Radar target recognition method based on two-dimensional self-selection weighted convolutional neural network

Abstract

This invention discloses a radar target recognition method based on a two-dimensional self-selected weighted convolutional neural network, comprising: first, constructing a two-dimensional data signal from the original one-dimensional time-domain signal obtained by the radar signal processing system using a Toeplitz matrix; then, generating training and test sets from the constructed two-dimensional data signal according to the categories of the original signal; constructing a coordinate attention module, and then constructing a convolutional neural network composed of self-selected weighting modules; inputting the training set data into the constructed network for training to obtain a trained network model; finally, using the test set to test the trained network model to complete the radar target recognition and classification. This invention can improve the accuracy of target recognition with very small computational parameter costs, and is suitable for use in situations where computational resources are very limited but high target recognition accuracy is required.

Description

TECHNICAL FIELD

[0001] The application relates to a radar target recognition technology, in particular to a radar target recognition method based on a two-dimensional self-selection weighted convolutional neural network. BACKGROUND

[0002] The radar target recognition technology is a technology of transmitting electromagnetic waves for target detection, analyzing the characteristics of a target by using radar echoes reflected by the target, and finally recognizing the target by using the characteristics. As an important function of a radar system, target recognition has played an important role in sea and land monitoring systems, weather observation and the like. In the modern complex military war environment and multi-element hostile targets and tasks, how to effectively and timely recognize targets under the conditions of a harsh environment and limited resources is the key to winning the war. As an important long-distance target detector, the radar plays an important role in the military, and it is very urgent to strengthen the research on the radar target recognition technology, which is the core of the modern military war.

[0003] At present, the traditional radar target recognition method capable of automatically extracting radar deep features includes a support vector machine, an extreme learning machine and the like. However, the above method has low target recognition accuracy and limited applicable environmental conditions. The intelligent method of radar target recognition is based on the deep learning gradually developed in recent years. The classifier obtained by the method has high performance and strong robustness, is suitable for many scenes, and is the main development direction of the future radar target recognition system.

[0004] Due to the limitation of the battlefield conditions, the computing and storage resources of the hardware equipment for recognizing the radar target are very limited, and many deep learning algorithms are difficult to implement. Therefore, how to complete the high-precision target recognition task by using the limited resource conditions is very important.

[0005] The patent with the application number 202011067886.3 discloses a radar target recognition method based on a two-dimensional weighted residual convolutional neural network. The method constructs a one-dimensional radar signal into a two-dimensional signal by using a Toeplitz matrix, and then uses a network composed of a weighted residual module to classify the target. The weight extracted by the weighted residual module only focuses on the attention of each receptive field, and does not consider the channel attention. In addition, the method contains a large-size convolution kernel and a large-size weight calculation full connection layer, and has high requirements for storage and computing resources.

[0006] The patent with application number 202011067886.3 discloses a radar target recognition method based on a multi-channel multiplex convolutional neural network. The method constructs a one-dimensional radar signal into a two-dimensional signal by using an odd-even alternating cyclic right shift matrix, and then uses a network composed of a multi-channel multiplex module to perform target classification. The weight extracted by the multi-channel multiplex module only focuses on the attention of each receptive field, and does not consider channel attention. SUMMARY

[0007] The purpose of the present application is to provide a radar target recognition method based on a two-dimensional self-selecting weighted convolutional neural network, which can improve the accuracy of target recognition at a very small computational parameter cost, and is suitable for use in situations where computational resources are very limited but high target recognition accuracy is required.

[0008] The technical solution for achieving the purpose of the present application is as follows: In a first aspect, the present application provides a radar target recognition method based on a two-dimensional self-selecting weighted convolutional neural network, comprising the following steps:

[0009] Step 1: Constructing a one-dimensional radar target time domain signal into a two-dimensional signal based on a Toeplitz matrix;

[0010] Step 2: Dividing the signal constructed in step 1 into a training set and a test set according to the category of the signal and a certain proportion;

[0011] Step 3: Constructing a coordinate attention module;

[0012] Step 4: Constructing a convolutional neural network based on a self-selecting weighting module using the coordinate attention module obtained in step 3;

[0013] Step 5: Inputting the training set data generated in step 2 into the network constructed in step 4 for training;

[0014] Step 6: Inputting the test set data generated in step 2 into the network trained in step 5 for accuracy testing.

