Radar High-Resolution Range Profile Denoising and Recognition Method Based on GAN and Self-Attention

Abstract

The present invention discloses a radar high-resolution range image noise reduction and recognition method based on GAN and self-attention, which mainly solves the problems in the existing radar high-resolution range image noise reduction and recognition technology that features with strong separability are regarded as noise and eliminated, and long-distance dependence information cannot be extracted. The implementation scheme is as follows: 1) Obtain the target radar echo to generate a training sample set and a test sample set; 2) Construct a radar high-resolution range image noise reduction and recognition network based on GAN and self-attention, which consists of a generator, a discriminator, and an identifier; 3) Use the training sample set to train the noise reduction and recognition network constructed in 2) by the backpropagation algorithm; 4) Input the test sample set into the trained radar high-resolution range image noise reduction and recognition network to obtain the radar high-resolution range image noise reduction and recognition results. The present invention improves the target recognition rate in a low signal-to-noise ratio environment and can be used to realize the recognition of aircraft and satellite targets in a low signal-to-noise ratio environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and further relates to an end-to-end radar high-resolution range profile (HRRP) noise reduction and target recognition method, which can be used to realize feature extraction and recognition of targets such as airplanes and satellites in a low signal-to-noise ratio environment. Background Art

[0002] The high-resolution range profile (HRRP) of a target is the amplitude waveform of the vector sum of the target scatterer sub-echoes obtained by a broadband radar signal projected in the radar line-of-sight direction. It contains important information such as target size, structure, and scatterer distribution. HRRP recognition is an important research direction in the field of radar automatic target recognition (RATR). In actual application scenarios, the HRRP of a target contains noise, and the presence of noise will interfere with the feature extraction process, thereby affecting the target recognition result.

[0003] For noisy HRRP, existing target recognition methods generally adopt the following three schemes:

[0004] 1) Direct recognition is performed without special processing on the noisy HRRP. The noisy HRRP is input into a target recognition model established for noiseless HRRP to obtain a recognition result. This type of method has simple steps, but has low adaptability and poor recognition performance.

[0005] 2) Robust features are extracted. According to the physical characteristics of HRRP, noise-robust features of HRRP are extracted, such as the bispectrum features of complex radar signals, the scattering coefficients and positions of the main target scatter centers, etc. Then, a classifier is designed based on these features for target recognition. However, this type of method requires manual design for feature extraction and its performance is limited when the environment changes.

[0006] 3) Noise reduction is performed first and then recognition. A noise reduction model and a recognition model are respectively constructed. First, the noisy HRRP is denoised by the noise reduction model to generate a high signal-to-noise ratio HRRP, and then the recognition model is used to perform feature extraction and recognition on the high signal-to-noise ratio HRRP. This type of method can utilize a variety of known recognition models, but the recognition result cannot guide the noise reduction process, which may cause the noise reduction process to lose effective recognition features.

[0007] C. Zhao, X. He, J. Liang, T. Wang, C. Huang proposed a semi-supervised multi-task recognition framework in their published paper "Radar HRRP target recognition via semi-supervised multi-task deep network" (IEEE Access, 2019) for the problem of HRRP target recognition in low signal-to-noise ratio environments. It contains two models, DUBDNet and RSRNet, which respectively achieve noise reduction and recognition of low signal-to-noise ratio HRRP. The noisy HRRP is input into DUBDNet to obtain the noise-reduced HRRP, and then RSRNet is used to complete target recognition. However, this method trains the noise reduction model first and then the recognition model, and the two processes do not affect each other. There may be a problem that some features with strong separability are eliminated as noise during the noise reduction process, thus affecting the recognition rate.

[0008] J. Nie, Y. Xiao, L. Huang, F. Lv proposed a recognition method based on CN-LSGAN, STFT, and CNN in their published paper "Time-Frequency Analysis and Target Recognition of HRRP" (2021, Complexity). CN-LSGAN is used to reduce the noise of HRRP with low signal-to-noise ratio. STFT is used to transform the noise-reduced HRRP to obtain a two-dimensional time-frequency diagram that can represent both the time-domain and frequency-domain information of the signal. Then CNN is used to extract features from the two-dimensional time-frequency diagram to obtain the target recognition result. However, this method also processes the noise reduction and recognition processes separately and is not an end-to-end model. There is still a problem that some features with strong separability are eliminated as noise.

