An Underwater Acoustic Target Intelligent Detection System and Method Based on a Full-Bias Mutual Learning Strategy

Abstract

The present application provides an underwater acoustic target intelligent detection system and method based on a full-bias mutual learning strategy. The system includes: an underwater acoustic feature extraction module for extracting underwater acoustic features from underwater acoustic signals; and a full-bias mutual learning network including two networks trained by mutual learning, one of which is a full-bias network. The full-bias mutual learning network is used to input underwater acoustic features and output the classification of underwater acoustic signals. The advantages of the present application are as follows: It can directly process the received underwater acoustic signal data, with high real-time performance and fast response speed; it can improve the classification capabilities of two or more models simultaneously; it enhances the classification capabilities of weak items in the model mutual learning process and improves the overall recognition accuracy of the model; the trained artificial intelligence model combines the advantages of different networks, has strong generalization ability and high recognition accuracy.

Description

Technical Field

[0001] This application belongs to the technical field of underwater acoustic target recognition, and particularly relates to an underwater acoustic target intelligent detection system and method based on a full-bias mutual learning strategy. Background Technique

[0002] Underwater acoustic target intelligent recognition technology plays a crucial role in identifying different types of targets by analyzing passive signals. With the continuous development and increasing intelligence of underwater equipment systems, the importance of underwater acoustic target recognition technology has increased significantly.

[0003] Traditional underwater acoustic target recognition mainly relies on human observers to analyze the received signals in real time to identify the target type. Various theories and modern signal processing techniques, such as power spectrum, DEMON spectrum, and LOFAR spectrum, are used to assist observers in making judgments. However, different from target detection, underwater acoustic target recognition requires accurate identification of the actual target type. As the number of underwater targets increases, the amount of information processing involved also increases. The manual recognition method, which is high-intensity, time-consuming, and cumbersome, is no longer applicable to modern target recognition systems. Therefore, new technologies must be adopted to assist observers in completing target recognition. Underwater acoustic target intelligent recognition technology analyzes and extracts features from the original acoustic data. Then, these features are input into intelligent algorithms to train efficient classifiers for automated target recognition. The design of the classifier greatly affects the accuracy of underwater acoustic target intelligent recognition. In the field of deep learning, a single independently trained model is usually limited by the locality of data and the imbalance of samples, which will result in unsatisfactory model performance. In order to make full use of the relevant information between different models, model mutual learning has gradually become a popular research direction. Model mutual learning aims to improve the performance of each model by collaboratively learning the representations or knowledge between multiple models.

[0004] Deep mutual learning is an extension of model mutual learning, which further explores the interaction between multiple deep learning models. Deep mutual learning not only focuses on knowledge sharing between models, but also includes aspects such as parameter sharing, task sharing, and feature sharing between models.

[0005] The Chinese invention patent with the publication number CN110390949A and the name of "Underwater Acoustic Target Intelligent Recognition Method Based on Big Data" discloses: collecting a large number of underwater acoustic signals emitted by underwater acoustic targets, extracting appropriate features, then marking the data with or without targets, and establishing a training sample set to train an artificial intelligence model. In actual use, collecting the surrounding underwater acoustic signals and extracting features, and inputting them into the trained artificial intelligence model can determine the presence and type of underwater acoustic targets, and can identify underwater acoustic targets at a longer distance.

[0006] The technical solution of this invention patent judges whether there is a target for the collected underwater acoustic signals and is not competent for the task of target classification in the case of multiple targets.

[0007] The Chinese invention patent with the publication number CN115115886A and the title "Semi-Supervised Object Detection Method Based on Teacher-Student Model" discloses: obtaining a semi-supervised object detection dataset D; according to the fully supervised object detection method on the labeled dataset DL, using the model to predict the labels of the samples to obtain the teacher model; using the teacher model, making predictions on the unlabeled data XU, generating pseudo-labels, and updating the labeled set YU with the pseudo-labels; performing data augmentation on the unlabeled sample set XU in the fifth step and training the student model; updating the weight parameter θs of the student model into the weight θt of the teacher model in the way of exponential moving average; performing several rounds of iteration, and using the final teacher model as the target model trained by the semi-supervised object detection method based on the teacher-student model.

