A method for constructing a semi-supervised classification model for skywave radar ground and sea clutter based on an improved generative adversarial network

Abstract

The present invention discloses a method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network. This method uses the intermediate layer network of the discriminator in the semi-supervised generative adversarial network to filter out features in a multi-layer network, and weights the feature matching losses of each layer to obtain a joint feature matching loss. Furthermore, the standard adversarial loss and the joint feature matching loss are linearly weighted to obtain a weighted loss. Based on the standard adversarial loss, the joint feature loss, and the weighted loss, a ground and sea clutter classification model WL-SSGAN based on an improved generative adversarial network is designed. This method solves the technical problem of how to avoid manually adding labels and improve the classification performance of the model, achieving the technical effect of saving the cost of adding labels and improving classification efficiency.

Description

Technical Field

[0001] The present invention belongs to spectrum data processing technology in the field of radar data processing, and relates to a semi-supervised classification method for skywave radar ground and sea clutter based on an improved generative adversarial network. Background Art

[0002] Skywave over-the-horizon radar (THR) is a crucial piece of equipment for beyond-horizon long-range detection, widely used in both military and civilian applications. However, electromagnetic waves propagating through the ionosphere create coordinate registration issues, severely impacting THR's target positioning accuracy. THR ground and sea clutter recognition is the process of identifying whether the background clutter in each range-azimuth unit of a radar echo originates from the ground or the sea. Using this recognition result, a ground / sea boundary or terrain contour is formed, which is then matched with prior geographic information to provide coordinate registration parameters for target positioning. THR has attracted increasing attention due to its advantages, such as requiring no additional auxiliary signal sources or detection equipment.

[0003] Skywave radar emits electromagnetic signals from its transmitter, which are refracted by the ionosphere onto the surface of the Earth and the sea. Echoes from targets or clutter return along their original paths to the receiver, forming ground and sea clutter. The first-order Bragg peak of sea clutter is formed by the Bragg resonant scattering of the high-frequency electromagnetic waves emitted by the skywave radar and sea waves, appearing as a double peak symmetrically around zero frequency. Ground clutter samples exhibit a single peak near zero frequency because the echo characteristics of land or islands lack significant fluctuations. Land-sea boundary clutter samples combine the characteristics of sea and ground clutter samples, exhibiting a triple peak near zero frequency.

[0004] In recent years, deep learning has achieved remarkable success in land and sea clutter classification. Compared to traditional methods that rely on manual extraction of land and sea clutter features, land and sea clutter classification methods based on deep learning can leverage the powerful feature extraction capabilities of neural networks to automatically extract high-level representations of data from a large number of land and sea clutter samples, achieving accurate predictions for subsequent classification tasks. However, existing land and sea clutter classification methods, as fully supervised classification frameworks, rely on a large number of labeled training samples. Because labeling land and sea clutter samples is extremely time-consuming and labor-intensive, and requires expert knowledge in the field, fully supervised land and sea clutter classification methods have certain limitations in practical applications.

[0005] Therefore, a new classification method needs to be proposed to solve the technical problem of how to reduce the cost of adding labels to samples and improve the efficiency of sample classification. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for constructing a semi-supervised classification model of skywave radar ground and sea clutter based on an improved generative adversarial network, so as to solve the problem of how to reduce the cost of adding labels to ground and sea clutter samples and improve the classification accuracy of ground and sea clutter samples.

[0007] The present invention adopts the following technical solutions:

[0008] A first embodiment of the present invention provides a method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network, characterized by comprising:

[0009] The training samples of ground and sea clutter are input into the semi-supervised classification model for training. The training samples include labeled samples, unlabeled samples and generated samples. The semi-supervised generative adversarial network includes a discriminator and a generator.

[0010] The first adversarial loss of the discriminator is calculated based on the information entropy of the labeled samples output by the discriminator. The first adversarial loss represents the full supervision loss of the discriminator.

