Acoustic scene classification method based on MobileNet and compression excitation network

Abstract

The invention relates to an acoustic scene classification method based on MobileNet and a compression excitation network. The method comprises the following steps: acquiring acoustic scene information to be detected; inputting the acoustic scene information to be detected into an acoustic scene classification model to obtain acoustic scene classification; wherein the acoustic scene classification model is constructed based on a MobileNet network and a SENet network. A network architecture provided by the method comprises a feature fusion mechanism of a CNN module and a SENet module, and the CNN module and the SENet module capture local and global features respectively, so that representation learning of the network is enhanced; the MobieNet network takes the feature map output by the feature fusion mechanism as the input to replace the traditional CNN and RNN modules to perform feature extraction, and compared with the traditional CNN and RNN, the model has the characteristic of low complexity, and can perform accurate classification on a plurality of acoustic scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of sound signal processing, and in particular to an acoustic scene classification method based on MobileNet and a compression excitation network. Background Art

[0002] Sound is the most natural information carrier and is used as an important tool for communication between humans and machines. The information conveyed by sound can be extracted and analyzed through signal processing methods to promote environmental perception and understanding. Acoustic scenes are composed of various sounds mixed and overlapped from real scenes. For example, the acoustic scene of a train station can include the sound of a passing train, the footsteps and voices of train attendants and passengers, etc. Acoustic scene classification is an important signal processing task that obtains environmental feature information from actual audio samples in order to accurately classify a series of acoustic environments. Acoustic scene classification focuses on audio signal processing and machine learning, and has application prospects in real life. Deep learning is widely used in acoustic scene classification tasks because of its superior performance when processing a large number of features. For example, Convolutional Neural Networks (CNN), Recurrent Convolutional Neural Networks (RCNN), Depthwise Separable Convolution Neural Networks (DSCNN), etc. have been widely used in the field of acoustic scene classification.

[0003] So far, scholars at home and abroad have proposed a variety of acoustic scene classification methods: Yang et al. developed a CNN model with two-stage multi-scale feature fusion and channel weighting, which can effectively integrate acoustic features in acoustic scene classification through a top-down approach. Byttebier et al. and Sugahara et al. used ResNet to design a scene classification system for scene prediction. Galindo et al. used an end-to-end CNN network to design a scene classification algorithm based on the lottery hypothesis. Horvath et al., Cai et al., and Dong et al. used a DSCNN network to design a scene classification system. Cui et al. combined consistency learning and data enhancement algorithms in the CNN model to improve the robustness of the model and reduce device dependence. In order to improve the classification performance, researchers have also adopted some auxiliary classification methods, such as: Anastacio et al. introduced data enhancement, Xin et al. introduced the attention mechanism, Olisaemeka et al. increased the receptive field of the convolutional layer, etc. Waldekar et al. reconstructed the scene classification framework and proposed a two-layer classification framework. First, the audio is generally estimated to obtain a coarse label, and then the combination of spectral features is used to perform a second layer of audio to obtain a fine classification label. In order to deal with the spatial complexity problem of ResNet, Cao et al. adopted separable convolution and attention mechanism on channel space. These classification methods have shown good classification performance.

[0004] Although the introduction of neural networks has improved the performance of acoustic scene classification systems, it has also brought about problems such as huge scale, high computational complexity, and excessive memory consumption when applied to smart devices. Therefore, it is particularly important to achieve low model complexity. Studies have found that CNN usually has some redundant parameters, such as weights or filters. These parameters only generate additional calculations and storage, and do not have much impact on the classification performance of acoustic scene classification tasks. However, pruning weights requires modifying the network structure, and the operation is extremely complicated. Based on this, Singh et al. proposed that redundant filters in CNN networks that do not provide discriminative information in different acoustic scene categories can be pruned. However, this method requires processing of the data set, has high complexity in practical applications, and is not suitable for situations where the data set is large. On the other hand, although CNN and its network variants have shown great success in completing acoustic scene classification tasks, these neural networks have an inherent limitation, that is, they cannot learn global information well, which leads to insufficient global representation capabilities in CNN-based classification models. Therefore, the present invention designs a network architecture with low model complexity and improves the accuracy of acoustic scene classification. Summary of the invention

