Passive acoustic monitoring method and system for white-head leafy monkeys based on residual neural network

By introducing extrusion-excitation module and transfer learning into the ResNet residual neural network, the problem of audio data recognition in the white-headed langur monkey in the prior art is solved, and high accuracy and efficient audio recognition effects are achieved.

CN118942465APending Publication Date: 2024-11-12GUANGXI NORMAL UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411008443.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing passive acoustic monitoring methods are difficult to accurately separate audio data from white-headed langurs. Due to the artificial discrimination ability and data processing ability, they cannot process large-scale data, and are weak in complex feature recognition.

Method used

Using the ResNet residual neural network method, by replacing the Bottleneck layer as the extrusion-excitation module, a recognition model is built, transfer learning and data enhancement are carried out, and data preprocessing and feature extraction are used to improve the representation of the model. Capability and efficiency.

Benefits of technology

It effectively improves the recognition accuracy of audio data of white-headed langur monkeys, solves the problem that it is difficult to deal with large-scale data by manual identification, and enhances the recognition capabilities of the model and the efficiency of deployment on resource-constrained devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942465A_ABST
    Figure CN118942465A_ABST
Patent Text Reader

Abstract

The invention discloses a passive acoustic monitoring method and system for white-head leafy monkeys based on a residual neural network. The method comprises the following steps: acquiring audio data and constructing a data set; taking a ResNet residual neural network as a basic framework, and replacing a residual layer with a Bottleneck layer to obtain an identification model; the Bottleneck layer is provided with an extrusion-excitation module; training the recognition model based on the data set to obtain a trained recognition model; and inputting real-time data into the trained identification model to obtain identification information. The system comprises a data set construction module, a model construction module, a model training module and an identification module. By using the method and the device, the audio of the white-head leaflet can be quickly and accurately identified from a large amount of audio data, and acoustic monitoring is realized. The method can be widely applied to the field of audio processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of audio recognition, and in particular to a method and system for passive acoustic monitoring of white-headed langurs based on a residual neural network. Background Art

[0002] White-headed langurs mainly live in limestone areas with steep cliffs and cliffs, making observation and monitoring difficult. At the same time, white-headed langurs are very sensitive to changes and potential threats in the outside world, and tend to remain highly alert and flee quickly. This makes it very difficult to get close to them and observe them in detail. They often jump and move between trees and have a large range of activities. This mobility makes it difficult to track their routes and habits. White-headed langurs are social animals and often move in groups. This makes it more difficult to observe and identify individuals, especially in dense tree canopies. And white-headed langurs are usually active at dawn and dusk, and rarely active during the day and night. This activity time limits the opportunity to observe them.

[0003] Passive Acoustic Monitoring (PAM) is a monitoring method that obtains information by listening to the sounds in the environment. Passive acoustic monitoring does not require active sound. It usually uses highly sensitive acoustic sensors to collect sound data in the environment, thereby relying on the existing sounds in the environment for monitoring and analysis. Currently, most of the sounds made by animals are distinguished from environmental sounds by manual identification. However, the calls of white-headed langurs are concentrated in the range of 0-3Khz, which is in the same frequency band as the surrounding environmental noise. It is difficult for manual identification to accurately identify these mixed features, which can easily lead to misjudgment. At the same time, manual processing capabilities are limited and cannot process large-scale data. Limited by professional knowledge and experience levels, the ability to identify complex features is weak, making it difficult to cope with certain special situations. Summary of the invention

[0004] In view of this, in order to solve the technical problem that the existing passive acoustic monitoring method cannot accurately separate the audio data of white-headed langurs, in the first aspect, the present invention proposes a passive acoustic monitoring method for white-headed langurs based on a residual neural network, the method comprising the following steps:

[0005] Get audio data and build a dataset;

[0006] Taking the ResNet residual neural network as the basic framework, the Bottleneck layer replaces the residual layer to obtain the recognition model;

[0007] The Bottleneck layer is provided with a squeezing-excitation module;

[0008] Training the recognition model based on the data set to obtain a trained recognition model;

[0009] The real-time data is input into the trained recognition model to obtain recognition information.