[0011] The second adversarial loss of the discriminator is calculated based on the information entropy of the unlabeled sample output by the discriminator and the information entropy of the generated sample. The second adversarial loss represents the semi-supervised loss of the discriminator;

[0012] Calculate the joint feature loss based on the generated sample features and unlabeled sample features in the multi-layer network in the discriminator;

[0013] Weight the second adversarial loss and the joint feature loss to generate the weighted loss of the generator;

[0014] By repeatedly training the semi-supervised classification model, the classification weight of the discriminator is updated based on the first adversarial loss, the second adversarial loss and the weighted loss, and a semi-supervised classification model that meets the preset classification accuracy is determined.

[0015] Optionally, the first adversarial loss is calculated as follows:

[0016]

[0017] Among them, p D is the information entropy of the labeled sample output by the discriminator, y is the sample label, K represents the K-th sample space, x is the sample feature, represents the expected operator of labeled samples, L supervised For the first fight against loss.

[0018] Optionally, the second adversarial loss is calculated as follows:

[0019]

[0020] Among them, log[1-p D (y=K+1|x)] is the information entropy of the sample features in the K+1th sample space of the unlabeled sample output by the discriminator, logp D(y=K+1|x) is the information entropy of the sample features in the K+1th sample space of the generated sample output by the discriminator, is the expectation operator of unlabeled samples, is the expected operator for generating labeled samples, L unsupervised For the second against loss.

[0021] Optionally, calculating the joint feature loss based on the generated sample features and the unlabeled sample features in the multi-layer network in the discriminator includes:

[0022] Extract the unlabeled sample features and generated sample features in each layer of the network from the preset number of intermediate layers of the discriminator;

[0023] Solve the Euclidean distance between the unlabeled sample features and the generated sample features in each layer of the network to obtain the feature matching loss of each layer of the network;

[0024] The feature matching loss of each layer of the network is weighted and summed to obtain the joint feature matching loss.

[0025] Optionally, the joint feature matching loss is calculated as follows:

[0026]

[0027] Among them, l max is the sum of the intermediate layer networks representing the discriminator, l mul Indicates the number of preset layers selected, and Ch (l) is the number of channels corresponding to the sample features of the lth layer, Le (l) is the length of the sample feature, is the feature matching loss of each layer of the network, Generate the expected sample features for the lth layer, is the expectation of the unlabeled sample features of the lth layer, L FM is the joint feature matching loss.

[0028] Optionally, the weighted loss is calculated as follows:

[0029] L WL-SSGAN =α(-L unsupervised )+βL FM ,

[0030] Among them, α is the first constant, β is the second constant, L FM is the joint feature matching loss, L WL-SSGAN is the weighted loss.

[0031] Optionally, the method further includes: when only labeled samples are used as input of the semi-supervised classification model, only the discriminator works, and the loss of the discriminator is the sum of the first adversarial loss and the second adversarial loss.

[0032] Optionally, the method further includes: when there are labeled samples and unlabeled samples as inputs of the semi-supervised classification model, both the discriminator and the generator work, and at this time, random noise is added to the generator to output the generated samples.

[0033] Optionally, the LeakyReLU activation function is connected after the first convolutional layer of the discriminator and after all convolutional layers.

[0034] Optionally, the first convolutional layer of the generator is followed by a ReLU activation function and the last deconvolutional layer is followed by a Tanh activation function.

[0035] The beneficial effects of the present invention are as follows: the intermediate layer network of the discriminator in the semi-supervised generative adversarial network filters out features in the multi-layer network, and the feature matching losses of each layer are weighted to obtain the joint feature matching loss. Furthermore, the standard adversarial loss and the joint feature matching loss are linearly weighted to obtain the weighted loss. Based on the standard adversarial loss, the joint feature loss and the weighted loss, a land and sea clutter classification model WL-SSGAN based on the improved generative adversarial network is designed. When the input of WL-SSGAN only contains labeled samples, it is regarded as a fully supervised land and sea clutter classification model based on the improved generative adversarial network. When the input of WL-SSGAN contains both labeled and unlabeled samples, it is regarded as a semi-supervised classification model for land and sea clutter based on an improved generative adversarial network. On the one hand, features that are beneficial to improving the classification performance of the classifier are extracted from different layers of the discriminator network, and a weighted loss is calculated by weighting the standard adversarial loss and the joint feature matching loss of the multi-layer network. Compared with the traditional method of simply using adversarial loss or feature matching loss of a single-layer network, the classification model based on weighted loss in the present invention has higher classification performance. On the other hand, for samples with strong random distribution, the discriminator and classifier can be started at the same time. The discriminator can extract useful features from unlabeled samples to generate samples, so that the distribution of the generated samples is the same as that of the initial samples, thereby avoiding manual labeling, saving labeling costs, and improving classification efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A diagram showing the steps of a method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to the first embodiment of the present invention;

