A radar high-resolution range profile out-of-database target rejection method based on an adversarial learning one-class classifier

Abstract

The application discloses a radar high-resolution range profile out-of-database target rejection method based on an adversarial learning one-class classifier, and comprises the following steps: step 10, pre-processing the radar high-resolution range profile original sample of an airplane to obtain a training set; step 20, constructing a generator G and a discriminator D based on a GAN network; step 30, constructing a first target function according to the generator G and the discriminator D; step 40, training the generator G and the discriminator D based on the training set and constructing a second target function to obtain a trained generator G and a trained discriminator D; step 50, coupling the trained generator G and an intermediate generator G old with the trained discriminator D to form an adversarial judgment model; and step 60, inputting a to-be-judged sample set into the adversarial judgment model for judgment. The application can make the judgment boundary clearer, and the introduction of a pseudo abnormal sample and the intermediate generator enables the discriminator to better distinguish high / low quality reconstructed samples, thereby improving the judgment performance.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a method for rejecting targets outside a radar high-resolution range image database based on an adversarial learning classifier. Background Technology

[0002] With the continuous improvement of intelligence and informatization in modern warfare, Radar Automatic Target Recognition (RATA) technology has received wider attention. High-resolution range profiles (HRRPs) can be acquired from targets observed by broadband radar. They reflect the distribution of target scattering centers along the radar line of sight and contain rich information on target structural features. Furthermore, HRRPs are easy to acquire, easy to store, and computationally inexpensive, making them important in RATA systems. However, in practical military applications, the targets that RATA systems need to identify are often non-cooperative targets. Obtaining these target samples is difficult, leading to incompleteness in the template database. In such cases, if a target belongs to a category outside the database, classifying it as any category within the database is unreasonable, and it should be rejected (i.e., not belonging to the database).

[0003] Typically, a single-class classifier is used to address the rejection problem. Commonly used single-class classifiers include One-Class Support Vector Machine (OCSVM) and Support Vector Data Description (SVDD). These methods suffer from difficulties in choosing the kernel function form and optimizing the kernel parameters. In addition, the Isolation Forest (IF) model, while computationally efficient, is not suitable for extremely high-dimensional data and is only sensitive to globally sparse points, which also limits its rejection performance.

[0004] Following the successful applications of deep learning in computer vision and pattern recognition, the radar field has also been inspired to apply deep learning to the HRRP target rejection task. For example, Wan J et al. proposed using a convolutional neural network (CNN) as the encoder, utilizing local structural information in the feature space to extract richer features from the data, while using a deconvolutional neural network as the decoder. By measuring the reconstruction error between the input and the decoder, they achieved rejection of out-of-database targets. Ruff et al. proposed the Deep SVDD framework, which also leverages the advantages of deep learning in handling complexly distributed samples, thereby improving rejection performance.

[0005] However, existing traditional methods such as OCSVM and Isolation Forest have problems such as difficulty in choosing kernel function form and optimizing kernel parameters, not being suitable for particularly high-dimensional data and only being sensitive to global sparse points. Moreover, these models are not end-to-end models and their mapping effect on data features is not good enough. Existing deep network-based methods such as Deep SVDD, although they are end-to-end models, only use the loss magnitude for classification and discrimination, and the rejection boundary may not be clear. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a method for rejecting targets outside a radar high-resolution range image database based on an adversarial learning-based classifier. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] A method for rejecting targets outside a radar high-resolution range image database based on an adversarial learning classifier includes the following steps:

[0008] Step 10: Preprocess the original samples of the aircraft's radar high-resolution range image to obtain the training set;

[0009] Step 20: Construct a generator G and a discriminator D based on the GAN network;

[0010] The generator G comprises N encoder layers and M decoder layers connected in sequence; each encoder layer comprises a convolutional layer, a batch normalization layer and an activation function layer connected in sequence; each of the M-1 decoder layers comprises a deconvolutional layer, a batch normalization layer and an activation function layer connected in sequence, and the Mth decoder layer comprises a convolutional layer and an activation function layer.

