Heterogeneous sar image processing method based on deep feature transformation and transfer learning

Abstract

The application relates to a heterogeneous SAR image processing method based on deep feature transformation and transfer learning, which comprises the following steps: acquiring measured SAR image data of different sensors, different resolutions and different scenes; the measured SAR image data comprises SAR images with target labels and SAR images without target labels; sample expansion is conducted on the SAR images with target labels to obtain a plurality of source domain SAR data sets; the SAR images without target labels are taken as a target domain SAR data set; a deep learning network based on domain self-adaptation is trained by using the plurality of source domain SAR data sets and the target domain SAR data set; and target labeling is conducted on the SAR images without target labels through deep feature transformation and transfer learning. The application realizes heterogeneous SAR image processing.

Description

Technical Field

[0001] This invention belongs to the field of SAR image processing technology, specifically relating to a heterogeneous SAR image processing method based on depth feature transformation and transfer learning. 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] Traditional machine learning algorithms, especially statistical learning, only demonstrate good performance when the training and test data have the same distribution. When the distributions of the test and training data differ, the performance of traditional machine learning algorithms is significantly affected, making them unsuitable for processing heterogeneous SAR images. 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 heterogeneous SAR image processing method based on deep feature transformation and transfer learning, comprising:

[0006] Acquire measured SAR image data from different sensors, at different resolutions, and in different scenarios; including SAR images with and without target marking.

[0007] Multiple source domain SAR datasets are obtained by sample augmentation of SAR images, including those with target marking.

[0008] Use unlabeled SAR images as the target domain SAR dataset;

[0009] A domain-adaptive deep learning network was trained using multiple source domain SAR datasets and target domain SAR datasets. Targets were labeled in SAR images that had not been labeled using deep feature transformation and transfer learning.

[0010] Furthermore, the multiple source domain SAR datasets include a first source domain SAR dataset, a second source domain SAR dataset, and a third source domain SAR dataset; the first source domain SAR dataset is generated by using the InfoGAN model for adversarial learning on SAR images that have been labeled with targets.

[0011] A second source domain SAR dataset is generated by using the SAGAN model for adversarial learning on SAR images that have been marked with targets.

[0012] Target sample simulation based on electromagnetic calculations generates a third source domain SAR dataset.

[0013] Furthermore, when generating the first source domain SAR dataset,

[0014] 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.

[0015] 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.

[0016] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator network, discriminator network, and classification network of the InfoGAN model.

[0017] When generating the second source domain SAR dataset

[0018] In the SAGAN model, the input to the generator network is a multidimensional noise signal;

[0019] 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.

[0020] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator and discriminator networks of the SAGAN model.

[0021] 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.

[0022] The discriminant and classification networks of the SAGAN model include a shared feature network and separate output networks.

[0023] 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.

[0024] 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.

[0025] Furthermore, the feature network shared by the discriminant network and the classification network includes a three-layer sequentially connected convolutional network, wherein,

[0026] 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.

[0027] 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.

[0028] 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.

[0029] Furthermore, the generator network of the SAGAN model consists of four connected transposed convolutional networks, wherein a self-attention network is 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;

[0030] 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.

[0031] Furthermore, in the four convolutional layers of the discriminant network of the SAGAN model,

[0032] 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.

[0033] 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.

[0034] 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.

[0035] 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.

[0036] Furthermore, a method for generating a third-source-domain SAR dataset based on target sample simulation using electromagnetic calculations includes:

[0037] 1) Establish a three-dimensional CAD geometric model of the target;

[0038] 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.

[0039] 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;

[0040] 4) Based on the target complex scattering field output, perform imaging processing to generate simulated SAR image data of the target.

[0041] Furthermore, the domain-adaptive deep learning network includes a feature generator F(·) and a classifier C(·);

[0042] The feature generator F(·) is composed of a ResNet-101 backbone network and is used to extract effective learnable features from heterogeneous SAR datasets. The extracted features are then fed into the subsequent classifier C(·) for target classification.

[0043] Source domain SAR dataset middle For SAR image data, For labeling data; n s The number of samples;

[0044] Target Domain SAR Dataset middle, For SAR image data, n t The number of samples;

[0045] The source domain SAR dataset is input into a deep learning network for training. By using labeled data from the source domain, the deep learning network can accurately classify the data in the source domain SAR dataset.

[0046] Unlabeled SAR images from the target domain SAR dataset are input into a deep learning network for training, with the output of classifier C(·) as the result. This is a pseudo-label for the SAR image;

[0047] In pseudo-labeling By combining the original target domain SAR image with the classifier, a new classification result is obtained.

[0048] The SAR image is then labeled with the new classification results.

[0049] Furthermore, the overall training objective function of domain-adaptive deep learning networks Where λ is a hyperparameter of a balanced loss function; n s The number of source domain images, Let cross-entropy be the loss function. E(·) is the standard information entropy function, and δ represents the output of the softmax layer of the network.

[0050] This invention can achieve at least one of the following beneficial effects:

[0051] This invention constructs a complex, diverse, high-quality, and large-scale heterogeneous SAR image dataset through various means; using the heterogeneous SAR image dataset, a domain-adaptive deep learning network is employed to achieve target recognition of SAR images through deep feature transformation and transfer learning.

