CN115035912B - Automatic annotation method of underwater acoustic signal samples based on MOC model - Google Patents

Abstract

The invention discloses an automatic labeling method of underwater sound signal samples based on MOC models, in particular relates to an automatic labeling method of underwater sound signal samples based on MOC models, which aims to solve the problems that the traditional underwater sound signal sample labeling adopts a manual method, is time-consuming and labor-consuming, has low economic benefit, is limited by professionals and has low labeling accuracy, and comprises the steps of collecting underwater sound signals as samples and calculating the acoustic characteristics of the underwater sound signal samples by utilizing acoustic models; establishing a MOC model, wherein the MOC model sequentially comprises a first convolution layer, a preferred convolution residual layer, a second convolution layer, a attention mechanism layer, a full connection layer and a classification layer, inputting acoustic features of underwater sound signal samples into the MOC model for training, and outputting marked underwater sound signal samples until loss converges to obtain a trained MOC model; and (3) carrying out the operation on the underwater sound signal sample to be marked to obtain the marked underwater sound signal sample. Belonging to the field of underwater sound signal labeling.

Description

Automatic annotation method of underwater acoustic signal samples based on MOC model

Technical Field

The invention relates to a labeling method, in particular to an automatic labeling method for underwater sound signal samples based on a MOC model, and belongs to the field of underwater sound signal labeling.

Background technique

With the widespread application of deep learning and reinforcement learning, it is generally accepted that deep learning and reinforcement learning are applied to the pattern recognition task of underwater acoustic signals. Then, a problem arises: how to process underwater acoustic signal samples. Since there are very few underwater acoustic signal data sets, the data volume of the data sets is not large enough, and the data sets are not accurate enough, this problem seriously affects the development of the underwater acoustic field. Through research, it is found that the main root of this problem is the slow development of underwater acoustic signal labeling. The traditional method of labeling underwater acoustic signal samples is to use manual labeling. This manual labeling method is not only time-consuming and labor-intensive, but also has low economic benefits. At the same time, due to the limitations of the professionalism of the labelers, the accuracy of the labeling often cannot meet the requirements.

Summary of the invention

In order to solve the problem that traditional underwater acoustic signal sample labeling adopts manual method, which is not only time-consuming and labor-intensive, but also has low economic benefits and is limited by professionalism and has low labeling accuracy, the present invention proposes an automatic labeling method for underwater acoustic signal samples based on MOC model.

The technical solution adopted by the present invention is:

It includes the following steps:

S1. Collecting an underwater acoustic signal as a sample, and calculating the acoustic characteristics of the underwater acoustic signal sample using an acoustic model;

S2, establishing a MOC model, which includes a convolutional layer 1, a preferred convolutional residual layer, a convolutional layer 2, an attention mechanism layer, a fully connected layer and a classification layer in sequence, inputting the acoustic features of the underwater acoustic signal sample into the MOC model for training, and outputting the labeled underwater acoustic signal sample until the loss converges to obtain a trained MOC model;

S3. Execute S1-S2 on the underwater acoustic signal sample to be labeled to obtain a labeled underwater acoustic signal sample.

Preferably, the acoustic model in S1 includes a Gaussian mixture model and a hidden Markov model.

Preferably, a MOC model is established in S2, and the MOC model includes a convolution layer 1, a preferably convolution residual layer, a convolution layer 2, an attention mechanism layer, a fully connected layer and a classification layer in sequence. The acoustic features of the underwater acoustic signal sample are input into the MOC model for training, and the labeled underwater acoustic signal sample is output until the loss converges to obtain a trained MOC model. The specific process is:

S21, inputting the acoustic features of the underwater acoustic signal sample into the convolution layer 1 of the MOC model, and outputting feature 1;

S22, input feature 1 output by S21 into the preferred convolution residual layer, and output feature 2;

S23, input feature 2 output by S22 into convolutional layer 2, and output feature 3;

S24, input feature 3 output by S23 into the attention mechanism layer, and output feature 4;

S25, input feature 4 output by S24 into the fully connected layer, and output feature 5;

S26, input feature five outputted from S25 into the classification layer, and output the labeled underwater acoustic signal samples.

Preferably, the preferred convolution residual layer in S22 includes a first convolution layer, a residual layer, a second convolution layer, and an attention mechanism preferred convolution layer in sequence.

