Remote sensing image generation method for generating diffusion model based on federal angle perception

Abstract

The invention discloses a remote sensing image generation method for generating a diffusion model based on federal angle perception, and belongs to the technical field of image generation, and the method specifically comprises the steps: constructing a federal learning framework, and enabling each client to locally store a private remote sensing image data set; in each communication round, the central server distributes the current parameters of the global model to each client; each client preprocesses a local remote sensing image by adopting a de-noising diffusion probability model to generate an initial remote sensing image; the client takes the received global model parameters as an initial state, carries out training on an angle perception generation diffusion network model based on the initial remote sensing image, and generates a remote sensing image with a specific azimuth angle; the client side adopts a local data knowledge distillation strategy, and local distillation is carried out through cross entropy loss and a small amount of distillation loss; and the client uploads the updated model parameters to the server, and the server obtains a final global model through multi-round iteration.

Description

Technical Field

[0001] The present invention relates to the technical field of image generation, and particularly to a method for generating remote sensing images based on a federated angle-aware generative diffusion model. Background Art

[0002] In the field of remote sensing image generation, traditional methods may be difficult to meet the demand for generating synthetic aperture radar target images with specific azimuth angles. Moreover, since remote sensing image data is often distributed among different clients, there are problems such as data privacy and security, and traditional centralized learning methods cannot be directly applied. And deep learning-based models usually require a large amount of data for training to ensure the effectiveness of the trained model. The generative model is no exception, and sufficient training data is a necessary condition for generating image diversity. However, for synthetic aperture radar images, the cost of data acquisition is high. Therefore, generating remote sensing images with a small number of training samples is still an urgent problem to be solved.

[0003] When existing federated learning is applied to remote sensing image generation, the private remote sensing image datasets locally stored by each client differ in the number and quality of categories. This can lead to some clients having difficulty fully learning the features and knowledge of specific categories, resulting in the "forgetting" phenomenon, causing the model's ability to master the knowledge learned previously to decline. As the number of training rounds increases, this problem becomes more serious, affecting the performance of the model in global tasks. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for generating remote sensing images based on a federated angle-aware generative diffusion model, and solve the following technical problems:

[0005] When applying federated learning to remote sensing image generation, the private remote sensing image datasets locally stored by each client differ in the number and quality of categories, resulting in some clients having difficulty fully learning the features and knowledge of specific categories, affecting the performance of the model in global tasks.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A method for generating remote sensing images based on a federated angle-aware generative diffusion model includes the following steps:

[0008] Construct a federated learning framework, including several clients and a central server, and each client locally stores a private remote sensing image dataset; in each communication round, the central server distributes the current parameters of the global model to each client;

[0009] Each client uses a denoising diffusion probabilistic model to preprocess the local remote sensing images to generate initial remote sensing images;

[0010] The client uses the received global model parameters as the initial state and trains the angle perception generation diffusion network model based on the initial remote sensing image. A remote sensing image with a specific azimuth angle is generated through the generator, discriminator, and angle transformation layer;

[0011] The client adopts the local data knowledge distillation strategy and conducts local distillation through cross-entropy loss and minority distillation loss;

[0012] The client uploads the updated model parameters to the server. The server adopts the adaptive weight adjustment strategy to aggregate the client model parameters, generates a new global model, and obtains the final global model through multiple rounds of iteration.

[0013] As a further solution of the present invention: The denoising diffusion probability model is used to preprocess the local remote sensing image to generate the initial remote sensing image. The specific steps are as follows:

[0014] In the forward diffusion stage, the client gradually adds Gaussian-distributed noise to the local remote sensing image according to the preset diffusion step sequence. For each diffusion step t, the corresponding variance β is selected from a pre-set variance time table t , and noise is added to the current image X t-1 in turn to obtain a new image X t . As the number of diffusion steps increases, until the image is completely noised;

[0015] Starting from the completely noised image, the noise prediction model is trained. In each reverse step, according to the noise predicted by the model, the image is denoised based on the random vector sampled from the standard Gaussian distribution, and the initial remote sensing image is gradually reconstructed.

