Precise recognition method for target in heterogeneous SAR image under small sample condition

Abstract

The present application relates to a kind of small sample conditions under the accurate identification method of heterogeneous SAR image target, comprising: sample expansion is carried out to small sample conditions under the heterogeneous SAR image and constitutes support set;The support set contains the support sample of target labeled of N categories;Target to be labeled SAR image is as query sample and constitutes query set;One query sample in query set and any optional sample in each category in support set constitute sample pair;N sample pairs are input to embedding network module, image pre-processing and feature extraction are carried out, and N feature pair data containing support sample feature and query sample feature are obtained;N feature pair data are input into similarity network module and similarity discrimination scoring is carried out, and the class of support sample in the highest feature pair data of similarity score is regarded as the class of query sample and target labeling is carried out.The present application realizes the accurate identification of heterogeneous SAR image target.

Description

Technical Field

[0001] This invention belongs to the field of SAR image processing technology, specifically relating to a method for accurate target identification in heterogeneous SAR images under small sample conditions. Background Technology

[0002] SAR images are sensitive to observation parameters. The distribution of images varies depending on the sensor, resolution, scene, and even the same sensor at different times or different regions within the same time period. SAR images measured by different sensors, at different resolutions, and in different scenes are SAR images with different characteristics, i.e., heterogeneous SAR images.

[0003] In many applications, the number of SAR image data samples acquired for known targets is limited. The unprecedented breakthroughs achieved by deep learning in various fields largely rely on a large amount of available labeled data, which requires significant investment in collection and annotation. Deep learning models that require large amounts of available labeled data struggle to solve the target recognition problem in heterogeneous SAR images under limited sample conditions. Summary of the Invention

[0004] Based on the above analysis, this invention aims to disclose a heterogeneous SAR image processing method based on deep feature transformation and transfer learning, thereby enabling heterogeneous SAR image processing.

[0005] This invention discloses a method for accurate target identification in heterogeneous SAR images under small sample conditions, comprising:

[0006] A support set is constructed by augmenting heterogeneous SAR images under small sample conditions; the support set contains labeled support samples of N target categories;

[0007] The SAR images of the target to be labeled are used as query samples to form a query set;

[0008] A sample pair is formed by combining a query sample from the query set with a randomly selected sample from each category in the support set; N categories form N sample pairs.

[0009] N sample pairs are input into the embedding network module for image preprocessing and feature extraction to obtain N feature pairs containing supporting sample features and query sample features;

[0010] The data of N feature pairs are input into the similarity network module for similarity discrimination and scoring. The category of the supporting sample in the data of the feature pair with the highest similarity score is used as the category of the query sample for target labeling.

[0011] Furthermore, the embedded network module includes an image preprocessing module;

[0012] The image preprocessing module includes a cropping module, a max pooling module, an average pooling module, and a channel connection module.

[0013] The cropping module is used to perform image cropping operations on single-channel input SAR images;

[0014] The max pooling module is used to perform max pooling operation on a single-channel input SAR image;

[0015] The average pooling module is used to perform average pooling operations on single-channel input SAR images;

[0016] The channel connection module is used to connect the output images of the cropping module, the max pooling module, and the average pooling module to form a three-channel preprocessed image.

[0017] Furthermore, the embedded network module includes a feature extraction module for extracting features from the preprocessed image of the three channels;

[0018] The feature extraction module includes four sequentially connected convolutional modules, with a pooling module between the first and second convolutional modules and a pooling module between the second and third convolutional modules.

[0019] Each convolutional module includes a convolutional layer, a batch normalization layer, and a non-linear activation layer; the pooling module employs max pooling.

[0020] Furthermore, the similarity network module includes a two-layer convolutional module, a flattening module, a first similarity module, a second similarity module, and a scoring module;

[0021] The input feature pairs are processed through two sequentially connected convolutional modules, and the output features are then fed into the flattening module.

[0022] The flattening module flattens the input features and outputs them to the first similarity module and the second similarity module respectively.

[0023] The first similarity module uses a cosine similarity metric function to calculate the distance between features;

[0024] The second similarity module uses the distance between the output features of two sequentially connected fully connected layers.

[0025] The scoring module scores the similarity between the input feature pairs based on the output results of the first similarity module and the second similarity module.