[0011] The discriminator D includes, in sequence, an input layer, a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer, a second batch normalization layer, a second activation layer, a third convolutional layer, a third batch normalization layer, a third activation layer, a fourth convolutional layer, a fourth batch normalization layer, a fourth activation layer, a fully connected layer, and a sigmoid layer.

[0012] Step 30: Construct a first objective function based on the generator G and the discriminator D;

[0013] Step 40: Train the generator G and the discriminator D based on the training set and construct a second objective function. Update the parameters of the generator G and the discriminator D using the first objective function and the second objective function to obtain the trained generator G and the trained discriminator D. Before the loss of the generator G reaches the preset reconstruction loss, randomly save one generator G as an intermediate generator G. old ;

[0014] Step 50: Combine the trained generator G and the intermediate generator G old Each is coupled with the trained discriminator D to form an adversarial decision model;

[0015] Step 60: Input the sample set to be judged into the adversarial judgment model for judgment;

[0016] If the output probability of the adversarial determination model is greater than the threshold, then the samples in the sample set to be determined are determined to be abnormal samples from outside the database.

[0017] In one embodiment of the present invention, step 10 includes:

[0018] Step 11: Extract temporal features from each original radar high-resolution range image sample of the aircraft to obtain HRRP temporal feature samples;

[0019] Step 12: Perform amplitude normalization and centroid alignment operations on each of the HRRP time-domain feature samples to obtain the training set.

[0020] In one embodiment of the present invention, the expression of the first objective function V1(G,D) is:

[0021]

[0022] Among them, X~P t This indicates that the input sample X was sampled from the training set. Represents noise samples Sampling is performed on a training set with added noise. E[·] represents the expectation operation, G(·) represents the output data of the generator G, and D(·) represents the output data of the discriminator D.

[0023] In one embodiment of the present invention, step 40 includes:

[0024] Step 41: Using the samples in the training set as input samples X, add noise to generate noise samples. The input generator G obtains the reconstructed samples

[0025] Step 42, reconstruct the sample And the input sample X is input to the discriminator D;

[0026] Step 43: Repeat steps 41 and 42 to iteratively train the generator G and discriminator D, and update the parameters of the generator G using the first objective function to obtain the trained generator G; wherein, before the loss of the generator G reaches the preset reconstruction loss, a generator G is randomly saved as an intermediate generator G. old ;

[0027] Step 44: Construct a second objective function V2(G,D) based on the discriminator D; the expression of the second objective function V2(G,D) is:

[0028]

[0029] Where α and β both represent compromise hyperparameters. This represents the reconstructed sample output by the generator G after inputting the input sample X. noise sample Input intermediate generator G old The intermediate samples output are E[·], which represents the expectation operation; G(·) represents the output data of the generator G; and D(·) represents the output data of the discriminator D. Indicates a pseudo-anomaly sample.

[0030] and For two randomly selected different input samples, and For the sample and samples Input intermediate generator G old (·) Output data, This is the generated pixel-level mean;

[0031] Step 45, using the trained generator G and the intermediate generator G old The discriminator D is iteratively trained and its parameters are updated using the second objective function to obtain the trained discriminator D; wherein, the input data of the discriminator D includes: the input sample X, the reconstructed sample X, and the input sample X. The pseudo-anomaly sample and the intermediate sample

[0032] In one embodiment of the present invention, the cross-entropy loss function is converged after steps 43 and 45 are completed.

[0033] In one embodiment of the present invention, the expression for the cross-entropy loss function is:

[0034]

[0035] Where H represents the cross-entropy loss function, Z pre (k) represents the predicted classification probability vector, Z train (k) represents the true probability vector of the samples in the library, k = 1, 2, ..., k represents the class number of the samples in the library, and log represents the logarithm operation to the base 10;

[0036] The v-class samples in the training set are in-database samples, and the wv-class samples in the training set are out-of-database samples.