[0008] The technical solution of this invention patent uses the pseudo-labels generated by the teacher model to improve the model prediction ability, and it is difficult to improve the data classification ability that is not sensitive to the teacher-student network itself, and the improvement ability of this method is limited.

[0009] The Chinese invention patent with the publication number CN114663848A and the title "Object Detection Method and Device Based on Knowledge Distillation" discloses: training a teacher network using a sample image set to obtain a target teacher network, introducing a loss function, using a differentiable group search method to search for the aggregation feature weights of each convolutional group of the target teacher network group by group; using the aggregation feature weights to extract the corresponding aggregation features from each convolutional group of the target teacher network; using the extracted aggregation features as knowledge to perform aggregation feature distillation on the student network to obtain a target student network, and then inputting the image to be detected into the target student network to detect the target object and its location in the image to be detected. This embodiment extracts more knowledge in the teacher network by coupling features from different layers of the teacher network, has less computational overhead, and improves the accuracy of the student network when detecting target objects.

[0010] The performance of the model provided by the technical solution of this invention patent depends to a large extent on the performance of the teacher model, and at the same time, the weak items of the teacher model will also be brought to the student network, and the two models cannot be improved simultaneously. Summary of the Invention

[0011] The purpose of this application is to overcome the defect of insufficient accuracy in judging underwater acoustic signals in the prior art.

[0012] To achieve the above object, the present application proposes an underwater acoustic target intelligent detection system based on a full-biased mutual learning strategy, and the system includes:

[0013] An underwater acoustic feature extraction module for extracting underwater acoustic features from underwater acoustic signals; the underwater acoustic features include one-dimensional features and two-dimensional spectrograms; the one-dimensional features include 91 high-order statistical functions obtained by combining 7 low-order descriptors and 13 statistical functions; the low-order descriptors include: zero-crossing rate, spectral skewness, spectral kurtosis, spectral clarity, spectral center point, the first coefficient of the Mel frequency cepstral coefficient, and the fifth coefficient of the Mel frequency cepstral coefficient; the statistical functions include: arithmetic mean, maximum value, minimum value, maximum-minimum deviation, standard deviation, skewness, kurtosis, first quartile, second quartile, third quartile, 1-2 interquartile range, 2-3 interquartile range, and 1-3 interquartile range;

[0014] A full-biased mutual learning network, including two networks trained by mutual learning, one of which is a full-biased network; the full-biased mutual learning network is used to input underwater acoustic features and output the classification of underwater acoustic signals.

[0015] As an improvement of the above system, when the underwater acoustic features are input into the full-biased mutual learning network, the two-dimensional spectrogram is input into the full-biased network, and the one-dimensional features are input into the other network.

[0016] As an improvement of the above system, the training process of the full-biased mutual learning network includes:

[0017] Input the training sample set to perform mutual learning training on the two networks, and stop training when the training required accuracy is reached or the maximum training number of times is reached;

[0018] Set one of the networks as a full-biased network and perform mutual learning training again.

[0019] As an improvement of the above system, setting one of the networks as a full-biased network includes: canceling the optimizer of this network so that it no longer performs training, or changing this network to imitate the network output with an artificially set probability distribution.

[0020] As an improvement of the above system, the process of the underwater acoustic feature extraction module extracting the underwater acoustic features of the underwater acoustic signal includes:

[0021] Split the long sound signal data into segments of 1 second each;

[0022] Calculate 7 low-order descriptors and 13 statistical functions of the split signal data, and combine them with each other to obtain one-dimensional features; draw a spectrogram to obtain a two-dimensional spectrogram.

[0023] As an improvement of the above system, the full bias mutual learning network includes one deep convolutional network and one DNN network; the DNN network is a full bias network.

