Semantic Segmentation Detection Method, Device and Medium for Marine Mammal Click Signals

Abstract

The method, device and medium for semantic segmentation detection of marine mammal click signals disclosed in this application relate to the field of speech recognition. The method includes: obtaining the original sound signal of a marine mammal; processing the original sound signal to obtain a time-frequency diagram; introducing Bayesian convolution and wavelet transform convolution into an encoder-decoder architecture to construct an acoustic image semantic segmentation network model; inputting the time-frequency diagram into the acoustic image semantic segmentation network model to obtain the semantic segmentation detection result of the marine mammal click signal. This application can detect the local and fine voiceprints contained in the marine mammal sound signal, avoid problems such as missed detection, and thus improve the accuracy of the semantic segmentation detection of the marine mammal click signal.

Description

Technical Field

[0001] The present application relates to the field of speech recognition, and particularly to a method, device and medium for semantic segmentation detection of click signals of marine mammals. Background Art

[0002] Marine mammals are a key part of the marine ecosystem and play an irreplaceable role in maintaining ecological balance and protecting biodiversity. The vocal signals of marine mammals are mainly divided into three categories: Click, Burst-pulse, and Whistle. Due to the overlap of environmental noise and the acoustic signals of marine mammals in frequency and time, problems such as misjudgment or missed detection are likely to occur during the signal extraction and recognition of the acoustic signals of marine mammals. Therefore, the detection of the acoustic signals of marine mammals faces extremely high requirements.

[0003] To address the challenges brought by marine noise interference to acoustic signal detection, the inventors have proposed various feature extraction and classification methods for the acoustic signals of marine mammals in the known technology. Traditional methods can effectively detect signals under high signal-to-noise ratios, but are prone to false detection when there is interference. Although methods using machine learning and the like are currently emerging for the detection of the acoustic signals of marine mammals, these methods have problems such as missed detection for local and fine acoustic patterns. Summary of the Invention

[0004] To solve the above problems existing in the prior art, the present application provides a method, device and medium for semantic segmentation detection of click signals of marine mammals.

[0005] To achieve the above object, the present application provides the following solutions:

[0006] In a first aspect, the present application provides a method for semantic segmentation detection of click signals of marine mammals, including:

[0007] Obtain the original acoustic signal of a marine mammal;

[0008] Process the original acoustic signal to obtain a time-frequency diagram;

[0009] Introduce Bayesian convolution and wavelet transform convolution into an encoder-decoder architecture to construct an acoustic image semantic segmentation network model;

[0010] Input the time-frequency diagram into the acoustic image semantic segmentation network model to obtain the semantic segmentation detection result of the click signal of the marine mammal.

[0011] Optionally, obtaining the original acoustic signal of a marine mammal includes:

[0012] The original acoustic signals of marine mammals are collected by a hydrophone at a set sampling rate.

[0013] Optionally, the original acoustic signals are processed to obtain a time-frequency diagram, including:

[0014] The original acoustic signals are filtered to obtain filtered signals;

[0015] Audio segments are obtained based on the filtered signals;

[0016] The audio segments are subjected to time-frequency conversion to obtain the time-frequency diagram.

[0017] Optionally, the original acoustic signals are filtered to obtain filtered signals, including:

[0018] The cut-off frequency and sampling rate of the original acoustic signals are obtained;

[0019] The normalized cut-off frequency is determined according to the cut-off frequency and the sampling rate;

[0020] Based on the normalized cut-off frequency, the ideal impulse response is obtained by using the inverse discrete-time Fourier transform;

[0021] The Hamming window is used to smoothly truncate the ideal impulse response to obtain the filter coefficients;

[0022] By using convolution operation, the filter coefficients are applied to the original acoustic signals to obtain the filtered signals.

[0023] Optionally, the audio segments are subjected to time-frequency conversion to obtain the time-frequency diagram, including:

[0024] Each audio segment is windowed by using the Hamming window;

[0025] The fast Fourier transform is performed on the filtered signals in each windowed audio segment to obtain the frequency-domain signals, so as to generate the time-frequency diagram.