[0037] The beneficial effects of this invention are:

[0038] This invention introduces a discriminator on top of a deep network model. The adversarial decision model of this invention can map HRRP features to a more separable space and enable the generator to obtain global vision feedback from the output of the discriminator, making the rejection boundary clearer. At the same time, pseudo-abnormal samples and intermediate generators are introduced in the training, which enables the discriminator to better distinguish between high / low quality reconstructed samples, thereby improving the rejection performance.

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0040] Figure 1 A flowchart illustrating a method for rejecting external targets in a high-resolution radar range image database based on an adversarial learning classifier, provided in an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the network structure of the adversarial decision-making model provided in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram illustrating the simulation results of the rejection performance of targets outside the database. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0044] Example 1

[0045] like Figure 1 and Figure 2 As shown, a method for rejecting targets outside a radar high-resolution range image database based on a classifier of adversarial learning is presented. This method is based on a classifier of generative adversarial networks (GANs) and includes the following steps:

[0046] Step 10: Preprocess the original radar high-resolution range images of the aircraft to obtain a training set; the number of categories in the training set is w. Specifically, the preprocessing steps for the original radar high-resolution range images of w types of aircraft include: Steps 11-13.

[0047] Step 11: Extract temporal features from each original radar high-resolution range image sample of class w aircraft to obtain HRRP temporal feature samples.

[0048] In this step, the original high-resolution radar range image is a complex vector, which is formed by arranging the echo envelopes of each range cell sequentially according to the cell number. The modulus of the data of each range cell is taken to obtain the HRRP time-domain feature sample x, that is: x=[|x1|,|x2|,...,|x P |] T P represents the total number of range cells in a single sample, and this time-domain feature contains the amplitude information of the radar high-resolution range image.

[0049] Step 12: Perform amplitude normalization and centroid alignment operations on each HRRP time-domain feature sample to obtain the training set.

[0050] The amplitude normalization operation is specifically performed on the sample set {x}. 1 ,x 2 ,...,x m Each HRRP time-domain feature sample in} Perform the operation as follows:

[0051]

[0052] Where m represents the number of original samples, and n represents the feature dimension of a single sample. It is the i-th feature in the j-th HRRP time-domain feature sample, and ||·||2 is the L2 norm operation; This represents the magnitude of the i-th dimension after the HRRP time-domain feature sample magnitude normalization process.

[0053] The centroid alignment operation is specifically performed on the sample set {x}. 1 ,x 2 ,...,x m Each HRRP time-domain feature sample in} The centroid feature position is obtained by rounding down the calculation using the centroid alignment formula. The sample is then shifted until the feature is at the center. The centroid alignment formula is:

[0054]

[0055] Among them, W b This represents the centroid of the b-th echo signal in the sample set. Let represent the a-th data in the b-th echo signal in the sample set, where a represents the sequence number of the data in the echo signal, 1≤a≤l, and l represents the length of the echo signal in the sample set.

[0056] Step 20: Construct a generator G and a discriminator D based on the GAN network.

[0057] The generator G consists of a denoising autoencoder, which is used to generate high-quality reconstructed samples. The denoising autoencoder includes encoding units and decoding units. The encoding units consist of N encoder layers connected in sequence, and the decoding units consist of M decoder layers connected in sequence. N and M are integers greater than 1.

[0058] Each encoder layer consists of a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence; each of the M-1 decoder layers consists of a deconvolutional layer, a batch normalization layer, and an activation function layer connected in sequence, and the Mth decoder layer consists of a convolutional layer and an activation function layer.

[0059] Preferably, taking N=4 and M=4 as an example, the parameters of each layer of the encoding unit are set as follows:

[0060] The first encoder layer: The input data passes through a convolutional layer, and then the convolutional data is passed to the activation function layer through batch normalization. The output of the activation function layer is then passed to the second encoder layer. The convolutional layer has 1 input channel, 64 filters, uses the ReLU activation function, and has a kernel size of (1*5).

[0061] The second encoder layer: similar to the parameter settings of the first layer, except that the number of input channels of the convolutional layer is set to 64 and the number of filters is set to 128, and the output is passed to the third encoder layer.