[0052] Furthermore, 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. Attached Figure Description

[0053] 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.

[0054] Figure 1 This is a flowchart of the heterogeneous SAR image processing method in an embodiment of the present invention;

[0055] Figure 2 This is a diagram of a domain-adaptive deep learning network structure in an embodiment of the present invention. Detailed Implementation

[0056] 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.

[0057] One embodiment of the present invention discloses a heterogeneous SAR image processing method based on deep feature transformation and transfer learning, such as... Figure 1 As shown, it includes:

[0058] Step S1: Acquire measured SAR image data from different sensors, at different resolutions, and in different scenes; including SAR images with and without target marking.

[0059] Step S2: Perform sample augmentation on SAR images, including those with target marking, to obtain multiple source domain SAR datasets;

[0060] Step S3: Use the SAR images without target labeling as the target domain SAR dataset;

[0061] Step S4: Train a domain-adaptive deep learning network using multiple source domain SAR datasets and target domain SAR datasets. Then, perform target labeling on SAR images that have not been labeled using deep feature transformation and transfer learning.

[0062] For target recognition tasks in SAR images, there is often a problem of insufficient sample size. With a small number of samples, it is not enough to train a network model to recognize targets in the image. Therefore, in this embodiment, in step S2, the SAR images are expanded using various methods to construct a complex, diverse, high-quality, and large-scale SAR image dataset.

[0063] Specifically, in step S2,

[0064] The InfoGAN model is used to perform adversarial learning on SAR images with target marking to generate the first source domain SAR dataset;

[0065] A second source domain SAR dataset is generated by using the SAGAN model for adversarial learning on SAR images that have been marked with targets.

[0066] Target sample simulation based on electromagnetic calculations generates a third source domain SAR dataset.

[0067] In generating the first source domain SAR dataset, the InfoGAN model is used, which 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), where z is noise and c is a latent variable.

[0068] 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:

[0069] I(X;Y)=H(X)-H(X|Y)=H(Y)-H(Y|X);

[0070] Where H(X) represents the information entropy of random variable X, and H(X|Y) represents "the uncertainty of X given Y".

[0071] 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:

[0072] min G max D V I (G,D)=V(G,D)-λI(c;G(z,c));

[0073] 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);

[0074] Among them, mutual information is the lower L I (G,Q)=E c～P(c),x～G(z,c) [logQ(c|x)]+H(c);

[0075] More specifically, in the InfoGAN model, the dimension of discrete latent variables and / or the dimension of 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.

[0076] 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.

[0077] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator network, discriminator network, and classification network of the InfoGAN model.

[0078] In a specific embodiment of this example, a preferred InfoGAN model structure is provided, in which,

[0079] The InfoGAN model's generator network 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.

[0080] The discriminant network and the classification network include a shared feature network and their respective output networks;

[0081] 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.

[0082] 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.

[0083] The feature network shared by the discriminant network and the classification network includes a three-layer sequentially connected convolutional network, wherein,

[0084] 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.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] The learning rate for the generator network, the discriminator network, and the classification network is 0.0002.

[0089] The generator network, discriminator network, and classification network all use the Adam optimizer, with optimizer parameters β1 = 0.5 and β2 = 0.999.

[0090] 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.

[0091] 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).

[0092] The fully connected network in the output network of the discrimination network outputs two discrimination results: if yes, the images are SAR images of the same target; otherwise, they are SAR images of different targets.

[0093] 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.

[0094] In generating the second source domain SAR dataset, the SAGAN model is used, which 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. That is, the network parameters trained each time can only be used to augment SAR images of one target class.

[0095] 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.

[0096] By setting the optimizer parameters, the Adam optimizer is used to optimize the generator and discriminator networks of the SAGAN model.

[0097] To achieve better image generation, this embodiment adjusts the position of the added attention network, thereby improving the quality of the generated image.

[0098] Specifically, the generator network of the SAGAN model consists of four layers of 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.

[0099] 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.

[0100] Specifically, in the four convolutional layers of the discriminative network of the SAGAN model,

[0101] 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.

[0102] 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.

[0103] 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.

[0104] 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.

[0105] 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.

[0106] Both the generator and discriminator networks were trained using a learning rate of 0.0002.

[0107] Both the generator network and the discriminator network use the Adam optimizer, with optimizer parameters β1 = 0.5 and β2 = 0.999.

[0108] 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.

[0109] 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.

[0110] 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.

[0111] 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.

[0112] Specifically, methods for generating third-source-domain SAR datasets based on electromagnetic calculations through target sample simulation include:

[0113] 1) Establish a three-dimensional CAD geometric model of the target;

[0114] 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.

[0115] 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.

[0116] 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;

[0117] 4) Based on the target complex scattering field output, perform imaging processing to generate simulated SAR image data of the target.

[0118] 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.

[0119] 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.

[0120] 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.

[0121] Multiple source domain SAR datasets with abundant sample quantities were obtained through sample augmentation using various methods, providing data support for subsequent domain-adaptive deep learning networks.