[0016] As a further solution of the present invention: Training the angle perception generation diffusion network model based on the initial remote sensing image to generate a remote sensing image with a specific azimuth angle includes:

[0017] The generator includes an encoder and a decoder. After receiving the initial remote sensing image, the encoder extracts features from the image through convolutional layers, pooling layers, and activation functions, and converts the image into a high-dimensional feature vector;

[0018] The angle transformation layer performs a spatial transformation on the feature vector extracted by the encoder according to the input azimuth angle information, encodes the input azimuth angle information, and converts it into a form that interacts with the feature vector; according to the encoded azimuth angle information, the angle parameters for feature transformation are calculated, and through matrix multiplication and addition operations, the feature vector is combined with the angle parameters to adjust the direction and scale of the feature vector to match the specific azimuth angle;

[0019] The decoder receives the feature vectors after angular transformation and reconstructs a remote sensing image with a specific azimuth angle through a transposed convolution layer, an upsampling layer, and an activation function.

[0020] As a further solution of the present invention: The angle-aware generative diffusion network is adversarially trained through a least squares generative adversarial network to optimize the remote sensing image. The generator and the discriminator jointly update the parameters by minimizing the reconstruction loss and the adversarial loss.

[0021] The discriminator uses the intermediate layer features of a pre-trained VGG network as supervision and distinguishes real and generated images through a least squares loss function. The spectral normalization technique is introduced to constrain the Lipschitz constant of the discriminator, and the gradient penalty term is used to prevent mode collapse.

[0022] As a further solution of the present invention: The specific steps of the angle transformation layer include:

[0023] The angle transformation layer converts the azimuth angle information into a numerical representation and processes the preprocessed azimuth angle information through a fully connected neural network. The fully connected network contains multiple hidden layers, and each hidden layer uses an activation function for non-linear transformation. After the calculation of the fully connected network, the angle parameters for feature transformation are output. The angle parameters include the rotation angle, the scaling ratio, and the translation amount, which are used to define the rotation, scaling, and translation of the feature vector in space.

[0024] Based on the angle parameters, a linear transformation is performed on the feature vector extracted by the encoder to adjust the direction and scale of the feature vector so that the feature vector matches a specific azimuth angle, including: multiplying the feature vector by a rotation matrix to perform a planar rotation on the feature vector; multiplying each element of the feature vector by the corresponding scaling ratio; adding the corresponding translation amount to each element of the feature vector.

[0025] The transformed feature vector is fused with the original feature vector, and the transformed feature vector and the original feature vector are added element-wise to generate a new feature vector.

[0026] As a further solution of the present invention: The local data knowledge distillation strategy is adopted, and local distillation is performed through the cross-entropy loss and the minority distillation loss, which specifically includes:

[0027] According to the local dataset and the model prediction results, the cross-entropy loss and the minority distillation loss are calculated respectively; for each sample in the local dataset, its cross-entropy loss is calculated, and the sum of the cross-entropy losses of all samples is obtained as the total cross-entropy loss L CE ; for the minority class samples, their minority distillation loss L MC is calculated, and a combination of the cross-entropy loss L CE and the minority distillation loss L MCThe loss function L is linearly combined, L = L CE + βL MC , where the hyperparameter β is used to represent the knowledge preservation strength in the local distribution, and the final local distillation loss is obtained;

[0028] Based on the backpropagation algorithm, the parameters of the generative diffusion network model are updated according to the local distillation loss, the gradients with respect to the model parameters are calculated, and according to the gradient descent algorithm, the parameters of the model are updated according to the set learning rate, so that the local distillation loss continuously decreases until convergence.

[0029] As a further solution of the present invention: in the local distillation, the client retains the output probability distribution of the previous round of the global model as teacher knowledge, and the current local model as the student model; new knowledge of local data is learned through the cross-entropy loss, and at the same time, the prediction distributions of the teacher model and the local model for minority classes are aligned through the KL divergence loss.

