HRRP sample generation method based on SACGAN model

Abstract

The present invention discloses a HRRP sample generation method based on the SACGAN model, comprising: generating a small sample training set based on radar echoes; wherein the small sample training set includes HRRP samples of multiple categories, each HRRP sample having a corresponding category label; constructing a SACGAN model including a generator, a discriminator, and an auxiliary classifier based on self-attention and spectral normalization; inputting the HRRP samples and category labels into the SACGAN model, and training the SACGAN model based on a hinge loss function; and using the trained SACGAN model to generate a sample expansion training set. This method not only solves the problem of poor quality of generated HRRP samples caused by gradient explosion or gradient vanishing in the discriminator when performing HRRP sample expansion using the existing CWGAN model, but also solves the problem of poor quality of generated samples when performing HRRP sample expansion using the CACGAN model.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a HRRP sample generation method based on a SACGAN model. Background Art

[0002] A high-resolution range profile (HRRP) is a radar image of a target in a wideband radar, where the target's echo signal can be viewed as the vector sum of all scattered echoes within each resolution range unit. HRRP is one-dimensional information, containing characteristic information such as the target's geometric structure and the energy distribution of scattering points. Compared to two-dimensional echo signals (SAR and ISAR), HRRP is easier to acquire, store, and process, making it highly valuable for radar target recognition and classification. However, establishing an HRRP recognition database for non-cooperative enemy targets makes it difficult for radars to detect and continuously track them, resulting in a lack of sufficient HRRP samples. When using samples from this incomplete HRRP recognition database as a training set for a recognition system, the small number of input HRRP samples leads to significant differences between samples within and outside the database. Consequently, the features extracted by the recognition system fail to represent the target's essential characteristics, thus impacting the recognition performance and generalization ability of the classification system.

[0003] In his paper "Research and Implementation of Radar One-Dimensional Range Profile Target Recognition Methods Based on Deep Learning" (Master's thesis at the University of Electronic Science and Technology of China, June 2019), Si Jinxiu disclosed a method for expanding the sample size of the HRRP recognition database based on the Conditional Wasserstein Generative Adversarial Network (CWGAN). This method first preprocesses the acquired HRRP data and divides it into training and test sets. Then, a CWGAN network structure consisting of two modules, the generator and the discriminator, is constructed. The loss functions of the generator and discriminator are then designed, and the loss function of the discriminator module is optimized using the weight clipping method. Finally, the CWGAN model is used to generate HRRP data to enhance the original dataset. The disadvantage of this method is that using the weight clipping method to optimize the loss function of the discriminator in the CWGAN model can cause the discriminator to experience gradient explosion or gradient vanishing, resulting in poor quality of the HRRP data generated by the CWGAN model.

[0004] In her patent application, "HRRP Recognition Database Sample Expansion Method Based on CACGAN" (Patent Application No.: 202110283773.5, Publication No.: 112784930A), Ma Peiwen discloses a method for expanding the HRRP recognition database using a generative adversarial network (CACGAN). This method first concatenates samples with their corresponding class labels and uses them as input to the CACGAN. Then, during the network architecture design phase, a conditional auxiliary classification generative network consisting of a generator, a discriminator, and an auxiliary classifier is constructed. A gradient penalty term is added to the discriminator's loss function, and the cross-entropy loss function is used to calculate the auxiliary classifier's loss. Finally, the CACGAN model is used to generate HRRP samples of different categories for database sample expansion. A drawback of this method is that, because the CACGAN model focuses more on the separability of generated samples, even when separability is good, the generated samples have low diversity and poor quality. This method fails to balance the separability and diversity of the generated samples, resulting in low recognition performance for classification systems trained using the generated HRRP data to expand the recognition database. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a HRRP sample generation method based on the SACGAN model to solve the problem that when the CWGAN model is used to expand the HRRP recognition database samples, the discriminator has gradient explosion or gradient vanishing, resulting in poor discriminator stability and poor quality of the generated HRRP samples; and when the CACGAN model is used to expand the HRRP recognition database samples, without balancing the separability and quality of the generated samples, the generated samples have strong separability but poor quality, and the recognition performance of the classifier trained after the recognition database samples are expanded is low. The technical problems to be solved by the present invention are achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides a method for generating HRRP samples based on a SACGAN model, comprising:

[0007] Step 1: Generate a small sample training set based on radar echoes; wherein the small sample training set includes HRRP samples of multiple categories, and each HRRP sample has a corresponding category label;

[0008] Step 2: Construct a SACGAN model consisting of a generator, a discriminator, and an auxiliary classifier based on self-attention and spectral normalization;

[0009] Step 3: Input the HRRP sample and the category label into the SACGAN model, and train the SACGAN model based on the hinge loss function;

[0010] Step 4: Use the trained SACGAN model to generate samples to expand the training set.