[0009] In recent years, deep learning has been widely applied to automatic target recognition ATR based on HRRP. Such methods perform feature extraction through data-driven and have obtained excellent recognition performance, receiving extensive attention in radar target recognition research.

[0010] The existing feature extraction methods based on deep learning can generally adopt the following several schemes:

[0011] 1) Autoencoder

[0012] The autoencoding model AE is a neural network that tries to reproduce the input signal as much as possible. It belongs to an unsupervised learning algorithm and consists of an encoder and a decoder. These two parts can be regarded as a feature extraction process. The output of the encoder can be used as data features for subsequent tasks such as classification or recognition. However, such methods have limited ability to extract the structured information of HRRP samples.

[0013] 2) Convolutional Neural Networks

[0014] Convolutional neural network (CNN) extracts local features and deep robust feature information from the original signal through a series of convolution and pooling operations. It has good generalization ability, but CNN cannot represent the order relationship between data segments.

[0015] 3) Recurrent Neural Networks

[0016] Recurrent neural networks (RNNs) are advantageous when modeling data with temporal relationships. Their internal structure, consisting of recursive connections between neurons, enables the memorization and transfer of historical information. Consequently, they are widely used in tasks such as classifying, predicting, and generating sequence data. However, this approach ignores the fact that target and noise regions in HRRP samples have different influences on recognition results, and its recognition capabilities still need to be improved.

[0017] 4) Attention Mechanism

[0018] Attention is a complex cognitive function in the human brain, referring to the ability to focus on some information while simultaneously ignoring others. When using neural networks to process large amounts of input, we can leverage the human brain's attention mechanism to focus on key inputs, thereby improving the efficiency of the network.

[0019] In their paper "Target-aware recurrent attentional network for radar HRRP target recognition" (Signal Processing, 2019), B.Xu, B.Chen, J.Wan, H.Liu, and L.Jin proposed an RNN-based attention model. The specific steps of this method are as follows: first, an RNN is used to encode the data segments of a single HRRP sample and extract the temporal correlation between each distance unit. Then, the attention mechanism is used to assign different weight coefficients to the output of each time step based on the contribution of each data segment to recognition. This coefficient is used to perform a weighted sum of the outputs of each time step as a separable feature. Finally, this separable feature is connected to a softmax classifier to output the recognition result. Because this method requires the use of an RNN to model HRRP, it cannot extract long-range dependency information and cannot process the data segments in HRRP in parallel.

[0020] Liu Jiaqi, Chen Bo, and Jie Qian proposed in the paper "Radar HRRP Target Recognition Based on Attention Mechanism and Bidirectional GRU Model" published in Radar Science and Technology in 2019 to combine bidirectional GRU with the attention mechanism. The specific steps of this method are as follows: The HRRP samples in the time domain are divided into forward and reverse sequences through a sliding window, and they are respectively passed through two independent GRU networks for feature extraction. Then, the features extracted at the same moment are concatenated, thereby utilizing the bidirectional temporal information of HRRP. The hidden layer features after weighted summation are obtained through the attention mechanism for target recognition and classification. Although this method utilizes the bidirectional temporal information of HRRP, there is still a problem that it cannot extract long-distance dependence information. Summary of the Invention

[0021] The purpose of the present invention is to propose a radar high-resolution range profile noise reduction and recognition method based on the generative adversarial network GAN and self-attention aiming at the deficiencies of the above-mentioned existing technologies, so as to reconstruct high signal-to-noise ratio HRRP, and improve the target recognition performance in a low signal-to-noise ratio environment by extracting long-distance dependence information in HRRP.

[0022] The technical idea of the present invention is: Generate high signal-to-noise ratio HRRP through GAN, process one-dimensional high signal-to-noise ratio HRRP samples using the sliding window method to generate sequence samples, and obtain the target recognition result of HRRP samples by constructing a deep neural network based on self-attention. Its implementation scheme includes the following:

[0023] (1) Divide and process the radar echoes of three types of aircraft targets to generate a training sample set S1 and a test sample set S2;

[0024] (2) Construct a radar high-resolution range profile noise reduction and recognition network:

[0025] Build a generator, which is composed of five convolutional layers and five deconvolutional layers connected in sequence. Each convolutional layer has a LeakyReLU operation, and each deconvolutional layer is followed by a LeakyReLU and a skip connection operation, which is used to reduce the noise of the noisy HRRP and generate high signal-to-noise ratio HRRP after noise reduction;

[0026] Build a discriminator, which is composed of five convolutional layers, a flatten layer, and two fully connected layers connected in sequence. Each convolutional layer is followed by a LeakyReLU operation, which is used to assist the generator in improving the noise reduction performance;

[0027] Build a recognizer, which includes a sliding window layer, a position encoding layer, three fully connected layers, five groups of feature extractors, a flatten layer, and a softmax classifier. The sliding window layer, the position encoding layer, the third fully connected layer, the five groups of feature extractors, the flatten layer, the fourth fully connected layer, the fifth fully connected layer, and the softmax classifier are connected in sequence, and are used to extract features from high signal-to-noise ratio HRRP and give the target recognition result;

[0028] Connect the generator to the discriminator and the recognizer respectively to form a radar high-resolution range image denoising and recognition network based on the generative adversarial network GAN and self-attention;

[0029] (3) Input the training sample set S1 into the denoising and recognition network constructed in (2), and perform adversarial training on it through the backpropagation algorithm to obtain a trained radar high-resolution range image denoising and recognition network based on GAN and self-attention;

[0030] (4) Input the test sample set S2 into the trained radar high-resolution range image denoising and recognition network based on GAN and self-attention for testing to obtain the target recognition result output by the network.

[0031] The present invention has the following advantages compared with the prior art:

[0032] First, since the recognizer of the present invention uses position encoding in the position encoding layer to represent the sequence relationship between HRRP data segments, it can make full use of the temporal information in the sequence, and uses the self-attention mechanism in the self-attention layer of the feature extractor to make the model pay more attention to the target regions with strong distinguishability and fully extract the dependence relationship between each data segment, overcoming the problem of losing long-distance dependence information in the prior art, and improving the feature extraction ability and the target recognition performance.

[0033] Second, since the present invention converts the noisy low signal-to-noise ratio HRRP into denoised high signal-to-noise ratio HRRP through the generator, and directly inputs the high signal-to-noise ratio HRRP into the recognizer to obtain the target category, it overcomes the problem in the prior art that denoising and recognition are carried out step by step, resulting in the inability to achieve end-to-end training.

[0034] Third, since the present invention couples the loss function of the recognizer with the loss function of the generator during training, the generator can retain the separable features useful for the recognition process during the denoising process, effectively reducing the problem of loss of recognition information caused by the denoising process in the prior art, and improving the target recognition performance in a low signal-to-noise ratio environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is the implementation flowchart of the present invention.

[0036] Figure 2It is the network framework diagram in the present invention. Specific embodiments

[0037] The embodiments and effects of the present invention will be further described below with reference to the accompanying drawings.

[0038] Refer to Figure 1 The implementation steps of this embodiment are as follows.

[0039] Step 1, generate a training sample set and a test sample set.

[0040] 1.1) Respectively obtain the original radar echoes of three types of aircraft targets, namely An-26 aircraft, Citation aircraft, and Yak-42 aircraft. Starting from the first echo of each segment of echo data, sample every 20th echo, and add Gaussian white noise with a fixed signal-to-noise ratio to the sampled original radar echoes respectively to obtain the noisy radar echoes of each type of target;

[0041] 1.2) Respectively perform Fourier transforms on the original radar echoes and the noisy radar echoes of each type of target after sampling to obtain the noise-free high-resolution range profiles (HRRPs) of the radar and the noisy HRRPs of the radar;

[0042] 1.3) Respectively perform preprocessing of 2-norm normalization and centroid alignment on the noise-free HRRPs and the noisy HRRPs obtained in 1.2);

[0043] 1.4) According to the flight tracks of the three types of aircraft, select the fifth and sixth segments of radar echoes of the An-26 aircraft, the sixth and seventh segments of radar echoes of the Citation aircraft, and the second and fifth segments of radar echoes of the Yak-42 aircraft. The noise-free HRRPs and the noisy HRRPs after preprocessing in 1.3) corresponding to the six segments of radar echo data segments are used as the training sample set S1; the noise-free HRRPs and the noisy HRRPs after preprocessing in 1.3) corresponding to the remaining radar echo data segments are used as the test sample set S2.