[0026] Optionally, Bayesian convolution and wavelet transform convolution are introduced in the encoder-decoder architecture to construct an acoustic image semantic segmentation network model, including:

[0027] In the encoder part of the encoder-decoder architecture, a Bayesian convolution pyramid and a multi-scale mapping module are constructed to obtain an initial encoder;

[0028] In the decoder part of the encoder-decoder architecture, wavelet decomposition convolution is constructed, and multi-scale feature alignment and fusion are realized by combining the inverse wavelet transform and the alignment filtering module to obtain an initial decoder;

[0029] Obtain the MobileNetV2 model, and combine the initial encoder and the initial decoder to form an initial semantic segmentation network model; the MobileNetV2 model is used to extract local features and global features of the input image;

[0030] Obtain a sample data set, and divide the sample data set into a training set and a test set according to a set ratio;

[0031] Use the training set to train the initial semantic segmentation network model, and use the test set to test the trained initial semantic segmentation network model during the training process until the trained initial semantic segmentation network model meets the set requirements, and obtain the trained initial semantic segmentation network model;

[0032] Use the trained initial semantic segmentation network model as the acoustic image semantic segmentation network model.

[0033] Optionally, the Bayesian convolutional pyramid is composed of multi-scale Bayesian convolutional layers with different convolutional kernel sizes.

[0034] Optionally, the process of inputting the time-frequency map into the acoustic image semantic segmentation network model to obtain the semantic segmentation detection result of the marine mammal click signal includes:

[0035] Input the time-frequency map into the MobileNetV2 model in the acoustic image semantic segmentation network model to obtain global features and local features;

[0036] Input the local features into the encoder in the acoustic image semantic segmentation network model. The Bayesian convolutional pyramid in the encoder performs multi-scale feature extraction on the local features to obtain first multi-scale features; after the multi-scale mapping module in the encoder performs feature mapping on the first multi-scale features, dimensionality reduction processing is performed using an image pooling layer to obtain encoded features;

[0037] Input the global features into the decoder in the acoustic image semantic segmentation network model. The wavelet decomposition convolution in the decoder decomposes the global features to obtain second multi-scale features; the second multi-scale features are processed using Bayesian convolution, and inverse wavelet transform is performed on the processed second multi-scale features to obtain reconstructed features; in the decoder, the reconstructed features and the global features processed by Bayesian convolution are fused to obtain first fusion features;

[0038] Input the first fusion features and the encoded features into the alignment filtering module in the decoder. After multi-scale fusion and alignment filtering processing, second fusion features are obtained; after the second fusion features are sequentially processed by Bayesian convolution and segmentation processing, output features are obtained;

[0039] Obtain the semantic segmentation detection result of the marine mammal click signal based on the output feature.

[0040] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the steps of the above-provided method for detecting semantic segmentation of marine mammal click signals.

[0041] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-provided method for detecting semantic segmentation of marine mammal click signals are implemented.

[0042] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0043] The method, device, and medium for detecting semantic segmentation of marine mammal click signals provided by the present application can effectively enhance the uncertainty modeling ability of the constructed model by introducing Bayesian convolution in the encoder-decoder architecture, and can fully exploit the context information in the marine mammal sound signal. By introducing wavelet transform convolution in the encoder-decoder architecture, it is possible to accurately depict the local details of the marine mammal sound signal, realize the detection of local fine sound patterns in the marine mammal sound signal, and avoid problems such as missed detection. Based on this, by using an acoustic image semantic segmentation network model to detect the semantic segmentation of marine mammal click signals, the accuracy of the detection of marine mammal click signal segmentation can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0045] Figure 1 It is a schematic flowchart of a method for detecting semantic segmentation of marine mammal click signals provided by an embodiment of the present application;

[0046] Figure 2 It is a schematic diagram of the implementation processing flow of a method for detecting semantic segmentation of marine mammal click signals provided by an embodiment of the present application;

[0047] Figure 3 It is a schematic diagram of an example of the result of manual marking provided by an embodiment of the present application;

[0048] Figure 4 Schematic diagram of the structure of the acoustic image semantic segmentation network model provided by an embodiment of the present application;

[0049] Figure 5 Schematic diagram of the wavelet decomposition convolution process provided by an embodiment of the present application;

[0050] Figure 6 Schematic diagram of the structure of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0051] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0052] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0053] In an exemplary embodiment, the present application provides a method for semantic segmentation detection of the click sound signal of marine mammals. This method is executed by a computer device, which can be specifically executed by a computer device such as a terminal or a server alone, or jointly executed by a terminal and a server. In the embodiments of the present application, this method is described by taking it as an example applied to a server. As Figure 1 shown, this method includes:

[0054] Step 100: Obtain the original sound signal of the marine mammal.