[0037] Figure 2 A schematic diagram of a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network provided in Example 1 of the present invention;

[0038] Figure 3 A spectrum diagram of a full-range segment of a beam of a skywave radar collected in real time, provided in the first embodiment of the present invention;

[0039] Figure 4 A schematic diagram of a range-azimuth unit sea clutter sample obtained by sampling a spectrum diagram provided in the first embodiment of the present invention;

[0040] Figure 5 A schematic diagram of a range-azimuth unit ground clutter sample obtained by sampling a spectrum diagram provided in the first embodiment of the present invention;

[0041] Figure 6 This is a schematic diagram of land-sea boundary clutter samples of a range-azimuth unit obtained by sampling a spectrum diagram provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] A first embodiment of the present invention provides a method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network, characterized by comprising:

[0044] Step S101: Inputting training samples of ground and sea clutter into a semi-supervised classification model for training, wherein the training samples include labeled samples, unlabeled samples, and generated samples, and the semi-supervised generative adversarial network includes a discriminator and a generator;

[0045] In one embodiment, before training the semi-supervised classification model, a training sample set is first established. It should be noted that in order to use test samples to test the classification performance of the classification model after the classification model training is completed, a test sample set needs to be established at the same time.

[0046] Specifically, the source of the ground-sea clutter dataset is the echo power spectrum of the skywave radar target or clutter. The skywave radar emits an electromagnetic wave signal from the transmitter, which is refracted by the ionosphere to the surface of the ground and sea. The echo signal of the target or clutter returns to the receiver along the original path to form ground-sea clutter. The ground-sea clutter samples of the distance-azimuth unit that meet the project requirements are screened and calibrated from the skywave radar ground-sea clutter range-Doppler spectrum database, including sea clutter samples, ground clutter samples and ground-sea boundary clutter samples. The distance-azimuth unit sea clutter samples that meet the project requirements appear as double peaks symmetrical to the zero frequency, the ground clutter samples appear as a single peak near the zero frequency, and the ground-sea boundary clutter samples appear as three peaks near the zero frequency. The ground-sea clutter samples are then divided into two parts: a training set and a test set;

[0047] For example, the total number of sea clutter, ground clutter, and ground-sea boundary clutter samples is 1000 each, of which 700 (70%) are used for training and 300 (30%) for testing. The training samples are randomly divided into two parts: one part consists of a small number of labeled samples whose labels are valid, and the other part consists of a large number of unlabeled samples whose labels can be ignored. In addition, the testing samples are only used to evaluate the performance of WL-SSGAN and do not participate in any training process. The semi-supervised classification model is trained only with the training samples.

[0048] Step S102: Calculate the first adversarial loss of the discriminator based on the information entropy of the labeled samples output by the discriminator, where the first adversarial loss represents the full supervision loss of the discriminator;

[0049] Step S103: Calculate the second adversarial loss of the discriminator based on the information entropy of the unlabeled sample output by the discriminator and the information entropy of the generated sample, where the second adversarial loss represents the semi-supervised loss of the discriminator;

[0050] In one embodiment, when the training samples only include labeled samples and unlabeled samples, the model is composed of a semi-supervised generative adversarial network (SSGAN), in which only the discriminator is active and the loss function regularization consists of two parts: supervised loss and semi-supervised loss:

[0051] L SSGAN =L supervised +L unsupervised ,

[0052] Among them, the first adversarial loss is the full supervision loss, which is calculated as follows:

[0053]

[0054] Among them, p D is the information entropy of the labeled sample output by the discriminator, y is the sample label, K represents the K-th sample space, x is the sample feature, represents the expected operator of labeled samples, L supervised For the first fight against loss.