[0062] The third encoder layer: similar to the parameter settings of the second layer, except that the number of input channels of the convolutional layer is set to 128 and the number of filters is set to 128, and the output is passed to the fourth encoder layer.

[0063] The fourth encoder layer: similar to the parameter settings of the third layer, it passes the output to the next layer (decoder layer);

[0064] The parameter settings for each layer of the decoding unit are as follows:

[0065] The first decoder layer: The input data passes through a deconvolution layer, and then the deconvolution-processed data is passed to the activation function through batch normalization. The output of the activation function layer is then passed to the second decoder layer. The convolutional layer has 128 input channels, 128 filters, uses the ReLU activation function, and has a kernel size of (1*5).

[0066] The second decoder layer: similar to the parameter settings of the first layer, it passes the output to the third decoder layer;

[0067] The third decoder layer: similar to the parameter settings of the second layer, except that the number of input channels of the convolutional layer is set to 128 and the number of filters is set to 64, and the output is passed to the fourth encoder layer.

[0068] The fourth encoder layer: similar to the parameter settings of the third layer, it passes through a convolutional layer as the final output. The difference is that the number of input channels of the convolutional layer is set to 64, the number of filters is set to 1, and the Tanh activation function is used.

[0069] A 16-layer discriminator D is constructed to distinguish between good and bad reconstruction quality. Discriminator D includes the following layers arranged in sequence: input layer, first convolutional layer, first batch normalization layer, first activation layer, second convolutional layer, second batch normalization layer, second activation layer, third convolutional layer, third batch normalization layer, third activation layer, fourth convolutional layer, fourth batch normalization layer, fourth activation layer, fully connected layer, and sigmoid layer.

[0070] The parameters of each layer of the discriminator D are set as follows: the number of convolutional kernels in the first convolutional layer is set to 64, the number of convolutional kernels in the second to fourth convolutional layers is set to 128, the kernel size is set to 1×5, and the kernel stride is set to 2; the first to fourth activation layers all use the ReLU activation function, and the fifth activation layer uses the Sigmoid activation function; the number of neurons in the fully connected layer is set to 2048.

[0071] Step 30: Construct the first objective function based on the generator G and discriminator D. The expression of the first objective function V1(G,D) is:

[0072]

[0073] Among them, X~P t This indicates that the input sample X was sampled from the training set. Represents noise samples Sampling is performed on a training set with added noise. E[·] represents the expectation operation, G(·) represents the output data of the generator G, and D(·) represents the output data of the discriminator D.

[0074] Step 40: Train the generator G and discriminator D based on the training set and construct a second objective function. Update the parameters of generator G and discriminator D through the first and second objective functions to obtain the trained generator G and discriminator D. Before the loss of generator G reaches the preset reconstruction loss, randomly save one generator G as an intermediate generator G. old .

[0075] In this step, samples of class v in the training set are selected as in-library (normal) samples and labeled as 0. The remaining samples of classes wv are selected as out-of-library (abnormal) samples and labeled as 1. The generator G and the discriminator D read and learn the distribution of in-library (normal) samples for training, and update the parameters through the objective function. The training is carried out in two stages. Specifically, step 40 includes steps 41-45.

[0076] Step 41: Using the samples in the training set as input samples X, add noise to generate noise samples. The input generator G obtains the reconstructed samples

[0077] Step 42, reconstruct the sample Input sample X and input discriminator D;

[0078] Step 43: Repeat steps 41 and 42 to iteratively train the generator G and discriminator D, and update the parameters of the generator G using the first objective function to obtain the trained generator G; wherein, before the loss of the generator G reaches the preset reconstruction loss, a generator G is randomly saved as an intermediate generator G. old ;

[0079] In this embodiment, the first stage includes steps 41-43, where the generator G attempts to generate fake data (reconstructed samples) that appear realistic. This sample, along with the real sample (input sample) X, is then fed into the discriminator D. The discriminator D learns to distinguish between real and fake data, and the result serves as a supervision signal for the generator G. This training is performed until the generator G begins to reconstruct samples with a low reconstruction loss (preset reconstruction loss). The first stage minimizes the loss function L, where L = L G+D +λL R L G+D Let V1(G,D) be the loss function for the first objective function. The loss is the sample reconstruction loss, and λ is the weighted hyperparameter. During the first stage of training, the true reference standard for the discriminator D's output probability is specifically:

[0080]

[0081] That is, when the input is a real sample X, the discriminator output probability should be close to 0; when the input sample is a reconstructed sample from the generator... At that time, the discriminator output probability should be close to 1.

