Multi-view hrrp radar air target recognition method based on dynamic network structure

Abstract

The application discloses a multi-view HRRP radar air target recognition method based on a dynamic neural network, first constructs a dynamic network DSPNN, including: a feature extraction module, a feature fusion recognition module, then generates multi-view HRRP data, inputs the multi-view HRRP data into the feature extraction module of the network for feature extraction, inputs the extracted features into the feature fusion recognition module for multi-view dynamic fusion, multi-view features are weighted and fused into final features, finally, multi-view features are compressed and activated to obtain a classification result, and target recognition is completed. The method of the application uses a multi-view recognition method based on a dynamic structure and dynamic parameters, effectively improves the adaptive ability of a high-resolution range profile recognition network of an air target, solves problems existing in a single-view static network, and further improves the recognition ability of the air target.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a multi-view HRRP radar air target recognition method based on a dynamic network structure. BACKGROUND

[0002] With the development of deep neural networks, more and more scholars have applied them to the field of high-resolution range profile target recognition, but most of the current researches of the scholars are based on high-resolution range profile target recognition under a single-view static network. Compared with a multi-view dynamic neural network, the single-view static network mainly has the following problems:

[0003] Firstly, the single-view network model has only one view of feature input. The recognition performance of a deep network is often directly linked to the number of input features, and the number of input features can be directly increased by increasing the number of input views. At the same time, in addition to the features of each view, the multi-view input also implies the features between different views, which can also improve the recognition performance of the network. Secondly, the static network model not only does not have the dynamic adaptive ability, but also causes waste of resources. Some of the current static network models can achieve good performance due to having enough parameters to fully express all targets. Considering the differences between sample targets and different views of the same sample target, it is difficult to extract high-quality features of all samples by using a unified model structure. At the same time, in the current complex recognition environment and target diversification background, if all recognition targets are considered, the parameters required by the static network will be more and more. SUMMARY

[0004] To solve the above technical problems, the application provides a multi-view HRRP (high-resolution range profile) radar air target recognition method based on a dynamic network structure, which combines dynamic structure and dynamic parameters to improve the adaptive ability of the network, and has better recognition performance in the high-resolution range profile recognition task of air targets compared with a static single-view network.

[0005] The technical scheme adopted by the application is as follows: A multi-view HRRP radar air target recognition method based on a dynamic network structure, and the specific steps are as follows:

[0006] S1, constructing a dynamic network DSPNN, including a feature extraction module and a feature fusion recognition module;

[0007] S2, generating multi-view HRRP data and inputting the data to the feature extraction module of the network for feature extraction;

[0008] S3, based on step S2, input the extracted features into a feature fusion recognition module for processing, first adjust the weights of the features of each view angle through an adaptive weight gate, weight and fuse the multi-view angle features into final features, then perform feature compression and activation on the multi-view angle features to obtain a classification result, and complete target recognition.

[0009] Further, the step S1 is specifically as follows:

[0010] S11, a feature extraction module;

[0011] The feature extraction module comprises 6 gated skip layer modules and 1 dynamic convolution kernel module, is used for dynamically extracting features of multi-view angle HRRP data, takes a residual network as a basic feature extraction unit, and realizes dynamic adjustment of the network in structure and parameters through a gated skip layer mechanism and a dynamic convolution kernel respectively.

[0012] The gated skip layer module comprises 1 global average pooling layer, 2 fully connected layers, 1 long short-term memory network (LSTM), 1 batch normalization (BN) layer, 1 Sigmoid activation function and 1 residual block, and is used for judging whether the current residual block is executed according to the high-resolution range profile features extracted by the previous layer.

[0013] The residual block comprises 2 convolution layers, 1 random dropout (Dropout) layer and 2 ReLu activation functions.

[0014] The output generation and transmission of the LSTM module in the working process are completed through a forgetting gate, an input gate, a cell state update and an output gate.

[0015] The forgetting gate is used for determining the reservation degree of the output information of the LSTM module in the previous skip layer gating mechanism, the input gate is used for determining the reservation degree of the input information of the LSTM module in the current skip layer gating structure, the cell state update is used for ensuring the sharing of parameters between different LSTM modules, and the output gate is used for outputting the output value of the LSTM in the current skip layer gating structure, which is also sent as a parameter to the LSTM module of the next skip layer gating module.