[0010] In some embodiments, the step of obtaining audio data and constructing a data set specifically includes:

[0011] Get audio data;

[0012] Performing format unification, noise reduction and labeling processing on the audio data to obtain pre-processed data;

[0013] The preprocessed data are divided into a training set and a test set.

[0014] Through this preferred step, the data is preprocessed and divided, making it more unified and standardized for model training.

[0015] In some embodiments, it also includes:

[0016] Performing data augmentation on the training set;

[0017] The data enhancement methods include volume enhancement, noise enhancement, speech rate disturbance enhancement and spectrum enhancement.

[0018] Through this preferred step, the data is enhanced in a variety of ways to expand the data set.

[0019] In some embodiments, the step of training the recognition model based on the data set to obtain a trained recognition model specifically includes:

[0020] Based on the transfer learning method, determining the initial weight of the recognition model;

[0021] Training the recognition model based on the training set to obtain a preliminary recognition model;

[0022] The test set is used to perform the preliminary test, disabling the BN layer and the Dropout layer, and selecting the model with the highest accuracy as the trained recognition model.

[0023] In some embodiments, the step of performing data enhancement on the training set specifically includes:

[0024] Setting a gain factor, and performing volume enhancement on the training set based on the gain factor;

[0025] Setting environmental noise, and performing noise enhancement on the training set based on the environmental noise;

[0026] Setting a speed factor, and modifying the audio frame based on the speed disturbance factor, to perform speech rate disturbance enhancement on the training set;

[0027] The spectrum area and time interval within the preset range are masked, and the locally missing information is introduced to perform spectrum enhancement on the training set.

[0028] In some embodiments, before the step of inputting the real-time data into the trained recognition model, the step further includes:

[0029] Perform short-time Fourier transform on real-time data to convert the original time-domain audio signal into the time-frequency domain to obtain a time-frequency domain matrix;

[0030] defining a triangular filter whose center frequencies are spaced using a Mel frequency scale;

[0031] Multiplying the time-frequency domain matrix by each of the filters element by element and summing the results to obtain a filter bank characteristic matrix;

[0032] The filter bank feature matrix is ​​logarithmically transformed to obtain final features.

[0033] In some embodiments, the step of inputting the real-time data into the trained recognition model to obtain recognition information specifically includes:

[0034] Inputting the final features corresponding to the real-time data into the trained recognition model;

[0035] Perform dimension conversion and convolution operations based on the convolution block to obtain the convolution block output;

[0036] Performing convolution, batch normalization, compression-excitation, and residual connection on the convolution block output based on the residual block to obtain a residual block output;

[0037] Based on the fully connected layer, the residual block output is nonlinearized and fully connected, and the multi-classification output is converted into probability through SoftMax to obtain recognition information.

[0038] In some embodiments, it also includes:

[0039] Generate training data using the trained recognition model and add it to the training set;

[0040] Training the trained recognition model based on the training set;

[0041] Iterate the steps of replenishing the training set and training the model.

[0042] In some embodiments, it also includes:

[0043] The trained recognition model is deployed on the edge computing device.

[0044] Through this step, real-time analysis of audio data is achieved.

[0045] The present invention also proposes a passive acoustic monitoring system for white-headed langurs based on a residual neural network, the system comprising:

[0046] The dataset construction module is used to obtain audio data and construct a dataset;

[0047] A model building module is used to replace the residual layer with the Bottleneck layer using the ResNet residual neural network as a basic framework to obtain a recognition model; the Bottleneck layer is provided with a squeeze-excitation module;

[0048] A model training module is used to train the recognition model based on the data set to obtain a trained recognition model;

[0049] The recognition module is used to input real-time data into the trained recognition model to obtain recognition information.

