A method for identifying true and false radar targets based on a multi-mode deep network

Abstract

The present invention belongs to the technical field of radar true and false target recognition, and specifically relates to a radar true and false target recognition method based on a multi-modal deep network. First, the invention preprocesses the one-dimensional range profile data of radar true and false targets as the input of the multi-modal deep network. The multi-modal deep network consists of 3 parallel one-dimensional convolutional sub-networks to extract multi-modal features from the one-dimensional range profile. Then, a stacked autoencoder with 3 layers is cascaded to further extract recognition features from the multi-modal features. Finally, a softmax classification layer is used to complete the target type recognition. Since multiple convolutional sub-networks with different convolutional kernel sizes are used for preliminary feature extraction, and then a stacked autoencoder is used to further extract recognition features, more refined and comprehensive non-linear features can be extracted, improving the target recognition performance. Simulation experiments are carried out on the one-dimensional range profile data of four types of simulated targets, and the experimental results verify the effectiveness of the method.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar true and false target recognition, and particularly relates to a radar true and false target recognition method based on a multi-modal deep network. Background Art

[0002] With the development of broadband radars, it is easy to obtain the one-dimensional range profile of a target. The one-dimensional range profile contains target structure shape information that is beneficial for recognition and can achieve real-time recognition. Therefore, the one-dimensional range profile can be used to recognize radar true and false targets.

[0003] Currently, the recognition of the one-dimensional range profile of radar true and false targets mainly uses conventional shallow machine learning methods for classification and recognition. It is necessary to preset recognition features manually. However, for the one-dimensional range profile of radar true and false targets, it is very sensitive to target attitude changes, resulting in the difficulty of artificially setting its recognition features. In recent years, recognition methods based on deep learning have been gradually introduced into the field of radar target recognition. By using a deep learning model to automatically learn high-order non-linear features beneficial for recognition from one-dimensional range profile data, good results have been achieved. Therefore, researching a method for recognizing the one-dimensional range profile of true and false targets based on a deep network model is expected to further improve the target recognition rate. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for recognizing the one-dimensional range profile of radar true and false targets based on a multi-modal deep network. First, a one-dimensional convolutional sub-network with multiple model parameters is used to extract target features from the one-dimensional range profile. Then, a concatenated stack auto-encoder is used to further extract recognition features, so that more refined and comprehensive non-linear features can be extracted, and the recognition performance of the target can be improved.

[0005] The technical solution of the present invention is as follows:

[0006] A radar true and false target recognition method based on a multi-modal deep network, comprising the following steps:

[0007] S1. Define the j-th training one-dimensional range profile of the i-th type of true and false target as an n-dimensional column vector x ij , 1 ≤ i ≤ g, 1 ≤ j ≤ N i , where N i is the number of training one-dimensional range profile samples of the i-th type of true and false target, and N is the total number of training one-dimensional range profile samples; perform the following preprocessing on the one-dimensional range profile x ij :

[0008]

[0009] where ||·|| represents the norm of a vector;

[0010] S2. Construct a multi-modal deep network model, including a multi-modal sub-network, a stacked autoencoder, and a classification layer. Among them, the multi-modal sub-network consists of three parallel one-dimensional convolutional sub-networks with different model parameters. The input of the multi-modal sub-network is After feature extraction by three parallel one-dimensional convolutional sub-networks, the three obtained features are concatenated and output as x (1) ; The stacked autoencoder consists of three cascaded autoencoders. The input of the stacked autoencoder is x (1) , and the output of the stacked autoencoder is x (2) ; x (2) is input into the classification layer for classification output;

[0011] S3. Use the obtained in S1 to train the multi-modal deep network model constructed in S2. Specifically, use the BP method to train the model parameters of the entire deep network. The loss function is the mean squared error function, and the optimization method is the steepest gradient descent method. The optimal number of iterations and learning rate are determined by experiments;

[0012] S4. Input the one-dimensional range images of the obtained true and false targets into the trained multi-modal deep network model, and use the label corresponding to the maximum component in the output vector of the softmax classification layer as the target recognition category.

