extraction by a second feature extractor composed of a plurality of two-dimensional convolutional layers connected in series to obtain mel-spectrogram features; The block embedding module is used to fuse the bright spot model graph features with the Mel-spectrogram features to obtain the fused features; The fused features are input into the classifier, and the classification result of whether there is an underwater target is output to complete the underwater target recognition.

2. The underwater target recognition method according to claim 1, characterized in that: Preprocessing the underwater acoustic target echo signal to obtain a corresponding bright spot model map includes: constructing a matched filter, performing matched filtering on the underwater acoustic target echo signal, and obtaining a corresponding bright spot model map.

3. The underwater target recognition method according to claim 1, characterized in that: The underwater acoustic target echo signal is preprocessed to obtain a corresponding Mel spectrogram, including: framing, windowing, and Fourier transforming the underwater acoustic target echo signal, and then using a Mel filter bank to perform discrete cosine transform and size transformation to obtain a Mel spectrogram.

4. The underwater target recognition method according to claim 1, characterized in that: The block segmentation module is used to segment the bright spot model image into multiple blocks, and then merge and position encode them to obtain sequence data, including: According to the preset side length, the bright spot model graph is divided into multiple small blocks of the same size, and a position encoding tensor is preset for each small block; After flattening multiple small blocks of the same size into one-dimensional vectors, they are projected to the preset embedding dimension through a linear layer to obtain the corresponding initial tensor; Add the preset position encoding tensor corresponding to each small block to the initial tensor to obtain the embedded tensor; The embedding tensors of all small blocks in the bright spot model graph are stacked to obtain the sequence data corresponding to the bright spot model graph.

5. The underwater target recognition method according to claim 1, characterized in that: The shift window converter comprises: A first feature extraction layer, comprising a first normalization layer, a window multi-head attention mechanism, a second normalization layer and a first feedforward network connected in series in a forward propagation direction; the output of the window multi-head attention mechanism is residually connected to the input of the first normalization layer as the input of the second normalization layer; the input of the second normalization layer is residually connected to the output of the first feedforward network as the output of the first feature extraction layer; The second feature extraction layer includes a third normalization layer, a sliding window multi-head attention mechanism, a fourth normalization layer and a second feedforward network which are sequentially connected in series along the forward propagation direction; the output of the first feature extraction layer is residually connected with the output of the sliding window multi-head attention mechanism as the input of the fourth normalization layer; the input of the fourth normalization layer is residually connected with the output of the second feedforward network as the output of the second feature extraction layer.

6. The underwater target recognition method according to claim 1, characterized in that: The second feature extractor includes an input layer, a first two-dimensional convolution layer, an activation function layer, a second two-dimensional convolution layer, an activation function layer, a maximum pooling layer and an output layer which are sequentially connected in series along a forward propagation direction.

7. The underwater target recognition method according to claim 1, characterized in that: The block embedding module is used to fuse the bright spot model graph features with the Mel spectrum graph features to obtain the fused features, including: Process the bright spot model graph features into feature graph blocks of a preset size, and convert them into bright spot vectors of a preset dimension; Process the Mel-spectrogram features into feature blocks of a preset size and convert them into Mel-spectrogram vectors of a preset dimension; Concatenate the bright spot vector in the bright spot model graph feature with its corresponding mel spectrum vector in the mel spectrum graph feature to obtain multiple block vectors; Concatenate all block vectors to obtain fused features.

8. The underwater target recognition method according to claim 1, characterized in that: The classifier is a binary classifier.

9. An underwater target recognition device, characterized in that: include: A signal acquisition module, used for acquiring echo signals of hydroacoustic targets in the water area to be identified; A signal preprocessing module, used to preprocess the underwater acoustic target echo signal to obtain a corresponding bright spot model map and a Mel spectrum map; The bright spot model graph feature acquisition module is used to input the bright spot model graph into the first feature extractor, use the block segmentation module to segment the bright spot model graph into multiple blocks, merge and position encode them, and obtain sequence data; use the shift window transformer to extract features from the sequence data to obtain the bright spot model graph features; A mel-spectrogram feature acquisition module is used to extract features from the mel-spectrogram through a second feature extractor composed of a plurality of two-dimensional convolutional layers connected in series, thereby acquiring mel-spectrogram features; A feature fusion module is used to fuse the bright spot model graph features with the Mel spectrum graph features using the block embedding module to obtain fused features; The classification module is used to input the fusion features into the classifier, output the classification result of whether there is an underwater target, and complete the underwater target recognition.

10. An underwater target recognition device, characterized in that: include: An underwater target recognition device, used to collect and, based on the echo signal of the hydroacoustic target in the water area to be recognized, execute the steps of the underwater target recognition method according to any one of claims 1 to 8 to obtain the classification result of the underwater target recognition; The display is communicatively connected with the underwater target recognition device and is used to obtain and display the classification results of the underwater target recognition.

Images