[0050] Based on the above scheme, the present invention provides a method and system for passive acoustic monitoring of white-headed langurs based on residual neural networks. By using a bottleneck layer and adding a squeeze-excitation mechanism module SE on the basis of the residual neural network ResNet architecture, the model extracts richer and more effective feature representations, pays more attention to the subtle features of the white-headed langur audio, improves the representation ability of the model, and reduces the number of parameters, which is conducive to improving the efficiency of the model and deploying it on resource-constrained devices. The present invention effectively solves the problem that manual identification cannot process large-scale passive acoustic audio data. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the steps of a method for passive acoustic monitoring of white-headed langurs based on residual neural network of the present invention;

[0052] Figure 2 It is a schematic diagram of the structure of the identification network of the present invention;

[0053] Figure 3 Schematic diagram of the structure of the Bottleneck layer of the present invention;

[0054] Figure 4 It is a schematic diagram of the structure of the squeeze-excitation module of the present invention;

[0055] Figure 5 The present invention is a structural block diagram of a white-headed langur passive acoustic monitoring system based on a residual neural network. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] It should be noted that, for the convenience of description, only the parts related to the relevant invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0058] It should be understood that the "system", "device", "unit" and / or "module" used in this application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0059] Unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0060] In the description of the embodiments of the present application, "plurality" means two or more than two. The following terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0061] In addition, flow charts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0062] Reference Figure 1 , is a flow chart of an optional example of a method for passive acoustic monitoring of white-headed langurs based on a residual neural network proposed in the present invention. The method can be applied to a computer device. The detection method proposed in this embodiment may include but is not limited to the following steps:

[0063] Step S1, obtaining audio data and constructing a data set;

[0064] Step S2, using the ResNet residual neural network as the basic framework, replacing the residual layer with the Bottleneck layer to obtain a recognition model; the Bottleneck layer is provided with a squeeze-excitation module;

[0065] Step S3: training the recognition model based on the data set to obtain a trained recognition model;

[0066] Step S4: input the real-time data into the trained recognition model to obtain recognition information.

[0067] In some embodiments, the step S1 specifically includes:

[0068] S1.1, obtain audio data;

[0069] S1.2, performing format unification, noise reduction and labeling processing on the audio data to obtain pre-processed data;

[0070] The original audio collected by multiple recording devices has different storage formats and data lengths, which is not convenient for subsequent processing. Therefore, it is necessary to first unify the audio format, perform noise reduction, and then manually annotate the audio.

[0071] S1.3, dividing the preprocessed data into a training set and a test set.

[0072] Specifically, after completing the audio preprocessing, the data set is randomly divided into a training set and a test set in a ratio of approximately 7:3, where the training set is used to train the network to determine the optimal parameters, and the test set is used to verify the reliability and feasibility of the network model.

[0073] The audio recording locations are all in the White-headed Langur Nature Reserve in Chongzuo City. These data come from multiple recording devices. These audios are divided into two categories. One is the manually recorded wav format audio, which has different durations, and the other is the 12s MP3 format audio recorded by passive acoustic monitoring equipment. First, the manually recorded audio is segmented. We use the sinc algorithm to resample the long audio from the original 44100hz to the 32000Hz of PAM recording, and change the audio format to MP3 format. Then, it is divided into several audio segments through a sliding window with a length of 12s and an overlap time of 0s. For the last segment, if the length is less than 12 seconds, it is padded with zero values ​​to make its length reach 12 seconds.

[0074] The audio data was subjected to short-time Fourier transform (STFT) to convert the audio signal into a spectrum representation. According to the frequency domain range of the white-headed langur's call, the corresponding frequency index was found, and the rows in the spectrum matrix that were greater than the maximum frequency index were set to zero to implement the frequency truncation operation, retaining only the frequency domain where the white-headed langur's call was located.

