A SAR Sample Augmentation Method Based on Semi-supervised Generative Adversarial Network
Through the method of semi-supervised generative adversarial network, the SAR sample set is expanded, and the overfitting problem caused by sparse data samples in the SAR field is solved. The generated new samples improve the model training effect and sample diversity.
Patent Information
- Application Number
- CN202110943782.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-17
- Publication Date
- 2025-04-29
- Estimated Expiration
- 2041-08-17
AI Technical Summary
In the SAR field, due to the sparse data samples, deep learning methods are prone to overfitting during the training stage of neural network models, resulting in poor detection performance.
The SAR sample expansion method based on the semi-supervised generative adversarial network is adopted. By establishing an embedded sample set and a fine-tuning sample set, pre-trained generator and discriminator loss functions are designed, and the semi-supervised generative adversarial network is trained using the Adam optimizer to generate new samples to expand the original sample set.
Effectively eliminate the background inconsistency of the input samples, the generated new samples are closer to the real samples, improving sample diversity and the training effect of deep learning models, and reducing the occurrence of overfitting.
Smart Images
Figure CN113792756B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of SAR sample expansion and deep learning, and in particular to a SAR sample expansion method based on a semi-supervised generative adversarial network. Background Art
[0002] Synthetic Aperture Radar (SAR), as an active microwave sensor, can achieve real-time long-distance monitoring of static time-sensitive targets (such as tanks, armored vehicles, ships, etc.) all day and all weather, and has important application value in military and civilian fields. Among them, SAR image target detection and recognition capability is an important performance indicator of SAR.
[0003] Due to the high scene complexity of SAR images, traditional detection algorithms (such as CFAR) generally perform signal processing on the data first, and then compare it with the detection threshold to obtain the detection result; however, this type of method is affected by the clutter estimation parameters, and the detection performance needs to be improved.
[0004] Deep learning methods have been widely used in intelligent target detection and recognition due to their good learning and prediction performance. Their accuracy and efficiency are better than other existing methods and can be applied to the SAR field. However, deep learning methods usually require a large number of training samples in the neural network model training stage. However, due to its niche and professional nature, the data samples in the SAR field are relatively scarce. There are only a few public sample sets (SSDD, AIR-SARShip, etc.), and the number of samples is in the order of 1k, which makes it impossible for the neural network model to be fully trained, thus causing overfitting, which brings difficulties to the application of deep learning methods in the SAR field. To prevent overfitting, one method is to expand the SAR samples. Summary of the invention
[0005] The purpose of the present invention is to provide a SAR sample expansion method based on a semi-supervised generative adversarial network to expand SAR samples so that deep learning methods can be better applied to the SAR field.
[0006] The technical solution adopted by the present invention is: a SAR sample expansion method based on a semi-supervised generative adversarial network, comprising the following steps:
[0007] Step 1: Create an embedded sample set and a fine-tuned sample set based on the original sample set.
[0008] Step 2: Select the generator G; design the pre-trained generator loss function, and use multiple different fine-tuning samples to pre-train the generator G; select the discriminator D, design the pre-trained discriminator loss function, and use multiple different original samples to pre-train the discriminator D.
[0009] Step 3: Connect the pre-trained generator G and the pre-trained discriminator D in sequence to establish a generative adversarial network, design a semi-supervised loss function, and obtain a semi-supervised generative adversarial network.
[0010] Step 4: Use the Adam optimizer to train the semi-supervised generative adversarial network, and obtain the optimal generator with the lowest loss in the training iteration based on the loss function value of the generator G.
[0011] Step 5: Input any embedded sample into the optimal generator, and the optimal generator outputs a new sample, which is used as an expanded sample of the original sample set.