[0024] This application also provides an underwater acoustic target intelligent detection method based on the full bias mutual learning strategy, which is implemented based on the above system. The method includes:

[0025] Extract the underwater acoustic features in the underwater acoustic signal by using the underwater acoustic feature extraction module, input them into the trained full bias mutual learning network, and output the classification of the underwater acoustic signal.

[0026] Compared with the prior art, the advantages of this application are as follows:

[0027] 1. The underwater acoustic target intelligent detection method based on the full bias mutual learning strategy proposed by the present invention can directly process the received underwater acoustic signal data, with high real-time performance and fast response speed.

[0028] 2. The underwater acoustic target intelligent detection method based on the full bias mutual learning strategy proposed by the present invention can improve the classification ability of two or more models simultaneously.

[0029] 3. The underwater acoustic target intelligent detection method based on the full bias mutual learning strategy proposed by the present invention combines the methods of knowledge distillation and learning with added bias, enhances the classification ability of the weak items in the model mutual learning process, and improves the overall recognition accuracy of the model.

[0030] 4. The underwater acoustic target intelligent detection method based on the full bias mutual learning strategy proposed by the present invention is based on the actual underwater acoustic target sound signal data. The trained artificial intelligence model combines the advantages of different networks, has strong generalization ability and high recognition accuracy. Description of the Drawings

[0031] Figure 1 Shown is a schematic diagram of the principle of the underwater acoustic target intelligent detection method based on the full bias mutual learning strategy;

[0032] Figure 2 Shown is a schematic diagram of the full bias mutual learning strategy. Detailed Embodiments

[0033] The technical solutions of this application will be described in detail below with reference to the drawings.

[0034] Similar to deep mutual learning, the full-biased mutual learning strategy largely follows the method of traditional mutual learning, using the imitation loss between models for information interaction to improve the classification ability of different models. The difference is that the full-biased mutual learning strategy adds a full-biased network that is completely biased towards one category. The purpose is to improve the recognition ability of the student model for a certain category without reducing the accuracy of other categories. In this way, the problem that the network is insensitive to data of a certain category can be largely solved, and the effect of mutual learning between models can be greatly improved.

[0035] This invention mainly aims at the problem in the underwater acoustic classification task that different models are sensitive to different types of data to varying degrees. By introducing deep mutual learning, it aims to enable the system to continuously learn and optimize from different models, improving the accuracy and robustness of target recognition. After introducing mutual learning, two or more classification models as student models are iteratively trained by calculating the imitation loss between each other, and their respective capabilities are improved. However, the existing deep mutual learning technology automatically discovers the correlation between features and then improves the performance of the classifier, lacking human intervention. Therefore, some deviations may occur during the training process, resulting in low classification ability for a certain category. To enable it to better adapt to the complex and changing underwater environment and improve the recognition ability for a certain weak category, we propose a mutual learning strategy based on a full-biased network. By introducing the full-biased network, we can artificially guide the network to develop in the direction we expect. The imitation loss between the full-biased model and the student model leads the student model to develop in the direction of the full-biased model. At the same time, the overall recognition accuracy of the model is improved, and it has strong generalization ability. Therefore, this invention aims to give full play to the advantages of deep mutual learning technology and bring a more intelligent and efficient solution to the field of underwater acoustic target classification.

[0036] The intelligent underwater acoustic target detection system based on the full-biased mutual learning strategy includes an underwater acoustic feature extraction module and a full-biased mutual learning network. The underwater acoustic feature extraction module processes the original underwater acoustic data and outputs one-dimensional feature data and two-dimensional spectrogram data, which are sent to the full-biased mutual learning network for underwater acoustic target detection.

[0037] (1) Underwater acoustic feature extraction module

[0038] Underwater acoustic target classification based on deep learning mainly relies on feature extraction. Extracting appropriate features is a prerequisite for obtaining good classification results. There are three major types of features in underwater acoustic signal analysis, namely time-domain waveform analysis, time-frequency analysis, and auditory feature analysis. We will first select the low-level descriptors (LLDs) corresponding to these four categories.