[0005] The present invention discloses a network architecture using MobileNet and a compressed excitation network for acoustic scene classification, which is used to solve the problems in the prior art such as huge network architecture scale, high computational complexity, and insufficient global representation capability in the CNN classification model.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] An acoustic scene classification method based on MobileNet and compressed excitation network, comprising:

[0008] Collecting acoustic scene information to be detected;

[0009] The acoustic scene information to be detected is input into an acoustic scene classification model to obtain an acoustic scene classification; wherein the acoustic scene classification model is constructed based on a MobileNet network and a SENet network.

[0010] Optionally, the acoustic scene classification model includes: a logMBE module, a CNN module, a SENet module, a fusion module, a MobileNet module and a feedforward network layer;

[0011] Inputting the acoustic scene information to be detected into the acoustic scene classification model includes:

[0012] Inputting the acoustic scene information to be detected into the logMBE module for feature extraction to extract an energy feature map;

[0013] Input the energy feature map into the CNN module and the SENet module respectively for feature extraction to obtain a CNN feature map and a SENet feature map;

[0014] Input the CNN feature map and the SENet feature map into the fusion module to perform feature fusion to obtain a fused feature map;

[0015] The fused feature map is input into the MobileNet module for feature extraction, and then input into the feedforward network layer to obtain the acoustic scene classification.

[0016] Optionally, inputting the acoustic scene information to be detected into the logMBE module for feature extraction includes:

[0017] Performing pre-emphasis filtering processing on the acoustic scene information to be detected;

[0018] Performing frame segmentation processing on each audio in the pre-emphasized acoustic scene information to segment it into a set of overlapping frames;

[0019] Multiply each frame by a Hamming window;

[0020] Using short-time Fourier transform, spectrum information is extracted from each frame after multiplying it by the Hamming window.

[0021] The extracted spectrum information is filtered using a bandpass filter, and the logarithmic energy is calculated to obtain a logarithmic Mel spectrum graph.

[0022] Optionally, the CNN module includes: a convolutional layer, a batch normalization layer and a Gaussian error linear unit;

[0023] The input of the convolution layer is a two-dimensional matrix of corresponding pixel values ​​of the image conversion. The feature map is obtained through convolution calculation of the convolution layer. The feature map is input to the batch normalization layer to further extract representative features and reduce the dimension of the feature map. Finally, it is nonlinearly mapped through the Gaussian error linear unit.

[0024] Optionally, the SENet module includes: a first feature map reorganization submodule, a compression excitation submodule, and a second feature map reorganization submodule;

[0025] Inputting the energy feature map into the SENet module for feature extraction includes:

[0026] Input the energy characteristic graph into the first characteristic graph reorganization submodule to perform the first characteristic graph reorganization to obtain a reorganized W×H×C characteristic graph X;

[0027] The reorganized W×H×C feature map X is input into the compression excitation submodule, and the convolution processing of the two-dimensional spatial kernel of the preset convolution operator is performed to obtain a new W×H×C feature map U; the feature map U is compressed into a 1×1×C feature vector Z using the global average pooling of the channel, and the feature vector Z is input into two fully connected layers. The first fully connected layer compresses the C channels into C / r, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and then uses Sigmoid activation to obtain a 1×1×C weight s. The feature map U is multiplied by the weight s and re-weighted to obtain the feature map finally output by the compression excitation submodule;

[0028] The second feature map reorganization submodule is used to reorganize the feature map finally output by the compression excitation submodule, and the three parameters WHC of the feature map are changed to obtain the feature map X'.

[0029] Optionally, inputting the CNN feature map and the SENet feature map into the fusion module to perform feature fusion includes:

[0030] Merge the CNN feature map and the SENet feature map;

[0031] The merged feature map is subjected to CNN convolution processing.