[0015] In a second aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of the first aspect.

[0016] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the steps of the method of the first aspect.

[0017] In a fourth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the method of the first aspect.

[0018] Compared with the prior art, the present application has the following advantages: 1) the original one-dimensional ground reconnaissance radar time-domain target signal is constructed into a two-dimensional target signal based on a Toeplitz matrix, the advantages of a two-dimensional convolution kernel are exerted, and the trainability and recognition accuracy of the ground reconnaissance radar target time-domain signal are improved; 2) a hole convolution is used in the self-selection weighting module to extract features under different receptive fields, compared with a large-size convolution kernel for extracting features, the same performance is achieved while the parameter quantity and calculation quantity of the network are reduced, and the training and recognition efficiency of the network is improved; 3) the self-selection weighting module first performs feature fusion on the feature maps extracted by different convolution kernels under different receptive fields and then passes through an attention module, so that information under different receptive fields can be paid attention to, and self-selection weighting under different receptive fields is completed; 4) coordinate attention is used to obtain channel attention of different receptive fields, compared with only extracting weights of different receptive fields, more features can be paid attention to, and the accuracy of target recognition is improved; the introduction of coordinate attention has a very small calculation parameter quantity cost, so that the recognition efficiency of the two-dimensional radar target signal is higher and more feasible. Meanwhile, the self-selection weighting module contains two coordinate attention modules of maximum pooling and average pooling, which can extract more rich target feature attention and have stronger learning ability.

[0019] The present application will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flow chart of a radar target recognition method based on a two-dimensional self-selection convolutional neural network is provided.

[0021] Figure 2 A one-dimensional graph of a ground reconnaissance radar target time-domain signal.

[0022] Figure 3 A two-dimensional signal graph constructed based on a Toeplitz matrix.

[0023] Figure 4 A coordinate attention module graph constructed by the present application.

[0024] Figure 5 A self-selection weighting module graph provided by the present application.

[0025] Figure 6 A network structure graph constructed by the present application.

[0026] Figure 7 A comparison graph of the correct rates of the embodiments 1, 2 and 3 of the present application in a test set. DETAILED DESCRIPTION

[0027] In combination Figures 1-6As shown, the application proposes a two-dimensional target recognition method based on a two-dimensional self-selected weighted convolutional neural network for one-dimensional ground reconnaissance radar time domain signals, including the following steps:

[0028] Step 1, for the original radar signal data:f n =[f1, f2, f3,..., f M ], (n=1, 2, 3,..., M)

[0029] Where f1, f2, f3,..., f M are radar time domain signals of ground reconnaissance radar signals, M=1024; the specific method for constructing the original one-dimensional target signal into a two-dimensional signal based on the Toeplitz matrix is as follows:

[0030]

[0031]

[0032] After construction, it becomes H Toeplitz , and the two-dimensional target signal after construction is as shown in Figure 3 .

[0033] Step 2, after step 1, all one-dimensional ground reconnaissance radar target time domain signals are obtained, and then the total two-dimensional data set is divided into a training set and a test set according to the category and certain proportion of the target.

[0034] Step 3, build a coordinate attention module, set the input feature map dimension as c x h x w, c, h, w represent the channel number, length and width of the feature map respectively, perform X and Y dimension pooling, fuse the obtained tensor, and then use 1 x 1 convolution to convert the channel number to d, d is defined as follows.

[0035] d=max(d0, c / r)

[0036] Where d0 is the lower threshold, and r is the attenuation factor.

[0037] The above fused tensor is divided into two tensors with dimensions d x h x 1 and d x 1 x w along the spatial dimension, and then 1 x 1 convolution is used to increase the channel number of the two tensors to 3c, and then 3c x h x 1 and 3c x 1 x w tensors are obtained. At this point, the tensors obtained by the average pooling coordinate attention module and the maximum pooling coordinate attention module are added for information fusion, and finally the three groups of tensors are multiplied to obtain the final three output weight values.