[0039] Time-domain waveform features are a very important feature representation method. By analyzing the time variation of underwater acoustic signals, various useful time-domain features can be extracted, such as peak amplitude, waveform energy, and waveform envelope. These time-domain features can reflect the acoustic behavior characteristics of the target at different time intervals and help distinguish different types of underwater targets.

[0040] In underwater target classification tasks, time-frequency signal analysis is an important supplement to single-domain signal analysis. Compared with only focusing on the time characteristics of the signal, time-frequency analysis provides more detailed information about how the signal changes in the time domain and frequency domain. This analysis method can capture the dynamic characteristics of the frequency change with time in the underwater acoustic signal, thereby revealing the unique time-frequency characteristics of different targets. Through time-frequency analysis, we can obtain important features such as instantaneous frequency, time-frequency spectrogram, and instantaneous energy distribution. These features effectively reflect the time-frequency characteristics of the underwater acoustic signal and provide important clues for identifying and classifying different types of underwater targets.

[0041] Auditory features simulate the perception of the human auditory system to better understand and interpret the information in underwater acoustic signals. Human perception of sound is based on a set of complex auditory features, such as frequency, intensity, spectrum, etc. Therefore, in underwater target recognition, extracting and analyzing features similar to the human auditory system can provide a deeper understanding of underwater acoustic signals and provide key information for target recognition. By simulating the perception features of the human auditory system, we can obtain important acoustic features such as pitch, timbre, rhythm, etc., which are crucial for distinguishing different types of underwater targets. In addition, integrating these auditory features into target recognition can make the recognition system closer to the human cognitive process, thereby improving the accuracy and reliability of recognition.

[0042] A total of 7 LLDs were selected in this application as shown in Table 1, and 13 statistical functions were applied as shown in Table 2. Finally, we applied 13 statistical functions to the 7 low-order descriptors to obtain 7 * 13 = 91 features called HSF (High-Level Statistical Functions) as one-dimensional features.

[0043] Table 1 Low-Order Descriptors

[0044]

[0045]

[0046] Table 2 Statistical Functions

[0047]

[0048] This application also draws a two-dimensional spectrogram as an underwater acoustic feature. The process of drawing a two-dimensional spectrogram includes:

[0049] Preprocessing: First, preprocess the underwater acoustic signal, which may include denoising, normalization, etc., to ensure signal quality.

[0050] Fourier transform: Use the Short-Time Fourier Transform (STFT) to convert the time-domain signal into a frequency-domain signal to obtain the spectrum of the signal.

[0051] Spectrum analysis: Analyze the spectrum to extract useful features, such as frequency components, amplitudes, etc.

[0052] Convert to spectrogram: Convert the spectral data into a spectrogram. Common spectrograms include Lofar spectrogram, Audio spectrogram, Demon spectrogram, Histogram spectrogram, etc.

[0053] Feature extraction: Extract features from the spectrogram, such as MFCC (Mel Frequency Cepstral Coefficients), etc.

[0054] Plot the spectrogram: Plot the spectrogram to display the spectral characteristics of the signal.

[0055] (2) Full Bias Mutual Learning Network

[0056] Deep Mutual Learning (DML) is a knowledge extraction method used to achieve mutual improvement between two or more networks. By using the KL divergence as a regularization term, the student networks can learn from each other's predictions and their relationships with the ground truth labels, thereby improving the recognition accuracy of the model.

[0057] In the initial DML, the training of the two networks is carried out independently. Each network performs forward inference to obtain logical values. Subsequently, the total loss is updated by calculating the cross-entropy loss between the predicted values and the labels of the current batch of data and the KL divergence between the models. The updated loss prompts the two networks to converge. It is generally believed that if a network takes a particularly bad step in an iteration period. For example, the model may become more inclined to predict certain input data as belonging to a certain category. This may lead to a significant increase in the total loss of the mutual learning model. This, in turn, may lead to reverse optimization of the training results, thereby prolonging the training duration. Although it may cause instability during the training process, in most cases, the final effectiveness of the trained model will not be significantly affected. Due to the prominence of the prediction of a specific category, it can even enhance the discrimination ability of a specific category during the mutual learning process.