[0032] Optionally, inputting the fused feature map into the MobileNet module for feature extraction includes:

[0033] Using the expansion layer, the number of channels of the fused feature map is increased;

[0034] Use the depthwise separable convolutional layer to extract features from the feature map with increased number of channels;

[0035] Use the projection layer to reduce the number of channels of the extracted feature map;

[0036] Repeat the expansion layer, depth-wise separable convolution layer, and projection layer processing several times to obtain the feature map finally extracted by the MobileNet module.

[0037] Optionally, the feature map extracted by the MobileNet module further includes:

[0038] Input the feature map extracted by the MobileNet module into the convolution layer and the average pooling layer to obtain a feature vector;

[0039] The feature vector is input into the feed-forward network layer.

[0040] The beneficial effects of the present invention are:

[0041] The network architecture proposed by the method of the present invention includes a feature fusion mechanism of a CNN module and a SENet module, which capture local and global features respectively, thereby enhancing the representation learning of the network;

[0042] In the network architecture proposed by the method of the present invention, the feature map of the feature fusion module is input into the MobieNet module, replacing the traditional CNN and RNN modules for feature extraction, which realizes the low complexity of the model and can accurately classify multiple acoustic scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0044] Figure 1 The network architecture of the proposed method in the embodiment of the present invention; wherein (a) is the network architecture of the proposed method, (b) is the network architecture of the SENet module, and (c) is the network architecture of the MobileNet module;

[0045] Figure 2 This is the confusion matrix of ten acoustic scenes using the proposed method on the DCASE 2020 dataset of an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] An acoustic scene classification method based on MobileNet and compression and excitation network, including:

[0049] Collecting acoustic scene information to be detected;

[0050] The acoustic scene information to be detected is input into the acoustic scene classification model to obtain the acoustic scene classification; wherein the acoustic scene classification model is constructed based on the MobileNet network and the SENet network.

[0051] Specifically, the acoustic scene classification model in this embodiment includes a sound signal processing (logMBE) module. The logMBE module measures the audio of different acoustic scenes through a recording device and re-encodes it, and extracts the energy characteristics of its logarithmic Mel spectrum. An input audio record is first divided into a set of frames, and the logarithmic Mel spectrum energy of each frame is calculated and regarded as the feature of this frame. The process of obtaining the logarithmic Mel spectrum energy of the audio record is: pre-emphasis, frame segmentation, windowing, spectrum analysis and bandpass filtering; the sound signal outputs a feature map through the logMBE module, which is input into the CNN module and the SENet module respectively to obtain two feature maps and then The fusion module includes two processes: feature map merging and CNN convolution. The CNN module consists of three parts, including convolution layer, batch normalization layer (BN) and Gaussian error linear unit (GELU). The fused feature map is further extracted through a series of MobileNet modules, and the classification task is realized through the feedforward network layer (FNN). FFN consists of three layers: an input layer, a hidden layer and an output layer. Finally, a SoftMax function is used to normalize the output of FFN to obtain the prediction probability of each category. By selecting the node with the maximum prediction probability, the proposed method obtains the final classification of multiple acoustic scenes.

[0052] Furthermore, the acoustic scene classification model includes: a logMBE module, a CNN module, a SENet module, a fusion module, a MobileNet module and a feedforward network layer;

[0053] Inputting the acoustic scene information to be detected into the acoustic scene classification model includes:

[0054] The acoustic scene information to be detected is input into the logMBE module for feature extraction to extract the energy feature map;

[0055] The energy feature map is input into the CNN module and the SENet module for feature extraction to obtain the CNN feature map and the SENet feature map;

[0056] Input the CNN feature map and the SENet feature map into the fusion module to perform feature fusion and obtain the fused feature map;

[0057] The fused feature map is input into the MobileNet module for feature extraction and then into the feedforward network layer to obtain the acoustic scene classification.