[0075] All audios are manually annotated to determine the audio content, and thus the audio is divided into positive samples and negative samples. The samples are then labeled in the format of "positive (negative) - type of white-headed langur call - recording location - recording date". In addition, when labeling, the type, number and order of calls contained in the audio segment are recorded.

[0076] In some embodiments, it also includes:

[0077] S1.4. Performing data enhancement on the training set; the data enhancement methods include volume enhancement, noise enhancement, speech rate disturbance enhancement and spectrum enhancement.

[0078] S1.4.1. Volume enhancement is achieved by randomly adjusting the gain factor of the audio signal to increase or decrease the overall volume of the audio.

[0079] S1.4.2. Noise enhancement is achieved by adding environmental noise to the original audio signal to generate new audio data, thereby achieving noise enhancement.

[0080] S1.4.3. Divide the audio signal into frames, calculate the amplitude spectrum and phase spectrum, calculate the target position according to the speed perturbation factor, and interpolate the target phase to synthesize the modified spectrum; then, through inverse phase locking transformation and overlap-addition of the modified audio frames, speech rate perturbation enhancement is achieved.

[0081] S1.4.4. Randomly select spectrum regions and time intervals within a certain range and mask them to introduce local missing information and simulate partial missing or discontinuous audio to achieve spectrum enhancement.

[0082] In some embodiments, before step S4, the method further includes:

[0083] Perform short-time Fourier transform on real-time data to convert the original time-domain audio signal into the time-frequency domain to obtain a time-frequency domain matrix;

[0084] The time-frequency domain matrix is ​​a two-dimensional matrix, which contains spectrum information at different times.

[0085] A set of triangular filters is defined, wherein the center frequencies of the filters are spaced using a Mel frequency scale;

[0086] Multiplying the time-frequency domain matrix by each of the filters element by element and summing the results to obtain a filter bank characteristic matrix;

[0087] The filter bank feature matrix is ​​logarithmically transformed to obtain final features.

[0088] In addition, the data processing process is also performed in step S3, that is, the data needs to be converted into matrix features before being input into the model.

[0089] In some embodiments, in step S2, wherein:

[0090] Reference Figure 2 ,The recognition model includes 1 convolution block, 4 residual blocks consisting of a Bottleneck layer and SE module, and a fully connected (FC) part. The convolution block consists of a Transpose layer, an Unsqueeze layer and a Convolution layer. The input log filter features are first dimensionally transformed by the Transpose layer and the Unsqueeze layer, and then are convolutionally transformed by the convolution layer.

[0091] The output of the convolutional block is fed into a residual block consisting of a Bottleneck layer and a Squeeze-Excitation (SE) module, see Figure 3 , which includes convolutional layers, batch normalization (BN) layers, ReLU functions, squeeze and excitation layers, and residual connections. First, the input features are convolved through the convolutional layer to extract feature information. Then the output of the convolutional layer is normalized through the BN layer, which helps to speed up the network training process and improve the stability of the model. Subsequently, the ReLU activation function is used to introduce nonlinearity, enhance the network's expressiveness, and help learn complex features. The ReLU function is as follows:

[0092]

[0093] The convolutional layer is applied again to further extract feature information, followed by the BN layer and the ReLU function, and finally the convolutional layer again. The output of the convolutional layer is normalized by the BN layer, and the features are processed by the SE module. First, the features are compressed by global average pooling, and then the excitation operation is performed through two fully connected layers to adjust the channel weights of the features, thereby enhancing the expressive power of the features. The structure of the squeeze-excitation (SE) module refers to Figure 4 Next, residual connections are used to add the input features to the features processed by convolution, BN, ReLU, and SE, so that information can be directly transferred between network layers, avoiding the gradient vanishing problem and helping to train deeper networks. Finally, the ReLU activation function is applied again to further enhance the nonlinear expression ability of the network and improve the distinguishability of features.