[0082] Randomly save an immature generator from the first training phase as an intermediate generator G. old intermediate generator G old The quality of the generated samples is lower than that of the trained generator G, and the intermediate generator G... old Used for training the second phase.

[0083] Among them, intermediate generator G oldThe system is used to generate low-quality reconstructed samples (intermediate samples) and consists of a denoising autoencoder. The denoising autoencoder includes an encoding unit and a decoding unit. The encoding unit consists of I encoder layers connected in sequence. Each encoder layer includes a convolutional layer, a batch normalization layer, and an activation function layer connected in sequence, where I is an integer greater than 1. The decoding unit consists of K decoder layers connected in sequence. The first K-1 decoder layers each include a deconvolutional layer, a batch normalization layer, and an activation function layer connected in sequence, and the Kth decoder layer includes a convolutional layer and an activation function layer, where K is an integer greater than 1.

[0084] Preferably, taking I=4 and K=4 as an example, the intermediate generator G old The parameter settings for each layer of the encoding unit are as follows:

[0085] The first encoder layer: The input data passes through a convolutional layer, and then the convolutional data is passed through batch normalization to the activation function. The output of the activation function layer is then passed to the second encoder layer. The convolutional layer has 1 input channel, 64 filters, uses the ReLU activation function, and has a kernel size of (1*5).

[0086] The second encoder layer: similar to the parameter settings of the first layer, except that the number of input channels of the convolutional layer is set to 64 and the number of filters is set to 128, and the output is passed to the third encoder layer.

[0087] The third encoder layer: similar to the parameter settings of the second layer, except that the number of input channels of the convolutional layer is set to 128 and the number of filters is set to 128, and the output is passed to the fourth encoder layer.

[0088] The fourth encoder layer: similar to the parameter settings of the third layer, it passes the output to the next layer (decoder layer).

[0089] Intermediate Generator G old The parameter settings for each layer of the decoding unit are as follows:

[0090] The first decoder layer: The input data passes through a deconvolution layer, and then the deconvolution-processed data is passed to the activation function through batch normalization. The output of the activation function layer is then passed to the second decoder layer. The convolutional layer has 128 input channels, 128 filters, uses the ReLU activation function, and has a kernel size of (1*5).

[0091] The second decoder layer: similar to the parameter settings of the first layer, it passes the output to the third decoder layer;

[0092] The third decoder layer: similar to the parameter settings of the second layer, except that the number of input channels of the convolutional layer is set to 128 and the number of filters is set to 64, and the output is passed to the fourth encoder layer.

[0093] The fourth encoder layer: similar to the parameter settings of the third layer, it passes through a convolutional layer as the final output. The difference is that the number of input channels of the convolutional layer is set to 64, the number of filters is set to 1, and the Tanh activation function is used.

[0094] Step 44: Before the start of the second phase of training, construct the second objective function V2(G,D) based on the discriminator D; the expression of the second objective function V2(G,D) is:

[0095]

[0096] Where α and β both represent compromise hyperparameters. This represents the reconstructed sample output by the generator G after inputting the input sample X. noise sample Input intermediate generator G old The intermediate samples output are E[·], which represents the expectation operation; G(·) represents the output data of the generator G; and D(·) represents the output data of the discriminator D. Indicates a pseudo-anomaly sample.

[0097] and For two randomly selected different input samples, and For the sample and samples Input intermediate generator G old (·) Output data, This represents the pixel-level mean value. Pseudo-anomaly samples can be generated using the pseudo-anomaly sample generation module.