[0058] Furthermore, the acoustic scene information to be detected is input into the logMBE module for feature extraction including:

[0059] Perform pre-emphasis filtering on the acoustic scene information to be detected;

[0060] Performing frame segmentation processing on each audio in the pre-emphasized acoustic scene information to segment it into a set of overlapping frames;

[0061] Multiply each frame by a Hamming window;

[0062] Using short-time Fourier transform, spectrum information is extracted from each frame after multiplying it by the Hamming window.

[0063] The extracted spectrum information is filtered using a bandpass filter, and the logarithmic energy is calculated to obtain a logarithmic Mel spectrum graph.

[0064] Specifically, in this embodiment, Figure 1 (a) shows the sound signal processing (logMBE) module. The logMBE module measures the audio re-encoding of different acoustic scenes through the recording device and extracts the energy features of its logarithmic Mel spectrum map. An input audio record is first divided into a set of frames. The logarithmic Mel spectrum energy of each frame is calculated and regarded as the feature of this frame. The following is a detailed description of the logarithmic Mel spectrum energy of the audio sample:

[0065] The first step is pre-emphasis.

[0066]

[0067] In formula (1), s[n] and Represent the samples of the input audio recording and the samples after passing through the pre-emphasis filter, respectively.

[0068] The second step is frame segmentation. Each audio recording is segmented into a set of overlapping frames, with a frame size of N samples and an overlap of M samples between each pair of consecutive frames.

[0069] The third step is windowing. Each frame is multiplied by a Hamming window.

[0070]

[0071] Where w[n] is the Hamming window function, and its expression is:

[0072] w[n]=0.54-0.46 cos[2πn / (N-1)],0≤n≤N-1 (3)

[0073] The fourth step is spectrum analysis. Short-time Fourier transform is used to obtain Extract the spectrum information.

[0074]

[0075] The fifth step is bandpass filtering. The amplitude spectrum is filtered by using a set of bandpass filters. After passing through the filter bank, the logarithmic Mel spectrum energy is obtained by calculating the logarithmic energy.

[0076]

[0077] In formula (5), J is the number of filters, A(k) is the amplitude of X[k], and M j (k) is the frequency response of the jth filter, E j is the output of the jth filter;

[0078] The sound signal outputs a feature map through the logMBE module, which is input into the CNN module and SENet module respectively. The two feature maps are then sent to the fusion module, which includes two processes: feature map merging and CNN convolution.

[0079] Furthermore, the CNN module includes: a convolutional layer, a batch normalization layer and a Gaussian error linear unit.

[0080] The input of the convolution layer is a two-dimensional matrix of corresponding pixel values ​​of the image conversion. The feature map is obtained through convolution calculation of the convolution layer. The feature map is input to the batch normalization layer to further extract representative features and reduce the dimension of the feature map. Finally, it is nonlinearly mapped through the Gaussian error linear unit.

[0081] Specifically, in this embodiment, the CNN module consists of three parts, including a convolutional layer, a batch normalization layer (BN), and a Gaussian error linear unit (GELU).

[0082] Further, the SENet module includes: a first feature map reorganization submodule, a compression excitation submodule, and a second feature map reorganization submodule;

[0083] Input the energy feature map into the SENet module for feature extraction including:

[0084] Input the energy feature map into the first feature map reorganization submodule to perform the first feature map reorganization to obtain the reorganized W×H×C feature map X; where W refers to the width of the feature map, H refers to the height of the feature map, and C refers to the number of channels of the feature map;

[0085] The reorganized W×H×C feature map X is input into the compression excitation submodule, and the convolution processing of the two-dimensional spatial kernel of the preset convolution operator is performed to obtain a new W×H×C feature map U; the feature map U is compressed into a 1×1×C feature vector Z using the global average pooling of the channel, and the feature vector Z is input into two fully connected layers. The first fully connected layer compresses the C channels into C / r, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and then uses Sigmoid activation to obtain a 1×1×C weight s. The feature map U is multiplied by the weight s and re-weighted to obtain the feature map finally output by the compression excitation submodule;

[0086] The second feature map reorganization submodule is used to perform feature reorganization on the feature map finally output by the compression excitation submodule to obtain the feature map X'.