[0055] The second adversarial loss is the semi-supervised loss, which is calculated as follows:

[0056]

[0057] Among them, log[1-p D (y=K+1|x)] is the information entropy of the sample features in the K+1th sample space of the unlabeled sample output by the discriminator, logp D (y=K+1|x) is the information entropy of the sample features in the K+1th sample space of the generated sample output by the discriminator, is the expectation operator of unlabeled samples, is the expected operator for generating labeled samples, L unsupervised For the second against loss.

[0058] Step S104: Calculate the joint feature loss based on the generated sample features and unlabeled sample features in the multi-layer network in the discriminator;

[0059] In one embodiment, optionally, calculating the joint feature loss based on the generated sample features and the unlabeled sample features in the multi-layer network in the discriminator includes:

[0060] Extract the unlabeled sample features and generated sample features in each layer of the network from the preset number of intermediate layers of the discriminator;

[0061] Solve the Euclidean distance between the unlabeled sample features and the generated sample features in each layer of the network to obtain the feature matching loss of each layer of the network;

[0062] The feature matching loss of each layer of the network is weighted and summed to obtain the joint feature matching loss.

[0063] Specifically, a network set with a preset number of layers is selected from the intermediate network of the discriminator, and unlabeled sample features and generated sample features are extracted from each layer of the network. The feature matching loss of each layer of the network is obtained by calculating the Euclidean distance between the unlabeled sample features and the generated sample features in each layer of the network:

[0064]

[0065] in, Generate the expected sample features for the lth layer, is the expectation of the unlabeled sample features of the lth layer, is the feature matching loss of the l-th layer network.

[0066] Furthermore, the number of channels corresponding to the jointly generated samples and the unlabeled samples and the length of the sample features are combined

[0067] The joint feature matching loss is calculated as follows:

[0068]

[0069] Among them, l max is the sum of the intermediate layer networks representing the discriminator, l mul Indicates the number of preset layers selected, and Ch (l) is the number of channels corresponding to the sample features of the lth layer, Le (l) is the length of the sample feature, is the feature matching loss of each layer of the network, Generate the expected sample features for the lth layer, is the expectation of the unlabeled sample features of the lth layer, L FM is the joint feature matching loss.

[0070] Step S105: weighting the second adversarial loss and the joint feature loss to generate a weighted loss of the generator;

[0071] In one embodiment, it should be noted that when the training samples input into the semi-supervised model include both labeled samples and unlabeled samples, both the discriminator and the generator in the model are in effect. In this case, the loss of the generator is a weighted loss, which is specifically calculated as follows:

[0072] L WL-SSGAN =α(-L unsupervised )+βL FM ,

[0073] Among them, α is the first constant, β is the second constant, L FM is the joint feature matching loss, L WL-SSGAN is the weighted loss, L unsupervised For the second against loss.

[0074] It should be noted that α and β can be set according to the actual situation, and α+β=1, α,β≥0. The proportion of weight factors α and β can be controlled to weigh the contribution of adversarial loss and feature matching loss to WL-SSGAN training. In addition, l mul In addition, in order to improve the generalization ability of the WL-SSGAN model, the signal features extracted from a single layer are not directly used. Instead of feature matching loss, multi-layer signals are weighted to obtain joint feature matching loss L FM , in order to obtain better semi-supervised classification performance.

[0075] Step S106: By repeatedly training the semi-supervised classification model, the classification weight of the discriminator is updated based on the first adversarial loss, the second adversarial loss and the weighted loss, and a semi-supervised classification model that meets the preset classification accuracy is determined.

[0076] In one embodiment, the semi-supervised model is trained multiple times by using training samples. Each time, the first adversarial loss, the second adversarial loss, and the weighted loss at the current training moment are back-propagated to update the classification weights of the classifier, and finally the classification type and probability are output to guide the classification accuracy of the classifier to meet the preset classification accuracy, thereby obtaining the final semi-supervised classification model.