[0030] As a further solution of the present invention: the server adopts an adaptive weight adjustment strategy to aggregate the client model parameters to generate a new global model, specifically including:

[0031] After receiving the model parameters uploaded by each client, the server uses the model uploaded by the client to predict the client's local dataset according to the model parameters uploaded by each client and the local dataset, calculates the accuracy and recall rate between the prediction result and the true label, and evaluates the performance of each client in this round of training;

[0032] According to the performance of the client, a corresponding weight coefficient is assigned to each client. The higher the accuracy and recall rate, the higher the weight. The accuracy and recall rate are normalized, and the weight coefficient is calculated through linear mapping;

[0033] The server performs weighted summation on the model parameters uploaded by each client according to the assigned weight coefficients to obtain new global model parameters.

[0034] As a further solution of the present invention: before distributing the global model parameters, the central server also performs an initialization operation on the global model, specifically including:

[0035] According to the remote sensing image generation task, the architecture of the angle-aware generative diffusion network model is set, including parameters such as the number of layers of the angle-conditioned encoder, decoder, and angle transformation layer, the number of neurons in each layer, and the convolutional kernel size; the random initialization method is used to assign initial values to the various parameters of the AGDN model.

[0036] The beneficial effects of the present invention:

[0037] The present invention uses a denoising diffusion probability model to generate high-fidelity remote sensing images from noise, and then uses an angle-conditioned encoder trained by a least squares generative adversarial network to guide the generation process and adapt to inherent perspective changes. Next, the server adopts an adaptive weight adjustment aggregation scheme to generate an appropriate global generative diffusion model. At the same time, a local data knowledge refinement strategy is adopted to alleviate the forgetting problem in joint learning by retaining previous global knowledge. While protecting data privacy and generating high-quality synthetic aperture radar images, it can also significantly improve the target recognition accuracy. The method proposed by the present invention provides a novel and secure data generation method for the synthetic aperture radar imaging field. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below with reference to the accompanying drawings.

[0039] Figure 1 It is a flowchart of the present invention.

[0040] Figure 2 It is a flowchart of the angle-aware generative diffusion network model of the present invention;

[0041] Figure 3 It is a schematic diagram of the generation of remote sensing images of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] Please refer to Figures 1 - 3 As shown, the present invention is an embodiment of a method for generating remote sensing images based on a federated angle-aware generative diffusion model.

[0044] 4.1 The purpose of the present invention is to train a model for generating remote sensing image data in a federated learning system composed of K clients and a central server. Each client k has a local remote sensing image dataset D k , and the entire dataset is In each communication round t, the server distributes the parameters w (t-1) of the global model to the clients K (t) ;

[0045] Immediately afterwards, each client k ∈ K (t) uses the local dataset D k to train the AGDN model and uses the LDKDS method to perform data distillation on the local data.

[0046] During the training phase, each client first uses DDPM to generate an initial remote sensing image, and then obtains the transformed representation of the input remote sensing image through the mapping function M g Specifically, given an input image x i and a target image x sampled from the remote sensing image t , the angle transformation layer (angletransformerlayer, ATL) calculates the transformation parameters, then generates a sampling grid, and finally uses the sampler to combine the input feature map with the sampling grid to generate the target feature image x t . At the same time, LDKDS performs anti-forgetting through the cross-entropy loss L CE and the minority distillation loss L MC . After the training ends, the model parameters w will be updated on the client k (t) and uploaded to the server side:

[0047]

[0048] where L is the loss function. At the end of the t-th round, the client uploads the locally updated parameters to the server and aggregates them in the way of parameter averaging, as follows:

[0049]

[0050] 4.2 In the framework of the present invention, first, DDPM denoises the remote sensing image data. This method enables the model to transform a completely noisy image into an image approximating a real synthetic aperture radar image through continuous iteration. The advantage of DDPM lies in its progressive denoising process, which can effectively capture the detailed changes and texture features in the image, making the generated image visually closer to the real image.