[0087] Specifically, in this embodiment, the SENet module includes three parts: a first feature map reorganization module, a compression excitation submodule, and a second feature map reorganization module. The compression excitation submodule convolves the reorganized W×H×C feature map X with several two-dimensional spatial kernels to obtain a new W×H×C feature map U, and then uses the global average pooling of the channel to directly compress the feature map containing global information into a 1×1×C feature vector Z. The feature vector Z is input into two fully connected layers. The first fully connected layer compresses C channels to C / r, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and then obtains a 1×1×C weight s through Sigmoid activation. The feature map U is multiplied by the weight s and re-weighted to obtain the final output feature map.

[0088] The working principle of SENet module is as follows Figure 1As shown in (b), the compression excitation module convolves the reorganized W×H×C feature map X with several two-dimensional spatial kernels to obtain a new W×H×C feature map U. tr :X→U, through F tr The feature map X becomes the feature map U. tr It can be regarded as a standard convolution operator. The formula is defined as follows:

[0089]

[0090] Among them, X is the input feature map, U is the output feature map, and V c refers to the filter parameters, Represents a two-dimensional spatial kernel and * represents a convolution operation.

[0091] Use global average pooling F over channels sq , the feature map containing global information is directly compressed into a 1×1×C feature vector Z, so that the generated channel-level statistics Z contains contextual information, alleviating the problem of channel dependency. The definition is as follows:

[0092]

[0093] In order to utilize the information gathered in the compression operation, the expansion operation is used to fully capture the channel dependency. The feature vector Z is input to two fully connected layers. The first fully connected layer compresses C channels to C / r, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels back to C channels, and then through Sigmoid activation to obtain a 1×1×C weight s, where r refers to the compression ratio.

[0094] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (8)

[0095] The last step is the weighting operation. The attention weights obtained previously are weighted to the features of each channel. The feature map U is multiplied by the weight s and re-weighted to obtain the final output feature map.

[0096]

[0097] The benefit is that the SENet module can extract global information from the feature map to solve the problem of insufficient global representation ability in the CNN classification model. Compared with the traditional Transformer model that extracts global information, the SENet module has lower model complexity and better effect.

[0098] Furthermore, the CNN feature map and the SENet feature map are input into the fusion module, and feature fusion is performed including:

[0099] Merge the CNN feature map and the SENet feature map;

[0100] The merged feature map is subjected to CNN convolution processing.

[0101] Furthermore, the fused feature map is input into the MobileNet module for feature extraction including:

[0102] Use the expansion layer to increase the number of channels of the fused feature map;

[0103] Use the depthwise separable convolutional layer to extract features from the feature map with increased number of channels;

[0104] Use the projection layer to reduce the number of channels of the extracted feature map;

[0105] Repeat the expansion layer, depth-wise separable convolution layer, and projection layer processing several times to obtain the feature map finally extracted by the MobileNet module.

[0106] Furthermore, the feature map extracted by the MobileNet module also includes the following before being input into the feedforward network layer:

[0107] The feature map extracted by the MobileNet module is input into the convolution layer and the average pooling layer to obtain the feature vector;

[0108] Input the feature vector into the feed-forward network layer.

[0109] Specifically, in this embodiment, Figure 1 As shown in (c), the working principle of MobileNet is that MobileNet contains a series of seven MoblieNet modules and adopts a feature pyramid structure. As the network depth increases, the resolution of the feature map decreases, but the number of channels increases. After seven MoblieNet modules, the feature map is obtained, and then the feature map is input into the convolution layer and the average pooling layer to obtain the feature vector as the input of the final feedforward network layer (FNN). The MobileNet module includes three processes: the expansion layer increases the number of feature map channels, the depth-separable convolution layer extracts features, and the projection layer reduces the number of feature map channels. The projection layer uses a 1×1 convolutional network structure, and its purpose is to map high-dimensional features to low-dimensional space. The function of the expansion layer is just the opposite. It uses a 1×1 convolutional network structure, and its purpose is to map low-dimensional space to high-dimensional space. The expansion layer has a hyperparameter that is the number of times the dimension is expanded, which can be adjusted according to the actual situation. The default value is 6, which means 6 times expansion. The benefit is that compared with the traditional CNN and RNN feature extraction methods, the MobileNet module structure is simpler and the computational complexity is lower.