[0077] Combine Figure 2 The implementation of the above steps S101 to S106 is described as follows:

[0078] The training set sample sizes for the first to tenth groups are as follows:

[0079] (1) The number of labeled samples is 15. Among them, the number of labeled sea clutter samples is 5, the number of labeled ground clutter samples is 5, and the number of labeled ground-sea boundary clutter samples is 5. The number of unlabeled samples is 2085. Among them, the number of unlabeled sea clutter samples is 695, the number of unlabeled ground clutter samples is 695, and the number of unlabeled ground-sea boundary clutter samples is 695;

[0080] (2) The number of labeled samples is 30. Among them, the number of labeled sea clutter samples is 10, the number of labeled ground clutter samples is 10, and the number of labeled ground-sea boundary clutter samples is 10. The number of unlabeled samples is 2070. Among them, the number of unlabeled sea clutter samples is 690, the number of unlabeled ground clutter samples is 690, and the number of unlabeled ground-sea boundary clutter samples is 690;

[0081] (3) The number of labeled samples is 45. Among them, the number of labeled sea clutter samples is 15, the number of labeled ground clutter samples is 15, and the number of labeled ground-sea boundary clutter samples is 15. The number of unlabeled samples is 2055. Among them, the number of unlabeled sea clutter samples is 685, the number of unlabeled ground clutter samples is 685, and the number of unlabeled ground-sea boundary clutter samples is 685; c

[0082] (4) The number of labeled samples is 60, of which 20 are labeled sea clutter samples, 20 are labeled ground clutter samples, and 20 are labeled ground-sea boundary clutter samples. The number of unlabeled samples is 2040, of which 680 are unlabeled sea clutter samples, 680 are unlabeled ground clutter samples, and 680 are unlabeled ground-sea boundary clutter samples.

[0083] (5) The number of labeled samples is 75. Among them, the number of labeled sea clutter samples is 25, the number of labeled ground clutter samples is 25, and the number of labeled ground-sea boundary clutter samples is 25. The number of unlabeled samples is 2025. Among them, the number of unlabeled sea clutter samples is 675, the number of unlabeled ground clutter samples is 675, and the number of unlabeled ground-sea boundary clutter samples is 675;

[0084] (6) The number of labeled samples is 90. Among them, the number of labeled sea clutter samples is 30, the number of labeled ground clutter samples is 30, and the number of labeled ground-sea boundary clutter samples is 30. The number of unlabeled samples is 2010. Among them, the number of unlabeled sea clutter samples is 670, the number of unlabeled ground clutter samples is 670, and the number of unlabeled ground-sea boundary clutter samples is 670;

[0085] (7) The number of labeled samples is 105. Among them, the number of labeled sea clutter samples is 35, the number of labeled ground clutter samples is 35, and the number of labeled ground-sea boundary clutter samples is 35. The number of unlabeled samples is 1995. Among them, the number of unlabeled sea clutter samples is 665, the number of unlabeled ground clutter samples is 665, and the number of unlabeled ground-sea boundary clutter samples is 665;

[0086] (8) The number of labeled samples is 120. Among them, the number of labeled sea clutter samples is 40, the number of labeled ground clutter samples is 40, and the number of labeled ground-sea boundary clutter samples is 40. The number of unlabeled samples is 1980. Among them, the number of unlabeled sea clutter samples is 660, the number of unlabeled ground clutter samples is 660, and the number of unlabeled ground-sea boundary clutter samples is 660;

[0087] (9) The number of labeled samples is 135. Among them, the number of labeled sea clutter samples is 45, the number of labeled ground clutter samples is 45, and the number of labeled ground-sea boundary clutter samples is 45. The number of unlabeled samples is 1965. Among them, the number of unlabeled sea clutter samples is 655, the number of unlabeled ground clutter samples is 655, and the number of unlabeled ground-sea boundary clutter samples is 655;

[0088] (10) The number of labeled samples is 150. Among them, the number of labeled sea clutter samples is 50, the number of labeled ground clutter samples is 50, and the number of labeled ground-sea boundary clutter samples is 50. The number of unlabeled samples is 1950. Among them, the number of unlabeled sea clutter samples is 650, the number of unlabeled ground clutter samples is 650, and the number of unlabeled ground-sea boundary clutter samples is 650.