[0055] Step 101: Process the original sound signal to obtain a time-frequency diagram.

[0056] Step 102: Introduce Bayesian convolution and wavelet transform convolution into the encoder-decoder architecture to construct an acoustic image semantic segmentation network model.

[0057] Step 103: Input the time-frequency diagram into the acoustic image semantic segmentation network model to obtain the semantic segmentation detection result of the click sound signal of the marine mammal.

[0058] By implementing the above steps 100 - step 103, the present application can detect the local and fine sound patterns contained in the sound signal of marine mammals, avoid problems such as missed detection, and thus can significantly improve the accuracy of the segmentation detection of the click sound signal of marine mammals.

[0059] In another exemplary embodiment of the present application, in order to comprehensively collect the original acoustic signals of marine mammals, in this embodiment, the above step 100 may be: using a hydrophone to collect the original acoustic signals of marine mammals at a set sampling rate.

[0060] Among them, the hydrophone used to collect the acoustic signals of marine mammals is generally deployed in seawater in the form of a mooring buoy and operates in a self-contained manner. Based on this, in the actual application process, the response frequency of the hydrophone used can be 20 Hz to 150 kHz, the sensitivity is -190 dB, and the sampling rate is 400 kHz. The hydrophone works continuously to achieve continuous collection of acoustic signals, so as to obtain the original acoustic signals of marine mammals.

[0061] In another exemplary embodiment of the present application, in order to provide a data basis for processing for the acoustic image semantic segmentation network model, in this embodiment, the implementation process of the above step 101 can be replaced by the following steps 200 - 202.

[0062] Step 200: Perform filtering processing on the original acoustic signal to obtain a filtered signal.

[0063] In the actual application process, taking the implementation of high-pass filtering processing as an example, the implementation manner of the above step 200 may include:

[0064] (1) Obtain the cut-off frequency and sampling rate of the original acoustic signal, and calculate the normalized cut-off frequency according to the cut-off frequency and sampling rate. Among them, the normalized cut-off frequency is expressed as:

[0065] .

[0066] In the formula, is the normalized cut-off frequency, is the cut-off frequency, is the sampling rate.

[0067] (2) Based on the normalized cut-off frequency, use the inverse discrete-time Fourier transform to obtain the ideal impulse response (which can also be called the infinite impulse response) of the ideal high-pass filter. Among them, the ideal impulse response is expressed as:

[0068] .

[0069] In the formula, is the ideal impulse response, is the discrete-time index, is the order of the filter, is the discrete-time unit impulse function.

[0070] (3) Use a Hamming window to smoothly truncate the ideal impulse response to obtain the filter coefficients (which can also be called filter coefficients). Among them, the filter coefficients are expressed as:

[0071] .

[0072] In the formula, is the filter coefficient, is the Hamming window function.

[0073] (4) Use convolution operation to apply the filter coefficients to the original acoustic signal to obtain the filtered signal. Among them, the filtered signal is expressed as:

[0074] .

[0075] In the formula, is the filtered signal, is the value of the impulse response at position k, is the value of the discrete signal at the summation variable position, is the length of the impulse response.

[0076] Step 201: Obtain audio segments based on the filtered signal.

[0077] Step 202: Perform time-frequency conversion on the audio segments to obtain a time-frequency diagram. Among them, each audio segment is windowed using a Hamming window. The windowed audio segment is subjected to a fast Fourier transform to convert it from the time domain to the frequency domain, thereby converting the acoustic signal into a time-frequency diagram.

[0078] In another exemplary embodiment of the present application, a high-quality dataset can be constructed based on a multi-parameter constrained tick detection method to support the excellent performance of the acoustic image semantic segmentation network (BayesWaveNet) model in a complex noise environment.

[0079] In the encoder-decoder architecture, the encoder enhances the uncertainty modeling ability through smooth Bayesian convolution, and the decoder accurately extracts local details through wavelet decomposition convolution, and combines an alignment filtering module to achieve multi-scale feature fusion. Based on this, in this embodiment, in combination with the process shown in Figure 2 , the specific construction process of the acoustic image semantic segmentation network model in step 102 is described, including:

[0080] Step 1: In the encoder part of the encoder-decoder architecture, construct a Bayesian convolution pyramid and a multi-scale mapping module to obtain an initial encoder. Among them, the Bayesian convolution pyramid consists of multi-scale Bayesian convolution layers with different convolution kernel sizes, and is used to capture multi-scale features of the image. Bayesian convolution uses a probability distribution to describe the weight parameters.