[0110] This example also tests an acoustic scene classification method based on MobileNet and a compressed excitation network:

[0111] In order to evaluate the effectiveness of the proposed method, the TAU Urban Acoustic Scenes 2020 Mobile, Development dataset (DCASE 2020) dataset was used. The datasets contain 13370 and 23035 audio records from ten different scene categories in the ASC task of the IEEE AASP challenge DCASE. The ten scene categories are airport, bus, subway, subway station, park, public square, shopping mall, street pedestrians, street traffic and tram. Each audio recording in both datasets lasts for 10 seconds. The classification performance of acoustic scene classification can be measured by accuracy, which indicates the number of correctly classified recordings in all test recordings. In addition, accuracy is the official evaluation indicator of DCASE for acoustic scene classification. The higher the accuracy value, the higher the classification accuracy of the acoustic scene.

[0112] Table 1 provides a comprehensive comparison of the average accuracy and various types of precision of the proposed method and other deep learning-based acoustic scene classification methods on the dataset. It can be seen that on the DCASE 2020 dataset, its classification results are compared with those of the DCASE2020 baseline, the DCASE2021 baseline, and the other three deep learning-based acoustic scene classification methods. Compared with this method, these five deep learning-based acoustic scene classification methods are all based on CNN or ResNet or their combination. As shown in Table 1, the proposed method has the highest average accuracy score of 78.8% on the DCASE 2020 dataset. Compared with the DCASE2020 baseline and the DCASE2021 baseline, the average accuracy values ​​of the proposed method are improved by 24.7% and 31.1%, respectively, and compared with the Cui, Vilouras, and Cao methods, they are improved by 9.1%, 8.5%, and 6.4%, respectively. In addition, in 80% of the acoustic scene categories, the classification accuracy of the proposed method is between 67.3% and 87.9%, ranking first or second among the proposed method and the other five deep learning-based acoustic scene classification methods. Overall, on the DCASE 2020 dataset, the proposed method achieved a high average accuracy rate for acoustic scenes, and its performance and reliability are superior to other deep learning-based acoustic scene classification methods.

[0113] Table 1

[0114]

[0115] Figure 2The confusion matrices of ten acoustic scenes using the proposed method on the DCASE 2020 dataset are shown. The confusion matrix can directly show the similarities between acoustic scenes. These confusion matrices provide a comprehensive view of the classification results of acoustic scenes, which helps to distinguish between true and predicted categories. In these confusion matrices, the rows and columns represent the true and predicted categories, respectively, and the values ​​in the main diagonal and off-diagonal positions represent the number of correct and incorrect classifications, respectively. The darker the red color of the diagonal position, the higher the classification accuracy of the acoustic scene. As shown Figure 2 As shown in the figure, the classification accuracy of the proposed method on the DCASE 2020 dataset ranges from 67.3% to 87.9%, and the accuracy of 4 acoustic scenes exceeds 80%. These statistics show that the proposed method achieves a higher accuracy rate for acoustic scenes. In general, the method proposed in this embodiment is a means to achieve accurate classification of multiple acoustic scenes, and its adaptability is also relatively strong.

[0116] The network architecture proposed in the method of this embodiment includes a feature fusion mechanism of a CNN module and a SENet module, which capture local and global features respectively, thereby enhancing the representation learning of the network; and the output feature map is used as the input of MobieNet, replacing the traditional CNN and RNN modules for feature extraction, which realizes the low complexity of the model and can accurately classify multiple acoustic scenes.

[0117] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.

Claims

1. An acoustic scene classification method based on MobileNet and compressed excitation network, characterized in that: include: Collecting acoustic scene information to be detected; Inputting the acoustic scene information to be detected into an acoustic scene classification model to obtain an acoustic scene classification; The acoustic scene classification model is built based on the MobileNet network and the SENet network.

2. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 1, characterized in that: The acoustic scene classification model includes: a logMBE module, a CNN module, a SENet module, a fusion module, a MobileNet module and a feedforward network layer; Inputting the acoustic scene information to be detected into the acoustic scene classification model includes: Inputting the acoustic scene information to be detected into the logMBE module for feature extraction to extract an energy feature map; Input the energy feature map into the CNN module and the SENet module respectively for feature extraction to obtain a CNN feature map and a SENet feature map; Input the CNN feature map and the SENet feature map into the fusion module to perform feature fusion to obtain a fused feature map; The fused feature map is input into the MobileNet module for feature extraction, and then input into the feedforward network layer to obtain the acoustic scene classification.

3. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 2, characterized in that: Inputting the acoustic scene information to be detected into the logMBE module for feature extraction includes: Performing pre-emphasis filtering processing on the acoustic scene information to be detected; Performing frame segmentation processing on each audio in the pre-emphasized acoustic scene information to segment it into a set of overlapping frames; Multiply each frame by a Hamming window; Using short-time Fourier transform, spectrum information is extracted from each frame after multiplying it by the Hamming window. The extracted spectrum information is filtered using a bandpass filter, and the logarithmic energy is calculated to obtain a logarithmic Mel spectrum graph.

4. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 2, characterized in that: The CNN module includes: a convolutional layer, a batch normalization layer and a Gaussian error linear unit; The input of the convolution layer is a two-dimensional matrix of corresponding pixel values ​​of the image conversion. The feature map is obtained through convolution calculation of the convolution layer. The feature map is input to the batch normalization layer to further extract representative features and reduce the dimension of the feature map. Finally, it is nonlinearly mapped through the Gaussian error linear unit.

5. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 2, characterized in that: The SENet module includes: a first feature map reorganization submodule, a compression excitation submodule, and a second feature map reorganization submodule; Inputting the energy feature map into the SENet module for feature extraction includes: Input the energy feature map into the first feature map reorganization submodule to perform the first feature map reorganization to obtain a reorganized W×H×C feature map X; wherein W refers to the width of the feature map, H refers to the height of the feature map, and C refers to the number of channels of the feature map; The reorganized W×H×C feature map X is input into the compression excitation submodule, and the convolution processing of the two-dimensional spatial kernel of the preset convolution operator is performed to obtain a new W×H×C feature map U; the feature map U is compressed into a 1×1×C feature vector Z using the global average pooling of the channel, and the feature vector Z is input into two fully connected layers. The first fully connected layer compresses the C channels into C / r, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and then uses Sigmoid activation to obtain a 1×1×C weight s. The feature map U is multiplied by the weight s and re-weighted to obtain the feature map finally output by the compression excitation submodule; The second feature map reorganization submodule is used to perform feature reorganization on the feature map finally output by the compression excitation submodule to obtain the feature map X'.

6. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 2, characterized in that: Inputting the CNN feature map and the SENet feature map into the fusion module, and performing feature fusion includes: Merge the CNN feature map and the SENet feature map; The merged feature map is subjected to CNN convolution processing.

7. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 2, characterized in that: Inputting the fused feature map into the MobileNet module for feature extraction includes: Using the expansion layer, the number of channels of the fused feature map is increased; Use the depthwise separable convolutional layer to extract features from the feature map with increased number of channels; Use the projection layer to reduce the number of channels of the extracted feature map; Repeat the expansion layer, depth-wise separable convolution layer, and projection layer processing several times to obtain the feature map finally extracted by the MobileNet module.

8. The acoustic scene classification method based on MobileNet and compressed excitation network according to claim 2, characterized in that: The feature map extracted by the MobileNet module also includes the following before being input into the feedforward network layer: Input the feature map extracted by the MobileNet module into the convolution layer and the average pooling layer to obtain a feature vector; The feature vector is input into the feed-forward network layer.

Images