[0016] The dynamic convolution kernel module comprises 1 attention mechanism module and l convolution kernels, the dynamic convolution kernel module sets a convolution kernel group, allows the network to dynamically allocate weights to the convolution kernel members of the convolution kernel group according to the high-resolution range profile features of different samples, and then obtains the convolution kernel for the sample according to the weighted sum of the weights.

[0017] The attention mechanism module comprises 1 average pooling layer, 2 fully connected layers, 1 ReLu activation function and a Softmax function.

[0018] S12, feature fusion recognition module;

[0019] The feature fusion recognition module comprises a dynamic feature fusion module and a recognition module; the dynamic feature fusion module comprises an adaptive weight module, each view feature is first adjusted by the adaptive weight module to have a respective weight; then, the multi-view features are combined by splicing and sent to the recognition module; finally, the classification result is obtained by feature compression and activation.

[0020] The adaptive weight module comprises two feature extraction blocks, one average pooling layer and one full connection layer, and is used for realizing weight calculation of different view feature maps.

[0021] The feature extraction block comprises one convolution layer, one BN layer and one Sigmoid activation function.

[0022] The recognition module comprises one convolution layer, one average pooling layer, one full connection layer and one Softmax function, and is used for realizing classification and recognition.

[0023] Further, the step S2 is specifically as follows:

[0024] azimuth angles in a specific angle range (67° to 112°) are selected, the target has an HRRP data under each angle when rotating every 1°, short-time Fourier transform is performed on each HRRP data to obtain time-frequency graphs of the HRRP data under different views, the time-frequency graphs of the views are arranged and combined together every m to obtain an HRRP time-frequency graph of m views, and then the generated multi-view HRRP data is input to the feature extraction module of the network.

[0025] After the output feature of the previous layer gate skip layer module is input into the next layer gate skip layer module, the global information of the output feature of the previous gate skip layer module is first extracted by the global average pooling layer and the full connection layer; then, the output feature is sent to the LSTM module with shared parameters, the LSTM module outputs the current feature result by synthesizing the information of the front gate function and continues to be sent to the full connection layer and the BN layer; finally, the output result in the interval (0, 1) is obtained by the Sigmoid activation function, and the output result is binarized as the output of the gate skip layer function, and the mathematical expression is as follows:

[0026] x i =S(i)F(x i-1 )+(1-S(i))x i-1 (1)

[0027] Wherein, i represents the number of the current residual block; x i-1 represents the input feature of the current layer, x i represents the output of the current layer; S(i) represents the output of the gate skip layer function; F(xi-1 ) represents the output of the processed residual block.

[0028] When the output of the gating skip function is 1, the input is sent to the residual block adjusted by the dynamic convolution kernel module for processing, and the output result is F(x i-1 ); when the output of the gating skip function is 0, the input feature will directly skip the current residual block, and the output result is x i-1 . When i is not equal to k, it means that there is still a residual block, and the output is sent to the next layer gating skip module; if it is equal to k, the output is sent to the feature fusion recognition module for weight adjustment and fusion with the features of other views.

[0029] where k = 6.

[0030] The implementation of the dynamic weight of the dynamic convolution kernel module is as follows:

[0031] First, the global information of the input feature is extracted by global average pooling; then, the nonlinear features of the input are learned by two fully connected and ReLU activation functions; finally, the feature weights of the corresponding convolution kernel are output by the Softmax function. The process expression of dynamic convolution is as follows:

[0032]

[0033] where z represents the input feature, y represents the output feature, C represents the dynamic convolution operation; W(z) represents the convolution kernel parameter generated by the sample self-adaption; L represents the number of generated weights; A l (z) represents the lth attention weight; W l represents the lth convolution kernel parameter.

[0034] Further, the step S3 is specifically as follows:

[0035] First, the input feature passes through two feature extraction blocks, an average pooling layer, and a fully connected network layer to obtain the weight of the view. Then, the multi-view is weighted and spliced according to the weight; then, the feature compression is performed through the convolution layer and the average pooling layer; finally, the feature is mapped to the target category through the fully connected network, and the Softmax function is activated to obtain the probability of different categories and then complete the recognition.

[0036] The method of the present application first constructs a dynamic network DSPNN, including: a feature extraction module, a feature fusion recognition module, then generates multi-view HRRP data, inputs the feature extraction module of the network for feature extraction, inputs the extracted features to the feature fusion recognition module for processing, first adjusts the weight of each view feature through the adaptive weight gate, weights and fuses the multi-view features into the final feature, finally, the multi-view features are compressed and activated to obtain the classification result, and the target recognition is completed. The method of the application uses a multi-view recognition method based on dynamic structure and dynamic parameters, effectively improves the adaptive ability of the high-resolution range profile recognition network of the air target, solves the problems existing in the single-view static network, and further improves the recognition ability of the air target. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the multi-view HRRP radar air target recognition method based on the dynamic network structure of the present application.