[0012] Compared with the prior art, the present invention has the following beneficial effects:
[0013] By utilizing the principle of generative adversarial networks, the generator and discriminator can spontaneously learn the detailed information of SAR samples in the adversarial game, which can effectively eliminate the inconsistency of the background of the input embedded samples and make the output samples closer to the real samples. By designing appropriate semi-supervised loss functions and alternating iterative training methods, the convergence efficiency of the generative adversarial network can be made faster and more stable. When expanding the samples, the number and distribution of targets in the input samples can be freely changed to make the output new samples diverse. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0015] Figure 1 A schematic diagram of synthesizing SSDD modified fine-tuned samples and SSDD Embedded embedded samples based on SSDD original samples;
[0016] Figure 1 (a) Schematic diagram of the process of synthesizing SSDD modified fine-tuning samples based on SSDD original samples;
[0017] Figure 1 (b) is a schematic diagram of the process of synthesizing SSDD Embedded samples based on the original SSDD samples;
[0018] Figure 2 It is a schematic diagram of the structure of the generator G;
[0019] Figure 3 Flowchart for semi-supervised generative adversarial networks;
[0020] Figure 4 This is a calculation diagram of the regional loss function loss_pixel in the generator G loss function;
[0021] Figure 4(a) is the original sample y of SSDD;
[0022] Figure 4 (b) is the SSDD Embedded sample x1;
[0023] Figure 4 (c) is the extended sample G(x1,y);
[0024] Figure 4 (d) is Box_1;
[0025] Figure 4 (e) is Box_ship;
[0026] Figure 4 (f) is Box;
[0027] Figure 4 (g) is Box_2;
[0028] Figure 4 (h) is ΣBox_ship&(~Box_2);
[0029] Figure 4 (i) is ΣBox^Box_2;
[0030] Figure 5 A comparison diagram of the input graph of the optimal generator and the output graph of the optimal generator;
[0031] Figure 5 (a) is the input graph of the optimal generator;
[0032] Figure 5 (b) is the output diagram of the optimal generator. DETAILED DESCRIPTION
[0033] The embodiments of the present invention will be described in detail below with reference to examples, but those skilled in the art will understand that the following examples are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention.
[0034] In one embodiment of the present invention, the expanded sample set is selected as the SSDD (SAR Ship Detection Dataset) sample set.
[0035] The technical solution adopted by the present invention is: a SAR sample expansion method based on a semi-supervised generative adversarial network, comprising the following steps:
[0036] Reference Figure 1 , which is a schematic diagram of synthesizing SSDD modified fine-tuned samples and SSDD Embedded samples based on SSDD original samples.
[0037] Step 1: Create the SSDD Embedded sample set and the SSDD modified fine-tuning sample set based on the SSDD original sample set.
[0038] The sub-steps of step 1 are:
[0039] Use Photoshop software to synthesize any target slice image in the original SSDD sample set with any background image in the original SSDD sample set to obtain the SSDD Embedded sample, such as Figure 1 As shown in (a), we build the SSDD Embedded sample set;
[0040] Use Photoshop software to fine-tune the background pixel brightness of any target in the original SSDD sample set to obtain the SSDD modified fine-tuning sample, such as Figure 1 As shown in (b), the SSDD modified fine-tuning sample set is established;
[0041] The SSDD Embedded sample is denoted as x1, the SSDD modified fine-tuned sample is denoted as x2, and the SSDD original sample is denoted as y.
[0042] Reference Figure 2 , which is a schematic diagram of the structure of the generator G.
[0043] Step 2: Select the generator G; design the pre-trained generator loss function, and use multiple different SSDDmodified fine-tuning samples to pre-train the generator G; select the discriminator D, design the pre-trained discriminator loss function, and use multiple different SSDD original samples to pre-train the discriminator D.
[0044] The sub-steps of step 2 are:
[0045] 2a) Select generator G. Generator G uses AttR2U-Net, which adds a cyclic residual convolution module (Recurrent ResNet) and an attention gate module (Attention Gate) on the basis of U-Net.
[0046] The U-Net includes an encoder and a decoder; the encoder is composed of a plurality of cascaded submodules, each of which includes a convolution layer and a pooling layer (max pooling downsampling layer). Each time the input image passes through a submodule, abstract features are extracted and the resolution is reduced; the decoder is composed of a plurality of cascaded submodules, each of which includes a convolution layer and an upsampling layer, and is used to restore the abstract features learned by the encoder to the resolution.
[0047] The cyclic residual convolution module effectively improves network performance by combining cyclic connections and residual connections; the attention gate module is used to readjust the output features of the encoder before splicing the features of each resolution of the encoder Encoder with the corresponding features in the decoder Decoder.
[0048] 2b) The process of pre-training the generator G is as follows: any SSDD modified fine-tuning sample x2 is input into the generator G, and all SSDD original samples y in the SSDD original sample set are input into the generator G as conditions. The generator G outputs the pre-training extended sample G(x2,y), and a round of training is completed;
[0049] The generator G is pre-trained for 20 rounds to obtain the pre-trained generator G.
