SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning

Abstract

A SAR true and false target identification and target recognition method based on decoupling and reconstruction learning belongs to the field of remote sensing image processing technology. The present invention mainly aims at the problem of difficulty in true and false target identification and target recognition, and constructs a SAR true and false target identification method based on decoupling and reconstruction learning, which realizes the decoupling of false correlation features and target category correlation features, thereby promoting the application of true and false target identification and target recognition. The present invention first obtains false targets by simulation, constructs a true and false target data set in combination with real target data, and then constructs a convolutional autoencoder as the backbone network for true and false target feature extraction and reconstruction learning, and constructs a decoupling and reconstruction learning model based on the backbone network. After model training, the decoupled false correlation features and category correlation features are obtained, ensuring the sparsity and integrity of the decoupled features, improving the robustness of the features to noise, and then achieving high-precision true and false target identification and target recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of SAR remote sensing image processing, in particular to a high-resolution SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging radar that uses pulse compression and synthetic aperture to acquire high-resolution SAR images. This radar operates around the clock and in all weather conditions, has a certain degree of penetration into vegetation, and holds significant application value in both military and civilian fields. With the advancement of SAR system technology, imaging modes have become increasingly diverse. A single SAR platform may have multiple receiving antennas. Errors such as platform attitude variations and amplitude and phase imbalances between sensor channels can introduce multi-channel effects, leading to the appearance of false targets. These false targets appear slightly darker than real targets in the image, but their geometrical appearance is similar to that of real targets, making it difficult to distinguish between true and false targets. The presence of these false targets poses significant challenges to SAR ground observation and target detection and recognition. Therefore, research on true and false target discrimination and recognition is crucial for improving the performance of SAR remote sensing applications.

[0003] For the critical applications of true and false target discrimination and recognition in SAR images, existing methods primarily extract features from true and false targets and then construct multi-task classifiers to achieve end-to-end model optimization for both true and false target discrimination and recognition. However, these methods ignore the coupling of false correlation features with target category correlation features, resulting in the model only learning false data associations. This leads to insufficient feature representation capabilities, which in turn can easily lead to performance degradation in practical applications and makes it difficult to meet the requirements of high-precision target recognition applications. Therefore, existing technologies for true and false target discrimination and recognition still need improvement and development. Summary of the Invention

[0004] Aiming at the difficulty of distinguishing true and false targets and identifying targets, the present invention constructs a SAR true and false target distinction and target identification method based on decoupling and reconstruction learning, realizes the decoupling of false correlation features and target category correlation features, and thus achieves high-precision true and false target distinction and target identification.

[0005] The present invention is achieved through the following technical solutions:

[0006] A SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning comprises the following steps:

[0007] Step 1: Prepare a dataset of real and fake SAR target slices. Use high-resolution SAR data as a real target template and use simulation to generate a fake target image corresponding to the template. The fake target label is 1 and the real target label is 0.

[0008] Step 2: Dataset division and preprocessing: Based on the prepared true and false SAR image dataset, the true and false target samples of each category are divided into training sets and test sets according to a certain ratio. The ratio of true and false target images to training sets and test sets is the same. All image slice centers are cropped to the same size and stored in a unified format.

[0009] Step 3: Construct a convolutional autoencoder model. The main structure of this model consists of an encoder and a decoder structure. The encoder structure is used to extract the latent features of the target, while the decoder structure reconstructs the latent features into the original image to ensure the integrity of the extracted features.

[0010] Step 4: Based on the constructed convolutional autoencoder model, a decoupling and reconstruction learning model is constructed. To extract false correlation features and category correlation features respectively, under the criteria of feature independence and information integrity, true and false target identification loss, target recognition loss, decoupling loss, and reconstruction loss are constructed. The training data is used to jointly optimize these four loss functions to achieve decoupling and reconstruction learning, and then train to obtain the optimal model weights.

[0011] Step 5: Use the trained model weights to achieve true and false target identification and target recognition. Discard the decoder structure and use the encoder structure to extract deep features of the input SAR image target. Then, perform feature decoupling on the extracted deep features to obtain false correlation features and category correlation features to achieve true and false target identification and target recognition. Finally, use the target recognition rate and false target recognition rate to evaluate the performance.

[0012] Furthermore, the step 1 is specifically as follows:

[0013] Step 1: Prepare a high-resolution SAR image target slice dataset as a real target template and label each target slice with a category.

[0014] Step 2: Generate a false target image corresponding to the template by adding a 1% quadratic phase error to the range-Doppler domain of the real target image, set the false target label to 1, and the real target label to 0.