[0038] Figure 2 The structure diagram of the gating skip layer module in the embodiment of the present application.

[0039] Figure 3 The structure diagram of the dynamic convolution kernel module in the embodiment of the present application.

[0040] Figure 4 The structure diagram of the feature fusion recognition module in the embodiment of the present application.

[0041] Figure 5 The recognition effect comparison diagram of different networks under different signal-to-noise ratios in the embodiment of the present application. DETAILED DESCRIPTION

[0042] The present application mainly adopts the method of simulation experiment for verification, all steps and conclusions are verified correct on the windows 10 operating system platform using PyTorch framework and Stochastic Gradient Descent (SGD) optimizer. The method of the present application will be further described below in combination with the drawings and embodiments.

[0043] As shown in Figure 1 The flowchart of the multi-view HRRP radar air target recognition method based on the dynamic network structure of the present application, the specific steps are as follows:

[0044] S1, a dynamic network (Deep Structure Parameter Neural Network, DSPNN) is constructed, including: a feature extraction module, a feature fusion recognition module;

[0045] S2, generate multi-view HRRP data, input to the feature extraction module of the network for feature extraction;

[0046] S3, based on step S2, input the extracted features to the feature fusion recognition module for processing, first adjust the weight of each view feature through the adaptive weight door, weight and fuse the multi-view features into the final feature, then perform feature compression and activation on the multi-view features to obtain the classification result, and complete target recognition.

[0047] In the embodiment, the step S1 is specifically as follows:

[0048] S11, a feature extraction module;

[0049] The feature extraction module comprises 6 gated skip layer modules and 1 dynamic convolution kernel module; used for dynamically extracting features of multi-view input HRRP data, taking a residual network as a basic feature extraction unit, and realizing dynamic adjustment of network structure and parameters through gated skip layer mechanism and dynamic convolution kernel respectively.

[0050] The structure of the gated skip layer module is as shown in Figure 2 The gated skip layer module comprises 1 global average pooling layer, 2 fully connected layers, 1 long short-term memory network (LSTM), 1 batch normalization (BN) layer, 1 Sigmoid activation function and 1 residual block; used for judging whether the current residual block is executed according to the high-resolution range image features extracted by the previous layer.

[0051] The residual block comprises 2 convolution layers, 1 random dropout (Dropout) layer and 2 ReLu activation functions.

[0052] In order to enhance the strategy of skip layer, an LSTM module is added in the network. The output generation and transmission of the LSTM module in the working process are completed through the forgetting gate, input gate, cell state update and output gate.

[0053] The forgetting gate is used to determine the retention degree of the output information of the LSTM module in the previous skip layer gating mechanism; the input gate is used to determine the retention degree of the input information of the LSTM module in the current skip layer gating structure; the cell state update is used to ensure the sharing of parameters between different LSTM modules; and the output gate is used to output the output value of the LSTM in the current skip layer gating structure, which is also sent to the LSTM module of the next layer gating skip layer module as a parameter.

[0054] The structure of the dynamic convolution kernel module is as shown in Figure 3As shown, the dynamic convolution kernel module includes 1 attention mechanism module and l convolution kernels; the dynamic convolution kernel module sets a convolution kernel group, and enables the network to dynamically assign weights to the convolution kernel members of the convolution kernel group according to the high-resolution range image features of different samples, and then obtains the convolution kernel for the sample through weighted summation according to the weights.

[0055] The attention mechanism module includes 1 average pooling layer, 2 fully connected layers, 1 ReLu activation function and one Softmax function.

[0056] S12, a feature fusion recognition module;

[0057] The feature fusion recognition module can dynamically evaluate the importance of the output features and give weights according to the importance, and complete fusion and recognition on this basis. The structure of the feature fusion recognition module is as shown in Figure 4 As shown, the feature fusion recognition module includes 1 dynamic feature fusion module and 1 recognition module; the dynamic feature fusion module includes an adaptive weight module, and each view feature first adjusts its own weight through the adaptive weight module; then, after being combined into multi-view features, the multi-view features are sent to the recognition module; finally, the classification result is obtained through feature compression and activation.