[0089] When only labeled samples are used as input to the semi-supervised classification model, only the discriminator works, and the loss of the discriminator is the sum of the first adversarial loss and the second adversarial loss.

[0090] At this time, the discriminator loss of WL-SSGAN is L D is the standard SSGAN loss:

[0091] L D =L supervised +L unsupervised ,

[0092] Among them, L supervised is the first adversarial loss, L unsupervised For the second against loss.

[0093] When labeled samples and unlabeled samples are used as inputs of the semi-supervised classification model, both the discriminator and the generator work. At this time, useful sample features are extracted from the unlabeled samples by adding random noise to the generator, and the generated samples are output. It should be noted that the labeled samples are obtained through manual annotation.

[0094] At this time, the generator loss of WL-SSGAN is the proposed weighted loss L WL-SSGAN :

[0095] L WL-SSGAN =αL adv +βL FM =α(-L unsupervised )+βL FM

[0096] Among them, α+β=1, α,β≥0.

[0097] The feature matching loss of a single-layer network is:

[0098]

[0099] The joint feature matching loss function is as follows:

[0100]

[0101] To facilitate understanding of L WL-SSGAN , replace L with D(x) unsupervis 1-p in D (y=K+1|x), so L unsupervis Convert to standard GAN loss form:

[0102]

[0103] Further, L WL-SSGAN can be rewritten as the following more verbose representation:

[0104]

[0105] Based on the established ground and sea clutter training set containing a small number of labeled samples and a large number of unlabeled samples, according to the discriminator loss part L D and the generator loss L supervised The parameters of the constructed land and sea clutter semi-supervised classification model based on the improved generative adversarial network are updated alternately to perform semi-supervised learning.

[0106] The test set is used as the input of the trained fully supervised classification model for land and sea clutter to test its classification performance.

[0107] Optionally, the LeakyReLU activation function is connected after the first convolutional layer of the discriminator and after all convolutional layers.

[0108] In one embodiment, the discriminator includes seven 1D convolutional layers and two fully connected layers, where all convolution kernels have a size of 4 and a stride of 2. The first convolutional layer is followed by a LeakyReLU activation function, all convolutional layers are followed by 1D batch normalization and a LeakyReLU activation function, and the first fully connected layer is followed by a LeakyReLU activation function. In addition, all convolutional layers are followed by residual blocks and a LeakyReLU activation function. The discriminator input is a 1×512-dimensional signal, and the output is a 3D classification result.

[0109] Optionally, the first convolutional layer of the generator is followed by a ReLU activation function and the last deconvolutional layer is followed by a Tanh activation function.

[0110] In one embodiment, the generator consists of eight 1D deconvolution layers, all with a kernel size of 4, a stride of 1 for the first layer, and 2 for the remaining layers. The first seven deconvolution layers are followed by 1D batch normalization and a ReLU activation function, while the last deconvolution layer is followed by a Tanh activation function. Additionally, the first seven deconvolution layers are followed by a residual block and a ReLU activation function. The generator takes as input 100D Gaussian random noise and outputs a 1×512 signal.

[0111] Based on the improved generative adversarial network semi-supervised classification model for skywave radar ground and sea clutter shown in the first embodiment of the present invention, the semi-supervised classification performance test of WL-SSGAN was performed.

[0112] In the experiment of this embodiment, the experimental environment and corresponding version numbers used to train the WL-SSGAN model are: System: Windows 10 (64-bit), GPU: NVIDIA GeForce RTX 3090, CUDA: 11.6, Python: 3.9.0 (in Anaconda 4.11.0), torch: 1.11.0, torchvision: 0.12.0, numpy: 1.22.3.