[0011] In a second aspect, the present invention provides a radar target recognition method, comprising:

[0012] Constructing an HRRP sample set using the method described in the above embodiment;

[0013] Using the HRRP sample set to train a network model of a radar target recognition system until the network converges;

[0014] The sample to be identified is input into the trained radar target recognition system network to obtain the recognition result.

[0015] Beneficial effects of the present invention:

[0016] In the HRRP sample generation method based on the SACGAN model provided by the present invention, a SACGAN model including a generator, a discriminator and an auxiliary classifier is first constructed based on self-attention and spectral normalization; on the one hand, the auxiliary classifier introduced in the model uses category labels to assist in training, so that when the number of training set samples is small, the network has less deviation in feature extraction and feature selection of the data, and the generated samples are of higher quality; on the other hand, a self-attention module is introduced in the generator and discriminator, which can capture the global information of the sample, focus on the important features of the data, effectively learn key information, and improve the efficiency and performance of the network, so that the sample quality generated by the model is higher; at the same time, the use of the spectral normalization method also makes the discriminator satisfy the Lipschitz constraint, thereby limiting the gradient of the discriminator to a constant range, and a better gradient can be sent back to the generator, solving the gradient vanishing and gradient exploding problems of the GAN model; finally, the hinge loss function is used in the network training process, and only samples that are not reasonably judged to be true or false will affect the gradient, which can make the model training process more stable. Therefore, compared with the existing HRRP sample generation method, the HRRP samples generated by the method proposed in the present invention are of higher quality, thereby improving the accuracy of target recognition of the radar system.

[0017] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 1 is a flow chart of a method for generating HRRP samples based on a SACGAN model provided by an embodiment of the present invention;

[0019] Figure 2 2 is a schematic diagram of the structure of the generator in the SACGAN model provided by an embodiment of the present invention;

[0020] Figure 3Schematic diagram of the structure of the discriminator and auxiliary classifier in the SACGAN model provided by an embodiment of the present invention;

[0021] Figure 4 The present invention provides a flow chart of a radar target recognition method. DETAILED DESCRIPTION

[0022] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0023] Example 1

[0024] See Figure 1 , Figure 1 1 is a flow chart of a method for generating HRRP samples based on the SACGAN model provided by an embodiment of the present invention, which includes:

[0025] Step 1: Generate a small sample training set based on radar echoes; the small sample training set includes HRRP samples of multiple categories, and each HRRP sample has a corresponding category label.

[0026] Specifically, step 1 includes:

[0027] 11) Extracting HRRP samples of multiple categories from the radar echo along the range dimension on the radar line of sight.

[0028] 12) Set a category label for each HRRP sample, and combine all processed samples and corresponding category labels to form a small sample training set.

[0029] Optionally, as an implementation, this embodiment sets a category label for the HRRP sample as follows:

[0030] The class label of each sample with the class number 1 in the small sample training set is recorded as y1, the class label of each sample with the class number 2 is recorded as y2, ..., and the class label of each sample with the class number U is recorded as y U , y1 takes the value of 1, y2 takes the value of 2, ..., y U The value is U, which represents the total number of category labels in the small sample training set.

[0031] Step 2: Based on self-attention and spectral normalization, a SACGAN (Self-attention Auxiliary Classifier Generative Adversarial Networks) is constructed, which includes a generator, a discriminator, and an auxiliary classifier.

[0032] 21) Build a generator consisting of a first fully connected layer, four first convolutional modules, two first self-attention modules, and a first output layer.

[0033] See Figure 2 , Figure 2 : This is a schematic diagram of the structure of the generator in the SACGAN model provided by an embodiment of the present invention; wherein, the first fully connected layer, the first first convolution module, the first first attention module, the second first convolution module, the second first attention module, the third first convolution module, and the fourth first convolution module are connected in sequence to form the feature extraction part of the generator.