[0026] The two convolutional modules have the same structure, and each convolutional module includes a convolutional layer, a batch normalization layer, a non-linear activation layer, and a pooling layer.

[0027] In the two fully connected layers sequentially connected by the second similarity module, the first fully connected layer uses the ReLU function as a non-linear activation; the second fully connected layer uses the Sigmoid function as a non-linear activation.

[0028] Furthermore, the heterogeneous SAR images under small sample conditions are augmented to form a support set, and the InfoGAN model is used to perform adversarial learning to augment the samples of SAR images with labeled targets.

[0029] In the InfoGAN model, the dimensions of discrete latent variables and / or continuous latent variables are added to the noise dimension of the input generator network.

[0030] At the set learning rate, the generator network, discriminator network, and classification network are trained in each round, enabling the InfoGAN model to augment the input SAR images of multiple target categories.

[0031] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator network, discriminator network, and classification network of the InfoGAN model.

[0032] Furthermore, the generator network of the InfoGAN model consists of a four-layer connected transposed convolutional network, which takes the input signal, including noise and latent variables, and passes it through the generator network to obtain generated image data with the same size as the SAR image input to the InfoGAN model.

[0033] The discriminant network and the classification network include a shared feature network and their respective output networks;

[0034] The output network of the discrimination network is a single-layer fully connected network; the discrimination network, through the feature network and the output network, discriminates the input SAR image data and the generated image data and outputs the discrimination result.

[0035] The output network of the classification network consists of a convolutional network and a fully connected network; the classification network obtains the classification result through the feature network and the output network; the classification result is the mean and variance of discrete latent variables and / or continuous latent variables.

[0036] Furthermore, the feature network shared by the discriminant network and the classification network includes a three-layer sequentially connected convolutional network, wherein,

[0037] The first convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0038] The second convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 4 kernels, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0039] The third convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 8 kernels, a stride of 1, and zero padding of 0; the non-linear activation layers use the ReLU activation function.

[0040] Furthermore, sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and the SAGAN model is used to perform adversarial learning to augment the samples of SAR images with labeled targets.

[0041] In the SAGAN model, the input to the generator network is a multidimensional noise signal;

[0042] At the set learning rate, the discrimination network is trained in each round, enabling the SAGAN model to augment SAR images of this type of target.

[0043] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator and discriminator networks of the SAGAN model.

[0044] The generator network of the SAGAN model consists of four layers of connected transposed convolutional networks, with a self-attention network added in the third transposed convolutional network; the multidimensional noise signal is passed through the generator network to obtain generated image data with the same size as the SAR image input to the SAGAN model.

[0045] The discrimination network comprises a four-layer convolutional network connected in sequence, wherein a self-attention network is added to the second layer of the convolutional network; it discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result; the discrimination network discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result.

[0046] Furthermore, in the four convolutional layers of the discriminant network of the SAGAN model,

[0047] The first convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0048] The second convolutional network consists of convolution + self-attention, batch normalization layers, and non-linear activation layers; the convolutional layers have 4 kernels, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0049] The third convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 8 kernels, a stride of 1, and zero padding of 0; the non-linear activation layers use the ReLU activation function.

[0050] The fourth convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 1, and zero padding of 0; the non-linear activation layers use the sigmoid function as the activation function.

[0051] Furthermore, sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and target sample simulation is conducted using electromagnetic calculations to augment the sample; specifically including:

[0052] 1) Establish a three-dimensional CAD geometric model of the target;

[0053] 2) Based on the simulated SAR imaging parameters, determine the radar incident wave, scattered wave propagation vectors, and transmission and reception polarization modes. Set the rotation center based on the target's three-dimensional CAD geometric model, and use it as the phase reference center of the target coordinate system. Determine the sweep bandwidth and target rotation angle range based on the resolution.

[0054] 3) Based on the frequency sampling interval and azimuth sampling interval, electromagnetic modeling is performed using target electromagnetic scattering theories, including physical optics, geometric diffraction theory, and physical diffraction theory, to complete the calculation of the target complex scattering field;

[0055] 4) Based on the target complex scattering field output, perform imaging processing to generate simulated SAR image data of the target.