Preferably, in S22, feature 1 outputted from S21 is inputted into a preferred convolution residual layer, and feature 2 is outputted. The specific process is as follows:

S221, input feature 1 output by S21 into the first convolutional layer, and output feature β;

S222, input the feature β output by S221 into the residual layer, and output the feature γ;

S223, input the feature γ output by S222 into the second convolutional layer, and output the feature δ;

S224, input the feature δ output by S223 into the preferred convolutional layer of the attention mechanism, and output the feature ξ;

S225. Multiply the feature ξ output by S224 by the feature one output by S21 to obtain feature two.

Preferably, the attention mechanism in S224 preferably includes, in sequence, a global average pooling layer (GAP), a one-dimensional convolution layer, preferably a convolution layer one, preferably a convolution layer two, and preferably a convolution layer three.

Preferably, in S224, the feature δ outputted from S223 is inputted into the attention mechanism preferably convolutional layer, and the feature ξ is outputted. The specific process is:

S2241, input the feature δ output by S223 into the global average pooling layer, and output feature a;

S2242, input feature a output by S2241 into a one-dimensional convolutional layer, and output feature b;

S2243, input feature b output by S2242 into the preferred convolutional layer 1, and output feature c;

S2244, input feature c output by S2243 into the preferred convolutional layer 2, and output feature d;

S2245, input feature c output by S2243 and feature d output by S2244 into the preferred convolutional layer 3, and output feature e;

S2246, aggregate feature c output by S2243, feature d output by S2244, and feature e output by S2245 to obtain aggregated features, process the aggregated features using Sigmoid, and output feature f;

S2247. Add the feature f output by S2246 and the feature δ output by S223 to obtain the feature ξ.

Preferably, in S2243, the feature b output by S2242 is input into the preferred convolutional layer 1 to output feature c. The specific process is as follows:

Ⅰ. Set A convolution kernels of different sizes in the preferred convolution layer 1, and set the maximum size of the convolution kernel in the preferred convolution layer 1 according to the dimension of the acoustic characteristics of the underwater acoustic signal sample described in S1:

Wherein, C represents the dimension of the acoustic feature of the underwater acoustic signal sample, and the acoustic feature is a Mel spectrum;

K represents the convolution kernel size, k is an odd number and an integer;

Ⅱ. Use the convolution kernel optimization algorithm to select the best convolution kernel among A convolution kernels. The specific process is as follows:

Calculate the outputs of A convolution kernels, use a similarity algorithm to obtain the two convolution kernels with the highest similarity, select the convolution kernel with a larger size among the two convolution kernels as the optimal convolution kernel of the preferred convolution layer 1, and obtain the preferred convolution layer 1 with the optimal convolution kernel;

III. Input feature b output by S2242 into the preferred convolution layer 1 with the optimal convolution kernel to obtain feature c.

Preferably, in S2244, the feature c outputted by S2243 is inputted into the preferred convolutional layer 2 to output feature d. The specific process is as follows:

The number of convolution kernels, the maximum size of the convolution kernel, and the optimal convolution kernel of the preferred convolution layer 2 are the same as those of the preferred convolution layer 1. The feature c obtained in III is input into the preferred convolution layer 2, and the feature d is output.

Preferably, in S2245, the feature c output by S2243 and the feature d output by S2244 are input into the preferred convolution layer 3 to output feature e. The specific process is as follows:

The number of convolution kernels, the maximum size of the convolution kernel, and the optimal convolution kernel of the preferred convolution layer three are the same as those of the preferred convolution layer one. The feature c obtained in III and the feature d output by S2244 are input into the preferred convolution layer three, and the feature e is output.

Beneficial effects:

The present invention first uses the existing acoustic model to calculate the acoustic features of the underwater acoustic signal sample, in preparation for the next step of extracting the feature embedding of the underwater sound signal sample; then the MOC model is established by using the method of optimal convolution, attention mechanism and multi-layer feature fusion, the MOC model includes convolution layer one, optimal convolution residual layer, convolution layer two, attention mechanism layer, full connection layer and classification layer in sequence; the preferred convolution residual layer includes the first convolution layer, residual layer, second convolution layer, attention mechanism preferred convolution layer in sequence; the attention mechanism preferred convolution layer includes global average pooling layer (GAP), one-dimensional convolution layer, optimal convolution layer one, optimal convolution layer two, and optimal convolution layer three in sequence. The feature embedding of the acoustic features of the underwater acoustic signal sample is extracted according to the MOC model, so that the extracted feature embedding is more comprehensive and more representative, and finally the classifier (classifier function Softmax) of the MOC model is used to classify and label the underwater acoustic signal sample to obtain whether the sound signal in the underwater acoustic signal sample belongs to a normal signal or a noise signal.