[0044] Step 2, construct a radar high-resolution range profile noise reduction and recognition network based on GAN and self-attention.

[0045] 2.1) Establish a generator: It is composed of five convolutional layers and five transposed convolutional layers connected in sequence. Each convolutional layer has a LeakyReLU operation, and each transposed convolutional layer is followed by a LeakyReLU and a skip connection operation, which is used to denoise the noisy HRRP and generate a denoised high signal-to-noise ratio HRRP;

[0046] The parameter settings of each layer of this generator are as follows:

[0047] The convolutional kernel sizes of the five convolutional layers and the five transposed convolutional layers are all 15;

[0048] The number of convolution kernels in the first convolutional layer and the fourth deconvolutional layer is 16 each;

[0049] The number of convolution kernels in the second convolutional layer, the third convolutional layer, the second deconvolutional layer, and the third deconvolutional layer is 32 each;

[0050] The number of convolution kernels in the fourth convolutional layer, the fifth convolutional layer, and the first deconvolutional layer is 64 each;

[0051] The number of convolution kernels in the fifth deconvolutional layer is 1;

[0052] 2.2) Build a discriminator: It is composed of five convolutional layers, a flatten layer, and two fully connected layers connected in sequence. There is a LeakyReLU operation after each convolutional layer to assist the generator in improving the noise reduction performance;

[0053] The parameter settings of each layer of this discriminator are as follows:

[0054] The size of the convolution kernels in the five convolutional layers is 15;

[0055] The number of convolution kernels in the sixth convolutional layer is 16;

[0056] The number of convolution kernels in the seventh convolutional layer and the eighth convolutional layer is 32 each;

[0057] The number of convolution kernels in the ninth convolutional layer and the tenth convolutional layer is 64 each;

[0058] The number of neurons in the first fully connected layer is 8;

[0059] The number of neurons in the second fully connected layer is 1;

[0060] 2.3) Build a recognizer: It includes a sliding window layer, a position encoding layer, three fully connected layers, five groups of feature extractors, a flatten layer, and a softmax classifier. The sliding window layer, the position encoding layer, the third fully connected layer, the five groups of feature extractors, the flatten layer, the fourth fully connected layer, the fifth fully connected layer, and the softmax classifier are connected in sequence to extract features from the high signal-to-noise ratio HRRP and give the target recognition result;

[0061] The functions and structural parameters of each layer of this recognizer are as follows:

[0062] The sliding window layer is used to perform sliding window on the high signal-to-noise ratio HRRP output by the generator with a window length d = 6 and a stride to convert it into a sequence form HRRP with a sequence length of T = 84 as the output;

[0063] The position encoding layer is used to generate T = 84 position encodings using the sine function, and its sine function is expressed as follows:

[0064] PE(pos,j) = cos(pos / 10 2j / T )

[0065] Among them, pos ∈ [1, T] represents the serial number of the currently calculated position encoding in the sequence, and j ∈ [1, d] represents that the j-th element of the position encoding is currently calculated. The sum of these T position encodings and the output of the sliding window layer item by item is the output of this position encoding layer;

[0066] The five groups of feature extractors are used to extract the dependency relationships between sequence data. Each group of feature extractors consists of a self-attention layer, a residual connection layer, a normalization layer, a feed-forward layer, a residual connection layer, and a normalization layer cascaded in sequence;

[0067] The input of this self-attention layer is For the first group of feature extractors, X represents the output of the position encoding layer. For the other four groups of feature extractors, X represents the output of the previous group of feature extractors. The output of this self-attention layer is In the formula: Among them, are all learnable parameter matrices, and set d model = d k = d v = 128;

[0068] This feed-forward layer contains two fully connected layers, and the number of neurons in them is 512 and 128 respectively;

[0069] The number of neurons in the third, fourth, and fifth fully connected layers is 128, 3072, and 3 respectively;

[0070] 2.4) Connect the generator to the discriminator and the recognizer respectively to form a radar high-resolution range image denoising and recognition network based on GAN and self-attention.

[0071] Step 3, train the radar high-resolution range image denoising and recognition network based on GAN and self-attention.

