A ground target recognition method based on deep learning

Abstract

The present invention belongs to the field of radar recognition technology, and specifically relates to a ground target recognition method based on deep learning. The method first preprocesses the HRRP data to eliminate its amplitude sensitivity and translation sensitivity; then uses the CWT to perform time-frequency analysis on the HRRP data. Through the multi-scale decomposition of the CWT, the time-frequency characteristics of the signal at different scales can be extracted and the robustness of the model to noise can be enhanced; the time-frequency data obtained after processing are then input into a convolutional neural network. A large number of annotated HRRP data sets are used to train the CNN. During the training process, the network weights are continuously adjusted through the backpropagation algorithm to optimize the network performance. After the training is completed, the trained CNN model is used to classify and identify the test HRRP signal, and its performance is evaluated. The proposed method can more comprehensively describe the HRRP signal, improve the feature expression ability, and has certain recognition performance advantages under low signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the field of radar recognition technology, and specifically relates to a ground target recognition method based on deep learning. Background Art

[0002] With the advancement of radar technology, radar automatic target recognition (RATR) has become increasingly important in both military and civilian applications. High-resolution range profile (HRRP), a key method for radar systems to acquire target information, has attracted widespread attention due to its ability to provide rich, detailed target features. HRRP analyzes the reflected signals from the target's scattering center at different range bins to generate a one-dimensional range profile of the target. These profiles contain important information such as the target's size, shape, and scattering characteristics.

[0003] Traditional RATR methods rely on manual feature extraction, but these methods often require extensive prior knowledge and parameter tuning, and struggle to adapt to practical challenges such as target pose variations and noise interference. In recent years, deep learning technology, with its powerful data processing and feature extraction capabilities, has achieved revolutionary progress in fields such as image recognition and speech processing. The application of deep learning methods in RATR can be broadly divided into three categories. The first category is based on autoencoders, which compress input data to extract effective features for recognition. Autoencoders can also be used to mitigate the effects of noise in non-cooperative target recognition. However, their generalization capabilities are limited by the feature extraction of HRRP data. The second category is convolutional neural networks (CNNs). Due to their excellent performance in image processing, they have naturally been introduced to HRRP target recognition and have shown great potential for processing high-dimensional radar data. However, due to their limited receptive field, they struggle to extract global features. The third category is time series models, such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and transformers. These time series models treat HRRP as sequential data. However, HRRP reflects a scattering distribution, so it is more appropriate to view it as a one-dimensional profile rather than sequential data.

[0004] Most existing deep learning-based RATR methods directly process HRRP, ignoring the time-frequency characteristics of radar signals. In recent research, signal time-frequency analysis has been primarily applied to extract target micro-motion features. References such as "Target recognition algorithm based on HRRP time-spectrogram feature and multi-scale asymmetric convolutional neural network" and "Convolutional neural networks for radar HRRP target recognition and rejection" use short-time Fourier transforms (STFTs) to analyze the time-frequency information of HRRP. However, because the time and frequency resolutions obtained by the STFTs are limited by the length of the window function, they cannot simultaneously provide high time and frequency resolution, limiting their performance when analyzing non-stationary signals. References such as "Towards Efficient Lesion Localization Based on Template Occlusion Strategy in Intelligent Diagnosis" propose an autoencoder with an RNN that uses discrete wavelet transforms (DWTs) to automatically learn a hierarchical representation of the signal. This approach successfully improves signal feature extraction capabilities by leveraging the multi-scale properties of the DWT. However, despite its advantages in signal processing, the DWT has limitations in terms of continuity and adaptability in time-frequency analysis. Summary of the Invention

[0005] To address the above issues, the present invention proposes a deep learning-based method for classifying and identifying high-resolution one-dimensional range profiles, namely, CWT-CNN. This method first preprocesses the HRRP data to eliminate its amplitude sensitivity and translation sensitivity. It then uses the CWT to perform time-frequency analysis on the HRRP data. Through the multi-scale decomposition of the CWT, it can extract the time-frequency features of the signal at different scales and enhance the model's robustness to noise. The processed time-frequency data is then input into a convolutional neural network. The CNN is trained using a large number of annotated HRRP datasets. During the training process, the network weights are continuously adjusted using a backpropagation algorithm to optimize network performance. After training, the trained CNN model is used to classify and identify test HRRP signals, and its performance is evaluated. The proposed method can more comprehensively describe HRRP signals, improve the expressiveness of features, and exhibit certain recognition performance advantages under low signal-to-noise ratios.