The present invention uses the MOC model to extract feature embedding of the acoustic features according to the acoustic features of the underwater acoustic signal samples, and then automatically classifies and labels the obtained feature embedding using the classifier function Softmax, replacing the traditional manual labeling method, which not only saves time and labor, but also improves economic benefits. At the same time, it is not restricted by professionalism and improves the accuracy of labeling.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a structural diagram of the MOC model;

FIG2 is a structural diagram of the preferred convolutional residual layer OCA-Res2Block;

Figure 3 is a structural diagram of the attention mechanism's preferred convolutional layer OCA-Block;

Detailed ways

Specific implementation method 1: This implementation method is described in conjunction with Figures 1 to 3. This implementation method describes an automatic annotation method for underwater acoustic signal samples based on the MOC model, which includes the following steps:

S1, collecting an underwater acoustic signal as a sample, and calculating the acoustic characteristics of the underwater acoustic signal sample using an acoustic model;

Underwater sound signals (referred to as underwater acoustic signals) are taken as samples, and the acoustic features of underwater sound signal samples are extracted using traditional acoustic models, in preparation for the next step of extracting feature embedding of underwater sound signal samples. The underwater acoustic signals include the sound signals generated when the ship is sailing, and the acoustic models include Gaussian mixture models and hidden Markov models;

S2. Establish a MOC model. The MOC model includes a convolutional layer 1, a preferred convolutional residual layer, a convolutional layer 2, an attention mechanism layer, a fully connected layer, and a classification layer in sequence. The acoustic features of the underwater acoustic signal sample are input into the MOC model for training, and the labeled underwater acoustic signal sample is output until the loss converges to obtain a trained MOC model. The specific process is as follows:

The MOC model is also known as the multi-layer optimal convolutional network model. It is a model that combines multi-layer feature fusion and attention mechanism. The structure diagram is shown in Figure 1, where Conv1D refers to one-dimensional convolution; BN stands for batch normalization; ReLU is the activation function; FC represents the fully connected layer; C and T are the input sizes respectively; k is the convolution kernel size; d is the expansion rate of the void convolution; S is the number of categories.

S21, input the acoustic features of the underwater acoustic signal sample into the convolution layer 1 of the MOC model, and output feature 1. The specific process is:

In the convolution layer 1, the acoustic features of the underwater acoustic signal sample are processed using an activation function to obtain processed acoustic features. Nonlinear factors are added using the activation function to improve the expressive power of the MOC model. The processed acoustic features are then batch normalized to obtain low-dimensional features (feature 1) after preliminary convolution processing. The low-dimensional features are nonlinear features. Batch normalization is used to limit the processed acoustic features to an undetermined range, which improves the generalization ability of the MOC model to a certain extent. Batch normalization is also called batch data normalization or batch standardization. The convolution kernel size of the convolution layer 1 is 5, and the expansion rate of the dilated convolution is 1.

S22, input feature 1 output by S21 into the preferred convolution residual layer, and output feature 2. The specific process is:

The preferred convolution residual layer (OCA-Res2Block) consists of a Res2Block layer and an attention-based preferred convolution layer (OCA-Block), which combines the attention mechanism with the residual idea. The convolution kernel size of the preferred convolution residual layer is 3, and the dilation rate of the dilated convolution is 2. The preferred convolution residual layer includes the first convolution layer, the residual layer (Res2Conv1D), the second convolution layer, and the attention mechanism preferred convolution layer (OCA-Block), as shown in Figure 2.

S221, input the feature 1 output by S21 into the first convolutional layer, and output feature β. The specific process is:

In the first convolutional layer, the activation function is used to process the low-dimensional features (feature one) output by S21 to obtain the processed low-dimensional features, and then the processed low-dimensional features are batch normalized to obtain feature β. The low-dimensional features are processed by the above convolution to increase the dimension of the low-dimensional features and reduce the parameters of the low-dimensional features.

S222, input the feature β output by S221 into the residual layer, and output the feature γ. The specific process is:

In the residual layer (Res2 Conv1D), the feature β output by S221 is processed by an activation function to obtain the processed feature β, which is a nonlinear feature. The processed feature β is then batch normalized to obtain a feature γ (feature γ) with different granularities.