[0072] Refer to Figure 2 , the specific implementation of this step is as follows:

[0073] 3.1), initialize all learnable parameters in the radar high-resolution range image denoising and recognition network based on GAN and self-attention;

[0074] 3.2) Input the noisy high-resolution range image x1 in the training sample set S1 into the generator, and calculate layer by layer along the network structure to generate the denoised high-resolution range image x2;

[0075] 3.3) Connect the noise-free high-resolution range image x3 and the noisy high-resolution range image x1 in the training sample set S1 to obtain a true sample r = [x3, x1]. Connect the denoised high-resolution range image x2 generated by the generator in 3.2) and the noisy high-resolution range image x1 to obtain a fake sample f = [x2, x1]. Input the true sample r and the fake sample f into the discriminator, and calculate layer by layer along the network structure to obtain the output of the discriminator, that is, the authenticity of the sample;

[0076] 3.4) Input the denoised high-resolution range image x2 generated by the generator in 3.2) into the recognizer, and calculate layer by layer along the network structure to obtain the output of the recognizer, that is, the target category;

[0077] 3.5) Set the loss function L G (D, G):

[0078]

[0079] where, denotes the expectation operation, D denotes the discriminator, G denotes the generator, G(x1) denotes the denoised high-resolution range image x2, λ is the regularization term coefficient, denotes the L1 regularization term, denotes the L2 regularization term, α is the proportionality coefficient of the L1 regularization term and the L2 regularization term, β is the recognition loss coefficient, denotes the recognizer loss function, t k denotes the k-th element in the true category vector of the target, y k denotes the k-th element in the output vector of the recognizer, K denotes the number of target categories;

[0080] 3.6) Set the loss function L D (D, G):

[0081]

[0082]

[0083] where, L GP denotes the gradient penalty term, λ GP denotes the coefficient of the gradient penalty term, denotes the gradient operation, is a point on the straight line, ε is a number randomly drawn from a uniform distribution on [0, 1];

[0084] 3.7) Update the learnable parameter w D in the discriminator, and the update formula is:

[0085]

[0086] in, is the parameter obtained after the current update of the discriminator, is the parameter of the discriminator before updating, η D is the discriminator learning rate, is the loss function L D (D,G) pair gradient;

[0087] 3.8) Use the updated parameters of the discriminator Repeat the calculation process from 3.2) to 3.5) and adjust the learnable parameters w in the generator G and the learnable parameters w in the recognizer R Update, the update formula is:

[0088]

[0089]

[0090] in, is the parameter obtained after the generator is currently updated, is the parameter obtained after the current update of the recognizer, are the parameters of the generator before updating, is the parameter of the recognizer before updating, η G are the learning rates of the generator and discriminator, is the loss function L G (D,G) pair The gradient, is the loss function L G (D,G) pair gradient;

[0091] 3.9) Use the updated parameters of the generator and the updated parameters of the recognizer Repeat 3.2) to 3.7) again, and perform multiple iterations until the loss function L rec After stable convergence, stop iteration and obtain the learnable parameter w in the generator G , the learnable parameters w in the discriminator D , the learnable parameter w in the recognizer R The optimal value of is obtained, and the trained radar high-resolution range image denoising recognition network based on GAN and self-attention is obtained.

[0092] Step 4: Input the test sample set S2 into the trained radar high-resolution range image denoising recognition network based on GAN and self-attention for testing to obtain the recognition result output by the network.

[0093] The effects of the present invention can be illustrated by the following simulation experiments.

[0094] 1. Simulation experiment conditions:

[0095] The data used in the simulation experiment of the present invention are the actual measured aircraft data of a radar of a domestic research institute. The center frequency of the radar is about 5.5 GHz, the signal bandwidth is 400 MHz, and the measured data includes a total of 3 types of target aircraft, namely the medium-sized propeller aircraft "An-26", the small jet aircraft "Jiangzhuang", and the medium-sized jet aircraft "Yak-42". The measured data is divided into different data segments according to the aircraft flight track. Among them, "An-26" and "Jiangzhuang" are each divided into 7 segments, and "Yak-42" is divided into 5 segments. The simulation experiment starts from the first echo of each data segment, samples every 20th echo, and finally selects 2,600 echoes from the 5th and 6th segments of "An-26", 2,600 echoes from the 6th and 7th segments of "Jiangzhuang", and 2,198 echoes from the 2nd and 5th segments of "Yak-42" as training data; 6,256 echoes from the 1st - 4th and 7th segments of "An-26", 6,500 echoes from the 1st - 5th segments of "Jiangzhuang", and 3,900 echoes from the 1st and 3rd - 4th segments of "Yak-42" as test data. All echoes are vectors of 1×256.