[0051] To further improve the angular diversity and accuracy of the generated image, the present invention introduces an angle-aware generation technology process as follows: The azimuth angle of the input remote sensing image is θ, denoted by x i ; the azimuth angle of the target remote sensing image is θ + △θ, denoted by x t . The present invention obtains the transformed representation of the input remote sensing image x g through the mapping function M i . Immediately afterwards, the encoder first inputs xi and maps it from the image to φ(x i ). Then, the decoder generates a pseudo-remote sensing image x g with the azimuth of θ + △θ through the reconstruction function M i φ(x f ), i.e., x g = δ(M i φ(x f ). In the initial state, since the network has not passed through the mapping function Mg Learning the angle transformation results in the generated remote sensing image x f and the real remote sensing image x t There are obvious differences between them. The image pairs {x i , x f} and {x i , x t} are then sent to the discriminator to determine whether each image pair is true or false. By minimizing the reconstruction loss (i.e., ) and the adversarial loss, the generator of the present invention is optimized. And piecewise interpolation is introduced in the stage of generating remote sensing images to further enhance the resolution and detail performance of the generated remote sensing images. Finally, given the real remote sensing image x r with the azimuth angle θ′, the AGDN generator of the present invention will generate the real remote sensing image x g with the azimuth angle θ + △θ′, that is, x g = δ(M g φ(x r ))

[0052] 4.2.1 The process of the DDPM of the present invention for generating the initial remote sensing image includes a forward process and a reverse process, as shown in the figure. In the forward process, the noise follows a standard Gaussian distribution. This process is based on the Markov chain hypothesis. It is known that the mean and variance of the noise added at each step are only affected by the current state, and its characteristic distribution is q(X t |X t-1 )

[0053] When the number of diffusion steps t is large enough, it is assumed that X t at this stage follows a standard Gaussian distribution. In the diffusion process, the diffusion step size is affected by the variable . Given a specific X t , and combined with the set diffusion step size. The diffusion process of each step is described as follows:

[0054]

[0055] Using the reparameterization method, let α t = 1 - β t and Z t ~ N(0, 1), there is

[0056] The final result of diffusion can be obtained by iteration Z ~ N(0, 1);

[0057] When :

[0058] The present invention can obtain the direct relationship between the real remote sensing image X0 at any time and the diffusion result X t ​

[0059]

[0060] The process of gradually removing a small amount of noise is called the inverse process, and finally reconstructs the distribution of remote sensing image data. Due to the properties of the Markov chain, as long as the distribution q(X t-1 |X t ) can be obtained, it is possible to gradually sample from the noise in the reverse process to obtain q(X0). This ultimately leads to data generation. The value of q(X t-1 |X t ) cannot be directly solved because it depends on the data distribution of the entire real sample. However, it can be expressed in terms of the current state X t and the initial state X0:

[0061]

[0062] When q(X t |X t-1 X0) = q(X t |X t-1 );

[0063]

[0064] The formula for the distribution that the training aims to fit can be expressed as

[0065]

[0066] In order to extract the features of remote sensing images and learn the angular transformation in the deep feature space, the generator of the present invention is designed according to the encoder-decoder architecture. Specifically, it mainly consists of three parts: a downsampling block (i.e., a convolutional layer, an instance normalization layer, and a ReLU activation layer), a residual block, and an upsampling block (i.e., a transposed convolutional layer, an instance normalization layer, and a ReLU activation layer). More importantly, the present invention integrates ATL into the encoder-decoder to achieve better angular transformation learning.

[0067] The present invention uses the intermediate layer of the pre-trained VGG model to guide the training of the encoder of the present invention. The VGG formula is as follows:

[0068]

[0069] Among them, P is a 1*1 convolutional layer adaptation, and E(I) defines the feature space. Φ5(I) represents the ReLU5_1 layer of the VGG network.

[0070] The decoder maximizes the distance between the real image angular features and the generated image, while the encoder E and the interpolator I are trained to minimize the distance between them. It is expressed as:

[0071]

[0072] F = E(I) is the angular feature of the initial remote sensing image extracted by the encoder, is the generated angular interpolation feature; || r and || x are the distributions of the real image and the interpolated image respectively, and L v is the defined gradient penalty term. Moreover, the interpolator I cooperates with the encoder E to generate a reasonable image. The pixel difference between the real target remote sensing image and the generated remote sensing image is minimized using the L1 norm, which ensures the consistency of the image content and is also convenient for the estimation of the mapping function M g . Therefore, the reconstruction term function is as follows

[0073]

[0074] where, || ||1 is the L1 norm, x i and x t are the input and target remote sensing images sampled from the distribution || r respectively. G(x i ) is the remote sensing image generated conditional on the input x i . By minimizing the reconstruction loss, the network is trained to estimate the mapping function M g . The final objective function of the encoder:

[0075]

[0076] where, L GAN , L recon , L v are defined in the formula respectively, and λ GAN , λ recon , λ v are set to 1 in the experiments of the present invention.