[0081] For example, as Figure 4 shown, the Bayesian convolutional pyramid contains Bayesian convolutional layers (BayesConv) with four convolutional kernel sizes of 1×1, 3×3, 5×5, and 7×7 to capture details and patterns of images at different scales and achieve comprehensive extraction of image features.

[0082] Step 2: In the decoder part of the encoder-decoder architecture, construct wavelet decomposition convolution, and combine inverse wavelet transform and alignment filtering module to achieve multi-scale feature alignment and fusion to obtain an initial decoder.

[0083] Based on Step 1 and Step 2, in the encoder part, a feature extraction network based on the Bayesian convolutional pyramid and multi-scale mapping module is constructed. Among them, a smooth Bayesian convolution is proposed in the encoder part and combined with the multi-scale mapping module. A wavelet decomposition convolution (WDConv) is proposed in the decoder part, combined with inverse wavelet transform (IWT) and alignment filtering module (MAF) to achieve multi-scale feature alignment and fusion.

[0084] Step 3: Obtain the MobileNetV2 model, and combine the initial encoder and initial decoder to form an initial semantic segmentation network model. The MobileNetV2 model is used to extract local and global features of the input image.

[0085] Step 4: Obtain a sample data set, and divide the sample data set into a training set and a test set according to a set ratio. In this step, the acoustic signals of marine mammals collected can be subjected to high-pass filtering and time-frequency conversion to obtain time-frequency sample maps, and then through multi-parameter screening and data annotation methods, a high-quality data set WhaleDataSet (i.e., the sample data set) is constructed.

[0086] Among them, the methods of multi-parameter screening and data annotation can be: judging whether there is a click signal of marine mammals in the time-frequency sample map according to the anti-interference real-time click sound detection method based on multi-parameter constraints. Based on this, all time-frequency sample maps in the data set WhaleDataSet are manually annotated, and the annotation results are as Figure 3 shown. Figure 3 In, part (a) is a schematic diagram of the manual annotation result of Neophocaena asiaeorientalis asiaeorientalis, part (b) is a schematic diagram of the manual annotation result of Sousa chinensis, and part (c) is a schematic diagram of the manual annotation result of noise. Figure 3 In the abscissa is the sampling time; the ordinate is the click signal of marine mammals, and the unit is decibel (dB).

[0087] Step 5: Train the initial semantic segmentation network model using the training set, and during the training process, use the test set to test the trained initial semantic segmentation network model until the trained initial semantic segmentation network model meets the set requirements (for example, reaches the maximum number of iterations or the model evaluation metric reaches the set value), then obtain the trained initial semantic segmentation network model.

[0088] For example, use the time-frequency sample map as the feature input of the initial semantic segmentation network model to train the initial semantic segmentation network model. The experimental environment for training and testing is implemented based on Python 3.6.5, and the hardware configuration uses an NVIDIA GTX 3090 graphics card, and the CUDA 10.0 toolkit is used to accelerate the calculation. The dataset WhaleDataset is randomly divided into a training set and a test set in a ratio of 7:3. The model training is carried out for 200 epochs (i.e., the set requirements), and the initial learning rate is set to 1×10 －4 . The learning rate adopts a fixed decay strategy, remaining unchanged for the first 100 epochs, and then decaying by 0.01 every 10 epochs. The optimizer selects the stochastic gradient descent method (SGD). The finally obtained BayesWaveNet model achieves an average intersection over union of 91.16% and an average precision of 95.27% in a complex noise environment, significantly superior to existing methods.

[0089] Step 6: Use the trained initial semantic segmentation network model as the acoustic image semantic segmentation network model. Among them, the obtained acoustic image semantic segmentation network model is as Figure 4 shown. Figure 4 In, the input global feature undergoes wavelet transform, that is, the global feature is decomposed into a low-frequency subband (L) and high-frequency subbands (H, V, D) through a set of specific filters. Among them, the low-frequency subband (L) is obtained through a low-pass filter, while the high-frequency subbands (H, V, D) correspond to the high-frequency information in the horizontal, vertical, and diagonal directions respectively, and are obtained through the high-pass filters f H 、f V and f D obtained.