[0015] Furthermore, the step 2 is specifically as follows:

[0016] Step 1: Dataset partitioning and preprocessing. Based on the prepared true and false SAR image dataset, the true and false target samples of each category are divided into training sets according to the sample ratio of 5%, 10%, and 20%. The ratio of true and false target images of each category is the same, and the remaining samples are used as the test set.

[0017] Step 2: Crop and scale all image slices to 128×128 in size and store them in a unified file format.

[0018] Furthermore, the step three is specifically as follows:

[0019] Step 1: Construct the encoder of the convolutional autoencoder: encoder f e It consists of 5 convolutional layers, 1 residual module and 1 fully connected layer. After 5 downsampling operations, the receptive field is expanded. Each convolutional layer consists of a convolution operation, an instance normalization layer and a ReLU activation function. The residual module consists of two convolutional layers and a residual connection. The number of input channels of the encoder is 3, and the number of output channels of each convolutional layer is 64, 128, 256, 256 and 256 respectively. The number of input and output channels of the residual module is 256. In order to reduce the amount of computation, the features output by the encoder are reduced in dimension using a fully connected layer. After dimensionality reduction, the ReLU activation function is used to enhance the nonlinear representation ability of the features. To avoid overfitting, the dropout rate is set to 0.5 after the fully connected layer. During training, 50% of the fully connected layer neurons are randomly discarded to enhance the expression ability of the fully connected layer.

[0020] Step 2: Construct the decoder of the convolutional autoencoder: The decoder structure is exactly the opposite of the encoder result. Transposed convolution is used instead of convolution operation to decode the latent variables. Similar to the encoder structure, the decoder structure consists of 5 transposed convolution layers, 1 residual module, and 1 fully connected layer. The function of the fully connected layer is to increase the dimension of the latent variables to meet the calculation requirements of the subsequent transposed convolution layer. After the dimension increase, the ReLU activation function is used to enhance the nonlinear representation ability of the latent variables. To avoid overfitting, the dropout rate is set to 0.5 after the fully connected layer. During the training process, 50% of the fully connected layer neurons are randomly dropped to enhance the expression ability of the fully connected layer. Each transposed convolution layer consists of a transposed convolution operation, an instance normalization layer, and a ReLU activation function. The structure of the residual module is consistent with that of the encoder. Corresponding to the encoder structure, the number of input channels of the decoder module is 256, the number of input and output channels of the residual module is also 256, and the number of output channels of the five transposed convolution layers is 256, 256, 128, 64, and 3, respectively, to achieve reconstruction of the input image.

[0021] Furthermore, the step 4 is specifically as follows:

[0022] Step 1: Using the encoder f of the convolutional autoencoder e The input image X is processed through the nonlinear mapping of the neural network to obtain the encoded feature vector Z0:

[0023] Z0=f e (X) (1)

[0024] Step 2: On this basis, use a multivariate independent Gaussian mixture model to characterize the false correlation feature Z fake and category-associated feature Z cls , each dimension of these two types of features is composed of independent latent variables, and each latent variable obeys the normal distribution. The neural network is used to calculate the mean vector μ and the variance vector σ 2 Make an estimate:

[0025]

[0026] Among them, f μ and f σ Represent the predicted mean vector μ and variance vector σ respectively 2 A fully connected neural network with multiple layers; further use KL divergence (Kullback-Leibler divergence) to constrain the latent variables to be as close as possible to the standard normal distribution N(0,I) and calculate the decoupling loss:

[0027] L Dec =KL(N(μ,σ 2 )||N(0,I)) (3)

[0028] Among them, the mean vector μ=[μ0,...,μ d-1 ], standard deviation vector σ=[σ0,...,σ d-1 ], since each variable is independent of each other, taking the KL divergence of the univariate normal distribution as an example, we can get:

[0029]

[0030] For a normal distribution with multiple independent variables, the decoupling loss can be calculated as:

[0031]

[0032] Among them, d is the number of latent variables, μ i and σ i are the i-th normal distribution N(μ i ,σ i 2 )’s latent variable mean and standard deviation, the i-th latent variable z can be obtained by reparameterization method i :

[0033] z i =μ i +ε i ×σ i (6)

[0034] Among them, ε i is a sample value from a standard normal distribution, the false correlation feature Z fake and category-associated feature Z cls Each feature contains d / 2 latent variables z i , take the independent latent variables z from the 0th to the d / 2-1th dimension i As a false correlation feature Z fake , take the independent latent variables z from the d / 2th to the d-1th dimension i As a false correlation feature Z cls :