[0050] 2c) Designing a pre-trained generator loss function: The pre-trained generator loss function uses the mean square error between the original SSDD sample y and the pre-trained extended sample G(x2,y), as shown in formula (1);
[0051] Loss G =E same [||yG(x2,y)||2]+αE diff [||yG(x2,y)||2] (1)
[0052] In formula (1), E(*) represents the expected value of the distribution function; same refers to the area where the SSDD modified fine-tuning sample x2 is the same as the SSDD original sample y, and diff refers to the area where the SSDD modified fine-tuning sample x2 is different from the SSDD original sample y, such as Figure 1 As shown in (a); α represents the weight coefficient of the diff area, and the value of α is 3.
[0053] 2d) Select a discriminator D, which uses a CNN recognition network, including C1-C5 layers and S1 layer. The specific structure and parameters are as follows:
[0054] C1 layer, the specific structure is The output feature map size is 208*208*32;
[0055] C2 layer, the specific structure is The output feature map size is 52*52*32;
[0056] C3 layer, the specific structure is The output feature map size is 52*52*32;
[0057] C4 layer, the specific structure is The output feature map size is 26*26*64;
[0058] C5 layer, the specific structure is The output feature map size is 13*13*1;
[0059] The S1 layer has the following structure: Output probability value p.
[0060] 2e) The process of pre-training the discriminator D is as follows: any SSDD original sample y or any of the pre-training extended samples G(x2,y) is input into the discriminator D, and the discriminator D outputs a probability value p, that is, a round of training is completed; the probability value p represents the possibility that the input to the discriminator D is a true sample. For the discriminator D, the SSDD original sample y is a true sample, and the pre-training extended sample G(x2,y) is a false sample;
[0061] The discriminator D is pre-trained for 20 rounds to obtain the pre-trained discriminator D.
[0062] 2f) Designing a pre-trained discriminator loss function: The pre-trained discriminator loss function adopts a cross entropy loss function, and the cross entropy loss function is shown in formula (2);
[0063] L=-E[y i ·logp i +(1-y i )·log(1-p i )] (2)
[0064] In formula (2), E(*) represents the expected value of the distribution function, y i Indicates the label of input sample i. When a true sample is input, the label is 1, and when a false sample is input, the label is 0. i represents the probability that sample i is predicted to be true;
[0065] For the discriminator D, the label corresponding to the pre-trained extended sample G(x2,y) is 0, and the label corresponding to the original sample y of SSDD is 1, so the loss function of the pre-trained discriminator is as shown in formula (3).
[0066]
[0067] refer to Figure 3 , is a flowchart of the semi-supervised generative adversarial network; reference Figure 4 , which is a calculation diagram of the regional loss function loss_pixel in the generator G loss function.
[0068] Step 3: Connect the pre-trained generator G and the pre-trained discriminator D in sequence to establish a generative adversarial network, design a semi-supervised loss function, and obtain a semi-supervised generative adversarial network.
[0069] The sub-steps of step 3 are:
[0070] 3a) The process of generating the adversarial network is as follows: any SSDD Embedded sample x1 is input into the generator G, and all SSDD original samples y are input into the generator G as conditions, and the generator G outputs the extended sample G(x1,y); any SSDD original sample y or any extended sample G(x1,y) is input into the discriminator D, and the discriminator D outputs the probability value p.
[0071] 3b) Design a semi-supervised loss function, which includes the loss function of the generator G and the loss function of the discriminator D.
[0072] The loss function of the generator G is divided into three parts:
[0073] The adversarial loss function loss_adv between the generator G and the discriminator D;
[0074] The regional loss function loss_region of the comparison between the output extended sample G(x1,y) and the input SSDD original sample y;
[0075] The pixel loss function loss_pixel of the output extended sample G(x1,y) compared with the original sample y of SSDD.
[0076] Design the adversarial loss function loss_adv, as shown in formula (4).