[0098] Step 45, using the trained generator G and intermediate generator G old The discriminator D is iteratively trained and its parameters are updated using a second objective function to obtain the trained discriminator D. The input data for discriminator D includes: input sample X, reconstructed sample X, and other parameters. False anomaly samples and intermediate samples

[0099] In this embodiment, the intermediate generator G is used in the second stage of training. oldThe discriminator D is updated by the generator G after the first stage of training. The input of the discriminator D consists of four parts: (1) real sample (input sample) X, (2) high-quality reconstructed sample generated by the generator G. (3) Pseudo-abnormal samples (4) By intermediate generator G old The generated low-quality reconstructed samples (intermediate samples) For the discriminator D, (1) and (2) are high-quality input samples, and (3) and (4) are low-quality input samples. The second stage provides the discriminator D with both high-quality and low-quality reconstruction examples, enabling it to understand the output types that the generator G will produce under abnormal inputs. During the second stage training, the true reference standard for the output probability of the discriminator D is specifically:

[0100]

[0101] That is, when the input is a real sample X or a high-quality sample reconstructed by the generator G. When the input sample is a pseudo-anomaly sample, the discriminator output probability should be close to 0; when the input sample is a pseudo-anomaly sample... Or generated by intermediate generator G old When reconstructing low-quality samples, the discriminator output probability should be close to 1.

[0102] In the training process across two training phases, after each iteration, the cross-entropy loss function is used to calculate the loss between the predicted classification label and the target true class label. Then, the backpropagation algorithm is used to iteratively update the network parameters until the cross-entropy loss function converges, completing the training and yielding the trained generator G and discriminator D. The cross-entropy loss function is as follows:

[0103] The expression for the cross-entropy loss function is:

[0104]

[0105] Where H represents the cross-entropy loss function, Z pre (k) represents the predicted classification probability vector, Z train (k) represents the true probability vector of the samples in the library, k = 1, 2, ..., k represents the class number of the samples in the library, and log represents the logarithm operation to the base 10.

[0106] Step 50, combine the trained generator G and the intermediate generator G old They are coupled with the trained discriminator D to form an adversarial decision model.

[0107] Step 60: Input the sample set to be judged into the adversarial judgment model for judgment;

[0108] If the output probability of the adversarial decision model is greater than the preset threshold, the sample to be decided is determined to be an abnormal sample from outside the database.

[0109] Specifically, step 60 involves inputting the sample set to be judged into the adversarial judgment model for judgment, including steps 61 and 62.

[0110] Step 61: Save the trained model parameters and load the model;

[0111] Step 62: The sample set to be judged includes normal samples from within the library with a label of 0 and abnormal samples from outside the library with a label of 1. The samples are processed by the trained generator G and the trained discriminator D to obtain the discrimination result. The discrimination result OCC is defined as follows:

[0112]

[0113] Where τ is a preset threshold, G(·) represents the output of the trained generator G, and D(·) represents the probability output by the trained discriminator D.

[0114] This invention employs a deep network model to address the challenges of kernel function selection and parameter optimization in existing OCSVM and IF algorithms, which are unsuitable for high-dimensional data. Building upon the deep network model, it incorporates adversarial learning principles by introducing a discriminator module. This module provides the generator with global feedback, resulting in clearer rejection boundaries. Furthermore, the training process is divided into two phases, with the second phase introducing pseudo-anomaly samples and an intermediate generator G. old This allows the discriminator to better distinguish between high and low quality reconstructed samples, thereby improving rejection performance.

[0115] Specifically, in practical applications of rejecting targets outside the radar high-resolution range image database, we are familiar with known samples within the database but not with those outside. Therefore, the adversarial decision-making model of this invention, OG-GAN (OldGenerator-Generative Adversarial Network), is based on a single-class classification system. It inherits the idea of ​​traditional single-class classifier algorithms that only learn features from samples within the database. By learning the descriptions of known class data, it designs a similarity metric to compare the similarity between new samples and that class of data with a set threshold. Compared to conventional kernel functions, using deep networks can better map radar HRRP data to a more separable feature space based on its characteristics, thereby improving rejection performance.