[0038] Step 4, build a self-selected weighted convolutional neural network, combined Figure 6The self-selection weighted convolutional neural network comprises 11 self-selection weighting modules, 5 convolutional down-sampling layers (7*7 convolutional down-sampling is adopted in the first layer, and 3*3 convolutional down-sampling is adopted in the following 4 layers), 1 global pooling layer, 2 fully connected layers and 1 softmax layer; the self-selection weighting module firstly uses 1*1 convolution for dimension reduction, reduces the channel number to 1 / 4 of the input, and then uses three hollow convolutions (3*3 convolution, 3*3 convolution with a hollow rate of 2 and 3*3 convolution with a hollow rate of 3) to realize a 3*3 size receptive field, a 5*5 size receptive field and a 7*7 size receptive field respectively; the feature maps obtained under the three receptive fields are fused and input into the coordinate attention module constructed in step 3 to obtain 3 output weight values; the features are fused according to the weight values to improve the convolution precision and reduce the calculation amount; since the feature maps under different receptive fields are fused and then input into the attention module, the information of each receptive field can be paid attention to, and the self-selection weighting of each receptive field is completed; meanwhile, the introduction of the residual structure can make the information propagation in the network more smooth, greatly reduce the network calculation complexity and effectively alleviate the gradient disappearance and gradient explosion problems in the neural network back propagation process, so that the network is more easily optimized; the self-selection weighting module can extract effective features while reducing the parameter amount of the network and improving the accuracy of the neural network in radar target signal recognition.

[0039] Step 5: input the training set data divided in step 2 into the network in step 4 for training; the optimizer used in the training is Adam Optimizer, and the loss function used is categorical_crossentropy; after the training, the trained model is saved;

[0040] Step 6: input the test set data divided in step 2 into the model trained in step 5 for testing.

[0041] The self-selection weighted convolutional block is used in the recognition of two-dimensional target signals; the feature structures under three different receptive fields are extracted by using the hollow convolution, the efficiency of the model is improved, the obtained feature maps are fused and input into the coordinate attention module to complete the self-selection weighting of each channel feature under each receptive field; the coordinate attention mechanism can embed the spatial position information into the channel attention, only needs very small calculation parameter cost, can pay attention to the important part of the target feature, obviously improves the accuracy of radar target recognition, and the introduction of the residual structure can make the information propagation in the network more smooth, greatly reduce the network calculation complexity and effectively alleviate the gradient disappearance and gradient explosion problems in the neural network back propagation process, so that the network is more easily optimized.

[0042] The application will be further described in detail below with reference to three embodiments.

[0043] Embodiment 1

[0044] 4000 samples are taken from the time domain target data of the ground reconnaissance radar with a length of 1024 as training data, and 10000 samples are taken as test data, a Toeplitz matrix method is used to construct a two-dimensional data set, and after being constructed into a training set and a test set respectively, a feature map under three receptive fields is extracted using a hollow convolution, Figure 5 Each convolution module adopts a convolution method to extract the weights of each receptive field for weighting, adopts the same network structure as Figure 6 the same network structure as in Embodiment 1, and the model after training is tested by the test set to obtain the result.

[0045] Embodiment 2

[0046] The training set and the test set used in Embodiment 1 are used, the network adopts the same structure as in Embodiment 1, Figure 5 The attention module in the convolution module adopts SE (Squeeze-and-Excitation) channel attention implementation, and the model obtained after training is tested by the test set to obtain the result.

[0047] Embodiment 3

[0048] The training set and the test set used in Embodiment 1 are used, and then the training set is used to train the network structure shown in Figure 4 , Figure 5 , Figure 6 , and the model after training is used to test the test set to obtain the result data.

[0049] Table 1 is the accuracy data of Embodiments 1, 2 and 3 in the test set, Table 2 is the parameter amount of the network of Embodiments 1, 2 and 3, and Table 3 is the classification time of Embodiments 1, 2 and 3 in the test set.