[0056] This invention can achieve at least one of the following beneficial effects:

[0057] This invention solves the problem of accurate target identification in heterogeneous SAR images under small sample conditions;

[0058] A complex, diverse, high-quality, and large-scale heterogeneous SAR image support set was constructed using various methods for target recognition. When generating source domain data using adversarial learning, the model's network was improved, the network size was reduced, the computational load was decreased, and the generation efficiency was increased.

[0059] Based on the idea of ​​meta-learning, an embedding network module and a similarity network module were constructed to label the target category in the query sample with a limited number of support samples. In the embedding network module, preprocessing was used to reduce invalid noise signals, reduce noisy features, and extract more effective features. In the similarity network module, the similarity between support samples and query samples was scored using both cosine distance and neural networks. Finally, the results of the two scoring methods were combined to obtain the final similarity score, which can more accurately determine the category of the target. Attached Figure Description

[0060] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0061] Figure 1 This is a flowchart of the method for accurate target identification of heterogeneous SAR images under small sample conditions in an embodiment of the present invention;

[0062] Figure 2 This is a diagram of the InfoGAN model network structure in an embodiment of the present invention;

[0063] Figure 3 This is a diagram of the SAGAN model network structure in an embodiment of the present invention;

[0064] Figure 4 This is a network structure diagram of the embedded network module in an embodiment of the present invention;

[0065] Figure 5 This is a network structure diagram of a similar network module in an embodiment of the present invention. Detailed Implementation

[0066] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0067] One embodiment of the present invention discloses a method for accurate target identification in heterogeneous SAR images under small sample conditions, such as... Figure 1 As shown, it includes:

[0068] Step S1: Expand the heterogeneous SAR images under small sample conditions to form a support set; the support set contains labeled support samples of N target categories;

[0069] Step S2: Use the SAR images of the target to be labeled as query samples to form a query set; the target category of the query sample is among the N target categories in the support set;

[0070] Step S3: Combine a query sample from the query set with a randomly selected sample from each category in the support set to form a sample pair; N categories form N sample pairs.

[0071] Step S4: Input N sample pairs into the embedding network module, perform image preprocessing and feature extraction, and obtain N feature pairs containing supporting sample features and query sample features;

[0072] Step S5: Input the N feature pairs into the similarity network module for similarity discrimination and scoring, and use the category of the supporting sample in the feature pair with the highest similarity score as the category of the query sample for target labeling.

[0073] For target recognition tasks in heterogeneous SAR images under small sample conditions, there are not enough support samples to provide for similarity comparison of query samples, so sample expansion is required to form a support set.

[0074] Based on this, in this embodiment, the heterogeneous SAR images of N categories of targets are expanded using various methods to construct a complex, diverse, high-quality, and large-scale SAR image support set.

[0075] The support set consists of sample augmentation of heterogeneous SAR images under small sample conditions, including sample augmentation using InfoGAN and SAGAN models for adversarial learning, and sample augmentation using electromagnetic computation for target sample simulation.

[0076] Specifically, sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and the InfoGAN model is used to perform adversarial learning to augment the samples on SAR images with labeled targets.

[0077] The InfoGAN model used includes three network structures: a generator network G, a discriminator network D, and a classification network Q; the generated data is represented as G(z, c); z is noise and c is a latent variable.

[0078] During model training, to prevent the generator network from ignoring the role of the latent variable c, mutual information I(X;Y) is used to constrain the latent variable c and the generated data G(z,c), measuring the relationship between the two random variables: the formula for calculating the mutual information between random variables X and Y is:

[0079] I(X;Y)=H(X)-H(X|Y)=H(Y)-H(Y|X);

[0080] Where H(X) represents the information entropy of random variable X, and H(X|Y) represents "the uncertainty of X given Y".