[0094] In the fully connected part, it first passes through 3 fully connected layers. The first and second fully connected layers are composed of 4096 neurons. After the fully connected operation, the activation function ReLU is used for nonlinear operation. The third connection layer is composed of 8 neurons. After completing the fully connected operation, the multi-classification output is finally converted into probability through SoftMax, as follows:

[0095]

[0096] Among them, z i is the output value of the i-th node, n is the number of output nodes, that is, the number of categories for classification and recognition. At this point, the residual neural network model ResNetSE based on the squeeze-excitation mechanism has been completed.

[0097] The trained network model ResNetSE has better recognition accuracy than the simple ResNet network architecture. The recognition accuracy of ResNetSE reaches 98.41%, and a BN (BatchNormalization) layer is added after the output of the convolutional layer to prevent gradient disappearance or gradient explosion, effectively solving key problems in white-headed langur audio recognition, such as insufficient audio samples, complex acoustic modeling, training stability, robustness and generalization ability, thereby greatly improving the performance of white-headed langur audio recognition.

[0098] In some feasible embodiments, the step S3 further includes:

[0099] The initial learning rate is adopted, the optimization method is the Adam method, and the learning rate decay function is the WarmupCosineSchedulerLR method. During the training process, the Dropout layer of the fully connected part needs to be enabled, and some neurons are randomly discarded to prevent the model from overfitting. At the same time, the accuracy and recall rate of the test set are used as the standard for evaluating the model. Then the transfer learning technology is introduced. ResNetSE is first pre-trained on the public UrbanSound8k dataset. The dataset is also randomly divided into a training set and a test set in a ratio of 7:3. Then the weights obtained through pre-training are applied as the initial weights to the white-headed langur call recognition task, and the recognition model is retrained. The training set is divided into several batches according to the amount of data in the training set, where the amount of data in each batch is 16 audio segments, and the audio of each batch is randomly shuffled during the training process. After the model training is completed, the reliability and feasibility of the model need to be further verified. In this step, the test set needs to be used to verify the reliability and feasibility of the network. When testing the network, the BN layer and Dropout layer are disabled to avoid changes in weights.

[0100] In some feasible embodiments, it also includes:

[0101] The trained recognition model is used to generate training data and added to a training set; the trained recognition model is trained based on the training set; and the steps of iteratively supplementing the training set and the training model are performed.

[0102] The results predicted by the recognition model are verified and corrected manually. The accuracy of these prediction results is checked manually, and possible incorrect labels are corrected. At the same time, the correct samples are labeled in the same label format, and the similarity between the false positive samples and the calls of the white-headed langurs is analyzed. Similar false positive audio and false negative audio are added to the original training set, and the recognition model is retrained on the expanded training set. Repeat the above steps, continuously select samples from the unlabeled data set, and perform a cycle of prediction-verification-adding to the training set-retraining until the performance of the test set no longer improves.

[0103] In some feasible embodiments, it also includes:

[0104] The optimized model is deployed on the edge computing device and integrated into the monitoring application. The edge computing device is then seamlessly integrated with the passive acoustic recording device to achieve real-time analysis of the recording.

[0105] like Figure 5 As shown, a passive acoustic monitoring system for white-headed langurs based on residual neural network includes:

[0106] The dataset construction module is used to obtain audio data and construct a dataset;

[0107] A model building module is used to replace the residual layer with the Bottleneck layer using the ResNet residual neural network as a basic framework to obtain a recognition model; the Bottleneck layer is provided with a squeeze-excitation module;

[0108] A model training module is used to train the recognition model based on the data set to obtain a trained recognition model;

[0109] The recognition module is used to input real-time data into the trained recognition model to obtain recognition information.

[0110] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0111] A passive acoustic monitoring device for white-headed langurs based on residual neural network:

[0112] at least one processor;

[0113] at least one memory for storing at least one program;

[0114] When the at least one program is executed by the at least one processor, the at least one processor implements the passive acoustic monitoring method of white-headed langurs based on residual neural network as described above.