[0096] Dataset 1: Perform Fourier transform, 2-norm normalization, and centroid alignment on the training data and test data to obtain the high-resolution range profiles HRRP without noise, which are used as the training sample set D1 and the test sample set D2 respectively;

[0097] Dataset 2: Add Gaussian white noise with a signal-to-noise ratio of 15 dB to the training data and test data. Perform Fourier transform, 2-norm normalization, and centroid alignment on the data with added noise to obtain the high-resolution range profiles HRRP with noise in a 15 dB environment. Combine them with the HRRP without noise in the training sample set D1 and the test sample set D2 respectively, as the training sample set D3 and the test sample set D4;

[0098] Dataset 3: Add Gaussian white noise with a signal-to-noise ratio of 10 dB to the training data and test data. Perform Fourier transform, 2-norm normalization, and centroid alignment on the data with added noise to obtain the high-resolution range profiles HRRP with noise in a 10 dB environment. Combine them with the HRRP without noise in the training sample set D1 and the test sample set D2 respectively, as the training sample set D5 and the test sample set D6;

[0099] Dataset 4: Add Gaussian white noise with a signal-to-noise ratio of 5 dB to the training data and test data. Perform Fourier transform, 2-norm normalization, and centroid alignment on the data with added noise to obtain the high-resolution range profiles HRRP with noise in a 5 dB environment. Combine them with the HRRP without noise in the training sample set D1 and the test sample set D2 respectively, as the training sample set D7 and the test sample set D8.

[0100] The hardware platform for the simulation experiment is an Intel Xeon E5-2683@2.00GHz CPU, 64GB RAM, and an NVIDIA GeForce GTX1080 Ti GPU. The software platform for the simulation experiment is Matlab 2016, Python 3.6, and TensorFlow 1.8.

[0101] 2. Content and Result Analysis of the Simulation Experiment:

[0102] Simulation Experiment 1: Using Dataset 1, the existing attention recurrent neural network method and the recognizer in the present invention are respectively applied to perform target recognition on the noiseless HRRP.

[0103] Simulation Experiment 2: Using Dataset 2, Dataset 3, and Dataset 4, the existing attention recurrent neural network method, the recognizer in the present invention, and the method of the present invention are respectively applied to perform target recognition on the noisy HRRP in environments with signal-to-noise ratios of 15 dB, 10 dB, and 5 dB.

[0104] The recognition accuracies of these three methods in the four environments of noiseless and signal-to-noise ratios of 15 dB, 10 dB, and 5 dB are calculated respectively through the following formula:

[0105]

[0106] Among them, a represents the recognition accuracy of the test sample set, N represents the number of samples in the test sample set, h(·) represents the recognition discrimination function, t (i) represents the true category of the i-th test sample in the test sample set, and y (i) represents the network output result corresponding to the i-th test sample in the test sample set. When t (i) and y (i) are equal, h(t (i) , y (i) ) is equal to 1; otherwise, h(t (i) , y (i) ) is equal to 0.

[0107] The specific calculation results are shown in Table 1:

[0108] Table 1 Comparison of the recognition rates of the three methods in environments with signal-to-noise ratios of 5 dB, 10 dB, 15 dB, and noiseless

[0109]

[0110] The following conclusions can be drawn from the simulation results:

[0111] The recognition rate of the recognizer proposed by the present invention has increased by 2.72% in a noise-free environment compared to the attention recurrent neural network method of the prior art, indicating that the recognizer proposed by the present invention uses a self-attention mechanism for feature extraction and can improve the recognition rate of radar high-resolution range profiles.

[0112] The recognition rate of the method of the present invention has increased by 1.77%, 2.12%, and 1.88% respectively in the environments with signal-to-noise ratios of 5 dB, 10 dB, and 15 dB compared to the recognizer in the present invention, indicating that in a low signal-to-noise ratio environment, the GAN-based noise reduction module can effectively improve the recognition rate of the network. Moreover, the method of the present invention uses a coupled loss function, enabling the generator to retain features useful for recognition when denoising noisy HRRPs, and achieving better recognition performance in a low signal-to-noise ratio environment.