[0081] Based on mutual information, the following constraint is imposed on the latent variable c and the generated sample G(z, c): the mutual information I(c; G(z, c)) between the latent variable c and the generated sample G(z, c) should be as large as possible. A penalty term related to mutual information is added to the loss function, resulting in a new objective function:

[0082] min G max D V I (G,D)=V(G,D)-λI(c;G(z,c));

[0083] When calculating mutual information I(c; G(z, c)), P(c|x) is approximated by the auxiliary probability distribution Q(c|x), and the resulting objective function is min G，Q max D V InfoGAN (G, D, Q) = V(G, D) - λL I (G, Q);

[0084] Among them, mutual information is the lower L I (G, Q) = E c～P(c)，x～G(z，c) [logQ(c|x)]+H(c);

[0085] In the InfoGAN model, the dimensions of discrete latent variables and / or continuous latent variables are added to the noise dimension of the input generation network. During sample expansion, the SAR images input to the discriminator network and the classification network for discrimination and classification of the generated image data can be SAR images of multiple target categories. That is, the network parameters trained each time can satisfy the expansion of SAR images of multiple target categories.

[0086] At a set learning rate, the generator network, discriminator network, and classification network are trained in each round, enabling the InfoGAN model to augment SAR images of multiple target categories.

[0087] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator network, discriminator network, and classification network of the InfoGAN model.

[0088] In one specific embodiment, a preferred InfoGAN model structure is provided, in which,

[0089] like Figure 2 As shown, the generator network of the InfoGAN model consists of four connected transposed convolutional networks. The generator network takes the input signal, including noise and latent variables, and passes it through the generator network to obtain generated image data with the same size as the SAR image input to the InfoGAN model.

[0090] The discriminant network and the classification network include a shared feature network and their respective output networks;

[0091] The output network of the discrimination network is a single-layer fully connected network; the discrimination network, through the feature network and the output network, discriminates the input SAR image data and the generated image data and outputs the discrimination result.

[0092] The output network of the classification network consists of a convolutional network and a fully connected network; the classification network obtains the classification result through the feature network and the output network; the classification result is the mean and variance of discrete latent variables and / or continuous latent variables.

[0093] The feature network shared by the discriminant network and the classification network includes a three-layer sequentially connected convolutional network, wherein,

[0094] The first convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0095] The second convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 4 kernels, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0096] The third convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 8 kernels, a stride of 1, and zero padding of 0; the non-linear activation layers use the ReLU activation function.

[0097] When generating an image from a SAR image with an input size of 32×32, the noise dimension of the generator network is 67+n_dis_c+n_con_c, where n_dis_c is the dimension of the discrete latent variables, i.e., the number of categories of the input data, which is set to 2, and n_con_c is the number of continuous latent variables, which is set to 1; thus forming 70-dimensional input data.

[0098] The learning rate for the generator network, the discriminator network, and the classification network is 0.0002.

[0099] The generator network, discriminator network, and classification network all use the Adam optimizer, with optimizer parameters β1 = 0.5 and β2 = 0.999.

[0100] In the generative network, a feature map of (1024, 1, 1) is obtained through the first transposed convolution; a feature map of (128, 8, 8) is obtained through the second transposed convolution; a feature map of (64, 16, 16) is obtained through the third transposed convolution; and a feature map of (1, 32, 32) is obtained through the fourth transposed convolution; thus, a generated image with the same size as the input SAR image is obtained.

[0101] In the feature network shared by the discriminant network and the classification network, the first convolutional network layer yields a feature map of (64, 16, 16), the second convolutional network layer yields a feature map of (128, 8, 8), and the third convolutional network layer yields a feature map of (1024, 1, 1).

[0102] The fully connected network in the output network of the discrimination network outputs two discrimination results: if yes, it is a SAR image of the same target; if no, it is a SAR image of different targets.

[0103] The convolutional network in the output network of the classification network transforms the feature map (1024, 1, 1) into (128, 1, 1); after passing through a fully connected network, the mean and variance of discrete and continuous latent variables are output; these are used to calculate the loss with the input latent variables, and the model parameters are trained with the minimum loss to obtain the final InfoGAN model that can generate augmented samples of the input SAR image.

[0104] Specifically, sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and the SAGAN model is used to perform adversarial learning to augment the samples on SAR images with labeled targets.

[0105] The SAGAN model used includes a generator network and a discriminator network. The generator network is fed with a multidimensional noise signal as input. During sample augmentation, the discriminator network is fed with SAR images of the same target class as the generated image data. In other words, the network parameters trained each time can only be used to augment SAR images of one target class.

[0106] At the set learning rate, the discrimination network is trained in each round, enabling the SAGAN model to augment SAR images of this type of target.