[0116] Unlike conventional autoencoders (AEs) that classify data solely based on the magnitude of the reconstruction loss, this deep network design employs an adversarial learning architecture. This involves introducing a discriminator module, allowing the generator module to receive feedback from its global perspective. Furthermore, the reconstructed samples generated by the generator can be used as pseudo-anomalies to better train the discriminator. In addition, the training is divided into two phases. The second phase introduces an intermediate generator, G, which performed poorly on sample reconstruction during the first phase iterations. old High-quality and low-quality input samples are defined and used to further optimize the discriminator D, so that the discriminator D can more clearly understand the output type that the generator G will produce under abnormal input conditions, thereby making the rejection boundary clearer.

[0117] Example 2

[0118] The effects of Example 1 are further described in conjunction with simulation experiments.

[0119] 1. Simulation conditions:

[0120] In the simulation experiment of this embodiment, the radar database used is the HRRP data of aircraft measured by a certain radar, which includes ten different types of aircraft: Airbus A319, Airbus A320, Airbus A321, Airbus A330-2, Airbus A330-3, Airbus A350-941, Boeing 737-7, Boeing 737-8, Boeing 747-89L and Bombardier CRJ-900. One type of each type is selected as an abnormal sample outside the database, and the remaining nine types are used as normal samples inside the database for the experiment. The training set contains 8,000 samples for each type of aircraft, and the training samples of the nine types of aircraft constitute the total training samples, with a total of 72,000 training samples. The test set contains 6,000 samples for each type of aircraft, and the test samples of the ten types of aircraft constitute the total training samples, with a total of 60,000 test samples.

[0121] The simulation software platform used in this embodiment is PyCharm 2021.3.3, CUDA 10.2, CUDNN 10.2 and PyTorch 1.12.

[0122] 2. Simulation content and result analysis:

[0123] This embodiment's simulation experiment: The simulation experiment is used to verify the out-of-library target rejection performance of the method proposed in this example.

[0124] Reference Figure 3 , Figure 3 This is a schematic diagram of simulation results provided by an example of the present invention for verifying the performance of rejecting targets outside the library. Figure 3The horizontal axis represents the label of the selected anomaly class, and the vertical axis represents the AUC value. The method proposed in this embodiment, along with two existing rejection methods—an OCSVM-based out-of-database target rejection method and an Isolation Forest-based out-of-database target rejection method—were used to reject targets on the aforementioned 10 types of high-resolution range image test samples. A commonly used quantitative indicator of rejection performance is the Area Under an OCR curve (AUC). A larger AUC value indicates a better rejection effect. The AUC values ​​obtained from experiments using the three methods are shown in Tables 1, 2, and 3.

[0125]

[0126] Table 1. Results of the off-database target rejection method of the present invention

[0127]

[0128] Table 2. Results of the OCSVM-based method for rejecting out-of-library targets

[0129]

[0130] Table 3. Results of the out-of-library target rejection method based on isoforest

[0131] Depend on Figure 3 The results show that, regardless of the aircraft category selected as the out-of-pipe anomaly sample, the AUC value of the out-of-pipe target rejection performance obtained by the OG-GAN method of this invention is higher than that of the out-of-pipe target rejection methods based on OCSVM and Isolation Forest, with several values ​​exceeding those of the other two methods by nearly 10 to 40 percentage points. Furthermore, the overall fluctuation of the rejection results of the OG-GAN method is also smaller than that of the other two methods. In summary, it is clear that the out-of-pipe target rejection performance of the method of this invention is significantly better than the other two models, and the out-of-pipe target rejection efficiency of the method of this invention is the highest.