[0035]

[0036] Step 3: To ensure the false correlation hidden feature Z fake Hidden feature Z associated with the category cls The information integrity of Z, that is, it contains all the information of the input image. The input image can be reconstructed using these two types of features; specifically, the false correlation latent feature Z fake Hidden feature Z associated with the category cls Then pass through the decoder f d Mapping the latent feature representation back to the image space yields the reconstructed image:

[0037] X′=f d (Concat(Z fake ,Z cls )) (8)

[0038] Then by constraining the reconstruction error between the input image and the reconstructed image to be the smallest, Z fake and Z cls Contains all information; this implementation intends to use the two norm to calculate the reconstruction error, and calculate the reconstruction loss as:

[0039] L rec =(Xf d (Concat(Z fake ,Z cls ))) 2 (9)

[0040] Step 4: In order to ensure that the category association features contain the key features of target recognition and the false association features contain information that is falsely related to the target, the target recognition loss and the true and false identification loss are used to inject effective information for target recognition and true and false target identification; the target recognition loss uses the cross entropy loss function to calculate the prediction probability of the i-th category and the i-th category label Error:

[0041]

[0042] Among them, C is the number of categories, and the true and false identification loss uses the binary cross entropy function to calculate the false target prediction probability p fake and true and false target labels y fake error:

[0043] L fake =-y fake logp fake -(1-y fake )log(1-p fake ) (11)

[0044] Step 5: In order to obtain the feature representation of distinguishable category associations and the feature representation of false associations, the above four types of losses are jointly optimized, and the overall loss function is

[0045] L=L cls +L fake +λ1L dec +λ2L rec (12)

[0046] Among them, λ1 and λ2 are the L dec and L rec The hyperparameters of loss are set to 1 and 0.4 in the experiment. The stochastic gradient descent algorithm (SGD algorithm) with momentum is used to optimize the loss function of formula (12) during training. In order to ensure the stability of the training process and speed up the convergence, the learning rate is adjusted using cosine annealing:

[0047]

[0048] Where N is the number of training epochs, η t ,η max and η min are the current, maximum and minimum learning rates respectively. In the experiment, η max =0.001,η min =0.0002, batch size is set to 4 during training, and a total of 50 epochs are trained.

[0049] Furthermore, the step five is specifically as follows:

[0050] Step 1: Use the trained model weights to achieve true and false target identification and target recognition. Discard the decoder structure and use the encoder structure to extract deep features of the SAR image targets in the test set. Then, decouple the extracted deep features to obtain false correlation features and category correlation features to achieve true and false target identification and target recognition. The true and false target identification confidence threshold is set to 0.5. If the prediction confidence is greater than 0.5, it is judged as a false target, otherwise it is a real target.

[0051] Step 2: Use target recognition rate and false target recognition rate as evaluation indicators. The evaluation indicators are defined as follows:

[0052]

[0053] Among them, N tp is the number of correctly identified samples, N gt is the total number of samples, Number of false targets correctly identified, N c is the total number of false samples.

[0054] The beneficial effects of the present invention compared to the prior art are:

[0055] 1. The present invention extracts the false correlation features and category correlation features of SAR targets through decoupling and reconstruction learning, so that the various dimensional features for true and false target identification and target recognition are independent of each other, the feature distribution is sparser, and it is more robust to noise disturbances, thereby making the model convergence process faster and the accuracy of true and false target identification and target recognition higher.

[0056] 2. The present invention can be applied to the field of SAR remote sensing image processing and can achieve high-precision true and false target discrimination and target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A framework diagram for the overall solution of decoupling and reconstruction learning;

[0058] Figure 2 Schematic diagram of the decoupling and reconstruction learning model;

[0059] Figure 3 Decoupling and reconstructing learning model diagrams for the testing process;

[0060] Figure 4 for Figure 2 Diagram of the encoder structure of the model’s convolutional autoencoder;

[0061] Figure 5 Schematic diagram of convolution operation;

[0062] Figure 6 Schematic diagram of the transposed convolution operation;

[0063] Figure 7 for Figure 2 The decoder structure diagram of the model's convolutional autoencoder;

[0064] Figure 8 This is a learning rate change curve based on cosine annealing;

[0065] Figure 9 Schematic diagram of optical images and true and false target SAR images of 10 types of targets;

[0066] Figure 10 This is a loss convergence curve diagram of the method of the present invention;