[0058] The adaptive weight module includes 2 feature extraction blocks, 1 average pooling layer and 1 fully connected layer, and is used to calculate the weights of different view feature maps. Figure 4 (a) is a feature fusion recognition module, and w1, w2, w m respectively represent the weights of the corresponding input features, Figure 4 (b) is an adaptive weight module, Figure 4 (c) is a recognition module.

[0059] The adaptive weight module includes 2 feature extraction blocks, 1 average pooling layer and 1 fully connected layer, and is used to calculate the weights of different view feature maps.

[0060] The feature extraction block includes 1 convolution layer, 1 BN layer and 1 Sigmoid activation function.

[0061] The recognition module includes 1 convolution layer, 1 average pooling layer, 1 fully connected layer and 1 Softmax function, and is used to realize classification and recognition.

[0062] In this embodiment, the step S2 is specifically as follows:

[0063] Select the azimuth angle in a certain angle range (67° to 112°), rotate every 1°, and there is an HRRP data for the target at each angle. In order to improve the recognition effect, a short-time Fourier transform is performed on each HRRP data to obtain a time-frequency graph of the HRRP data at different viewing angles. Then, the time-frequency graphs at the m viewing angles are arranged and combined every m to obtain an HRRP time-frequency graph at m viewing angles. Then, the generated multi-view HRRP data is input into the feature extraction module of the network.

[0064] In the embodiment, m is set to 3.

[0065] After the output feature of the previous layer of the gated skip connection module enters the next layer of the gated skip connection module, the global information of the output feature of the previous gated skip connection module is first extracted through the global average pooling layer and the full connection layer; then the global information is sent to the LSTM module with shared parameters, the LSTM module outputs the current feature result by synthesizing the information of the front gating function and continues to be sent to the full connection layer, the BN layer; finally, the output result in the interval (0, 1) is obtained by sending to the Sigmoid activation function, and the output result is binarized as the output of the gated skip connection function, and the mathematical expression is as follows:

[0066] x i =S(i)F(x i-1 )+(1-S(i))x i-1 (1)

[0067] Wherein, i represents the number of the current residual block; x i-1 represents the input feature of the current layer, x i represents the output of the current layer; S(i) represents the output of the gated skip connection function; F(x i-1 ) represents the output processed by the residual block.

[0068] When the output of the gated skip connection function is 1, the input is sent to the residual block adjusted by the dynamic convolution kernel module for processing, and the output result is F(x i-1 ); when the output of the gated skip connection function is 0, the input feature will directly skip the current residual block, and the output result is x i-1 . When i is not equal to k, it means that there is still a residual block, and the output is sent to the next layer of the gated skip connection module; if it is equal to k, the output is sent to the feature fusion recognition module for weight adjustment and fusion with the features of other viewing angles.

[0069] Wherein, k = 6.

[0070] The dynamic weight distribution of the dynamic convolution kernel module is as follows:

[0071] First, the global information of the input feature is extracted by global average pooling; then, the nonlinear features of the input are learned by two fully connected and ReLU activation functions; finally, the feature weights of the corresponding convolution kernel are output by the Softmax function. The process expression of dynamic convolution is as follows:

[0072]

[0073] Wherein, z represents the input feature, y represents the output feature, C represents the dynamic convolution operation; W(z) represents the convolution kernel parameter generated by the sample self-adaption; L represents the number of generated weights; A l (z) represents the lth attention weight; W l represents the lth convolution kernel parameter.

[0074] In the embodiment, the step S3 is specifically as follows:

[0075] First, the input feature passes through two feature extraction blocks, an average pooling layer and a fully connected network layer to obtain the weight of the view angle. Then, the multi-view angles are weighted and spliced according to the weight; then, the feature compression is performed through a convolution layer and an average pooling layer; finally, the feature is mapped to the target category through a fully connected network, and the different categories are activated by a Softmax function to obtain the probability of different categories and then the recognition is completed.

[0076] In the embodiment, after the target recognition is completed, the performance of the method of the application is evaluated, which is specifically as follows:

[0077] The performance evaluation indexes include the accuracy and the precision, and the total number of samples is N. The mathematical expression for calculating the accuracy is as follows:

[0078]

[0079] Given a certain category j, the precision of the jth category represents the proportion of the samples actually belonging to the jth category (i.e. the correctly predicted samples) in the samples predicted by the model to belong to the jth category. The mathematical expression for calculating the precision is as follows:

[0080]

[0081] Wherein, n +j represents the total number of samples predicted to belong to the jth category; n jj represents the actual total number of the jth samples; Rec represents the proportion of the correctly predicted samples in all the samples predicted to belong to the jth category; and Num represents the number of sample categories.