[0058] In traditional knowledge distillation methods, the pre-trained teacher network uses the softmax function during the inference process to soften the output target probabilities. Subsequently, the softened probability distribution is transferred as knowledge to the student network to facilitate its learning process. Similarly, in mutual learning networks, we constrain the output probabilities of one network and use the imitation loss between the two models as a regularization term to guide the network training in the direction we specify.

[0059] In this application, we propose a technical solution to add a fully biased network to the mutual learning model to improve its performance. As Figure 2 shown, similar to the original DML, the fully biased mutual learning network still uses the sum of the cross-entropy loss of the network itself and the KL divergence between the models as the total loss. However, the difference is that during the mutual learning process, one of the student networks sacrifices its performance, resulting in its predictions being completely biased towards a specific category. The fully biased network helps the network pay more attention to the specific category during the training process. This, in turn, enhances the network's discrimination ability for the specific category. The methods for the student network to be completely biased towards a specific category include canceling the optimizer of the model so that it no longer undergoes training, or canceling the model and changing it to an artificially set probability distribution to imitate the network output.

[0060] The following conducts relevant data analysis based on marine experimental data, which contains 3 types of target signals:

[0061] 1. Extract acoustic signal features, mainly including:

[0062] (1) Split the long sound signal data into one-second segments;

[0063] (2) Calculate the features of the signal to obtain a feature set;

[0064] 2. Label the data with corresponding categories.

[0065] 3. Combine the labeled data features into a training sample set and a test sample set. Take 2000s of data as the initial training sample set and 500s of data as the test set.

[0066] 4. Build an artificial intelligence model. Use a deep convolutional network and a DNN network built with multiple convolutional kernels as classifiers. The DNN parameters are set as follows: 91 input neurons, 1 hidden layer, 60 hidden neurons, the activation function is the sigmoid transfer function, 4 output neurons, the training function is the gradient descent algorithm training function, the loss function is the cross-entropy loss function, the training required accuracy is 10-9, the maximum number of training times is 200 times, and the learning rate is 0.001. This embodiment uses the deep convolutional network and the DNN network as examples for illustration. In fact, the system of this application is also applicable to other types of neural network models.

[0067] 5. Input the training sample set to perform mutual learning training on the deep convolutional network and the DNN network. When the training required accuracy is reached or the maximum number of training times is reached, stop the training.

[0068] 6. Set the DNN network among them as a fully biased network, bias towards the third category with relatively weak classification (which can be set manually according to the results of step 5), and perform mutual learning training again.

[0069] 7. Input the test sample set for testing. The accuracy of the deep convolutional network mutual learning model is 94.61%, and the accuracy of the deep convolutional network under full bias mutual learning is 95.97%, which is increased by 1.36% compared to the former.

[0070] This application also provides an underwater acoustic target intelligent detection method based on the full bias mutual learning strategy, which is implemented based on the above system. The method includes:

[0071] Extract the underwater acoustic features in the underwater acoustic signal by using the underwater acoustic feature extraction module, input them into the trained full bias mutual learning network, and output the classification of the underwater acoustic signal. Among them, when inputting the underwater acoustic features into the full bias mutual learning network, input the two-dimensional spectrogram into the full bias network, and input the one-dimensional features into the other network.

[0072] This application can also provide a computer device, including: at least one processor, a memory, at least one network interface, and a user interface. Each component in the device is coupled together through a bus system. It can be understood that the bus system is used to realize the connection and communication between these components. In addition to the data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0073] Among them, the user interface can include a display, a keyboard, or a pointing device. For example, a mouse, a trackball, a touchpad, or a touch screen, etc.

[0074] It can be understood that the memory in the disclosed embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable ROM (PROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DRRAM). The memories described herein are intended to include but not be limited to these and any other suitable types of memories.

[0075] In some embodiments, the memory stores the following elements, executable modules or data structures, or subsets or supersets thereof: an operating system and application programs.