[0067] Figure 11 This is the loss convergence curve of the comparison method;

[0068] Figure 12 This is the feature distribution diagram after decoupling. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0070] The present invention mainly addresses the problem of difficulty in distinguishing true and false targets and identifying targets, and constructs a SAR true and false target identification method based on decoupling and reconstruction learning, which realizes the decoupling of false correlation features and target category correlation features, thereby promoting the application of true and false target identification and target identification. The present invention first obtains false targets through simulation, constructs a true and false target data set in combination with real target data, and then constructs a convolutional autoencoder as the backbone network for true and false target feature extraction and reconstruction learning, and constructs a decoupling and reconstruction learning model based on the backbone network. After model training, the decoupled false correlation features and category correlation features are obtained, ensuring the sparsity and integrity of the decoupled features, improving the robustness of the features to noise, and thus achieving high-precision true and false target identification and target recognition.

[0071] Example 1:

[0072] Combined with attachment Figures 1-8 The present invention proposes a SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning, comprising the following steps:

[0073] Step 1: Prepare a dataset of true and false SAR target slices. This embodiment uses the MSTAR target dataset as the true target template data with a resolution of 0.3m×0.3m. It uses 10 types of targets at a 17-degree side viewing angle, including 2S1 self-propelled howitzers, BMP2 infantry fighting vehicles, BRDM-2 armored reconnaissance vehicles, BTR-70 armored transport vehicles, BTR-60 armored transport vehicles, D7 bulldozers, T62 tanks, T72 tanks, ZIL131 freight trucks, and ZSU-23 / 4 self-propelled anti-aircraft guns. Using high-resolution SAR data as the true target template, a simulation algorithm is used to add a 1% quadratic phase error to generate a false target image corresponding to the template. The false target label is 1 and the true target label is 0. The true images and false images of each category in the training set are as follows: Figure 9 shown.

[0074] Step 2: Dataset division and preprocessing: Based on the prepared dataset of real and fake SAR images, the training set is divided into 5%, 10%, and 20% of the samples, and the rest is the test set. All image slices are cropped to 128×128 in the center and stored in a unified file format.

[0075] Step 3: Use the backbone structure of the convolutional autoencoder to decouple and reconstruct the category-related features and false-related features. The model structure is as follows: Figure 2 As shown in the figure. The constructed convolutional autoencoder model is an unsupervised neural network. The main structure of the model includes an encoder and a decoder structure. By optimizing this generative model, it is possible to learn a latent representation with as little information loss as possible, so that the original input can be reconstructed through the latent representation. The encoder structure is used to extract the latent features of the target, while the decoder structure reconstructs the latent features into the original image to ensure that the extracted latent variables have information integrity. The structural diagram of the convolutional autoencoder is shown in the figure. Figure 2 As shown, the encoder and decoder of the convolutional autoencoder are stacked using convolutional neural networks. e It consists of 5 convolutional layers, 1 residual module and 1 fully connected layer. After 5 downsampling operations, the receptive field is expanded. The specific structure is as follows Figure 4As shown, each convolutional layer consists of a convolution operation, an instance normalization layer and a ReLU activation function; the residual module consists of two convolutional layers and a residual connection. The number of encoder input channels is 3, and the number of output channels of each convolutional layer is 64, 128, 256, 256 and 256 respectively. The number of input and output channels of the residual module is 256. In order to reduce the amount of calculation and enhance the computational efficiency, a fully connected layer is used to reduce the dimensionality of the high-dimensional features output by the encoder. After dimensionality reduction, the ReLU activation function is used to enhance the nonlinear representation ability of the features. In order to reduce overfitting, the Dropout rate is set to 0.5 after the fully connected layer. During the training process, 50% of the fully connected layer neurons are randomly discarded to enhance the expression ability of the fully connected layer. The decoder structure is exactly the opposite of the encoder result. Transposed convolution is used instead of convolution operation to decode the latent variables. The schematic diagrams of forward convolution and transposed convolution are shown as follows. Figure 5 and Figure 6 Similar to the encoder structure, the decoder structure consists of 5 transposed convolutional layers, 1 residual module and 1 fully connected layer. The detailed structure is shown in Figure 7 As shown in the figure, the fully connected layer increases the feature dimension to meet the computational requirements of the subsequent transposed convolutional layer. It also restores the feature dimensions reduced in the encoder's fully connected layer to meet the requirements of reconstruction learning. After dimensionality increase, the ReLU activation function is used to enhance the nonlinear representation of the features. To reduce overfitting, a dropout rate of 0.5 is set after the fully connected layer, and 50% of the fully connected layer neurons are randomly dropped during training to enhance the representational power of the fully connected layer. Each transposed convolutional layer consists of a transposed convolution operation, an instance normalization layer, and a ReLU activation function. The residual module has the same structure as the encoder. Corresponding to the encoder structure, the decoder module has 256 input channels, and the residual module has 256 input and output channels. The output channels of the five transposed convolutional layers are 256, 256, 128, 64, and 3, respectively. The encoder and decoder structures form the backbone network for decoupling and reconstruction learning, used to extract target features. The goal is to learn sparse and complete latent variables through the reconstruction learning process of the convolutional autoencoder to represent the target's characteristics.