[0006] The technical solution adopted in the present invention is:

[0007] A ground target recognition method based on deep learning includes the following steps:

[0008] S1. Obtain the radar’s high-resolution range image HRRP, expressed as r = [r1, r2, …, r L ] T , where r i represents the sub-echo of the i-th range unit, L represents the total number of range units; the amplitude of the complex HRRP data is taken to obtain the real HRRP data x=|r|=[x1,x2,…,x l ] T ;

[0009] S2. Preprocessing the obtained real HRRP data, including:

[0010] Use L2 norm normalization:

[0011]

[0012] Calculate the centroid position of the HRRP signal:

[0013]

[0014] The center of the normalized data is shifted to the obtained centroid, and the shifted data is used as training data;

[0015] S3. Build a target recognition model, including a wavelet transform module, a convolution module, a pooling and fully connected module, and an output module;

[0016] The wavelet transform module performs wavelet transform on the training data to obtain time-frequency data X;

[0017] The convolution module extracts features from the time-frequency data X. The convolution module includes two convolution layers, each of which consists of a convolution layer with a kernel size of 3×3. There is a batch normalization layer and an activation function after the convolution layer. The convolution module model is expressed as follows:

[0018]

[0019] Among them, w ijl is the weight of each position of the convolution kernel, x ijl is the value of each position in the corresponding convolution area, (k1, k2, d) is the convolution kernel size, b is the corresponding offset, the superscript k represents the kth dimension of the data, μ (k) and σ (k) are the mean and standard deviation of the current batch of input data, and the parameter γ (k) and β (k)Used to scale and shift the normalized value, ∈ is a small number used to ensure that the denominator is non-zero, and f(·) is the activation function;

[0020] The pooling and fully connected module includes a pooling layer and two fully connected layers. The pooling layer performs a downsampling operation on the input feature map, and the output of the fully connected layer is expressed as:

[0021]

[0022] in, and b i They are the weight matrix, input data, and offset matrix of the i-th fully connected layer respectively; the first fully connected layer is followed by the Dropout function, and the activation function of the second fully connected layer uses the Sigmoid function;

[0023] The output module adopts softmax classifier for classification;

[0024] S4. Using the obtained training data to train the constructed target recognition model to obtain a trained target recognition model;

[0025] S5. After preprocessing, the acquired high-resolution range image of the target is input into the trained target recognition model to obtain the recognition result.

[0026] The beneficial effects of this invention are as follows: the method combines wavelet transform and convolutional neural network to identify ground targets. The continuous wavelet transform can extract the time-frequency information of HRRP data while effectively filtering out the influence of noise, making the model more robust to noise. At the same time, the multi-scale characteristics of the convolutional neural network (CWT) are utilized to effectively extract local features in the HRRP data. It has high recognition accuracy in small sample and high noise scenarios, and has certain advantages under different signal-to-noise ratios and different sample set sizes. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a network structure diagram of the present invention.

[0028] Figure 2 Schematic diagram of data preprocessing results.

[0029] Figure 3 Schematic diagram of model comparison under different signal-to-noise ratios. DETAILED DESCRIPTION

[0030] The present invention will be described in detail below with reference to the accompanying drawings and embodiments:

[0031] like Figure 1 As shown, the method of the present invention comprises the following steps:

[0032] S1. Obtain the high-resolution range image of the radar r = [r1, r2, ..., r L ] T , where r i represents the sub-echo of the ith range unit, and L represents the total number of range units. The amplitude of the complex HRRP data is taken to obtain the real HRRP data x=|r|=[x1,x2,…,x l ] T .

[0033] S2. Data preprocessing:

[0034] In view of amplitude sensitivity, this paper adopts L2 norm normalization to process the real HRRP data, namely:

[0035]

[0036] HRRP data is processed using center of mass alignment to reduce the impact of target pose variations. The center of mass of the HRRP signal is calculated using the following formula, and the calculated center of mass is aligned to the center of the signal.