[0034] Optionally, in this embodiment, the first convolution module includes an upsampling layer, a convolution layer, a spectral normalization layer, and a ReLU layer in sequence; wherein the sliding step size of the upsampling layer is set to 2, the convolution kernel size is set to 1×3, and the number of feature map channels is 512, 256, 128, and 64, respectively. The convolution kernel size used by QKV in the first self-attention module is all 1×1. The first output layer consists of a fully connected layer and Tanh, and the number of nodes in the fully connected layer is set to 4096.

[0035] 22) Build a discriminator consisting of a second fully connected layer, five second convolutional modules, two second self-attention modules, and a second output layer.

[0036] See Figure 3 , Figure 3 This is a structural diagram of the discriminator and auxiliary classifier in the SACGAN model provided by an embodiment of the present invention; wherein, the first second convolution module, the second second convolution module, the third second convolution module, the first second attention module, the fourth second convolution module, the second second attention module and the fifth second convolution module are connected in sequence to form the feature extraction part of the discriminator.

[0037] Optionally, in this embodiment, the second convolution module includes a convolution layer, a spectral normalization layer, and a LeakyReLU layer in sequence. The convolution kernel size is set to 1×3, and the number of feature map channels is 1, 32, 64, 128, and 256, respectively. The convolution kernel size used by QKV in the second self-attention module is 1×1. The second output layer consists of a fully connected layer and a Tanh layer, and the number of nodes in the fully connected layer is set to 4096.

[0038] 23) Build an auxiliary classifier, which includes a feature extraction part and a third output layer; the third output layer includes a fully connected layer and a softmax.

[0039] Optionally, in this embodiment, the feature extraction part of the auxiliary classifier is shared with the discriminator, such as Figure 3 shown.

[0040] 24) Combine the generator, discriminator, and auxiliary classifier to obtain the SACGAN model.

[0041] The SACGAN model constructed in the present invention, on the one hand, introduces an auxiliary classifier and uses category labels to assist in training, so that when the number of training set samples is small, the network's feature extraction and feature selection of the data are less biased, and the generated samples are of higher quality. On the other hand, a self-attention module is introduced in the generator and discriminator, which can capture the global information of the sample, focus on the important features of the data, effectively learn key information, and improve the efficiency and performance of the network, so that the sample quality generated by the model is higher. At the same time, the use of the spectral normalization method also makes the discriminator satisfy the Lipschitz constraint, thereby limiting the gradient of the discriminator to a constant range, and a better gradient can be transmitted back to the generator, solving the gradient vanishing and gradient exploding problems of the GAN model.

[0042] Step 3: Input the HRRP samples and category labels into the SACGAN model and train the SACGAN model based on the hinge loss function.

[0043] It is understandable that before training the constructed SACGAN model, it is necessary to set the training parameters.

[0044] Preferably, in this embodiment, the learning rate of the generator is set to 0.0002 and the learning rate of the discriminator is set to 0.0005 to train the SACGAN model.

[0045] Specifically, step 3 includes:

[0046] 31) Randomly generate a number of labeled embedded noise samples based on Gaussian distribution and use them as input to the generator in the SACGAN model to obtain a generated sample set of the specified category.

[0047] First, N noise samples are randomly sampled from the Gaussian distribution, N category labels are randomly generated, and the N category labels are encoded and embedded into the noise samples to obtain embedded noise samples.

[0048] Then, the embedded noise sample is used as the input of the generator to generate samples of the specified category to form a generated sample set.

[0049] 32) The generated samples in the generated sample set and the HRRP samples in the small sample training set, that is, the real samples, are input into the discriminator of the SACGAN model to output the probability of each sample being judged as a real sample; at the same time, the generated samples in the generated sample set and the HRRP samples in the small sample training set are input into the auxiliary classifier of the SACGAN model to output the probability of each sample being classified into each category.

[0050] 33) Calculate the loss function value of the discriminator and the loss function value of the generator in the SACGAN model at the current iteration respectively.

[0051] Specifically, the calculation formula of the loss function value of the discriminator is:

[0052]

[0053] Among them, L D Represents the loss value of the discriminator in the current iteration, which is composed of the discriminant loss and the classification loss L c Composition, x represents a set of real samples sampled from the small sample training set, p r represents the distribution of real data, p z Represents Gaussian distribution, y represents the category label corresponding to each sample, z represents a set of noise samples sampled from the Gaussian distribution, C(y|x) represents the probability that the auxiliary classifier correctly classifies the input real sample, G(z|y) represents a set of generated samples generated by the generator, C(y|G(z|y)) represents the probability that the classifier correctly classifies the input generated sample, D(x|y) represents the probability that the discriminator judges the input real sample as true, D(G(x|y)) represents the probability that the discriminator judges the input generated sample as true, and E(·) represents the expectation operation;

[0054] The calculation formula of the generator's loss function value is:

[0055]

[0056] Among them, L G represents the loss value of the generator in the current iteration process, z represents a set of noise samples sampled from the Gaussian distribution, y represents the category label corresponding to each sample, and p z Represents Gaussian distribution, G(z|y) represents a set of generated samples generated by the generator, C(y|G(y|z)) represents the probability that the classifier correctly classifies the input generated samples, and E(·) represents the expectation processing.