[0090] In another exemplary embodiment of the present application, in order to accurately identify the click signals of marine mammals using the finally obtained acoustic image semantic segmentation network model, the original acoustic signal outside the dataset can be processed according to the above steps 200 - 202 to obtain a time-frequency map, and then input it into the acoustic image semantic segmentation network model. The acoustic image semantic segmentation network model gives the corresponding recognition result to achieve the prediction of the original acoustic signal (i.e., new data). Based on this, the implementation process of the above step 103 can include:

[0091] Step 1): Input the time-frequency map into the MobileNetV2 model in the acoustic image semantic segmentation network model to obtain global features and local features. This step mainly uses the MobileNetV2 model to extract global and local features from the input time-frequency map.

[0092] Step 2): Input the local features into the encoder in the acoustic image semantic segmentation network model. The Bayesian convolutional pyramid in the encoder performs multi-scale feature extraction on the local features to obtain the first multi-scale features. After the multi-scale mapping module in the encoder performs feature mapping on the first multi-scale features, a dimensionality reduction process is carried out using an image pooling layer to obtain encoded features for subsequent global feature extraction.

[0093] Step 3): Input the global features into the decoder in the acoustic image semantic segmentation network model. The wavelet decomposition convolution in the decoder decomposes the global features to obtain the second multi-scale features. The second multi-scale features are processed using Bayesian convolution, and an inverse wavelet transform is performed on the processed second multi-scale features to obtain reconstructed features. The reconstructed features and the global features processed by Bayesian convolution are fused in the decoder to obtain the first fused features.

[0094] Step 4): Input the first fused features and the encoded features into the alignment filtering module in the decoder. After multi-scale fusion and alignment filtering processing, the second fused features are obtained. After the second fused features are sequentially processed by Bayesian convolution and segmentation processing, the output features are obtained.

[0095] Step 5): Obtain the semantic segmentation detection result of the marine mammal click signal based on the output features.

[0096] Based on the above description, in steps 3)-5), the data processing process of the acoustic image semantic segmentation network model can be described as follows: First, the MobileNetV2 model is used to extract global and local features from the input time-frequency map. Then, the global features are decomposed into multi-scale features through wavelet transform, and then Bayesian convolution is applied to process the features, and the signal is reconstructed through inverse wavelet transform (for example, the reconstructed signal is as shown in Figure 5 ). The wavelet decomposition convolution performs wavelet decomposition on the global features of the MobileNetV2 model, decomposes these feature maps into components of different frequencies, and performs Bayesian convolution operations on the sub-bands, obtaining a large receptive field and multi-frequency responses, thereby further extracting the local features and texture information of the time-frequency map. Subsequently, the global features and the processed local features are jointly input into the alignment filter (MAF) in the alignment filtering module. The alignment filter precisely aligns and deeply fuses features of different scales to ensure the information consistency between multi-scale features. Finally, refined detection is performed on the fused feature map to generate high-precision detection and classification results of marine mammal sound signals.

[0097] In summary, the method provided by this application constructs an acoustic image semantic segmentation network (BayesWaveNet) model based on Bayesian convolution and wavelet transform for the detection and classification of marine mammal acoustic signals, which can provide an efficient and reliable technical means for marine exploration and protection.

[0098] The BayesWaveNet model adopts an encoder-decoder architecture. In the encoder part, a smooth Bayesian convolution is proposed to effectively enhance the uncertainty modeling ability of the BayesWaveNet model, and it is combined with a multi-scale mapping module to fully exploit the context information of global features. In the decoder part, this application uses wavelet decomposition convolution to accurately depict and extract local details, and realizes the precise alignment and fusion of multi-scale features by introducing an alignment filtering module, supporting the BayesWaveNet model to achieve high-precision detection and classification of marine mammal acoustic signals in complex noise environments.

[0099] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 6 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store marine mammal click signal semantic segmentation detection data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for detecting and segmenting the semantics of marine mammal click signals.

[0100] Those skilled in the art can understand that Figure 6 the structure shown in

[0101] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0102] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0103] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0104] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0105] Those of ordinary skill in the art can understand that all or part of the processes of implementing the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, database, or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0106] In each of the embodiments provided in the present application, the database involved may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., and is not limited thereto. In each of the embodiments provided in the present application, the processor involved may be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., and is not limited thereto.