[0037]

[0038] The HRRP data after L2 normalization and centroid alignment is as follows Figure 2 As shown in Figure 2, the amplitude range of the processed HRRP samples is [0,1], and the centroid is close to the center point.

[0039] S3. Input the recognition model and perform the following processing on the recognition model:

[0040] Wavelet transform module

[0041] The continuous wavelet transform provides a method to locate signal features in both the time domain and the frequency domain by analyzing the signal under different scales and translation parameters. This transform uses a wavelet function as a basis and convolves it with the original signal by changing its scale and position to obtain the time-frequency representation of the signal. The corresponding calculation formula is:

[0042]

[0043] Where x(t) is the original signal; ψ * is the complex conjugate of the wavelet function; a is the scale parameter, which is used to adjust the scale or frequency range of the wavelet function on the time axis; b is the translation parameter.

[0044] The complex Gaussian wavelet is a complex number consisting of the derivatives of the Gaussian function and can be expressed as the P-order derivative of the Gaussian function, where P is an integer between 1 and 8. The mathematical expression of the complex Gaussian wavelet is as follows:

[0045]

[0046] Among them, C n Is a normalization constant that ensures that the wavelet function has unit energy. Compared with other types of wavelets, complex Gaussian wavelets have localized properties in both the time and frequency domains, which enables them to provide detailed information about the signal at specific time and frequency.

[0047] Convolutional module

[0048] In order to extract the feature information corresponding to the time-frequency data, a feature extraction module is designed. This module consists of two 2-D convolution blocks, and each convolution layer consists of a convolution layer with a kernel size of 3×3. Set the input data It is multi-channel, assuming that the weight of each position of the convolution kernel is w ijl , the corresponding convolution area has a value of x at each position ijl , the convolution kernel size is (k1, k2, d), then the output value of each convolution area after convolution is

[0049]

[0050] Where b is the corresponding offset.

[0051] Batch normalization processing module

[0052] In order to improve the stability of the model, a batch normalization layer is added after the convolution layer. The BN operation is obtained by Equation 6.

[0053]

[0054] Among them, BN is performed independently on each dimension of a batch of data, the superscript k represents the kth dimension of the data, μ (k) and σ (k) are the mean and standard deviation of the current batch of input data respectively. After normalizing the data, the parameter γ is used (k) and β (k) Scale and shift the normalized values.

[0055] Activation Function

[0056] Regarding the choice of nonlinear activation function, considering the problem of neuron "death" when using the ReLU function, the LeakyReLU function is selected to avoid the problem of gradient aliasing. At this point, the convolution block can be modeled as:

[0057]

[0058] Where f(·) is the activation function.

[0059] The Dropout function is added after the convolutional layer and the first fully connected layer. By adjusting the parameters of the Dropout function, some neurons in the network are randomly disabled.

[0060] Pooling layer and fully connected layer

[0061] The pooling layer extracts key features by downsampling the input feature map, improving the computational efficiency of the model and reducing overfitting. It also suppresses noise and reduces information redundancy. The formula is as follows:

[0062]

[0063] Among them, y ijl Represents the rectangular area related to the kth feature map The maximum pooling output value, x kpq Represents a rectangular area The element at (p,q) in .

[0064] Next, the feature map is passed to the fully connected layer, and the output can be obtained by formula 2-7:

[0065]

[0066] in, and b i are the weight matrix and bias matrix of the i-th fully connected layer respectively. The activation function of the second fully connected layer uses the Sigmoid function, and the output layer generally uses the softmax classifier for classification. The loss function is:

[0067]

[0068] Where Y is the output of the network, and λ is a regular term to prevent overfitting. If the value of the weight decay coefficient λ is too small, it may lead to overfitting, and vice versa.

[0069] S4. Using the obtained training data to train the constructed target recognition model to obtain a trained target recognition model;

[0070] S5. After preprocessing, the acquired high-resolution range image of the target is input into the trained target recognition model to obtain the recognition result.