[0077] The sensitivity of target features in remote sensing images to the azimuth angle is an important factor directly affecting the performance of the AGDN model. However, existing GAN-based methods rely on a global discriminator that maps the entire remote sensing image to a single scalar output. To improve the fidelity of the generated targets, a least-squares loss function is adopted. This can mitigate the instability of model training, promote a more harmonious interaction between the generator and the discriminator, and more accurately reflect the authenticity of the samples. This method can not only classify samples as "true" or "false", but also further optimize the performance of the generated samples. This method can not only classify samples as "true" or "false", but also improve the quality of the generated samples. In addition, the present invention also adopts spectral normalization (SN) technology to enhance the stability of the network training process. This method can control the Lipschitz constant of the discriminator, thereby accelerating the convergence speed.

[0078] 4.2.2 To capture the correlation between the target at different azimuth angles, the present invention designs a rotation matrix:

[0079]

[0080] where is the input image x i relative to the target image x t angle. (u, v) are the coordinates of x i , and (u′, v′) are the coordinates of the transformed remote sensing image . The input image is deformed according to the obtained affine matrix to reconstruct the target remote sensing image. However, for remote sensing images, the gap between the input and the target cannot be bridged by simply distorting the pixels with each other. Therefore, ATL is prone to generating blurred images. So the present invention does not directly use as the final generated result, but uses ATL as a guide for feature representation. Therefore, the proposed generator and the designed ATL can complement each other to estimate M g . In this way, both strategies can maximize their respective advantages, generate remote sensing images with controllable azimuth angles, and retain target details.

[0081] The workflow of ATL includes three main steps: First, the localization network receives the input feature map and calculates the affine transformation parameters θ = f(I) through a series of hidden layers, which describe the method of spatial transformation of the input image; Second, the grid generator G = {G i} uses these parameters to create a sampling grid; Finally, the sampler combines the input feature map and the generated sampling grid to generate the target feature map The main purpose of ATL is to capture the target correlation between different azimuth angles by dynamically adjusting the spatial layout of the input image, thereby improving the quality and accuracy of the generated remote sensing image and ensuring that the generated image is closer to the real target in terms of shape, edge, and details.

[0082] 4.3 The forgetting problem that exists when applying FL to the generation of remote sensing image data and its solution, local data knowledge distillation (LDKDS).

[0083] In the FL environment, there are not only quantitative differences but also qualitative differences in the remote sensing image data among various clients. Such differences may lead to a scarcity of image samples of certain categories or low-quality samples on some clients, making it impossible for these clients to fully learn and master the features and knowledge of these categories. This phenomenon will affect the performance of the model in global tasks, especially when identifying categories that are rare in local data. This phenomenon is called "forgetting", that is, the model gradually loses the ability to remember the knowledge learned previously. Moreover, as the number of training rounds increases, the forgetting problem becomes more serious. The forgetting problem is the main reason for the poor performance of applying FL to the field of remote sensing image data generation. Among them, the categories with a small number are called minority classes. The existence of minority classes will lead to a decline in the performance of the model in global tasks, especially when these categories need to be identified.

[0084] As is well known, there is a forgetting problem in continual learning. In continual learning, a direct way to observe the forgetting problem is to check whether the fitting of the new distribution and the old distribution is consistent. However, in the problem setting of the present invention, there are many types of remote sensing image data of local clients, considering that the local distributions among different clients are different. Therefore, the present invention establishes the local distribution P(D) and its global distribution to systematically analyze the forgetting phenomenon in local training.