[0077] loss_adv=-E[log(D(G(x1,y)))] (4)
[0078] The adversarial loss function loss_adv uses a cross entropy loss function, and the cross entropy loss function is shown in formula (2);
[0079] L=-E[y i ·logp i +(1-y i )·log(1-p i )] (2)
[0080] In formula (2), E(*) represents the expected value of the distribution function, y i Indicates the label of input sample i. When a true sample is input, the label is 1, and when a false sample is input, the label is 0. i represents the probability that sample i is predicted to be true;
[0081] For the generator G, the label corresponding to the extended sample G(x1,y) is 1, and the label corresponding to the original sample y is 0, so the adversarial loss function loss_adv is as shown in formula (4).
[0082] The regional loss function loss_region is designed as shown in formula (5).
[0083]
[0084] The regional loss function loss_region is the number of pixels that should not change but actually change in the output image of generator G compared to the input image of generator G. The calculation steps are as follows:
[0085] First, calculate the input Figure 4 The SSDD Embedded sample x1 shown in (b) is relative to the corresponding Figure 4 The area where the original SSDD sample y remains unchanged as shown in (a) is represented by a binary image Box_1, as shown in Figure 4 (d) as shown.
[0086] Then use a simple threshold detection method to detect the input Figure 4 The ship hull in the SSDD Embedded sample x1 shown in (b) is represented by a binary image Box_ship, as shown in Figure 4 (e) as shown.
[0087] Then take the union or operation of the binary graph Box_1 and the binary graph Box_ship to obtain the binary graph Box, such as Figure 4 (f) as shown.
[0088] Then calculate the output as Figure 4 The extended sample G(x1,y) shown in (c) is relative to the input Figure 4 The SSDD Embedded shown in (b) embeds the unchanged area of sample x1, which is represented by a binary image Box_2, as shown in Figure 4 (g) as shown.
[0089] Then perform XOR operation on the binary image Box and the binary image Box_2 to obtain ∑Box^Box_2, as Figure 4 (i) as shown; at the same time, the ship should be constrained not to be changed, and the binary graph ∑Box_ship&(~Box_2) is obtained, as Figure 4 (h) shown.
[0090] Finally, different weights are assigned to ∑Box^Box_2 and ∑Box_ship&(~Box_2), the weight coefficient of ∑Box^Box_2 is 1, and the weight coefficient of ∑Box_ship&(~Box_2) is 10, and the regional loss function loss_region is as shown in formula (5).
[0091] The pixel loss function loss_pixel is designed as shown in formula (6).
[0092] loss_pixel=E(||GR_flat-GT_flat||1+20·||GR_flat_ship-Imag_flat_ship||1) (6)
[0093] The pixel loss function loss_pixel is the magnitude of the change in the pixel points of the output image of generator G compared to the input image of generator G. The calculation steps are as follows:
[0094] First, the pixel loss function GR_flat of the output extended sample G(x1,y) outside the ship area and the pixel loss function GT_flat of the input SSDD original sample y outside the ship area are calculated respectively, as shown in formula (9).
[0095]
[0096] Then calculate the pixel loss function GR_flat_ship of the ship area of the extended sample G(x1,y) and the pixel loss function Input_flat_ship of the ship area of the original SSDD sample y, as shown in formula (10).
[0097]
[0098] Finally, the pixel loss of the output image of the outer area of the ship compared with the input image and the pixel loss of the output image of the ship area compared with the input image are calculated, and different weights are assigned to the two; the weight coefficient of the former is 1, and the weight coefficient of the latter is 20, and the pixel loss function loss_pixel is obtained, as shown in formula (6).
[0099] In summary, the loss function of the generator G is the weighted sum of the adversarial loss function loss_adv, the regional loss function loss_region and the pixel loss function loss_pixel, as shown in formula (7).
[0100]
[0101] In formula (7), α, β, and γ are the weighting coefficients of the adversarial loss function loss_adv, the regional loss function loss_region, and the pixel loss function loss_pixel, respectively, and α=1, β=1, and γ=50.
[0102] The loss function of the discriminator D is shown in formula (8).
[0103]
[0104] The loss function of the discriminator D uses a cross entropy loss function, and the cross entropy loss function is shown in formula (2);
[0105] L=-E[y i ·logp i +(1-y i )·log(1-p i )] (2)
[0106] In formula (2), E(*) represents the expected value of the distribution function, y i Indicates the label of input sample i. When a true sample is input, the label is 1, and when a false sample is input, the label is 0. i represents the probability that sample i is predicted to be true;
[0107] For the discriminator D, the label corresponding to the extended sample G(x1,y) is 0, and the label corresponding to the original sample y is 1, so the loss function of the discriminator D is as shown in formula (8).