[0113] The parameter configuration used to train WL-SSGAN is: Batch size: 64, Learning Rate: 0.0001, LeakyReLU: 0.2, Adma Optimizer: beta1-0.5, beta2-0.999, Data Normalization: -1~1, Weight Initialization.

[0114] To evaluate the performance of WL-SSGAN, the number of labeled samples γ in the training set is 15, 30, 45, 60, 75, 90, 105, 120, 135, 150, and the rest are the corresponding number of unlabeled samples. Next, consider the number of labeled samples γ, the proportion of α and β, and l mui The influence of these three factors on the classification performance of WL-SSGAN is compared with the fully supervised classifier WL-SSGAN trained with only a small number of labeled samples.

[0115] First, it is assumed that all intermediate layers of the discriminator contribute to the loss of the generator, i.e., l mui = {1, 2, 3, 4, 5, 6, 7}, fixing the joint feature matching loss function and measuring the impact of different weights of α and β on the classification performance of WL-SSGAN. The experimental results are shown in Table 1. Here, base represents the classification accuracy of the fully supervised WL-SSGAN classifier trained with only a small number of labeled samples.

[0116] Table 1 Classification accuracy of WL-SSGAN under different values of α and β

[0117]

[0118] The following conclusions can be drawn from Table 1: (1) The classification performance of WL-SSGAN is higher than that of the fully supervised classifier, which shows that the model can extract potential features related to land and sea clutter classification from a large number of unlabeled samples by adding a generator. Therefore, WL-SSGAN can improve the classification performance of the fully supervised model when there are only a small number of labeled samples. (2) As the number of land and sea clutter samples γ decreases, the improvement of WL-SSGAN classification performance becomes more obvious. This is because when the number of training samples is small, the fully supervised classifier is prone to overfitting. (3) When the joint feature matching loss L FM The number of layers l mul When fixed, the different weights of α and β obviously have an impact on the classification accuracy, and only using the adversarial loss L adv Or joint feature matching loss L FM , that is, the classification accuracy obtained when (α, β) = (1.0, 0.0) or (α, β) = (0.0, 1.0) is not the highest. Therefore, the proposed weighted loss function L WL-SSGAN Better than L adv and L FM .

[0119] Secondly, we assume that the adversarial loss and the joint feature matching loss have the same contribution to the generator loss, i.e., (α, β) = (0.5, 0.5). By fixing the contribution of the adversarial loss and the feature matching loss to WL-SSGAN, we can measure the l mulThe impact of different choices of on the classification performance of WL-SSGAN. The experimental results are shown in Table 2.

[0120] Table 21 mul Classification accuracy of WL-SSGAN under different values

[0121]

[0122] From Table 2, we can draw the following conclusions: When the adversarial loss L adv and joint feature matching loss L FM When the contributions α and β are fixed, l mul The choice of has an obvious impact on the classification accuracy, and using feature matching loss containing only a single layer Or the joint feature matching loss of all layers, i.e. l mul ={1} or l mul = {1-7} is not the highest classification accuracy. This shows that: (1) The proposed joint feature matching loss L FM Outperforms traditional feature matching loss (2) It is not necessary to combine all the intermediate layer features of the discriminator to provide the best semi-supervised classification performance. On the contrary, the features extracted by some layers may suppress the classification performance. Therefore, the optimal classification performance depends on the appropriate combination of layers. Therefore, the proposed WL-SSGAN can achieve semi-supervised classification of land and sea clutter, and the proposed weighted loss is better than the adversarial loss and feature matching loss.

[0123] like Figure 3 As shown in the figure, it can be seen that the ground clutter samples are skywave radar echo signals, and their data distribution is highly random. Therefore, it is extremely difficult to train the SSGAN generator using the standard adversarial loss. On the contrary, the improved generative adversarial network ground clutter semi-supervised classification method WL-SSGAN proposed in this invention can utilize this randomness to enable the feature matching loss to capture diverse sample features, such as Figure 4-Figure 6 As shown, this alleviates mode collapse in the generator and overfitting in the discriminator to a certain extent. Therefore, we propose a joint feature matching loss by weighting the intermediate layers of the discriminator in a semi-supervised generative adversarial network (SSGAN). Furthermore, we linearly weight the standard adversarial loss and the joint feature matching loss to propose the WL-SSGAN weighted loss. This achieves semi-supervised classification of land and sea clutter samples, and the proposed weighted loss outperforms either the adversarial loss or the feature matching loss alone.