[0076] Among them, the operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., and is used to implement various basic services and process hardware-based tasks. The application programs include various application programs, such as a media player and a browser, etc., and are used to implement various application services. The program for implementing the method of the disclosed embodiments of the present application can be included in the application programs.

[0077] In the above embodiments, by calling the programs or instructions stored in the memory, specifically, the programs or instructions stored in the application programs, the processor is configured to:

[0078] Execute the steps of the above method.

[0079] The above method can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, the steps of the above method can be completed by the integrated logic circuit of the hardware in the processor or instructions in the form of software. The above-mentioned processor may be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed above. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. Combining the steps of the above-disclosed method can be directly embodied as being completed by the execution of a hardware decoding processor, or completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0080] It can be understood that these embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or a combination thereof.

[0081] For software implementation, the technology of this application can be implemented by executing the functional modules of this application (such as procedures, functions, etc.). The software code can be stored in the memory and executed by the processor. The memory can be implemented inside or outside the processor.

[0082] The present application can also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, each step in the above method embodiments can be implemented.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them. Although the present application has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present application does not depart from the spirit and scope of the technical solutions of the present application, and they should all be covered within the scope of the claims of the present application.

Claims

1. An intelligent underwater acoustic target detection system based on a full-offset mutual learning strategy, characterized in that The system includes: An underwater acoustic feature extraction module for extracting underwater acoustic features from underwater acoustic signals; the underwater acoustic features include one-dimensional features and two-dimensional spectrograms; the one-dimensional features include 91 high-order statistical functions obtained by combining 7 low-order descriptors and 13 statistical functions; the low-order descriptors include: zero-crossing rate, spectral skewness, spectral kurtosis, spectral clarity, spectral center point, the first coefficient of the Mel frequency cepstral coefficient, and the fifth coefficient of the Mel frequency cepstral coefficient; the statistical functions include: arithmetic mean, maximum value, minimum value, maximum-minimum deviation, standard deviation, skewness, kurtosis, first quartile, second quartile, third quartile, 1-2 interquartile range, 2-3 interquartile range, and 1-3 interquartile range; and A fully-biased mutual learning network, including two networks trained by mutual learning, where one network is a fully-biased network; the fully-biased mutual learning network is used to input underwater acoustic features and output the classification of underwater acoustic signals. The training process of the fully-biased mutual learning network includes: Inputting a training sample set to perform mutual learning training on the two networks, and stopping training when the training required accuracy is reached or the maximum number of training times is reached. Setting one of the networks as a fully-biased network and performing mutual learning training again. The setting one of the networks as a fully-biased network includes: canceling the optimizer of this network so that it no longer performs training, or changing this network to imitate the output of a network with a manually set probability distribution.

2. The underwater acoustic target intelligent detection system based on the full-offset mutual learning strategy according to claim 1, wherein When the underwater acoustic features are input into the fully-biased mutual learning network, the two-dimensional spectrogram is input into the fully-biased network, and the one-dimensional features are input into the other network.

3. The underwater acoustic target intelligent detection system based on the full-offset mutual learning strategy according to claim 1, characterized in that, The process of the underwater acoustic feature extraction module extracting underwater acoustic features from underwater acoustic signals includes: Splitting the long sound signal data into segments of 1 second each. Calculating 7 low-order descriptors and 13 statistical functions of the split signal data, and combining them with each other to obtain one-dimensional features; drawing a spectrogram to obtain a two-dimensional spectrogram.

4. The underwater acoustic target intelligent detection system based on the full-offset mutual learning strategy according to claim 1, wherein, The fully-biased mutual learning network includes 1 deep convolutional network and 1 DNN network; the DNN network is a fully-biased network.

5. An intelligent underwater acoustic target detection method based on a fully-biased mutual learning strategy, implemented based on the system according to any one of claims 1-4, the method includes: Using the underwater acoustic feature extraction module to extract underwater acoustic features from underwater acoustic signals, inputting them into the trained fully-biased mutual learning network, and outputting the classification of underwater acoustic signals.

Images