[0013] The beneficial effects of the present invention are as follows: First, the present invention preprocesses the one-dimensional range image data of radar true and false targets as the input of the multi-modal deep network. The multi-modal deep network consists of three parallel one-dimensional convolutional sub-networks to extract multi-modal features from the one-dimensional range image. Then, a three-layer stacked autoencoder is cascaded to further extract recognition features from the multi-modal features. Finally, a softmax classification layer is used to complete target type recognition. Since multiple convolutional sub-networks with different kernel sizes are used for preliminary feature extraction, and then a stacked autoencoder is used to further extract recognition features, more refined and comprehensive non-linear features can be extracted, improving the target recognition performance. Simulation experiments are carried out on the one-dimensional range image data of four types of simulated targets, and the experimental results verify the effectiveness of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a schematic diagram of the overall process of the present invention;

[0015] Figure 2 is a structural block diagram of the multi-modal sub-network;

[0016] Figure 3 is a structural block diagram of the stacked autoencoder. DETAILED DESCRIPTION OF THE INVENTION

[0017] The following is a simulation to prove the effectiveness and progress of the present invention:

[0018] As Figure 1 shown, in the recognition process of the present invention, first, the one-dimensional range images of the true and false radar targets are preprocessed, and then input into a multi-modal subnet to extract multi-modal features. The multi-modal subnet includes 3 parallel one-dimensional convolutional subnets with different model parameters. The sizes of the convolutional kernels are 1×3, 1×5, and 1×7 respectively. Each convolutional subnet is composed of 3 cascaded convolutional pooling layers, as Figure 2 shown. Other network model parameters are determined by experiments. From Figure 2 it can be seen that is the preprocessed one-dimensional range image, which is used as the input of the 3 convolutional pooling subnets. x (1) is the output of the multi-modal subnet, which is composed of the outputs of the 3 convolutional pooling subnets and is used as the input of the stacked autoencoder. The activation function of the convolutional layer is ReLU. Then, through the stacked autoencoder, recognition features are further extracted from the multi-modal features. The stacked autoencoder is composed of 3-layer autoencoders, as Figure 3 shown. From Figure 3 it can be seen that x (1) is the input of the stacked autoencoder, and x (2) is the output of the stacked autoencoder. The activation function of each autoencoder is ReLU, and the number of nodes is determined by experiments. Finally, a softmax classification layer is used to complete the recognition of the target type

[0019] The simulation experiment designs four types of point targets: true target, debris, light decoy, and heavy decoy. The bandwidth of the radar transmitted pulse is 1000MHZ (the range resolution is 0.15m, and the radar radial sampling interval is 0.075m). The target is set as a uniformly scattered point target. The number of scattering points of the true target is 7, and the number of scattering points of the other three targets is 11. In the one-dimensional range images at intervals of 0.1° within the target attitude angle range of 0° to 90°

[0020] range, the one-dimensional range images selected by taking every other one are used for training, and the one-dimensional range images at other attitude angles are used as test data. Then, each type of target has 450 test samples.

[0021] For the four types of targets (true target, debris, light decoy, and heavy decoy), within the attitude angle range of 0° to 90°, the method for recognizing true and false targets based on the multi-modal deep network in this paper is used for recognition. The average correct recognition rate of the one-dimensional range images of the 4 types of true and false targets is 97%. Among them, the length of the one-dimensional range image is 50, the number of iterations is 2500 times, the number of kernels in the convolutional layer is 6, 12, and 6 respectively, the size of the pooling layer filter is 1×2, the number of nodes in the autoencoder is 250, 200, and 100 respectively, and the learning rate is 0.1. In addition, noise is added to the samples, and the signal-to-noise ratio is 20dB.

Claims

1. A method for identifying true and false radar targets based on a multi-modal deep network, characterized in that, Including the following steps: S1. Define the j-th training one-dimensional range profile of the i-th type of true or false target obtained as an n-dimensional column vector x ij , where 1 ≤ i ≤ g and 1 ≤ j ≤ N i , where N i is the number of training one-dimensional range profile samples of the i-th type of true or false target, and N is the total number of training one-dimensional range profile samples; perform the following preprocessing on the one-dimensional range profile x ij as follows: where ||·|| represents the norm of a vector; S2. Construct a multi-modal deep network model, including a multi-modal sub-network, a stacked auto-encoder, and a classification layer; among them, the multi-modal sub-network consists of 3 parallel one-dimensional convolutional sub-networks with different model parameters, and the input of the multi-modal sub-network is After feature extraction by 3 parallel one-dimensional convolutional sub-networks, the 3 obtained features are concatenated and output as x (1) ; The stacked auto-encoder consists of 3 cascaded auto-encoders, and the input of the stacked auto-encoder is x (1) , and the output of the stacked auto-encoder is x (2) ; x (2) is input into the classification layer for classification output; S3. Using what is obtained in S1 Train the multi-modal deep network model constructed in S2. Specifically, use the BP method to train the model parameters of the entire deep network. The loss function is the mean squared error function, and the optimization method is the steepest gradient descent method. The optimal number of iterations and learning rate are determined by experiments; S4. Input the one-dimensional range profiles of the true and false targets obtained into the pre-trained multi-modal deep network model, and use the label corresponding to the maximum component in the output vector of the softmax classification layer as the target recognition category.

Images