Claims

1. A radar high-resolution range profile noise reduction and recognition method based on GAN and self-attention, characterized in that, Including: (1) Divide and process the radar echoes of three types of aircraft targets to generate a training sample set S1 and a test sample set S2; (2) Construct a radar high-resolution range profile noise reduction and recognition network: Build a generator, which is composed of five convolutional layers and five deconvolutional layers connected in sequence. Each convolutional layer has a LeakyReLU operation, and each deconvolutional layer is followed by a LeakyReLU and skip connection operation, which is used to reduce the noise of the noisy HRRP and generate a high signal-to-noise ratio HRRP after noise reduction; Build a discriminator, which is composed of five convolutional layers, a flatten layer and two fully connected layers connected in sequence. Each convolutional layer is followed by a LeakyReLU operation, which is used to assist the generator to improve the noise reduction performance; Build a recognizer, which includes a sliding window layer, a position encoding layer, three fully connected layers, five groups of feature extractors, a flatten layer and a softmax classifier. The sliding window layer, the position encoding layer, the third fully connected layer, the five groups of feature extractors, the flatten layer, the fourth fully connected layer, the fifth fully connected layer, and the softmax classifier are connected in sequence, which is used to extract features from the high signal-to-noise ratio HRRP and give the target recognition result; Connect the generator to the discriminator and the recognizer respectively to form a radar high-resolution range profile noise reduction and recognition network based on the generative adversarial network GAN and self-attention; (3) Input the training sample set S1 into the noise reduction and recognition network constructed in (2), and perform adversarial training on it through the backpropagation algorithm to obtain a trained radar high-resolution range profile noise reduction and recognition network based on GAN and self-attention; The adversarial training of the radar high-resolution range profile noise reduction and recognition network based on GAN and self-attention through the backpropagation algorithm is realized as follows: (6a) Initialize all learnable parameters in the radar high-resolution range profile noise reduction and recognition network based on GAN and self-attention; (6b) Input the noisy high-resolution range profile x1 in the training sample set S1 into the generator, and calculate layer by layer along the network structure to generate a high-resolution range profile x2 after noise reduction; (6c) Connect the noise-free high-resolution range profile x3 in the training sample set S1 with the noisy high-resolution range profile x1 to obtain a real sample r = [x3, x1]. Connect the high-resolution range profile x2 after noise reduction in (6b) with the noisy high-resolution range profile x1 to obtain a fake sample f = [x2, x1]. Input the real sample r and the fake sample f into the discriminator, and calculate layer by layer along the network structure to obtain the output of the discriminator; (6d) Input the high-resolution range profile x2 after noise reduction in (6b) into the recognizer, and calculate layer by layer along the network structure to obtain the output of the recognizer, that is, the target category; (6e) Set the loss function L of the generator G (D, G): Among them, denotes the expectation operation, D denotes the discriminator, G denotes the generator, and λ is the regularization term coefficient. denotes the L1 regularization term. denotes the L2 regularization term, α is the proportionality coefficient between the L1 regularization term and the L2 regularization term, and β is the recognition loss coefficient. denotes the recognition loss function, and t k denotes the k-th element in the true class vector of the target, and y k denotes the k-th element in the output vector of the recognizer, and K denotes the number of target categories. (6f) Set the loss function L of the discriminator D (D, G): Among them, L GP represents the gradient penalty term, and λ GP represents the coefficient of the gradient penalty term, is a point on the straight line, and ε is a number randomly drawn from a uniform distribution on [0, 1]; (6g) Update the learnable parameter w in the discriminator D The update formula is as follows: Among them, are the parameters obtained after the current update of the discriminator, are the parameters of the discriminator before the update, and η D is the learning rate of the discriminator, is the loss function L D (D, G) pair gradient; (6h) Use the updated parameters of the discriminator Perform the calculation process from (6b) to (6e) again, and then update the learnable parameter w in the generator G and the learnable parameter w in the recognizer R The update formula is as follows: Among them, are the parameters obtained after the current update of the generator, are the parameters obtained after the current update of the discriminator, are the parameters of the generator before the update, are the parameters of the discriminator before the update, η G is the learning rate of the generator and the discriminator, is the loss function L G (D, G) pair gradient, is the loss function L G (D, G) pair gradient; (6i) Parameters updated by the generator and the parameters updated by the recognizer Repeat steps (6b) to (6g) multiple times for iterative updates until the loss function L rec converges stably. Then stop the iteration to obtain the learnable parameter w G in the generator, the learnable parameter w D in the discriminator, and the learnable parameter w R in the recognizer. The optimal values are obtained to get the trained radar high-resolution range profile denoising and recognition network based on GAN and self-attention. (4) Input the test sample set S2 into the trained radar high-resolution range profile noise reduction and recognition network based on GAN and self-attention for testing to obtain the target recognition result output by the network.