[0132] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0133] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0134] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0135] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0136] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0137] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for rejecting targets outside a radar high-resolution range image database based on an adversarial learning classifier, characterized in that, Includes the following steps: Step 10: Preprocess the original samples of the aircraft's radar high-resolution range image to obtain the training set; Step 20: Construct a generator based on the GAN network G and discriminator D ; The generator G It includes N encoder layers connected in sequence and M decoder layers connected in sequence; each encoder layer includes a convolutional layer, a batch normalization layer and an activation function layer connected in sequence. Each of the M-1 decoder layers includes a deconvolution layer, a batch normalization layer, and an activation function layer connected in sequence, and the Mth decoder layer includes a convolution layer and an activation function layer. The discriminator D It includes, in sequence, an input layer, a first convolutional layer, a first batch normalization layer, a first activation layer, a second convolutional layer, a second batch normalization layer, a second activation layer, a third convolutional layer, a third batch normalization layer, a third activation layer, a fourth convolutional layer, a fourth batch normalization layer, a fourth activation layer, a fully connected layer, and a sigmoid layer; Step 30, according to the generator G and the discriminator D Construct the first objective function; Step 40, adjust the generator based on the training set. G and the discriminator D The generator is trained and a second objective function is constructed, and the generator is updated using the first objective function and the second objective function. G and the discriminator D The parameters are used to obtain the trained generator. G and the trained discriminator D ; wherein, the generator G Before the loss reaches the preset reconstruction loss, a generator is randomly saved. G As an intermediate generator ; Step 50, the trained generator G and intermediate generator Compared with the trained discriminator D Coupled composition adversarial decision-making model; Step 60: Input the sample set to be judged into the adversarial judgment model for judgment; If the output probability of the adversarial determination model is greater than the threshold, then the samples in the sample set to be determined are determined to be abnormal samples from outside the database. The first objective function The expression is: in, Indicates input sample X Sampling from the training set, Representing noise samples Sampling from a training set with added noise. This indicates the operation of finding the expected value. Represents generator G The output data, Discriminator D The output data; Step 40 includes: Step 41: Use the samples in the training set as input samples. X The noise sample generated after adding noise Input generator G Obtain reconstructed samples ; Step 42, reconstruct the sample and the input sample X Input discriminator D ; Step 43, repeat steps 41 and 42 on the generator. G and discriminator D Perform iterative training and update the generator using the first objective function. G The parameters are used to obtain the trained generator. G ; wherein, the generator G Before the loss reaches the preset reconstruction loss, a generator is randomly saved. G As an intermediate generator ; Step 44, based on the discriminator D Constructing the second objective function The second objective function The expression is: in, and Both represent compromise hyperparameters. Indicates input sample X Input generator G The reconstructed sample output later, noise sample Input intermediate generator The intermediate output samples This indicates the operation of finding the expected value. Represents generator G The output data, Discriminator D The output data, Indicates a pseudo-anomaly sample. , , and For two randomly selected different input samples, and For the sample and samples Input intermediate generator The output data, This is the generated pixel-level mean; Step 45, using the trained generator G and the intermediate generator For discriminator D Perform iterative training and update the discriminator using the second objective function. D The parameters are used to obtain the trained discriminator. D Among them, the discriminator D The input data includes: the input sample X The reconstructed sample The pseudo-anomaly samples and the intermediate sample ; After completing steps 43 and 45, the cross-entropy loss function is brought to converge. The expression for the cross-entropy loss function is: in, H Represents the cross-entropy loss function. This represents the predicted classification probability vector. This represents the true probability vector of samples in the database. k =1,2,…, k This represents the category number of the sample in the library, and log represents the logarithmic operation to the base 10. The training set v The class of samples refers to the samples in the library, and the training set refers to the samples in the training set. wv The sample is an external sample.

2. The method for rejecting targets outside a radar high-resolution range image database based on an adversarial learning classifier according to claim 1, characterized in that, Step 10 includes: Step 11: Extract temporal features from each original radar high-resolution range image sample of the aircraft to obtain HRRP temporal feature samples; Step 12: Perform amplitude normalization and centroid alignment operations on each of the HRRP time-domain feature samples to obtain the training set.

Images