S223, input the feature γ output by S222 into the second convolutional layer, and output the feature δ. The specific process is:

In the second convolutional layer, the activation function is used to process the feature γ (feature γ) with different granularities output by S222 to obtain the processed feature γ with different granularities, and then the processed feature γ with different granularities is batch normalized to obtain a nonlinear high-dimensional feature δ (feature δ).

S224, input the feature δ output by S223 into the preferred convolutional layer of the attention mechanism, and output the feature ξ. The specific process is:

According to the SE-Block model recorded in the article Squeeze-and-Excitation Networks, the SE-Block model obtains the weights of different channels by using squeezing and excitation operations, so that the SE-Block model can autonomously determine which feature embeddings are important and which are not important. However, the SE-Block model will cause the loss of some important feature embeddings, reducing the performance of the SE-Block model. Therefore, the present invention uses a one-dimensional convolutional layer to replace the first fully connected layer in SE-Block, which not only reduces the huge amount of parameters brought by the use of the fully connected layer, but also can effectively reduce the amount of parameter calculation, while improving the calculation speed of the SE-Block model and reducing the operation time. Three layers of preferred convolutional layers (OC1d-Layer) are added in sequence after the one-dimensional convolutional layer, and an attention mechanism preferred convolutional layer (OCA-Block) is proposed. That is, the attention mechanism preferred convolutional layer includes a global average pooling layer (GAP), a one-dimensional convolutional layer, a preferred convolutional layer one, a preferred convolutional layer two, and a preferred convolutional layer three in sequence. As shown in Figure 3, k is the convolution kernel size and n is an odd number.

The present invention sets a plurality of convolution kernels of different sizes in each preferred convolution layer (OC1d-Layer), sets the maximum convolution kernel size of each OC1d-Layer according to the dimension of the acoustic features of the underwater acoustic signal sample after being processed by the acoustic model, and uses the convolution kernel optimization algorithm to select the convolution kernel (optimal convolution kernel) that can best represent the characteristics of the underwater acoustic signal in each OC1d-Layer. The total number of convolution kernels, the maximum size of the convolution kernel, and the optimal convolution kernel of each preferred convolution layer are all consistent, and then the existing preferred convolution operation is used to perform cross-channel interaction in a small range in the acoustic features, so as to obtain better and more representative feature embedding. At the same time, considering that the characteristic information of the underwater acoustic signal obtained after different OC1d-Layer processing is different, the multi-layer feature fusion idea is adopted to aggregate the outputs of the multi-layer OC1d-Layer.

S2241, input the feature δ output by S223 into the global average pooling layer, and output feature a;

S2242. Input feature a output by S2241 into a one-dimensional convolutional layer to output feature b with fewer parameters; the convolution kernel size in the one-dimensional convolutional layer is 1.

S2243, input feature b output by S2242 into the preferred convolutional layer 1, and output feature c. The specific process is as follows:

Ⅰ. Set A convolution kernels of different sizes in the preferred convolution layer 1, and set the maximum size of the convolution kernel in the preferred convolution layer 1 according to the dimension of the acoustic characteristics of the underwater acoustic signal sample described in S1:

Wherein, C represents the dimension of the acoustic feature of the underwater acoustic signal sample, and the acoustic feature is a Mel spectrum;

K represents the convolution kernel size, k is an odd number and an integer;

The present invention preferably has a convolutional layer 1. convolution kernels, the acoustic feature of the underwater acoustic signal is the Mel spectrum, and its frequency domain dimension is 80 dimensions, so according to formula (1), the maximum size of the convolution kernel is k = 7. Then the convolution kernel size in the preferred convolution layer 1 is k = 1, 3, 5, 7.

Ⅱ. Use the convolution kernel optimization algorithm to select the best convolution kernel among A convolution kernels. The specific process is as follows:

Calculate the outputs of A convolution kernels, use a similarity algorithm to obtain the two convolution kernels with the highest similarity, select the convolution kernel with a larger size among the two convolution kernels as the optimal convolution kernel of the preferred convolution layer 1, and obtain the preferred convolution layer 1 with the optimal convolution kernel;

The present invention calculates convolution kernels of different sizes. The output of the convolution kernels is obtained by using a similarity algorithm to find the two convolution kernels with the best similarity. The convolution kernel with the larger size among the two convolution kernels is selected as the convolution kernel (optimal convolution kernel) that can best represent the characteristics of the underwater acoustic signal in the preferred convolution layer 1, because the convolution kernel with the larger size generates fewer data parameters after convolution. The same is true for each subsequent OC1d-Layer. The similarity algorithm of the present invention adopts cosine similarity, Hamming distance, and Spearman correlation coefficient.