[0057] 34) Based on the Adam method, the network parameters are updated in sequence using the discriminator loss value and the generator loss value in the SACGAN model at the current iteration until convergence. After training is completed, the network parameters of the generator in the SACGAN model are saved.

[0058] This embodiment uses the hinge loss function during the network training process. Only samples that are not reasonably judged to be true or false will affect the gradient, which can make the model training process more stable.

[0059] Step 4: Use the trained SACGAN model to generate samples to expand the training set.

[0060] 41) Use the trained network parameters to initialize the generator in the SACGAN model.

[0061] 42) Randomly generate noise samples from the Gaussian distribution with the same number of targets of each category in the small sample training set, and input them into the generator together with the corresponding category labels to generate a generated sample set.

[0062] 43) Combine the generated sample set with the small sample training set to form an expanded sample training set.

[0063] In the HRRP sample generation method based on the SACGAN model provided by the present invention, a SACGAN model including a generator, a discriminator and an auxiliary classifier is first constructed based on self-attention and spectral normalization; on the one hand, the auxiliary classifier introduced in the model uses category labels to assist in training, so that when the number of training set samples is small, the network has less deviation in feature extraction and feature selection of the data, and the generated sample quality is higher; on the other hand, a self-attention module is introduced in the generator and discriminator, which can capture the global information of the sample, focus on the important features of the data, effectively learn key information and improve the efficiency and performance of the network, so that the sample quality generated by the model is higher; at the same time, the spectral normalization method is used to make the discriminator meet the Lipschitz constraint, thereby limiting the gradient of the discriminator to a constant range, and a better gradient can be returned to the generator, solving the gradient vanishing and gradient explosion problems existing in the GAN model; finally, the hinge loss function is used in the network training process, and only samples that are not reasonably judged to be true or false will affect the gradient, which can make the model training process more stable. Therefore, compared with the existing HRRP sample generation method, the HRRP sample generated by the method proposed in the present invention is of higher quality.

[0064] Example 2

[0065] Based on the above embodiment 1, this embodiment provides a radar target recognition method. Figure 4 , Figure 4 1 is a flow chart of a radar target recognition method provided by an embodiment of the present invention, which includes:

[0066] S1: Construct HRRP sample set.

[0067] S2: Use the HRRP sample set to train the network model of the radar target recognition system until the network converges;

[0068] S3: Input the sample to be identified into the trained radar target recognition system network to obtain the recognition result.

[0069] Specifically, step S1 of this embodiment uses the method provided in the first embodiment to construct an HRRP sample set, which will not be described in detail in this embodiment. Therefore, the HRRP samples generated by this method are of higher quality, thereby improving the accuracy of target recognition in the radar system.

[0070] Example 3

[0071] The effects of the present invention are further illustrated below through simulation experiments.

[0072] 1. Simulation conditions:

[0073] The hardware platform of the simulation experiment of the present invention is: Intel i7-10700 2.9GHz, memory 16GB, operating system is Windows 10, and Python version is 3.9.

[0074] 2. Simulation content and result analysis:

[0075] This experiment uses the present invention and the existing CACGAN-based HRRP recognition database sample expansion method to generate HRRP data respectively, and then uses the generated HRRP data to expand the samples in the small sample set to obtain the expanded training set after the expansion of the present invention and the expanded training set after the expansion of the CACGAN model.

[0076] Next, we constructed a CNN classifier recognition system to verify the quality of HRRP samples in the expanded training set of the present invention and the training set expanded using the CACGAN model. We input samples from the small sample training set, the training set expanded using the present invention's method, and the training set expanded using the CACGAN model into the CNN classifier recognition system to obtain a trained CNN classifier. We then input samples from the test set generated by the present invention's simulation experiment into the trained CNN classifier, outputting the predicted category for each sample in the test set.