[0122] Specifically, such as Figure 2 As shown, the domain-adaptive deep learning network in this embodiment includes a feature generator F(·) and a classifier C(·);

[0123] The feature generator F(·) is composed of a ResNet-101 backbone network and is used to extract effective learnable features from heterogeneous SAR datasets. The extracted features are then fed into the subsequent classifier C(·) for target classification.

[0124] Source domain SAR dataset middle For SAR image data, For labeling data; n s The number of samples;

[0125] Target Domain SAR Dataset middle, For SAR image data, n t The number of samples;

[0126] The source domain SAR dataset is input into a deep learning network for training. By using labeled data from the source domain, the deep learning network can accurately classify the data in the source domain SAR dataset.

[0127] Unlabeled SAR images from the target domain SAR dataset are input into a deep learning network for training, with the output of classifier C(·) as the result. This is a pseudo-label for the SAR image;

[0128] In pseudo-labeling By combining the original target domain SAR image with the classifier, a new classification result is obtained.

[0129] The SAR image is then labeled with the new classification results.

[0130] The overall training objective function of adaptive deep learning networks Where λ is a hyperparameter of a balanced loss function; n s The number of source domain images, Let cross-entropy be the loss function. E(·) is the standard information entropy function, and δ represents the output of the softmax layer of the network.

[0131] In summary, the embodiments of this invention construct complex, diverse, high-quality, and large-scale heterogeneous SAR image datasets through various means. Using these datasets, a domain-adaptive deep learning network is employed to achieve target recognition in SAR images through deep feature transformation and transfer learning. Furthermore, when generating source domain data using adversarial learning, the model's network is improved, its size reduced, computational load decreased, and generation efficiency increased.

[0132] 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 heterogeneous SAR image processing method based on deep feature transformation and transfer learning, characterized in that, include: Acquire measured SAR image data from different sensors, at different resolutions, and in different scenarios; This includes SAR images with and without target marking; Multiple source domain SAR datasets are obtained by sample augmentation of SAR images, including those with target marking. Use unlabeled SAR images as the target domain SAR dataset; A domain-adaptive deep learning network was trained using multiple source domain SAR datasets and target domain SAR datasets. Target labeling was performed on SAR images that had not been labeled using deep feature transformation and transfer learning. Domain-adaptive deep learning networks include feature generators. and classifier ; The feature generator Composed of a ResNet-101 backbone network, it is used to extract effective learnable features from heterogeneous SAR datasets; The extracted features are then fed into the subsequent classifier. In the middle, target classification is performed; Source domain SAR dataset middle For SAR image data, For labeling data; The number of samples; Target Domain SAR Dataset middle, For SAR image data, The number of samples; The source domain SAR dataset is input into a deep learning network for training, using labeled data from the source domain; This enables deep learning networks to accurately classify data in the source domain SAR dataset; Unlabeled SAR images from the target domain SAR dataset are input into a deep learning network for training, to create a classifier. Output This is a pseudo-label for the SAR image; In pseudo-labeling By combining the original target domain SAR image with the classifier, a new classification result is obtained. The SAR image is then labeled with the new classification results; The overall training objective function of domain-adaptive deep learning networks ;in, For a balanced loss function, the hyperparameters are denoted as . ; ; The number of source domain images, Let cross-entropy be the loss function. , It is the standard information entropy function. This represents the output of the softmax layer in the network.

2. The heterogeneous SAR image processing method according to claim 1, characterized in that, The multiple source domain SAR datasets include a first source domain SAR dataset, a second source domain SAR dataset, and a third source domain SAR dataset; wherein... The InfoGAN model is used to perform adversarial learning on SAR images with target marking to generate the first source domain SAR dataset; A second source domain SAR dataset is generated by using the SAGAN model for adversarial learning on SAR images that have been marked with targets. Target sample simulation based on electromagnetic calculations generates a third source domain SAR dataset.

3. The heterogeneous SAR image processing method according to claim 2, characterized in that, When generating the first source domain SAR dataset 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. When generating the second source domain SAR dataset 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.

4. The heterogeneous SAR image processing method according to claim 3, characterized in that, The InfoGAN model's generator network 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. 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.

5. The heterogeneous SAR image processing method according to claim 4, 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.

6. The heterogeneous SAR image processing method according to claim 2, characterized in that, 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 discriminant network of the SAGAN model comprises a four-layer convolutional network connected in sequence, wherein a self-attention network is added to the second layer of the convolutional network; the network discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result; the discriminant network discriminates between the input target-labeled SAR image data and the generated image data and outputs a discrimination result.

7. The heterogeneous SAR image processing method according to claim 5, 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.

8. The heterogeneous SAR image processing method according to claim 2, characterized in that, Methods for generating third-source-domain SAR datasets based on electromagnetic computation through target sample simulation include: 1) Establish a three-dimensional CAD geometric model of the target; 2) Based on the simulated SAR imaging parameters, determine the radar incident wave and scattered wave propagation vectors, as well as the transmission and reception polarization modes. Set the rotation center according to 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 according to 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