III. Input feature b output by S2242 into the preferred convolution layer 1 with the optimal convolution kernel to obtain feature c.

S2244, input the feature c output by S2243 into the preferred convolutional layer 2, and output feature d. The specific process is as follows:

The number of convolution kernels, the maximum size of the convolution kernel, and the optimal convolution kernel of the preferred convolution layer 2 are the same as those of the preferred convolution layer 1. The feature c obtained in III is input into the preferred convolution layer 2, and the feature d is output.

S2245, input the feature c output by S2243 and the feature d output by S2244 into the preferred convolution layer 3, and output the fused feature e. The specific process is:

The number of convolution kernels, the maximum size of the convolution kernel, and the optimal convolution kernel of the preferred convolution layer three are the same as those of the preferred convolution layer one. The feature c obtained in III and the feature d output by S2244 are input into the preferred convolution layer three, and the feature e is output.

S2246, aggregate feature c output by S2243, feature d output by S2244, and feature e output by S2245 to obtain aggregated features, process the aggregated features using Sigmoid, and output feature f with multiple characteristics;

S2247. Add the feature f output by S2246 and the feature δ output by S223 to obtain the feature ξ.

S225, multiply the feature ξ output by S224 by the feature 1 output by S21 to obtain feature 2. The specific process is:

Multiply the feature ξ output by S2246 by the feature one output by S21 to obtain feature two.

S23, input feature 2 output by S22 into convolution layer 2 for convolution and activation processing, and output feature 3;

S24, input feature 3 output by S23 into the attention mechanism layer, and output feature 4 with different weights. The specific process is as follows:

Feature three output by S23 is input into the attention mechanism layer. Through attention statistical pooling and batch normalization in the attention mechanism layer, the mean and standard deviation of the final frame-level feature (feature three) are calculated to obtain feature four with different weights.

S25, input feature 4 output by S24 into the fully connected layer for full connection and batch normalization, and output feature 5 with a score;

S26, input feature five outputted from S25 into the classification layer, and output the labeled underwater acoustic signal samples.

The feature 5 with the score output by S25 is input into the classification layer, and the feature 5 with the score is classified and labeled using the classifier function (Softmax) according to the attributes of the underwater acoustic signal sample. The attributes of the underwater acoustic signal sample include whether the acoustic signal in the underwater acoustic signal sample belongs to a normal signal or a noise signal. That is, the labeled information is the attribute of the underwater acoustic signal sample.

The present invention utilizes the MOC model to extract feature embedding of the acoustic features of the underwater acoustic signal, utilizes the method of optimizing convolution, attention mechanism and multi-layer feature fusion to extract comprehensive and representative features from the underwater acoustic signal, and finally uses a classifier to classify and label the underwater acoustic signal to obtain whether the acoustic signal in the underwater acoustic signal sample belongs to a normal signal or a noise signal.

S3. Execute S1-S2 on the underwater acoustic signal to be labeled to obtain a labeled underwater acoustic signal sample.

Claims (1)

1. The automatic underwater sound signal sample labeling method based on the MOC model is characterized by comprising the following steps of: it comprises the following steps:

S1, collecting an underwater sound signal as a sample, and calculating the acoustic characteristics of the underwater sound signal sample by using an acoustic model;

S2, establishing a MOC model, wherein the MOC model sequentially comprises a first convolution layer, a preferred convolution residual layer, a second convolution layer, an attention mechanism layer, a full connection layer and a classification layer, inputting acoustic features of the underwater sound signal samples into the MOC model for training, and outputting marked underwater sound signal samples until loss converges to obtain a trained MOC model;

S3, executing the S1-S2 on the underwater sound signal sample to be marked to obtain a marked underwater sound signal sample;

the acoustic model in the S1 comprises a Gaussian mixture model and a hidden Markov model;

establishing a MOC model in the step S2, wherein the MOC model sequentially comprises a first convolution layer, a preferred convolution residual layer, a second convolution layer, an attention mechanism layer, a full connection layer and a classification layer, acoustic features of the underwater sound signal samples are input into the MOC model for training, the marked underwater sound signal samples are output until loss converges, and the trained MOC model is obtained, wherein the specific process is as follows:

S21, inputting the acoustic characteristics of the underwater sound signal sample into a first convolution layer of an MOC model, and outputting a first characteristic;