[0107] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator and discriminator networks of the SAGAN model.

[0108] To achieve better image generation, this embodiment adjusts the position of the added attention network, thereby improving the quality of the generated image.

[0109] In one specific embodiment, a preferred SAGAN model structure is provided, in which,

[0110] In the SAGAN model, the input to the generator network is a multidimensional noise signal, and the input to the discriminator network is a SAR image of the same type of target.

[0111] At the set learning rate, the discrimination network is trained in each round, enabling the SAGAN model to augment SAR images of this type of target.

[0112] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator and discriminator networks of the SAGAN model.

[0113] like Figure 3 As shown, the generator network of the SAGAN model consists of four connected transposed convolutional networks, with a self-attention network added in the third transposed convolutional network; the multidimensional noise signal is passed through the generator network to obtain generated image data with the same size as the SAR image input to the SAGAN model.

[0114] The discrimination network comprises a four-layer convolutional network connected in sequence, wherein a self-attention network is added to the second layer of the convolutional network; it discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result; the discrimination network discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result.

[0115] Specifically, in the four convolutional layers of the discriminative network of the SAGAN model,

[0116] The first convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0117] The second convolutional network consists of convolution + self-attention, batch normalization layers, and non-linear activation layers; the convolutional layers have 4 kernels, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function.

[0118] The third convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 8 kernels, a stride of 1, and zero padding of 0; the non-linear activation layers use the ReLU activation function.

[0119] The fourth convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 1, and zero padding of 0; the non-linear activation layers use the sigmoid function as the activation function.

[0120] When generating an image from an input SAR image with a size of 32×32, the noise dimension of the input generation network is 128.

[0121] Both the generator and discriminator networks were trained using a learning rate of 0.0002.

[0122] Both the generator network and the discriminator network use the Adam optimizer, with optimizer parameters β1 = 0.5 and β2 = 0.999.

[0123] In the generative network, a feature map of (256, 4, 4) is obtained through the first transposed convolution; a feature map of (128, 8, 8) is obtained through the second transposed convolution; a feature map of (64, 16, 16) is obtained through the third transposed convolution; and a feature map of (1, 32, 32) is obtained through the fourth transposed convolution; thus, a generated image with the same size as the input SAR image is obtained.

[0124] In the discriminant network, the first convolutional network layer produces a feature map of (32, 16, 16); the second convolutional network layer produces a feature map of (64, 8, 8); the third convolutional network layer produces a feature map of (128, 4, 4); and the fourth convolutional network layer outputs the discriminant result.

[0125] Adding a self-attention network to the third layer of the transposed convolutional network of the generative network and the second layer of the convolutional network of the discriminative network as a supplement to the convolution helps the model learn long-distance correlations across image regions.

[0126] Furthermore, during the training process of the SAGAN model, a hinged loss function is used to achieve classification, which ultimately enables the generation of SAGAN models that supplement input SAR images with expanded samples.

[0127] Specifically, sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and target sample simulation is conducted using electromagnetic calculations to augment the sample; this includes:

[0128] 1) Establish a three-dimensional CAD geometric model of the target;

[0129] When creating a 3D CAD geometric model, CAD data interaction, polyhedral geometric modeling, or laser 3D scanning modeling can be used to create the 3D geometric model of the target.

[0130] 2) Based on the simulated SAR imaging parameters, determine the radar incident wave, scattered wave propagation vectors, and transmission and reception polarization modes. Set the rotation center based on the target's three-dimensional CAD geometric model, and use it as the phase reference center of the target coordinate system. Determine the sweep bandwidth and target rotation angle range based on the resolution.

[0131] 3) Based on the frequency sampling interval and azimuth sampling interval, electromagnetic modeling is performed using target electromagnetic scattering theories, including physical optics, geometric diffraction theory, and physical diffraction theory, to complete the calculation of the target complex scattering field;

[0132] 4) Based on the target complex scattering field output, perform imaging processing to generate simulated SAR image data of the target.