[0077] The existing technology used in the simulation experiment is a CACGAN-based HRRP recognition database sample expansion method, which refers to the HRRP recognition database sample expansion method based on the generative adversarial network CACGAN disclosed in the patent document "HRRP recognition database sample expansion method based on CACGAN" applied by Ma Peiwen (patent application number: 202110283773.5, application publication number: 112784930A). This method is used to generate HRRP data, add the generated HRRP data to the recognition database, and complete the sample expansion of the recognition database. It is referred to as the HRRP recognition database sample expansion method based on CACGAN.

[0078] The recognition database and training set used in the simulation experiments of this invention are HRRP electromagnetic simulation data for three types of aircraft. The small sample training set contains 200 HRRP data points for type 1, 200 HRRP data points for type 2, and 200 HRRP data points for type 3. The test sample set contains 1600 HRRP data points for type 1, 1600 HRRP data points for type 2, and 1600 HRRP data points for type 3. Each HRRP sample contains 256 range cells.

[0079] In the simulation experiment, the present invention used the sample expansion method to generate HRRP data to obtain the generated dataset of the present invention. The generated dataset contained 1400 HRRP data of category 1, 1400 HRRP data of category 2, and 1400 HRRP data of category 3. The generated dataset of the present invention and the CACGAN generated dataset were then used to complete sample expansion of the small sample training set, obtaining the expanded training set after the expansion of the present invention and the expanded training set after the expansion using the CACGAN model.

[0080] A five-layer CNN classifier recognition system was built, with the following structure: convolutional layer 1, convolutional layer 2, convolutional layer 3, fully connected layer 1, and fully connected layer 2. The number of feature maps in convolutional layers 1 to 3 was set to 32, 64, and 128, respectively. The convolution kernel size was set to 1×9, the convolution kernel sliding step size was set to 1, the pooling downsampling kernel size was set to 1×2, and the downsampling kernel sliding step size was set to 2. The input dimensions of fully connected layers 1 and 2 were 4096 and 128, respectively, and the output dimensions were 128 and 3, respectively.

[0081] The small sample training set, the expanded training set of the present invention, and the expanded training set of CACGAN are input into the CNN classifier respectively. After 300 iterations of training, three trained CNN classifiers are obtained. The three CNN classifiers are used to predict the category of each sample in the test set respectively. Then, the ratio of the number of test samples whose categories predicted by the three CNN classifiers for each sample in the test set are consistent with the category of the sample to the total number of test samples is calculated to obtain three target recognition accuracy rates. The higher the target recognition accuracy rate, the higher the recognition performance of the CNN classifier, and the more complete the azimuth angle of the HRRP sample in the expanded training set.

[0082] The results of the above three target recognition accuracy are shown in Table 1.

[0083] Table 1 Target recognition accuracy list

[0084] training set Recognition rate (%) Small sample training set 89.37 CACGAN expanded training set 90.04 The expanded training set of the present invention 91.33

[0085] The simulation results in Table 1 show that the recognition performance of the CNN trained with sample expansion using the present invention is superior to that of the CNN trained with sample expansion using the existing method. This demonstrates that the HRRP samples generated by the present invention are of higher quality and that the classification system of the CNN trained with sample expansion using the generated HRRP has high recognition performance for small sample training sets.

[0086] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A HRRP sample generation method based on the SACGAN model, characterized in that: include: Step 1: Generate a small sample training set based on radar echoes; wherein the small sample training set includes HRRP samples of multiple categories, and each HRRP sample has a corresponding category label; Step 2: Based on self-attention and spectral normalization, a SACGAN model consisting of a generator, a discriminator, and an auxiliary classifier is constructed, including: 21) Build a generator consisting of a first fully connected layer, four first convolutional modules, two first self-attention modules, and a first output layer; The first fully connected layer, the first first convolution module, the first first attention module, the second first convolution module, the second first attention module, the third first convolution module and the fourth first convolution module are connected in sequence to form a feature extraction part of the generator; 22) Build a discriminator consisting of a second fully connected layer, five second convolutional modules, two second self-attention modules, and a second output layer; The first second convolution module, the second second convolution module, the third second convolution module, the first second attention module, the fourth second convolution module, the second second attention module and the fifth second convolution module are connected in sequence to form a feature extraction part of the discriminator; 23) Building an auxiliary classifier, which includes a feature extraction part and a third output layer; the third output layer includes a fully connected layer and a softmax layer; 24) combining the generator, the discriminator, and the auxiliary classifier to obtain a SACGAN model; Step 3: Input the HRRP sample and the category label into the SACGAN model, and train the SACGAN model based on the hinge loss function; Step 4: Use the trained SACGAN model to generate samples to expand the training set.