[0115] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0116] A storage medium stores processor-executable instructions, which are used to implement the above-mentioned method for passive acoustic monitoring of white-headed langurs based on residual neural network when executed by the processor.

[0117] The contents of the above method embodiments are all applicable to the present storage medium embodiments. The functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0118] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A passive acoustic monitoring method for white-headed langurs based on residual neural network, characterized in that: The following steps are involved: Get audio data and build a dataset; Taking the ResNet residual neural network as the basic framework, the Bottleneck layer replaces the residual layer to obtain the recognition model; The Bottleneck layer is provided with a squeezing-excitation module; Training the recognition model based on the data set to obtain a trained recognition model; The real-time data is input into the trained recognition model to obtain recognition information.

2. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 1, characterized in that: The step of obtaining audio data and constructing a data set specifically includes: Get audio data; Performing format unification, noise reduction and labeling processing on the audio data to obtain pre-processed data; The preprocessed data are divided into a training set and a test set.

3. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 2 is characterized in that: Also includes: Performing data augmentation on the training set; The data enhancement methods include volume enhancement, noise enhancement, speech rate disturbance enhancement and spectrum enhancement.

4. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 2 is characterized in that: The step of training the recognition model based on the data set to obtain a trained recognition model specifically includes: Based on the transfer learning method, determining the initial weight of the recognition model; Training the recognition model based on the training set to obtain a preliminary recognition model; The test set is used to perform the preliminary test, disabling the BN layer and the Dropout layer, and selecting the model with the highest accuracy as the trained recognition model.

5. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 3 is characterized in that: The step of performing data enhancement on the training set specifically includes: Setting a gain factor, and performing volume enhancement on the training set based on the gain factor; Setting environmental noise, and performing noise enhancement on the training set based on the environmental noise; Setting a speed factor, and modifying the audio frame based on the speed disturbance factor, to perform speech rate disturbance enhancement on the training set; The spectrum area and time interval within the preset range are masked, and the locally missing information is introduced to perform spectrum enhancement on the training set.

6. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 1, characterized in that: Before the step of inputting the real-time data into the trained recognition model, the method further includes: Perform short-time Fourier transform on real-time data to convert the original time-domain audio signal into the time-frequency domain to obtain a time-frequency domain matrix; defining a triangular filter whose center frequencies are spaced using a Mel frequency scale; Multiplying the time-frequency domain matrix by each of the filters element by element and summing the results to obtain a filter bank characteristic matrix; The filter bank feature matrix is ​​logarithmically transformed to obtain final features.

7. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 6, characterized in that: The step of inputting the real-time data into the trained recognition model to obtain recognition information specifically includes: Inputting the final features corresponding to the real-time data into the trained recognition model; Perform dimension conversion and convolution operations based on the convolution block to obtain the convolution block output; Performing convolution, batch normalization, compression-excitation, and residual connection on the convolution block output based on the residual block to obtain a residual block output; Based on the fully connected layer, the residual block output is nonlinearized and fully connected, and the multi-classification output is converted into probability through SoftMax to obtain recognition information.

8. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 1, characterized in that: Also includes: Generate training data using the trained recognition model and add it to the training set; Training the trained recognition model based on the training set; Iterate the steps of replenishing the training set and training the model.

9. The method for passive acoustic monitoring of white-headed langurs based on residual neural network according to claim 1, characterized in that: Also includes: The trained recognition model is deployed on the edge computing device.

10. A passive acoustic monitoring system for white-headed langurs based on residual neural network, characterized in that: include: The dataset construction module is used to obtain audio data and construct a dataset; A model building module is used to replace the residual layer with the Bottleneck layer using the ResNet residual neural network as a basic framework to obtain a recognition model; the Bottleneck layer is provided with a squeeze-excitation module; A model training module is used to train the recognition model based on the data set to obtain a trained recognition model; The recognition module is used to input real-time data into the trained recognition model to obtain recognition information.