[0076] Step 4: Based on the constructed convolutional autoencoder model, in order to extract false correlation features and category correlation features respectively, under the principle of feature independence and information integrity, construct true and false target identification loss, target recognition loss, decoupling loss and reconstruction loss, use training data to jointly optimize these four types of loss functions to achieve decoupling and reconstruction learning, and then train to obtain the optimal model weights. The model structure is as follows: Figure 2 As shown. The encoder f using convolutional autoencoder e The input image X is passed through the nonlinear mapping of the neural network to obtain the encoded feature vector Z0:

[0077] Z0=f e (X) (1)

[0078] On this basis, in order to obtain the eigenvectors of each decoupled factor, the neural network is further used to decouple the eigenvector Z0 under the constraint of decoupling loss to obtain mutually independent latent variables, and the decoupled latent variables are used to distinguish true and false targets and identify targets. Specifically, assuming that the false correlation feature Z fake and category-associated feature Z cls Each contains d / 2 independent latent variables, each latent variable is an independent random variable, and each random variable obeys the normal distribution N(μ,σ 2 ). Then, the probability distribution of each random variable can be used to represent the potential factors related to a certain characteristic of the target, and then the false correlation feature Z fake and category-associated feature Z cls It can be characterized by a multivariate independent Gaussian mixture model. In order to estimate the distribution of each variable, a neural network is used to calculate the mean vector μ and the variance vector σ. 2 Make an estimate, so

[0079]

[0080] Among them, f μ and f σ Represents the fully connected neural network of the predicted mean and variance respectively. Once the mean and variance of the random variable distribution are determined, the distribution of the random variable can be determined, and the false correlation latent feature Z is obtained. fake Hidden feature Z associated with the category cls Furthermore, to facilitate optimization, the latent variables obtained by decoupling are constrained to be as close to the standard normal distribution as possible, and the KL divergence (Kullback-Leibler divergence) between the probability distribution of the latent variables and the standard normal distribution N(0,I) is then constrained to be as small as possible. The decoupling loss can then be calculated using the KL divergence:

[0081] L Dec =KL(N(μ,σ 2 )||N(0,I)), (3)

[0082] where the mean vector μ=[μ0,...,μ d-1 ], standard deviation vector σ=[σ0,...,σ d-1 ], since each variable is independent of each other, taking the KL divergence of the univariate normal distribution as an example, we can get

[0083]

[0084] Therefore, for a normal distribution with multiple independent variables, the decoupling loss can be obtained:

[0085]

[0086] Among them, d is the number of latent variables, μ i and σ i are the i-th normal distribution N(μ i ,σ i 2 )’s latent variable mean and standard deviation, the i-th latent variable z can be obtained by reparameterization method i :

[0087] z i =μ i +ε i ×σ i (6)

[0088] Among them, ε i is a sample value from a standard normal distribution, the false correlation feature Z fake and category-associated feature Z cls Each feature contains d / 2 latent variables z i , take the independent latent variables z from the 0th to the d / 2-1th dimension i As a false correlation feature Z fake , take the independent latent variables z from the d / 2th to the d-1th dimension i As a false correlation feature Z cls :

[0089]

[0090] In order to ensure that the false correlation latent feature Z fake Hidden feature Z associated with the category cls The information integrity of , that is, it contains all the information of the input image. It is hoped that the input image can be reconstructed by using these two types of features. Specifically, the false correlation latent feature Z fake Hidden feature Z associated with the category cls Then pass through the decoder f d Mapping the latent feature representation back to the image space yields the reconstructed image:

[0091] X′=f d (Concat(Z fake ,Z cls )). (8)

[0092] Then by constraining the reconstruction error between the input image and the reconstructed image to be the smallest, Z fake and Z clsThis embodiment uses the bi-norm to calculate the reconstruction error, and the reconstruction loss is calculated as:

[0093] L rec =(Xf d (Concat(Z fake ,Z cls ))) 2 (9)

[0094] Furthermore, in order to ensure the effectiveness of the decoupled features, ensure that the category-related features contain the key features for target recognition, and the false-related features contain information related to false targets, target recognition loss and true-false discrimination loss are used to inject effective information for target recognition and true-false target discrimination. cls and the true and false target identification module fc fake It consists of two fully connected layers. Based on the category association feature Z cls The predicted probability of the i-th target category is Based on the false correlation feature Z fake The false target prediction probability can be expressed as p fake =sigmoid(fc fake (Z fake )), where the activation function i=1,...,C,activation function sigmoid(x)=1 / (1+e -x ), where C is the number of categories.

[0095] After the activation function, the predicted probability is obtained in the range of [0,1]. During the training process, for the optimization process of multi-category target recognition problem, the target recognition loss uses the cross entropy loss function to calculate the prediction probability of the i-th category. and the i-th category label The target recognition loss can be calculated as:

[0096]

[0097] For the problem of distinguishing true and false targets in binary classification, the true and false target identification loss uses the binary cross entropy function to calculate the false target prediction probability p fake and true and false target labels y fake error:

[0098] L fake =-y fake logp fake -(1-y fake )log(1-p fake ) (11)

[0099] Among them, p fakeis the probability of predicting a false target, y fake are the labels of true and false targets. In order to obtain the feature representation of distinguishable category associations and false associations, the classification loss and the variational convolutional autoencoder loss are jointly optimized, and the overall loss function is

[0100] L=L cls +L fake +λ1L dec +λ2L rec (12)

[0101] Among them, λ1 and λ2 are used to balance L dec and L rec The loss hyperparameters are set to 1 and 0.4 respectively. During the training process, the stochastic gradient descent algorithm with momentum (SGD algorithm) is used to optimize the loss function of formula (12), with momentum set to 0.9 and weight decay set to 5×10 -4 Under the constraint of momentum, the model converges faster. In order to ensure stability during training and speed up convergence, cosine annealing is used to adjust the learning rate:

[0102]

[0103] Where N is the number of training epochs, η t ,η max and η min are the current, maximum and minimum learning rates respectively. In the experiment, η max =0.001,η min =0.0002, the learning rate change curve is as follows Figure 8 As shown in the figure, during the training process, the batch size is set to 4 and a total of 50 epochs are trained.

[0104] Step 5: Use the trained model weights to achieve true and false target identification and target recognition, discard the decoder structure, and the model structure is as follows Figure 3 As shown in the figure, the encoder structure is used to extract deep features of SAR image targets in the test set, and then the extracted deep features are decoupled to obtain false correlation features and category correlation features, thereby realizing the identification of true and false targets and target recognition. The true and false target identification confidence threshold is set to 0.5. If the prediction confidence is greater than 0.5, it is judged as a false target, otherwise it is a real target. This embodiment uses target recognition rate and false target recognition rate as evaluation indicators. The evaluation indicators are defined as follows:

[0105]

[0106] Among them, N tp is the number of correctly identified samples, N gt is the total number of samples, Number of false targets correctly identified, N c is the total number of false samples.

[0107] The true and false target discrimination and target recognition performance of the present invention can be further illustrated by the following experiments.

[0108] This experiment uses MSTAR data as a template to simulate false samples with a phase error of 1% to obtain a dataset of true and false targets. The dataset example is as follows: Figure 9 The number of samples in each category is shown in Table 1. This experiment used 5% training samples, 10% training samples and 20% training samples for performance testing. The comparison method was to directly use the encoder for feature extraction without using the decoupling and reconstruction learning method, and then use the extracted features to respectively use the same classifier as the proposed method to achieve true and false target identification and target recognition. In order to ensure the fairness of the experiment, the input data and training methods of the two methods are exactly the same during the experiment. The experimental results are shown in Figure 10 , Figure 11 , Figure 12 and as shown in Table 2.

[0109] Table 1 Experimental sample categories and quantities

[0110] category 2S1 BMP2 BRDM-2 BTR-70 BTR-60 D7 T62 T72 ZIL131 ZSU-23 / 4 Real goal 299 196 298 196 256 299 299 196 299 299 False Target 299 196 298 196 256 299 299 196 299 299

[0111] contrast Figure 10 As can be seen, the proposed method converges quickly after 20 epochs of training, achieving a faster convergence rate than the comparison method. The loss changes more smoothly during training, indicating that the proposed object recognition method based on feature decoupling and reconstruction learning is more stable during training.