[0133] Preferably, in the acquisition of SAR image samples based on high-precision electromagnetic simulation calculations, electromagnetic modeling of the coupling between the target and the environment is also considered. When the target is placed in a specific environment, especially when there are angular structures on the target and the environment surface, there is a significant multipath scattering effect between the target and the environment, and the coupling calculation between the target and the environment is relatively complex. This embodiment uses different processing methods for coupling with different environmental characteristics.

[0134] When the target is located in a flat, slightly rough environment, such as an airport or highway, electromagnetic modeling is performed using multipath scattering caused by specular reflection between the target and the environment.

[0135] When the target is located on a large-scale random micro-rough surface, such as the sea surface, electromagnetic modeling is performed using multipath scattering caused by surface undulations between the target and the environment.

[0136] By expanding the sample size through various methods, a support set with a sufficient number of samples was obtained, providing data support for subsequent sample queries.

[0137] In this embodiment, the embedding network module and the similarity network module for target classification are built based on the idea of ​​meta-learning. Specifically, the embedding network module extracts support sample features and query sample features to form feature pairs; the similarity network module scores the feature pairs and uses the category of the support sample in the feature pair with the highest similarity score among N categories as the target category of the query sample for target labeling.

[0138] To better extract features, the embedded network module in this embodiment performs image preprocessing before feature extraction to reduce invalid noise signals, reduce noisy features, and extract more effective features.

[0139] Specifically, such as Figure 4 As shown, the embedded network module includes an image preprocessing module and a feature extraction module;

[0140] The image preprocessing module includes a cropping module, a max pooling module, an average pooling module, and a channel connection module.

[0141] The cropping module is used to perform image cropping operations on single-channel input SAR images. In the obtained heterogeneous SAR radar images, the targets are all located in the center of the image, and most of the surrounding area is invalid noise signal. By cropping the middle region, the invalid noise signal around it is removed, thereby reducing the noise signal significantly.

[0142] The max pooling module is used to perform max pooling operations on single-channel input SAR images; max pooling helps to extract more effective features.

[0143] The average pooling module is used to perform average pooling operations on single-channel input SAR images; average pooling helps to reduce the characteristics of noise signals superimposed on the target signal.

[0144] The channel connection module is used to connect the output images of the cropping module, the max pooling module, and the average pooling module to form a three-channel preprocessed image.

[0145] Specifically, after inputting a single-channel H×W×1 SAR image, it undergoes three processing steps: intermediate region cropping, max pooling, and average pooling, to obtain a three-channel image. The processed image.

[0146] The feature extraction module extracts features from the preprocessed image of the three channels and concatenates the features of a support sample and a query sample to obtain a feature pair data.

[0147] Specifically, the feature extraction module includes four sequentially connected convolutional modules, a pooling module between the first and second convolutional modules, and a pooling module between the second and third convolutional modules.

[0148] Each convolutional module includes a convolutional layer, a batch normalization layer, and a non-linear activation layer; the pooling module employs max pooling.

[0149] More specifically, the convolutional layer uses a 3×3 Convd kernel with a stride of 2, zero padding of 1, and 64-bit filters; the nonlinear activation layer uses the ReLU function for nonlinear activation.

[0150] Specifically, such as Figure 5 As shown, the similarity network module includes a two-layer convolutional module, a flattening module, a first similarity module, a second similarity module, and a scoring module;

[0151] The input feature pairs are processed through two sequentially connected convolutional modules, and the output features are then fed into the flattening module.

[0152] The flattening module flattens the input features and then outputs them to the first similarity module and the second similarity module respectively.

[0153] The first similarity module uses a cosine similarity metric function to calculate the distance between features;

[0154] The second similarity module uses the distance between the output features of two sequentially connected fully connected layers.

[0155] The scoring module scores the similarity between the input feature pairs based on the outputs of the first similarity module and the second similarity module. Specifically, in this embodiment, the scoring module adds the outputs of the first similarity module and the second similarity module as the similarity score.

[0156] The two convolutional modules have the same structure. Each convolutional module includes a convolutional layer, a batch normalization layer, a non-linear activation layer, and a pooling layer. The convolutional layer uses a 3×3 Convd kernel with a stride of 2, zero padding of 1, and 64-bit filters. The non-linear activation layer uses the ReLU function. The pooling layer uses max pooling.