2. The HRRP sample generation method based on the SACGAN model according to claim 1, characterized in that: Step 1 includes: 11) extracting HRRP samples of multiple categories from radar echoes along the range dimension on the radar line of sight; 12) Setting a category label for each of the HRRP samples, and combining all processed samples and corresponding category labels to form a small sample training set.

3. The HRRP sample generation method based on the SACGAN model according to claim 1, characterized in that: The first convolution module includes an upsampling layer, a convolution layer, a spectral normalization layer, and a ReLU layer in sequence; wherein the sliding step size of the upsampling layer is set to 2, the convolution kernel size is set to 1×3, and the number of feature map channels is 512, 256, 128, and 64, respectively; The convolution kernel size used by QKV in the first self-attention module is 1×1.

4. The HRRP sample generation method based on the SACGAN model according to claim 1, characterized in that: The second convolution module includes a convolution layer, a spectral normalization layer, and a LeakyReLU layer in sequence; wherein the size of the convolution kernel is set to 1×3, and the number of feature map channels is 1, 32, 64, 128, and 256 in sequence; The convolution kernel size used by QKV in the second self-attention module is 1×1.

5. The HRRP sample generation method based on the SACGAN model according to claim 1, characterized in that: After step 2 and before step 3, it also includes: The learning rate of the generator is set to 0.0002, and the learning rate of the discriminator is set to 0.0005.

6. The HRRP sample generation method based on the SACGAN model according to claim 1, characterized in that: Step 3 includes: 31) Randomly generate a number of labeled embedded noise samples based on Gaussian distribution and use them as input to the generator in the SACGAN model to obtain a generated sample set of the specified category; 32) Inputting the generated samples in the generated sample set and the HRRP samples in the small sample training set into the discriminator of the SACGAN model to output the probability of each sample being judged as a real sample; at the same time, inputting the generated samples in the generated sample set and the HRRP samples in the small sample training set into the auxiliary classifier of the SACGAN model to output the probability of each sample being classified into each category; 33) Calculate the loss function value of the discriminator and the loss function value of the generator in the SACGAN model at the current iteration respectively; 34) Based on the Adam method, the network parameters are updated sequentially using the discriminator loss value and the generator loss value in the SACGAN model at the current iteration until convergence. After training is completed, the network parameters of the generator in the SACGAN model are saved.

7. The HRRP sample generation method based on the SACGAN model according to claim 6, characterized in that: The calculation formula of the loss function value of the discriminator is: ； in, Indicates the loss value of the discriminator in the current iteration, which is composed of the discriminant loss and the classification loss composition, represents a set of real samples sampled from a small sample training set, represents the distribution of real data, represents a Gaussian distribution, Indicates the category label corresponding to each sample, represents a set of noise samples sampled from a Gaussian distribution, represents the probability that the auxiliary classifier correctly classifies the input real sample, represents a set of generated samples generated by the generator, represents the probability that the classifier correctly classifies the input generated sample, It represents the probability that the discriminator judges the input real sample as true, represents the probability that the discriminator judges the input generated sample as true, Indicates the expected operation; The calculation formula of the loss function value of the generator is: ； in, Represents the loss value of the generator during the current iteration, represents a set of noise samples sampled from a Gaussian distribution, Indicates the category label corresponding to each sample, represents a Gaussian distribution, represents a set of generated samples generated by the generator, represents the probability that the classifier correctly classifies the input generated sample, Indicates that the desired processing is requested.

8. The HRRP sample generation method based on the SACGAN model according to claim 6, characterized in that: Step 4 includes: 41) Initialize the generator in the SACGAN model using the trained network parameters; 42) Randomly generate noise samples from the Gaussian distribution with the same number of targets of each category in the small sample training set, and input them into the generator together with the corresponding category labels to generate a generated sample set; 43) Combining the generated sample set with the small sample training set to form an expanded sample training set.

9. A radar target recognition method, characterized in that: include: Constructing an HRRP sample set using the method described in any one of claims 1 to 8; Using the HRRP sample set to train a network model of a radar target recognition system until the network converges; The sample to be identified is input into the trained radar target recognition system network to obtain the recognition result.

Images