[0112] In order to quantitatively analyze the feature distribution of decoupled category-related features and false-related features, t-SNE is used to visualize the decoupled features, such as Figure 12 As shown. As can be seen from the figure, the two types of features after decoupling are independent of each other, which verifies the effectiveness of feature decoupling. Subsequently, category-related features will be used for target recognition, and false-related features will be used to distinguish true and false targets. The experimental results are shown in Table 2. As can be seen from the table, the target recognition performance of the comparison method can only exceed 79% under the condition of 20% training samples, while the method used in the present invention can achieve 89% recognition performance under the condition of 10% training samples, and even the target recognition performance exceeds 90% under the condition of 20% training samples. It shows that whether it is under the conditions of a small number of samples or under the conditions of sufficient training samples, the present invention has a greater advantage in target recognition accuracy while ensuring a higher accuracy in distinguishing true and false targets.

[0113] Table 2 Recognition results of different sample ratios under 1% quadratic phase simulation error

[0114]

Claims

1. A SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning, characterized by: The method comprises the following steps: Step 1: Prepare a dataset of real and fake SAR target slices. Use high-resolution SAR data as a real target template and use simulation to generate a fake target image corresponding to the template. The fake target label is 1 and the real target label is 0. Step 2: Dataset division and preprocessing: Based on the prepared true and false SAR image dataset, the true and false target samples of each category are divided into training sets and test sets according to a certain ratio. The ratio of true and false target images to training sets and test sets is the same. All image slice centers are cropped to the same size and stored in a unified format. Step 3: Construct a convolutional autoencoder model. The main structure of this model consists of an encoder and a decoder structure. The encoder structure is used to extract the latent features of the target, while the decoder structure reconstructs the latent features into the original image to ensure the integrity of the extracted features. Step 4: Based on the constructed convolutional autoencoder model, a decoupling and reconstruction learning model is constructed. To extract false correlation features and category correlation features respectively, under the criteria of feature independence and information integrity, true and false target identification loss, target recognition loss, decoupling loss, and reconstruction loss are constructed. The training data is used to jointly optimize these four loss functions to achieve decoupling and reconstruction learning, and then train to obtain the optimal model weights. Step 5: Use the trained model weights to achieve true and false target identification and target recognition. Discard the decoder structure and use the encoder structure to extract deep features of the input SAR image target. Then, perform feature decoupling on the extracted deep features to obtain false correlation features and category correlation features to achieve true and false target identification and target recognition. Finally, use the target recognition rate and false target recognition rate to evaluate the performance.

2. The SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning according to claim 1 is characterized by: The step 1 is specifically as follows: Step 1: Prepare a high-resolution SAR image target slice dataset as a real target template and label each target slice with a category. Step 2: Generate a false target image corresponding to the template by adding a 1% quadratic phase error to the range-Doppler domain of the real target image, set the false target label to 1, and the real target label to 0.

3. The SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning according to claim 1 is characterized by: The step 2 is specifically as follows: Step 1: Dataset partitioning and preprocessing. Based on the prepared true and false SAR image dataset, the true and false target samples of each category are divided into training sets according to the sample ratio of 5%, 10%, and 20%. The ratio of true and false target images of each category is the same, and the remaining samples are used as the test set. Step 2: Crop and scale all image slices to the center size and stored in a unified file format.

4. The SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning according to claim 1 is characterized by: The step three is specifically as follows: Step 1: Construct the encoder of the convolutional autoencoder: encoder It consists of 5 convolutional layers, 1 residual module, and 1 fully connected layer. It expands the receptive field through 5 downsampling operations. Each convolutional layer consists of a convolution operation, an instance normalization layer, and a ReLU activation function. The residual module consists of two convolutional layers and a residual connection. The encoder has 3 input channels, and the output channels of each convolutional layer are 64, 128, 256, 256, and 256, respectively. The input and output channels of the residual module are both 256. The encoder output features are reduced in dimension using a fully connected layer. After dimensionality reduction, the ReLU activation function is used to enhance the nonlinear representation ability of the features. The dropout rate is set to 0.5 after the fully connected layer. During training, 50% of the fully connected layer neurons are randomly discarded to enhance the expression ability of the fully connected layer. The second step is to build the decoder of the convolutional autoencoder. The decoder structure is exactly the opposite of the encoder result. Transposed convolution is used instead of convolution operation to decode the latent variables. Similar to the composition of the encoder structure, the decoder structure consists of 5 transposed convolution layers, 1 residual module and 1 fully connected layer. The function of the fully connected layer is to increase the dimension of the latent variables to meet the calculation requirements of the subsequent transposed convolution layer. After the dimension increase, the ReLU activation function is used to enhance the nonlinear representation ability of the latent variables. The Dropout rate is set to 0.5 after the fully connected layer. During the training process, 50% of the fully connected layer neurons are randomly discarded to enhance the expression ability of the fully connected layer. Each transposed convolution layer consists of a transposed convolution operation, an instance normalization layer and a ReLU activation function. The structure of the residual module is consistent with that of the encoder structure. Corresponding to the structure of the encoder, the number of input channels of the decoder module is 256, the number of input and output channels of the residual module is 256, and the number of output channels of the 5 transposed convolution layers are 256, 256, 128, 64, and 3 respectively to achieve reconstruction of the input image.