[0082] In order to fully verify the identification performance of the model, the following aspects are analyzed respectively. In the experiment, the network is compared with Alex network (AlexNet), VGG16, residual network (ResNet), nested residual network (M-ResNet), efficient network V2 (EfficientNetV2), spatial convolutional neural network (SCNN) and dynamic structure neural network (DSNN), and the identification results of the proposed method under different signal-to-noise ratios are as shown in Figure 5 It can be seen from the table that the multi-view network recognition effect is better than that of the single-view network, and the network proposed in the application has better adaptive ability, can better extract sample features, and thus has better recognition performance.

[0083] In order to verify the contribution of the dynamic convolution kernel module, the gating skip layer module and the adaptive weight module in the network proposed in the method of the application to the overall identification, the ablation experiment is carried out under the condition that the signal-to-noise ratio is 10dB, and the experimental results are as shown in Table 1.

[0084] Table 1

[0085] Dynamic convolution kernel module Gated skip layer module Adaptive weight module Recognition accuracy 0 0 0 85.7％ 0 0 1 95.4％ 0 1 0 94.2％ 1 0 0 93.4％ 1 1 0 97.3％ 1 0 1 95.8％ 0 1 1 96.5％ 1 1 1 98.2％

[0086] From Table 1, it can be seen that based on the recognition rate of 85.7% without adding the module, the improvement of the adaptive weight module, the gating skip layer module and the dynamic convolution kernel module is 9.7%, 8.5% and 7.7% respectively, and the model with all the three modules has the best performance.

[0087] In order to further verify the superiority of the network proposed in the method of the application, the overall floating point operation amount of the model under the condition of 10dB signal-to-noise ratio is evaluated, and the ablation experiment of FLOPs is carried out for different modules involved, and the experimental results are as shown in Table 2.

[0088] Table 2

[0089]

[0090] The results in Table 2 show that the network with the gating skip layer module can adaptively adjust its structure according to the input sample, and the dynamic convolution kernel can further improve the adaptive ability of the model and the identification efficiency of the model on the basis of the gating skip layer module, and the proposed network has excellent operation efficiency.

[0091] To sum up, the method of the application uses a multi-view recognition method based on dynamic structure and dynamic parameters, a feature extraction module is responsible for dynamic feature extraction on the multi-view input, and a feature fusion recognition module is responsible for dynamic fusion and recognition of the features extracted from different views, effectively improving the adaptive ability of the high-resolution range image recognition network of the aerial target, solving the problems existing in the single-view static network, and further improving the recognition ability of the aerial target.

[0092] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the principles of the application and should not be construed as limiting the scope of the application to such specifically enumerated embodiments and examples. Various other specific adaptations and combinations of features that are not explicitly described can be made by those skilled in the art based on the teachings of the present application, without departing from the spirit and scope of the application, which should be determined by the appended claims.

Claims

1. A multi-view HRRP radar air target identification method based on a dynamic network structure, the specific steps of which are as follows: S1. Construct the dynamic network DSPNN, including: Feature extraction module, feature fusion and recognition module; The feature extraction module includes: 6 gated skip-layer modules and 1 dynamic convolutional kernel module; it is used to perform dynamic feature extraction on HRRP data with multi-view input, using the residual network as the basic feature extraction unit, and dynamically adjusting the network structure and parameters through the gated skip-layer module and the dynamic convolutional kernel respectively; The gated skip layer module includes: one global average pooling layer, two fully connected layers, one Long Short-Term Memory (LSTM) network, one batch normalized (BN) layer, and one... An activation function and one residual block are used to determine whether the current residual block should be executed based on the high-resolution range image features extracted from the previous layer. The feature fusion and recognition module includes: one dynamic feature fusion module and one recognition module; the dynamic feature fusion module includes an adaptive weight module, whereby the weights of each view feature are first adjusted by the adaptive weight module; then, they are combined into multi-view features by splicing and sent to the recognition module; finally, the classification result is obtained by feature compression activation. S2. Generate multi-view HRRP data and input it into the feature extraction module of the network for feature extraction; S3. Based on step S2, the extracted features are input into the feature fusion and recognition module for processing. First, the weights of the features from each perspective are adjusted by the adaptive weight module, and the multi-view features are weighted and fused into the final features. Then, the multi-view features are compressed and activated to obtain the classification results, thus completing the target recognition.