[0085] The core idea of LDKDS is to retain the global view for minority classes. More specifically, LDKDS performs local distillation through a linear combination loss function L between the cross-entropy loss L CE and the minority distillation loss L MC :

[0086]

[0087] where the hyperparameter β represents the knowledge preservation intensity on the local distribution. The minority class loss function L MC is the KL-Divergence between the predicted probability distribution of the model for minority classes and the predicted probability distribution of the global model:

[0088]

[0089] Among them, c≠y is the index of the minority classes of the remote sensing image, and y is the index of other classes of the remote sensing image. Minimize L MC Preserving the global knowledge outside the local distribution enables better alignment of the predictions of the minority classes of the remote sensing image with those of the global model. There are N groups of data in the local dataset D, x i probability distribution between cumulative divergence loss L KL is:

[0090]

[0091] By splitting the sum of the minority classes and other classes, the term becomes:

[0092]

[0093] Since and it can be obtained that

[0094] Next, the minority class part of the Kullback-Leibler divergence is derived by using double summation:

[0095] It can be obtained that:

[0096]

[0097] It shows that the KL divergence losses of the minority classes and other classes are equivalent, which means that in the optimization process, it is equally important to pay attention to the losses of the minority classes and other classes. By combining these two losses, LDKDS can maintain the memory of old knowledge while learning new knowledge.

[0098] Specifically, in the LDKDS loss function, the present invention tracks the true class signals of the labeled data in the local dataset through L CE to obtain new knowledge about the local distribution. At the same time, by adhering to the perspective of the global model, the present invention retains the prior knowledge about the global distribution. The hyperparameter β plays a crucial role in this process, which balances learning new knowledge and retaining previous insights. This method ensures that when the model encounters new minority classes, it can maintain the memory of past knowledge while updating and adapting. Through this dual strategy, LDKDS effectively enhances the robustness and adaptability of the model, ensuring excellent performance in multi-class tasks.

[0099] The above has described in detail an embodiment of the present invention, but the above content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the application of the present invention shall still fall within the scope covered by the patent of the present invention.

Claims

1. A remote sensing image generation method based on a federated angle-aware generation diffusion model, characterized in that: The following steps are involved: Construct a federated learning framework, including several clients and a central server. Each client stores a private remote sensing image dataset locally. In each communication round, the central server distributes the current parameters of the global model to each client. Each client uses a denoising diffusion probability model to preprocess the local remote sensing image to generate an initial remote sensing image; The client uses the received global model parameters as the initial state, trains the angle-aware generative diffusion network model based on the initial remote sensing image, and generates a remote sensing image with a specific azimuth through the generator, discriminator, and angle transformation layer; The client adopts the local data knowledge distillation strategy and performs local distillation through cross entropy loss and minority distillation loss; The client uploads the updated model parameters to the server. The server uses an adaptive weight adjustment strategy to aggregate the client model parameters, generate a new global model, and obtain the final global model through multiple rounds of iterations.

2. According to claim 1, a remote sensing image generation method based on a federated angle-aware generation diffusion model is characterized in that: The denoising diffusion probability model is used to preprocess the local remote sensing image to generate the initial remote sensing image. The specific steps are as follows: In the forward diffusion stage, the client gradually adds noise that conforms to the Gaussian distribution to the local remote sensing image according to the preset diffusion step sequence. For each diffusion step t, the corresponding variance β is selected from a preset variance time table. t , to the current image X t-1 Add noise to get the new image X t , as the number of diffusion steps increases, until the image is completely noisy; Starting from a completely noisy image, the noise prediction model is trained. In each reverse step, the image is denoised based on the noise predicted by the model and the random vectors sampled from the standard Gaussian distribution are used to gradually reconstruct the initial remote sensing image.

3. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 1 is characterized in that: Based on the initial remote sensing image, the angle-aware generative diffusion network model is trained to generate remote sensing images with specific azimuth angles, including: The generator includes an encoder and a decoder. After receiving the initial remote sensing image, the encoder extracts features of the image through a convolution layer, a pooling layer and an activation function, and converts the image into a high-dimensional feature vector. The angle transformation layer performs spatial transformation on the feature vector extracted by the encoder according to the input azimuth information, encodes the input azimuth information, and converts it into a form that interacts with the feature vector; according to the encoded azimuth information, the angle parameters used for feature transformation are calculated, and the feature vector is combined with the angle parameters through matrix multiplication and addition operations, and the direction and scale of the feature vector are adjusted to match the specific azimuth; The decoder receives the feature vector after angle transformation and reconstructs the remote sensing image with a specific azimuth through the deconvolution layer, upsampling layer and activation function.

4. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 3 is characterized in that: The angle-aware generative diffusion network is trained adversarially through the least squares generative adversarial network to optimize the remote sensing image. The generator and discriminator jointly update the parameters by minimizing the reconstruction loss and adversarial loss. The discriminator uses the intermediate layer features of the pre-trained VGG network as supervision and distinguishes real and generated images through the least squares loss function; the spectral normalization technology is introduced to constrain the Lipschitz constant of the discriminator, and the gradient penalty term is used to prevent mode collapse.

5. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 1 is characterized in that: The specific steps of the angle conversion layer include: The angle transformation layer converts the azimuth information into numerical representation, and processes the preprocessed azimuth information through a fully connected neural network, wherein the fully connected network includes multiple hidden layers, each of which uses an activation function for nonlinear transformation. After calculation by the fully connected network, the angle parameters for feature transformation are output, and the angle parameters include rotation angle, scaling ratio and translation amount, which are used to define the rotation, scaling and translation of the feature vector in space; Based on the angle parameter, a linear transformation is performed on the feature vector extracted by the encoder, and the direction and scale of the feature vector are adjusted so that the feature vector matches the specific azimuth, including: constructing a rotation matrix, multiplying the feature vector with the rotation matrix, and performing a plane rotation on the feature vector; multiplying each element of the feature vector by a corresponding scaling ratio; and adding a corresponding translation amount to each element of the feature vector; The transformed feature vector is fused with the original feature vector, and the transformed feature vector and the original feature vector are added element by element to generate a new feature vector.

6. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 1 is characterized in that: The local data knowledge distillation strategy is adopted to perform local distillation through cross entropy loss and minority distillation loss, including: According to the local dataset and model prediction results, the cross entropy loss and minority distillation loss are calculated respectively; for each sample in the local dataset, its cross entropy loss is calculated, and the cross entropy losses of all samples are summed to obtain the total cross entropy loss L CE ; For minority class samples, calculate their minority distillation loss L MC , construct the cross entropy loss L CE and a small amount of distillation loss L MC The loss function L formed by the linear combination, L = L CE +βL MC , the hyperparameter β is used to represent the knowledge preservation strength on the local distribution, and the final local distillation loss is obtained; Based on the back-propagation algorithm, the parameters of the generated diffusion network model are updated according to the local distillation loss, and the gradient relative to the model parameters is calculated. According to the gradient descent algorithm, the parameters of the model are updated according to the set learning rate, so that the local distillation loss continues to decrease until convergence.

7. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 6 is characterized in that: In the local distillation, the client retains the output probability distribution of the previous round of global model as the teacher knowledge, and the current local model as the student model; new knowledge of local data is learned through cross entropy loss, and the predicted distribution of the teacher model and the local model for minority categories is aligned through KL divergence loss.

8. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 1 is characterized in that: The server uses an adaptive weight adjustment strategy to aggregate client model parameters and generate a new global model, including: After receiving the model parameters uploaded by each client, the server uses the model uploaded by the client to predict the client's local dataset based on the model parameters and local dataset uploaded by each client, calculates the accuracy and recall rate between the prediction result and the true label, and evaluates the performance of each client in this round of training; According to the performance of the client, a corresponding weight coefficient is assigned to each client. The higher the accuracy and recall rate, the higher the weight. The accuracy and recall rate are normalized and the weight coefficient is calculated through linear mapping. The server performs weighted summation of the model parameters uploaded by each client according to the assigned weight coefficients to obtain new global model parameters.

9. The remote sensing image generation method based on the federated angle perception generation diffusion model according to claim 1, characterized in that: Before distributing the global model parameters, the central server also performs initialization operations on the global model, including: According to the remote sensing image generation task, the architecture of the angle-aware generative diffusion network model is set, including parameters such as the number of angle conditional encoders, decoders and angle transformation layers, the number of neurons in each layer, and the size of the convolution kernel. The random initialization method is used to assign initial values ​​to each parameter of the AGDN model.

Images