5. The SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning according to claim 1 is characterized by: The step 4 is specifically as follows: Step 1: Using the Convolutional Autoencoder Encoder Input picture The encoded feature vector is obtained through nonlinear mapping of the neural network : （1） Step 2: Use multivariate independent Gaussian mixture models to characterize spurious correlation features and category-associated features , each dimension of these two types of features is composed of independent latent variables, and each latent variable obeys the normal distribution. The neural network is used to calculate the mean vector and variance vector Make an estimate: （2） in, and Represent the predicted mean vectors and variance vector Fully connected neural network with multiple layers; KL divergence (Kullback-Leibler divergence) is used to constrain latent variables to be as close to the standard normal distribution as possible. , calculate the decoupling loss: （3） Among them, the mean vector , the standard deviation vector , since each variable is independent of each other, taking the KL divergence of the univariate normal distribution as an example, we can get: （4） For a normal distribution with multiple independent variables, the decoupling loss can be calculated as: （5） in, is the number of latent variables, and Respectively Normal distribution The mean and standard deviation of the latent variables can be obtained by reparameterization method. Hidden variables : （6） in, is a sample value from a standard normal distribution, the false correlation feature and category-associated features Features include Hidden variables , take the 0th to dimensional independent latent variables As a false correlation feature , take the To dimensional independent latent variables As a false correlation feature : （7） Step 3: To ensure false correlation hidden features Hidden features associated with categories The information integrity of the input image is that it contains all the information of the input image. The input image can be reconstructed using these two types of features. Specifically, the false correlation latent features Hidden features associated with categories Then pass through the decoder Mapping the latent feature representation back to the image space yields the reconstructed image: （8） Then by constraining the reconstruction error between the input image and the reconstructed image to be the smallest, and Contains all information; the reconstruction error is calculated using the two norm, and the reconstruction loss is calculated as: （9） Step 4: In order to ensure that the category-related features contain the key features for target recognition and the false-related features contain the false-related information of the target, target recognition loss and true-false discrimination loss are used to inject effective information for target recognition and true-false target discrimination; target recognition loss is calculated using the cross entropy loss function. Class prediction probability With the Category labels Error: （10） in, is the number of categories, and the true and false identification loss uses the binary cross entropy function to calculate the false target prediction probability and true and false target labels error: （11） Step 5: In order to obtain the feature representation of distinguishable category associations and the feature representation of false associations, the above four types of losses are jointly optimized, and the overall loss function is （12） in, and During balance training and The loss hyperparameters are used. During the training process, the stochastic gradient descent algorithm with momentum is used to optimize the loss function of formula (12). In order to ensure the stability during the training process and speed up the convergence, the cosine annealing is used to adjust the learning rate: （13） in, is the number of training epochs, 、 and are the current, maximum and minimum learning rates respectively.

6. The SAR true and false target discrimination and target recognition method based on decoupling and reconstruction learning according to claim 1, characterized in that: The step five is specifically as follows: Step 1: Use the trained model weights to achieve true and false target identification and target recognition. Discard the decoder structure and use the encoder structure to extract deep features of the SAR image targets in the test set. Then, decouple the extracted deep features to obtain false correlation features and category correlation features to achieve true and false target identification and target recognition. The true and false target identification confidence threshold is set to 0.

5. If the prediction confidence is greater than 0.5, it is judged as a false target, otherwise it is a real target. Step 2: Use target recognition rate and false target recognition rate as evaluation indicators. The evaluation indicators are defined as follows: （14） in, is the number of correctly identified samples, is the total number of samples, The number of false targets correctly identified, is the total number of false samples.

Images