2. The multi-view HRRP radar air target identification method based on a dynamic network structure according to claim 1, characterized in that: The residual block includes: two convolutional layers, one dropout layer, and two ReLU activation functions; The LSTM generates and transmits its output during operation through a relay of forget gate, input gate, cell state update, and output gate. Among them, the forget gate is used to determine the degree to which the output information of the LSTM in the previous gated skip module is retained; the input gate is used to determine the degree to which the input information of the LSTM in the current gated skip module is retained; the cell state update is used to ensure the sharing of parameters between different LSTMs; and the output gate is used to output the output value of the LSTM in the current gated skip module, which is also sent as a parameter to the LSTM of the next gated skip module. The dynamic convolution kernel module includes: one attention mechanism module and one convolution kernel; the dynamic convolution kernel module sets up a convolution kernel group, and the network dynamically assigns weights to the convolution kernel members of the convolution kernel group according to the high-resolution distance image features of different samples, and then obtains the convolution kernel for the sample by weighted summation according to the weights; The attention mechanism module includes: one average pooling layer, two fully connected layers, one ReLU activation function, and one Softmax function; The adaptive weighting module includes: two feature extraction blocks, one average pooling layer, and one fully connected layer; used to calculate the weights of feature maps from different perspectives. The feature extraction block includes: one convolutional layer, one batch normalization (BN) layer, and one... Activation function; The recognition module includes: one convolutional layer, one average pooling layer, one fully connected layer, and one... Functions used to implement classification and recognition.

3. The multi-view HRRP radar air target identification method based on a dynamic network structure according to claim 1, characterized in that, Step S2 is as follows: A specific azimuth angle range of 67° to 112° is selected, and the target is rotated every 1°. HRRP data exists for each angle. A short-time Fourier transform is performed on each HRRP data point to obtain the time-frequency map of the HRRP data under different viewing angles. Then, these time-frequency maps from various viewing angles are sorted according to... Combining these permutations and combinations yields... The HRRP time-frequency plots from different perspectives are then generated and input into the feature extraction module of the network. After the output features of the previous gated jump layer module enter the next gated jump layer module, they first pass through a global average pooling layer and a fully connected layer to extract global information of the output features of the previous gated jump layer module; Then it is fed into an LSTM module with shared parameters. The LSTM module synthesizes the information from the pre-gating function, outputs the current feature result, and continues to feed it into the fully connected layer and the Batch Normalization (BN) layer; finally, it is fed into... The activation function produces an output in the (0,1) interval. This output is binarized and used as the output of the gated skip function. Its mathematical expression is as follows: ； in, Indicates the number of the current residual block; Indicates the input features of the current layer. Indicates the output of the current layer; This represents the output of the gated skip function; This indicates the output after processing by the residual block; When the output of the gated skip function is 1, the input is fed into the residual block after the parameters of the dynamic convolution kernel module are adjusted for processing, and the output result is... When the output of the gated skip function is 0, the input features will directly skip the current residual block, and the output result will be... ;when Not equal to If the value is equal to 1, it indicates that a residual block still exists, and the output is sent to the next gated jump module; if it is equal to 1, it indicates that a residual block still exists. The output will then be sent to the feature fusion and recognition module for weight adjustment and fusion with features from other perspectives; in, =6; The implementation of the dynamic weights in the dynamic convolution kernel module is as follows: First, global information about the input features is extracted using global average pooling; then, through two fully connected layers and... The activation function learns the non-linear features of the input; finally, through... The function outputs the feature weights corresponding to the convolution kernel; the expression for the dynamic convolution process is as follows: ； in, Indicates input features, Indicates output features, Indicates dynamic convolution operation; This represents the parameters of the convolution kernel generated adaptively for the samples; Indicates the number of weights generated; Indicates the first Each attention weight; Indicates the first Each convolution kernel parameter.

4. The multi-view HRRP radar air target identification method based on a dynamic network structure according to claim 1, characterized in that, Step S3 is as follows: First, the input features are processed through two feature extraction blocks, an average pooling layer, and a fully connected network layer to obtain the weights for each viewpoint. Then, the multiple viewpoints are weighted and concatenated according to their weights. Next, feature compression is performed through convolutional layers and average pooling layers. Finally, a fully connected network is used to establish a mapping between the features and the target category, and... The function is activated to obtain the probabilities of different categories, thereby completing the recognition.

Images