[0107] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0108] Specific examples are used herein to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for semantic segmentation detection of click signals of marine mammals, characterized in that, Including: Obtain the original acoustic signal of marine mammals; Process the original acoustic signal to obtain a time-frequency map; In the encoder part of the encoder-decoder architecture, construct a Bayesian convolutional pyramid and a multi-scale mapping module to obtain an initial encoder; the Bayesian convolutional pyramid consists of multi-scale Bayesian convolutional layers with different convolutional kernel sizes; In the decoder part of the encoder-decoder architecture, construct a wavelet decomposition convolution, and combine the inverse wavelet transform and the alignment filtering module to achieve multi-scale feature alignment and fusion to obtain an initial decoder; Obtain the MobileNetV2 model, and combine the initial encoder and the initial decoder to form an initial semantic segmentation network model; the MobileNetV2 model is used to extract local features and global features of the input image; Obtain a sample dataset, and divide the sample dataset into a training set and a test set according to a set ratio; Use the training set to train the initial semantic segmentation network model, and use the test set to test the trained initial semantic segmentation network model during the training process until the trained initial semantic segmentation network model meets the set requirements, and obtain the trained initial semantic segmentation network model; Take the trained initial semantic segmentation network model as the acoustic image semantic segmentation network model; Input the time-frequency map into the acoustic image semantic segmentation network model to obtain the semantic segmentation detection result of the click signal of marine mammals.

2. The method for semantic segmentation detection of marine mammal click signals according to claim 1, wherein, Obtain the original acoustic signal of marine mammals, including: Collect the original acoustic signal of marine mammals using a hydrophone at a set sampling rate.

3. The method for semantic segmentation detection of the click signal of marine mammals according to claim 1, wherein Process the original acoustic signal to obtain a time-frequency map, including: Perform filtering processing on the original acoustic signal to obtain a filtered signal; Obtain audio segments based on the filtered signal; Perform time-frequency conversion on the audio segments to obtain the time-frequency map.

4. The method for semantic segmentation detection of the click signal of marine mammals according to claim 3, wherein, Perform filtering processing on the original acoustic signal to obtain a filtered signal, including: Obtain the cut-off frequency and sampling rate of the original acoustic signal; Determine the normalized cut-off frequency according to the cut-off frequency and the sampling rate; Based on the normalized cut-off frequency, use the inverse discrete-time Fourier transform to obtain the ideal impulse response; Use a Hamming window to smoothly truncate the ideal impulse response to obtain the filter coefficient; Perform convolution operation, apply the filter coefficient to the original acoustic signal to obtain the filtered signal.

5. The method for semantic segmentation detection of the click signal of marine mammals according to claim 3, wherein Perform time-frequency conversion on the audio segments to obtain the time-frequency map, including: Use a Hamming window to window each audio segment; Perform a fast Fourier transform on the filtered signal in each windowed audio segment to obtain a frequency-domain signal to generate the time-frequency map.

6. The method for semantic segmentation detection of the click signal of marine mammals according to claim 1, wherein The process of inputting the time-frequency map into the acoustic image semantic segmentation network model to obtain the semantic segmentation detection result of the click signal of marine mammals, including: Input the time-frequency map into the MobileNetV2 model in the acoustic image semantic segmentation network model to obtain global features and local features; Input the local features into the encoder in the acoustic image semantic segmentation network model. The Bayesian convolutional pyramid in the encoder performs multi-scale feature extraction on the local features to obtain first multi-scale features; after the multi-scale mapping module in the encoder performs feature mapping on the first multi-scale features, a dimensionality reduction process is performed using an image pooling layer to obtain encoded features; Input the global features into the decoder in the acoustic image semantic segmentation network model. The wavelet decomposition convolution in the decoder decomposes the global features to obtain second multi-scale features; the second multi-scale features are processed using Bayesian convolution, and an inverse wavelet transform is performed on the processed second multi-scale features to obtain reconstructed features; in the decoder, the reconstructed features and the global features processed by Bayesian convolution are fused to obtain first fusion features; Input the first fusion features and the encoded features into the alignment filtering module in the decoder. After performing multi-scale fusion and alignment filtering processing, second fusion features are obtained; after sequentially performing Bayesian convolution processing and segmentation processing on the second fusion features, output features are obtained; Based on the output features, obtain the semantic segmentation detection result of the marine mammal click signal.

7. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the method for detecting semantic segmentation of marine mammal click signals according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for detecting semantic segmentation of marine mammal click signals according to any one of claims 1-6.

Images