[0071] Example

[0072] The parameters used in this example are: the learning rate of the network model is 0.001, the batch size of the training data is 64, and the network training is performed using the batch gradient descent method. The technical solution of the present invention is used in sequence:

[0073] 1.According to step "S1", obtain the high-resolution range image r = [r1, r2, ..., r L ] T , where r i represents the sub-echo of the ith range unit, and L represents the total number of range units. The amplitude of the complex HRRP data is taken to obtain the real HRRP data x=|r|=[x1,x2,…,x l ] T ;

[0074] 2. Perform L2 norm normalization on the obtained real HRRP data according to step "S2" to obtain data S;

[0075] 3. According to step "S3", calculate the corresponding center of gravity W in the radar high-resolution image S:

[0076]

[0077] The center of the corresponding radar high-resolution range image S is shifted to the center of gravity W to obtain the moved data x′=[x′1,x′2,…,x′ l ] T

[0078] 4. According to step "S4", the HRRP data is input into the wavelet transform module, and the wavelet basis function is set to cgau8 and the length is 32 to obtain the time-frequency data X corresponding to the HRRP, where the size of X is 31×L.

[0079] 5. Following steps S5-S7, each convolutional layer consists of a 3×3 kernel, with 32 and 64 output channels, a stride of 1, and zero padding of 1. The resulting time-frequency data X is fed into the convolution module to extract feature information. The corresponding output data dimensions are 64×32×31×L and 64×64×31×L, respectively.

[0080] 6. Following step S8, a Dropout function is added after the convolutional layer and the first fully connected layer, and the deactivation rate of its neurons is set to 0.5.

[0081] 7. According to step "S9", the pooling layer uses the maximum pooling method, the dimension of the pooling domain is 4×4, and the moving stride is 2. By downsampling the input feature map, the key features are extracted, and the corresponding output data Y pool The dimension of Y is 64×64×15×L. The number of neurons in the first fully connected layer is 1024, and the number of neurons in the second fully connected layer is the number of target classification categories. pool Input two fully connected layers and get the output data y=[y1,y2,y i ,…,y l ] T, corresponding to the probability that the HRRP sample is identified as the i-th class.

[0082] like Figure 3 As shown, the present invention has better effects than the existing methods.

Claims

1. A ground target recognition method based on deep learning, characterized in that: The following steps are involved: S1. Obtain the radar’s high-resolution range image HRRP, expressed as r = [r1, r2, …, r L ] T , where r i represents the sub-echo of the i-th range unit, L represents the total number of range units; the amplitude of the complex HRRP data is taken to obtain the real HRRP data x=|r|=[x1,x2,…,x l ] T ; S2. Preprocessing the obtained real HRRP data, including: Use L2 norm normalization: Calculate the centroid position of the HRRP signal: The center of the normalized data is shifted to the obtained centroid, and the shifted data is used as training data; S3. Build a target recognition model, including a wavelet transform module, a convolution module, a pooling and fully connected module, and an output module; The wavelet transform module performs wavelet transform on the training data to obtain time-frequency data X; The convolution module extracts features from the time-frequency data X. The convolution module includes two convolution layers, each of which consists of a convolution layer with a kernel size of 3×3. There is a batch normalization layer and an activation function after the convolution layer. The convolution module model is expressed as follows: Among them, w ijl is the weight of each position of the convolution kernel, x ijl is the value of each position in the corresponding convolution area, (k1, k2, d) is the convolution kernel size, b is the corresponding offset, the superscript k represents the kth dimension of the data, μ (k) and σ (k) are the mean and standard deviation of the current batch of input data, and the parameter γ (k) and β (k) Used to scale and shift the normalized value, ∈ is a small number used to ensure that the denominator is non-zero, and f(·) is the activation function; The pooling and fully connected module includes a pooling layer and two fully connected layers. The pooling layer performs a downsampling operation on the input feature map, and the output of the fully connected layer is expressed as: in, and b i They are the weight matrix, input data, and offset matrix of the i-th fully connected layer respectively; the first fully connected layer is followed by the Dropout function, and the activation function of the second fully connected layer uses the Sigmoid function; The output module adopts softmax classifier for classification; S4. Using the obtained training data to train the constructed target recognition model to obtain a trained target recognition model; S5. After preprocessing, the acquired high-resolution range image of the target is input into the trained target recognition model to obtain the recognition result.

Images