2. The method according to claim 1, wherein in (1), the radar echoes of three types of aircraft targets are processed and divided to generate a training sample set S1 and a test sample set S2, which is realized as follows: The original radar echoes of three types of aircraft targets, namely An-26 aircraft, Citation aircraft, and Yak-42 aircraft, are obtained separately, and Gaussian white noise with a fixed signal-to-noise ratio is added to each of them to obtain the noisy radar echoes of each type; The Fourier transform is performed on the original radar echoes and the noisy radar echoes of each type respectively to obtain the noise-free high-resolution range profiles (HRRPs) of the radar and the noisy HRRPs of the radar. Then, the 2-norm normalization and centroid alignment preprocessing are performed on the noise-free HRRPs and the noisy HRRPs respectively; According to the flight tracks of the three types of aircraft, the 5th and 6th segments of the "An-26" aircraft, the 6th and 7th segments of the "Citation" aircraft, and the 2nd and 5th segments of the "Yak-42" aircraft are selected. The preprocessed noise-free HRRPs and noisy HRRPs corresponding to the six segments of radar echoes are used as the training sample set S1; the preprocessed noise-free HRRPs and noisy HRRPs corresponding to the remaining radar echo data segments are used as the test sample set S2.

3. The method according to claim 1, wherein, The parameter settings of each layer of the generator in (2) are as follows: The convolutional kernel sizes of the five convolutional layers and the five deconvolutional layers are both 15; The number of convolutional kernels in the first convolutional layer and the fourth deconvolutional layer are both 16; The number of convolutional kernels in the second convolutional layer, the third convolutional layer, the second deconvolutional layer, and the third deconvolutional layer are both 32; The number of convolutional kernels in the fourth convolutional layer, the fifth convolutional layer, and the first deconvolutional layer are both 64; The number of convolutional kernels in the fifth deconvolutional layer is 1.

4. The method according to claim 1, wherein The parameter settings of each layer of the discriminator in (2) are as follows: The convolutional kernel sizes of the five convolutional layers are both 15; The number of convolutional kernels in the sixth convolutional layer is 16; The number of convolutional kernels in the seventh convolutional layer and the eighth convolutional layer are both 32; The number of convolutional kernels in the ninth convolutional layer and the tenth convolutional layer are both 64; The number of neurons in the first fully connected layer is 8; The number of neurons in the second fully connected layer is 1.

5. The method according to claim 1, wherein The functions and structural parameters of each layer of the recognizer in (2) are as follows: The sliding window layer is used to perform sliding window on the high signal-to-noise ratio HRRP output by the generator according to a window length d = 6 and a step size to convert it into a sequence form HRRP with a sequence length of T = 84 as the output; The position encoding layer is used to generate T = 84 position encodings using the cosine function, and its cosine function is expressed as follows: PE(pos,j) = cos(pos / 10 2j / T' ) where pos ∈ [1, T'] represents the serial number of the currently calculated position encoding in the sequence, and j ∈ [1, d] represents that the currently calculated is the jth element of the position encoding. The sum of these T' position encodings and the output of the sliding window layer by corresponding terms is the output of this position encoding layer; The five groups of feature extractors are used to extract the dependency relationships between sequence data. Each group of feature extractors consists of a self-attention layer, a residual connection layer, a normalization layer, a feed-forward layer, a residual connection layer, and a normalization layer cascaded in sequence; The input of this self-attention layer is For the first group of feature extractors, X represents the output of the position encoding layer. For the other four groups of feature extractors, X represents the output of the previous group of feature extractors. The output of this self-attention layer is In the formula: Where are all learnable parameter matrices. Set d model = d k = d v = 128; This feed-forward layer contains two fully connected layers, and the number of neurons in them are 512 and 128 respectively; The number of neurons in the third, fourth, and fifth fully connected layers are 128, 3072, and 3 respectively.

Images