[0050] Table 1

[0051]

[0052]

[0053] Table 2

[0054] Example 1 Example 2 Example 3 Model parameter quantity 7,486,139 2,017,966 2,034,826

[0055] Table 3

[0056] Example 1 Example 2 Example 3 Classification time 124 s 12 ms 83 s 8 ms 91 s 9 ms

[0057] Figure 7is a comparison chart of the correct rates of the test set of embodiments 1, 2 and 3, table 2 is a model parameter amount table of embodiments 1, 2 and 3, and table 3 is a test set classification time of embodiments 1, 2 and 3. From the experimental results, it can be seen that the parameter amount of embodiment 2 is close to one fourth of that of embodiment 1, but after the model training converges, the accuracy is greatly improved due to the introduction of channel attention; embodiment 3 increases only a small amount of calculation and parameter amount at the cost of increasing the classification accuracy of the model by nearly 0.2 percentage points. It is shown that the self-selected weighted convolutional neural network proposed in the present application is very suitable for the target recognition and classification task of the ground reconnaissance radar signal.

[0058] The above-described embodiments only express several embodiments of the present application, and the description is relatively specific and detailed, but it should not be understood as limiting the scope of the patent of the present application. It should be pointed out that for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A radar target recognition method based on a two-dimensional self-selected weighted convolutional neural network, characterized in that, Comprising the following steps: Step 1, based on Toeplitz matrix, one-dimensional radar target time domain signal is constructed into two-dimensional signal; Step 2, the signal constructed in step 1 is divided into training set and test set according to the category of the signal and a certain proportion; Step 3, a coordinate attention module is constructed; assuming that the input feature map dimension is c x h x w, c, h and w represent the channel number, length and width of the feature map respectively, first, X and Y dimensions are pooled, the obtained tensor is fused, and then 1x1 convolution is used to convert the channel number to d, d is defined as follows: ； wherein is a lower threshold value, r is a decay factor; The above fused tensor is divided into two tensors with dimensions of d x h x 1 and d x 1 x w along the spatial dimension, and then 1x1 convolution is used to increase the channel number of the two tensors to 3c, and then 3c x h x 1 and 3c x 1 x w tensors are obtained, the tensors obtained by the average pooling coordinate attention module and the maximum pooling coordinate attention module are added for information fusion, and finally the three groups of tensors are multiplied to obtain the final three output weight values; Step 4, using the coordinate attention module obtained in step 3, a convolutional neural network based on self-selected weighting module is constructed; the constructed self-selected weighting module and self-selected weighting convolutional neural network, wherein the total network structure includes 11 self-selected weighting modules, 5 convolutional down-sampling layers, 1 global pooling layer, 2 fully connected layers and 1 softmax layer; the self-selected weighting module first uses 1x1 convolution for dimension reduction, reducing the channel number to 1 / 4 of the input, then uses three hollow convolutions for feature extraction in different receptive fields, so that the module can extract feature structures in different receptive fields, then the obtained feature maps are fused and input into the coordinate attention module constructed in step 3 to obtain three output weight values, and the self-selected weighting of each receptive field is completed by weighted summation; Step 5, input the training set data generated in step 2 into the network constructed in step 4 for training; Step 6, input the test set data generated in step 2 into the network trained in step 5 for accuracy test.

2. The radar target recognition method based on two-dimensional self-selected weighted convolutional neural network according to claim 1, characterized in that, The two-dimensional signal construction method based on Toeplitz matrix in step 1 is as follows: Let the original radar signal data with length M be: ; wherein , , , is a radar time domain signal of the ground surveillance radar signal, M = 1024; After Toeplitz matrix arrangement, the two-dimensional data is: ； 。 3. The radar target recognition method based on a two-dimensional self-selected weighted convolutional neural network according to claim 1, characterized in that, The 5 convolutional down-sampling layers, wherein the first layer uses 7x7 convolutional down-sampling, and the last 4 layers use 3x3 convolutional down-sampling.

4. The radar target recognition method based on a two-dimensional self-selected weighted convolutional neural network according to claim 1, characterized in that, Step 5 is: input the training set data obtained in step 2 into the network in step 4 for training, the optimizer of the training network is Adam Optimizer, and the loss function used is categorical crossentropy.

5. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method of any one of claims 1-4.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method of any one of claims 1-4.

7. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-4. The computer program is executed by the processor to implement the steps of the method of any one of claims 1-4.

Images