[0108] When calculating the loss function of the generator G, only the input graph, the annotation information of a small amount of known samples and the prediction information of the discriminator D are used. The output graph has no standard reference label, so the generative adversarial network is semi-supervised.
[0109] Step 4: Use the Adam optimizer to train the semi-supervised generative adversarial network, and obtain the optimal generator with the lowest loss in the training iteration based on the loss function value of the generator G.
[0110] The sub-steps of step 4 are:
[0111] During optimization, select "train" as the mode, set batch size to 2, epoch nums to 150, GPU number to 1, Adam optimizer parameters to beta1 to 0.5, beta2 to 0.999, and learning rate to 0.002; set the training method to alternating iterative training with an alternating interval of 5 epochs.
[0112] Then put the pre-trained generator G and the pre-trained discriminator D into the ". / models / " folder; then put the SSDD original sample set and the SSDD Embedded embedded sample set into the ". / datasets / train / " folder; and start training.
[0113] Finally, based on the loss function value of the generator G, the optimal generator with the lowest loss in the training iteration is obtained.
[0114] refer to Figure 5 , which is a comparison diagram of the input graph of the optimal generator and the output graph of the optimal generator.
[0115] Step 5: Figure 5 Any SSDD Embedded sample shown in (a) is input into the optimal generator, and the optimal generator outputs a new sample, such as Figure 5 As shown in (b), the new samples are used as expanded samples of the original sample set of SSDD.
[0116] Although the present invention has been described in detail in general terms and in specific embodiments in this specification, it is obvious to those skilled in the art that some modifications or improvements may be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.
Claims
1. A SAR sample expansion method based on a semi-supervised generative adversarial network, characterized in that: The following steps are involved: Step 1: Create an embedded sample set and a fine-tuned sample set based on the original sample set; Step 2, select generator G; Design the pre-trained generator loss function and use multiple different fine-tuning samples to pre-train the generator G; Select the discriminator D, design the pre-trained discriminator loss function, and use multiple different original samples to pre-train the discriminator D; Step 3, connect the pre-trained generator G and the pre-trained discriminator D in sequence to establish a generative adversarial network, design a semi-supervised loss function, and obtain a semi-supervised generative adversarial network; Step 4, using the Adam optimizer to train the semi-supervised generative adversarial network, based on the loss function value of the generator G, to obtain the optimal generator with the lowest loss in the training iteration; Step 5, input any embedded sample into the optimal generator, the optimal generator outputs a new sample, and the new sample is used as an expanded sample of the original sample set; In step 3, the semi-supervised loss function includes the loss function of the generator G and the loss function of the discriminator D; The loss function of the generator G is divided into three parts: The adversarial loss function loss_adv between the generator G and the discriminator D; The regional loss function loss_region of the comparison between the output extended sample G(x1,y) and the input original sample y; The pixel loss function loss_pixel of the output extended sample G(x1,y) compared with the original sample y; The loss function of the generator G is the weighted sum of the adversarial loss function loss_adv, the regional loss function loss_region and the pixel loss function loss_pixel, as shown in formula (1); In formula (1), α, β, and γ are the weighting coefficients of the adversarial loss function loss_adv, the regional loss function loss_region, and the pixel loss function loss_pixel, respectively, and α=1, β=1, and γ=50; The loss function of the discriminator D uses the cross entropy loss function, as shown in formula (2); The cross entropy loss function is shown in formula (3); L=-E[y i ·logp i +(1-y i )·log(1-p i )] (3) In formula (3), E(*) represents the expected value of the distribution function, y i Indicates the label of input sample i. When a true sample is input, the label is 1, and when a false sample is input, the label is 0. i represents the probability that sample i is predicted to be true; For the discriminator D, the label corresponding to the extended sample G(x1,y) is 0, and the label corresponding to the original sample y is 1, so the loss function of the discriminator D is as shown in formula (2).
2. The SAR sample expansion method based on a semi-supervised generative adversarial network according to claim 1 is characterized in that: The specific sub-steps of step 1 are: using Photoshop software to synthesize any target slice image in the original sample set with any background image in the original sample set to obtain an embedded sample, so as to establish an embedded sample set; Use Photoshop software to fine-tune the background pixel brightness of any target in the original sample set to obtain a fine-tuning sample to establish a fine-tuning sample set; The embedded sample is denoted as x1, the fine-tuned sample is denoted as x2, and the original sample is denoted as y.