[0157] In the two fully connected layers sequentially connected by the second similarity module, the first fully connected layer uses a 64→8 fully connected network with ReLU function as the non-linear activation; the second fully connected layer uses an 8→1 fully connected network with Sigmoid function as the non-linear activation.

[0158] Preferably, in the learning and training process of the embedded network module and the similarity network module in this embodiment, classification experiments are conducted on multiple categories in the support set and query set. Samples from each category are divided into the training set, validation set, and test set in a 70:15:15 ratio. The mean squared error loss function is used; the learning rate is 0.001; and the Adam optimizer is used with parameters β1 = 0.9 and β2 = 0.999. Through learning and training, the embedded network module can better obtain feature pairs between the support samples and query samples; and the scoring of the similarity network module becomes more accurate.

[0159] In summary, the embodiments of this invention solve the problem of accurate target identification in heterogeneous SAR images under small sample conditions. They construct complex, diverse, high-quality, and large-scale heterogeneous SAR image support sets for target identification through various means. When generating source domain data using adversarial learning, the model's network is improved, reducing its size, computational load, and generation efficiency. Based on the idea of ​​meta-learning, an embedding network module and a similarity network module are constructed, enabling the labeling of target categories in query samples with limited support samples. Furthermore, the embedding network module reduces invalid noise signals, lowers noise features, and extracts more effective features through preprocessing. The similarity network module scores the similarity between support samples and query samples using both cosine distance and neural networks. Finally, the results of the two scoring methods are combined to obtain the final similarity score, thereby more accurately determining the target category.

[0160] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for accurate target identification in heterogeneous SAR images under small sample conditions, characterized in that, include: Support sets are constructed by augmenting heterogeneous SAR images under small sample conditions. The support set contains labeled support samples for N categories of targets; The SAR images of the target to be labeled are used as query samples to form a query set; A sample pair is formed by combining a query sample from the query set with a randomly selected sample from each category in the support set; N categories form N sample pairs. N sample pairs are input into the embedding network module for image preprocessing and feature extraction to obtain N feature pairs containing supporting sample features and query sample features; The data of N feature pairs are input into the similarity network module for similarity discrimination and scoring. The category of the supporting sample in the data of the feature pair with the highest similarity score is used as the category of the query sample for target labeling. The similarity network module includes a two-layer convolutional module, a flattening module, a first similarity module, a second similarity module, and a scoring module; The input feature pairs are processed through two sequentially connected convolutional modules, and the output features are then fed into the flattening module. The flattening module flattens the input features and outputs them to the first similarity module and the second similarity module respectively. The first similarity module uses a cosine similarity metric function to calculate the distance between features; The second similarity module uses the distance between the output features of two sequentially connected fully connected layers. The scoring module scores the similarity between the input feature pairs based on the output results of the first similarity module and the second similarity module. The two convolutional modules have the same structure, and each convolutional module includes a convolutional layer, a batch normalization layer, a non-linear activation layer, and a pooling layer. In the two fully connected layers sequentially connected by the second similarity module, the first fully connected layer uses the ReLU function as a non-linear activation; the second fully connected layer uses the Sigmoid function as a non-linear activation.

2. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 1, characterized in that, The embedded network module includes an image preprocessing module; The image preprocessing module includes a cropping module, a max pooling module, an average pooling module, and a channel connection module. The cropping module is used to perform image cropping operations on single-channel input SAR images; The max pooling module is used to perform max pooling operation on a single-channel input SAR image; The average pooling module is used to perform average pooling operations on single-channel input SAR images; The channel connection module is used to connect the output images of the cropping module, the max pooling module, and the average pooling module to form a three-channel preprocessed image.

3. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 1, characterized in that, The embedded network module includes a feature extraction module for extracting features from the preprocessed image in three channels; The feature extraction module includes four sequentially connected convolutional modules, with a pooling module between the first and second convolutional modules and a pooling module between the second and third convolutional modules. Each convolutional module includes a convolutional layer, a batch normalization layer, and a non-linear activation layer; the pooling module employs max pooling.

4. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 1, characterized in that, Sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and the InfoGAN model is used to perform adversarial learning to augment the samples on SAR images with labeled targets. In the InfoGAN model, the dimensions of discrete latent variables and / or continuous latent variables are added to the noise dimension of the input generator network. At the set learning rate, the generator network, discriminator network, and classification network are trained in each round, enabling the InfoGAN model to augment the input SAR images of multiple target categories. By setting the optimizer parameters, the Adam optimizer is used to optimize the generator network, discriminator network, and classification network of the InfoGAN model.

5. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 4, characterized in that, The generator network of the InfoGAN model consists of four connected transposed convolutional networks. The input signal, including noise and latent variables, is passed through the generator network to obtain generated image data with the same size as the SAR image input to the InfoGAN model. The discriminant network and the classification network include a shared feature network and their respective output networks; The output network of the discrimination network is a single-layer fully connected network; the discrimination network, through the feature network and the output network, discriminates the input SAR image data and the generated image data and outputs the discrimination result. The output network of the classification network consists of a convolutional network and a fully connected network; the classification network obtains the classification result through the feature network and the output network; the classification result is the mean and variance of discrete latent variables and / or continuous latent variables.

6. The method for accurate target identification in heterogeneous SAR images under small sample conditions according to claim 5, characterized in that, The feature network shared by the discriminant network and the classification network includes a three-layer sequentially connected convolutional network, wherein, The first convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function. The second convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 4 kernels, a stride of 2, and zero padding of 1; the non-linear activation layers use the ReLU activation function. The third convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 8 kernels, a stride of 1, and zero padding of 0; the non-linear activation layers use the ReLU activation function.

7. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 4, characterized in that, Sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and the SAGAN model is used to perform adversarial learning to augment the samples on SAR images with labeled targets. In the SAGAN model, the input to the generator network is a multidimensional noise signal; At the set learning rate, the discrimination network is trained in each round, enabling the SAGAN model to augment SAR images of this type of target. By setting the optimizer parameters, the Adam optimizer is used to optimize the generator and discriminator networks of the SAGAN model. The generator network of the SAGAN model consists of four layers of connected transposed convolutional networks, with a self-attention network added in the third transposed convolutional network; the multidimensional noise signal is passed through the generator network to obtain generated image data with the same size as the SAR image input to the SAGAN model. The discrimination network comprises a four-layer convolutional network connected in sequence, wherein a self-attention network is added to the second layer of the convolutional network; it discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result; the discrimination network discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result.

8. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 7, characterized in that, In the four-layer convolutional network of the discriminant network of the SAGAN model, The first convolutional network consists of convolutional layers and non-linear activation layers; the convolutional kernel of the convolutional layer has 4 kernels, a stride of 2, and zero padding of 1. The activation function used in the nonlinear activation layer is the ReLU function; The second convolutional network includes convolution + self-attention, batch normalization layer and non-linear activation layer; the convolutional layer has 4 kernels, a stride of 2, and zero padding of 1. The activation function used in the nonlinear activation layer is the ReLU function; The third convolutional network consists of convolutional layers, batch normalization layers, and non-linear activation layers; the convolutional layers have 8 kernels, a stride of 1, and zero padding of 0; the non-linear activation layers use the ReLU activation function. The fourth convolutional network consists of convolutional layers and non-linear activation layers; the convolutional layers have a kernel size of 4, a stride of 1, and zero padding of 0; the non-linear activation layers use the sigmoid function as the activation function.

9. The method for accurate target identification of heterogeneous SAR images under small sample conditions according to claim 1, characterized in that, Sample augmentation is performed on heterogeneous SAR images under small sample conditions to form a support set, and target sample simulation is conducted using electromagnetic calculations to further augment the samples; specifically, this includes: 1) Establish a three-dimensional CAD geometric model of the target; 2) Based on the simulated SAR imaging parameters, determine the radar incident wave, scattered wave propagation vectors, and transmission and reception polarization modes. Set the rotation center based on the target's three-dimensional CAD geometric model, and use it as the phase reference center of the target coordinate system. Determine the sweep bandwidth and target rotation angle range based on the resolution. 3) Based on the frequency sampling interval and azimuth sampling interval, electromagnetic modeling is performed using target electromagnetic scattering theories, including physical optics, geometric diffraction theory, and physical diffraction theory, to complete the calculation of the target complex scattering field; 4) Generate simulated SAR image data of the target by performing imaging processing based on the target complex scattering field output.

Images