S22, inputting the first feature output in the S21 into a preferable convolution residual layer, and outputting a second feature;

S23, inputting the second characteristic output by the S22 into the second convolution layer, and outputting a third characteristic;

s24, inputting the third feature output in the S23 into the attention mechanism layer, and outputting the fourth feature;

S25, inputting the feature IV output in the S24 into the full-connection layer, and outputting a feature V;

s26, inputting the characteristics output in the S25 into the classification layer, and outputting marked underwater sound signal samples;

the preferred convolution residual layer in the step S22 sequentially comprises a first convolution layer, a residual layer, a second convolution layer and an attention mechanism preferred convolution layer;

In the step S22, the first feature output in the step S21 is input into a preferred convolution residual layer, and the second feature is output, which specifically includes:

S221, inputting the feature I output in the S21 into a first convolution layer, and outputting a feature beta;

s222, inputting the characteristic beta output by the S221 into a residual layer, and outputting a characteristic gamma;

S223, inputting the characteristic gamma output by the S222 into a second convolution layer, and outputting the characteristic delta;

S224, inputting the feature delta output in the S223 into a preferably convolution layer of the attention mechanism, and outputting a feature xi;

S225, multiplying the characteristic xi output by the S224 with the characteristic I output by the S21 to obtain a characteristic II;

the attention mechanism preferred convolution layer in S224 includes, in order, a global average pooling layer (GAP), a one-dimensional convolution layer, a preferred convolution layer one, a preferred convolution layer two, and a preferred convolution layer three;

In the step S224, the feature δ output in the step S223 is input into a preferably convolution layer, and the specific process is as follows:

S2241, inputting the feature delta output in the S223 into the global average pooling layer, and outputting a feature a;

s2242, inputting the feature a output in the S2241 into the one-dimensional convolution layer, and outputting the feature b;

s2243, inputting the characteristic b output in the S2242 into the first preferable convolution layer and outputting the characteristic c;

S2244, inputting the characteristic c output in the S2243 into the second preferable convolution layer, and outputting the characteristic d;

s2245, inputting the feature c output by S2243 and the feature d output by S2244 into the preferable convolution layer III, and outputting the feature e;

S2246, aggregating the feature c output by S2243, the feature d output by S2244 and the feature e output by S2245 to obtain an aggregated feature, processing the aggregated feature by using Sigmoid, and outputting a feature f;

S2247, adding the characteristic f output by S2246 and the characteristic delta output by S223 to obtain a characteristic xi;

In the step S2243, the feature b output in the step S2242 is input into the preferred convolution layer one, and the feature c is output, which specifically includes:

setting A convolution kernels with different sizes in a preferred convolution layer I, and setting the maximum size of the convolution kernels in the preferred convolution layer I according to the dimension of the acoustic characteristic of the underwater sound signal sample in S1:

Wherein C represents a dimension of an acoustic feature of the hydroacoustic signal sample, the acoustic feature being a mel spectrum;

K represents the convolution kernel size, K is an odd number and is an integer;

II, selecting an optimal convolution kernel in the A convolution kernels by utilizing a convolution kernel optimization algorithm, wherein the specific process is as follows:

Calculating the output of the A convolution kernels, obtaining two convolution kernels with highest similarity by using a similarity algorithm, and selecting the convolution kernel with large size in the two convolution kernels as an optimal convolution kernel of the optimal convolution layer I to obtain the optimal convolution layer I with the optimal convolution kernel;

III, inputting the characteristic b output by the S2242 into a preferable convolution layer I with an optimal convolution kernel to obtain a characteristic c;

in the step S2244, the feature c output in the step S2243 is input into the second preferred convolution layer, and the feature d is output, which specifically includes:

The number of convolution kernels, the maximum size of convolution kernels and the optimal convolution kernels of the second preferred convolution layer are the same as those of the first preferred convolution layer, and the characteristic c obtained in the III is input into the second preferred convolution layer to output the characteristic d;

In the step S2245, the feature c output in the step S2243 and the feature d output in the step S2244 are input into the preferred convolution layer three, and the feature e is output, which specifically includes the steps of:

and the number of convolution kernels, the maximum size of the convolution kernels and the optimal convolution kernels of the preferred convolution layer III are the same as those of the preferred convolution layer I, the maximum size of the convolution kernels and the optimal convolution kernels, the characteristics c obtained in III and the characteristics d output by S2244 are input into the preferred convolution layer III, and the characteristics e are output.