3. The SAR sample expansion method based on semi-supervised generative adversarial network according to claim 1 is characterized in that: In step 2, select generator G, which uses AttR2U-Net, which adds a cyclic residual convolution module and an attention gate module to U-Net. The U-Net includes an encoder and a decoder; the encoder is composed of a plurality of sub-modules in cascade, each of which includes a convolution layer and a pooling layer. Each time the input image passes through a sub-module, abstract features are extracted and the resolution is reduced; the decoder is composed of a plurality of sub-modules in cascade, each of which includes a convolution layer and an upsampling layer, and is used to restore the abstract features learned by the encoder to the resolution; The cyclic residual convolution module effectively improves network performance by combining cyclic connections and residual connections; the attention gate module is used to readjust the output features of the encoder before splicing the features of each resolution of the encoder Encoder with the corresponding features in the decoder Decoder.
4. The SAR sample expansion method based on a semi-supervised generative adversarial network according to claim 1, characterized in that: In step 2, the process of pre-training generator G is as follows: any fine-tuning sample x2 is input into generator G, and all original samples y in the original sample set are input into generator G as conditions. Generator G outputs pre-training extended sample G(x2,y), and a round of training is completed. The generator G is pre-trained for 20 rounds to obtain the pre-trained generator G.
5. The SAR sample expansion method based on semi-supervised generative adversarial network according to claim 4 is characterized in that: The loss function of the pre-trained generator uses the mean square error between the original sample y and the pre-trained extended sample G(x2,y), as shown in formula (4); Loss G =E same [||y-G(x2,y)||2]+αE diff [||y-G(x2,y)||2] (4) In formula (4), E(*) represents the expected value of the distribution function; same refers to the area where the fine-tuned sample x2 is the same as the original sample y, and diff refers to the area where the fine-tuned sample x2 is different from the original sample y; α represents the weight coefficient of the diff area, and α is 3.
6. The SAR sample expansion method based on a semi-supervised generative adversarial network according to claim 1, characterized in that: In step 2, a discriminator D is selected. The discriminator D adopts a CNN recognition network, including C1-C5 layers and S1 layer. The specific structure and parameters are as follows: C1 layer, the specific structure is The output feature map size is 208*208*32; C2 layer, the specific structure is The output feature map size is 52*52*32; C3 layer, the specific structure is The output feature map size is 52*52*32; C4 layer, the specific structure is The output feature map size is 26*26*64; C5 layer, the specific structure is The output feature map size is 13*13*1; The S1 layer has the following structure: Output probability value p.
7. The SAR sample expansion method based on semi-supervised generative adversarial network according to claim 1, characterized in that: In step 2, the process of pre-training the discriminator D is as follows: any original sample y or any of the pre-training extended samples G(x2,y) is input into the discriminator D, and the discriminator D outputs a probability value p, thus completing one round of training; The probability value p represents the possibility that the input to the discriminator D is a true sample. For the discriminator D, the original sample y is a true sample, and the pre-trained extended sample G(x2,y) is a false sample; The discriminator D is pre-trained for 20 rounds to obtain the pre-trained discriminator D.
8. The SAR sample expansion method based on semi-supervised generative adversarial network according to claim 7 is characterized in that: The pre-trained discriminator loss function adopts the cross entropy loss function, and the cross entropy loss function is shown in formula (5); L=-E[y i ·logp i +(1-y i )·log(1-p i )] (5) In formula (5), E(*) represents the expected value of the distribution function, y i Indicates the label of input sample i. When a true sample is input, the label is 1, and when a false sample is input, the label is 0. i represents the probability that sample i is predicted to be true; For the discriminator D, the label corresponding to the pre-trained extended sample G(x2,y) is 0, and the label corresponding to the original sample y is 1, so the loss function of the pre-trained discriminator is shown in formula (6); 9. The SAR sample expansion method based on a semi-supervised generative adversarial network according to claim 1, characterized in that: In step 3, the process of generating the adversarial network is as follows: input any embedded sample x1 into the generator G, and input all the original samples y as conditions into the generator G, and the generator G outputs the extended sample G(x1,y); input any original sample y or any extended sample G(x1,y) into the discriminator D, and the discriminator D outputs the probability value p.