Claims

1. A method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network, characterized in that: include: Inputting training samples of ground and sea clutter into a semi-supervised classification model for training, wherein the training samples include labeled samples, unlabeled samples, and generated samples, and the semi-supervised generative adversarial network includes a discriminator and a generator; Calculating a first adversarial loss of the discriminator according to the information entropy of the labeled samples output by the discriminator, where the first adversarial loss represents the full supervision loss of the discriminator; Calculating a second adversarial loss of the discriminator according to the information entropy of the unlabeled sample output by the discriminator and the information entropy of the generated sample, wherein the second adversarial loss represents the semi-supervised loss of the discriminator; Calculating a joint feature loss based on the generated sample features and the unlabeled sample features in the multi-layer network in the discriminator; weighting the second adversarial loss and the joint feature loss to generate a weighted loss of the generator; Repeatedly training the semi-supervised classification model, updating the classification weight of the discriminator based on the first adversarial loss, the second adversarial loss, and the weighted loss, and determining a semi-supervised classification model that meets a preset classification accuracy; Calculating the joint feature loss based on the generated sample features and the unlabeled sample features in the multi-layer network in the discriminator includes: Extracting unlabeled sample features and generated sample features in each layer of the network from the preset number of intermediate layers of the discriminator; Solve the Euclidean distance between the unlabeled sample features and the generated sample features in each layer of the network to obtain the feature matching loss of each layer of the network; Performing weighted summation on the feature matching losses of each layer of the network to obtain the joint feature matching loss; The joint feature matching loss is calculated as follows: Among them, l max is the sum of the intermediate layer networks representing the discriminator, l mul represents the selected preset number of layers, and Ch (l) is the number of channels corresponding to the sample features of the lth layer, Le (l) is the length of the sample feature, is the feature matching loss of each layer of the network, Generate the expected sample features for the lth layer, is the expectation of the unlabeled sample features of the lth layer, L FM is the joint feature matching loss.

2. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using a generative adversarial network according to claim 1, wherein: The first adversarial loss is calculated as follows: Among them, p D is the information entropy of the labeled sample output by the discriminator, y is the sample label, K represents the K-th sample space, x is the sample feature, represents the expected operator of labeled samples, L supervised For the first fight against loss.

3. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to claim 1, wherein: The second adversarial loss is calculated as follows: Among them, log[1-p D (y=K+1|x)] is the information entropy of the sample features in the K+1th sample space of the unlabeled sample output by the discriminator, logp D (y=K+1|x) is the information entropy of the sample features in the K+1th sample space of the generated sample output by the discriminator, is the expectation operator of unlabeled samples, is the expected operator for generating labeled samples, L unsupervised For the second counter loss.

4. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to claim 1, wherein: The weighted loss is calculated as follows: L WL-SSGAN =α(-L unsupervised )+βL FM , Among them, α is the first constant, β is the second constant, L FM is the joint feature matching loss, L WL-SSGAN is the weighted loss.

5. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to claim 1, wherein: The method further comprises: When only the labeled sample is used as the input of the semi-supervised classification model, only the discriminator works, and the loss of the discriminator is the sum of the first adversarial loss and the second adversarial loss.

6. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to claim 1, wherein: The method further comprises: When the labeled samples and the unlabeled samples are used as inputs of the semi-supervised classification model, both the discriminator and the generator work. At this time, the generated samples are output by adding random noise to the generator.

7. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to claim 1, wherein: The LeakyReLU activation function is connected to the back of the first convolutional layer of the discriminator and all the convolutional layers.

8. The method for constructing a semi-supervised classification model for skywave radar ground and sea clutter using an improved generative adversarial network according to claim 1, wherein: The first convolutional layer of the generator is connected to the ReLU activation function, and the last deconvolutional layer is connected to the Tanh activation function.

Images