Multi-client collaborative sar image target recognition method based on semi-decentralized architecture

Abstract

The application relates to a multi-client cooperative SAR image target recognition method based on a semi-decentralized architecture, which comprises the following steps: collecting SAR image data and performing pretreatment, generating corresponding target recognition labels, and obtaining a labeled data set; a semi-decentralized federated learning framework is built, and global model parameters are initialized; each client uses local private data to perform model training, and obtains local parameter consensus through inter-device communication; a server randomly selects a client to perform parameter transmission, aggregates model parameters, updates global model parameters, and broadcasts the updated global model parameters to all clients; iterative training is performed until the global model parameters reach a convergence state, and a trained global model is obtained; the trained global model is tested, and a global model that passes the test is selected to perform SAR image target recognition. Compared with the prior art, the application can effectively reduce the communication overhead between the server and the client under the premise of guaranteeing model performance and data privacy protection.

Description

Technical Field

[0001] This invention relates to the field of target recognition applications in remote sensing synthetic aperture radar, and in particular to a multi-client cooperative SAR image target recognition method based on a semi-decentralized architecture. Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses all-weather, all-time operational capabilities and has important applications in military surveillance, disaster monitoring, and environmental protection. However, SAR data acquisition is costly and difficult to obtain, while data sharing raises privacy concerns regarding sensitive information. Effective implementation of SAR Automatic Target Recognition (ATR) faces significant challenges related to data scarcity and privacy protection. To address these issues, Federated Learning (FL) technology has been proposed. By transferring model parameters among participating parties instead of raw data, it enables multiple institutions to collaboratively train models without sharing data, thus ensuring data privacy.

[0003] However, traditional FL architectures still suffer from high communication overhead. Traditional FL architectures employ a star topology, with a master server connecting multiple clients. Each training round includes local updates and global parameter aggregation. This centralized FL architecture has limitations such as high latency, high bandwidth utilization, and high power consumption, leading to high communication overhead. Moreover, if the central server fails, the entire training process will be severely affected.

[0004] With advancements in communication technology, the cost of inter-satellite communication has been significantly reduced, enabling highly reliable and low-latency inter-satellite connections. This makes it possible to replace as much inter-satellite-to-server communication as possible with inter-satellite communication, thereby significantly reducing the overall communication cost of the system. Therefore, while ensuring data privacy, combining inter-satellite communication technology to reduce the communication cost between servers and satellites and achieve joint model training has become a key challenge. Summary of the Invention

[0005] The purpose of this invention is to overcome the high communication overhead between the server and the satellite in the existing FL architecture and to provide a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A multi-client cooperative SAR image target recognition method based on a semi-decentralized architecture includes the following steps:

[0008] S1: Acquire SAR image data and preprocess it to generate corresponding target recognition labels, thus obtaining a labeled dataset;

[0009] S2: Build a semi-decentralized federated learning framework, which includes a server and multiple clients in the cluster. Clients within the cluster communicate with each other. Construct a neural network model for target recognition and initialize global model parameters.

[0010] S3: Allocate local private data to each client based on the labeled dataset. Each client uses the local private data to train the model. The local parameters are agreed upon through inter-device communication. The updated local model parameters are then prepared to be uploaded to the server.

[0011] S4: The server randomly selects a client to transmit parameters, aggregates the model parameters and updates the global model parameters, and broadcasts the updated global model parameters to all clients;

[0012] S5: Repeat steps S3 and S4 until the global model parameters reach convergence and the trained global model is obtained.

[0013] S6: Test the trained global model and select the global model that passes the test for SAR image target recognition.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] S101: Acquire SAR image data via satellite;

[0016] S102: Preprocess the acquired SAR image data, including denoising, correction and enhancement;

[0017] S103: Generate corresponding target identification labels based on the preprocessed SAR image data.

[0018] Furthermore, step S2 specifically includes the following steps:

[0019] S201: Establish a semi-decentralized federated learning framework, determine the participating clients and servers, and the number of clients; clients communicate with each other via devices, and the server communicates with each client via satellite link and is responsible for aggregating global model parameters;

[0020] S202: Pre-build a neural network model for target recognition and initialize global model parameters;

[0021] S203: The server distributes the initialized global model parameters to each client.

[0022] Furthermore, the neural network model for target recognition includes a backbone network for feature extraction and a classifier composed of fully connected layers for classification;

[0023] The backbone network extracts high-level features from the input SAR image step by step through a series of convolutional layers;

[0024] The classifier maps the high-level features extracted by the backbone network to the final output category.

[0025] Furthermore, step S3 specifically includes the following steps:

[0026] S301: Each client holds a local dataset allocated based on the labeled dataset. Each client performs local training based on the allocated global model parameters, optimizing the local model parameters to minimize the local cross-entropy loss function. The optimization calculation expression for the local model parameters is as follows:

[0027]

[0028] In the formula, These are the local model parameters for client n at iteration k. For the i-th SAR image of client n, D is the label for the i-th target of client n. (n) For local datasets;

[0029] S302: Each client communicates with other clients in its cluster. (k) In round-by-round D2D communication, local parameters are iteratively updated. In the k-th iteration, client n exchanges parameters with its neighbor client m. The update expression is:

[0030]

[0031] In the formula, For the local parameters of client n after round t+1 update, Let m be the consensus parameter between client n and neighboring client m in the k-th iteration. Let n be the consensus parameter of client n in the k-th iteration. ζ is the local parameter of client n in round t. (k) (n) is the set of neighbors of client n in the k-th iteration. For client m, the local parameters in round t;

[0032] S303: Consensus parameters aggregated locally by each client Ready to upload to the server.

[0033] Further, in step S303, the optimized client-side model parameter aggregation expression is used for parameter updating. The optimized client-side model parameter aggregation expression is:

[0034]

[0035] In the formula, d (k) This represents the maximum number of client neighbors in the cluster during the k-th iteration.

[0036] Furthermore, step S4 specifically includes the following steps:

[0037] S401: The server randomly selects a client m for parameter transmission;

[0038] S402: The server obtains the local parameter consensus for the current iteration number k from the selected client m.

[0039] S403: The server will receive local parameters for consensus. As new global model parameters This global model parameter The calculation expression is:

[0040]

[0041] In the formula, These are the global model parameters after the k-th round update;

[0042] S404: The server will update the global model parameters. The broadcast is sent to all clients, and the update expression for each client's local parameters is:

[0043]

[0044] In the formula, These are the model parameters for client n at the start of round k+1.

[0045] Furthermore, step S5 specifically includes the following steps:

[0046] S501: Each client receives the global model parameters sent by the server and updates its local model parameters.

[0047] S502: Each client uses a local dataset for model training to minimize the local loss function, the expression for which the local loss function is calculated is:

[0048]

[0049] In the formula, L n (θ) is the local loss function of client n, D n Given a local dataset, l is the cross-entropy loss function, and f(x) i ;θ) represents the model's predicted value, y i Identify labels for the target;

[0050] S503: Each client uses the gradient descent algorithm to update the local model parameters. The corresponding update expression is:

[0051]

[0052] In the formula, These are the local model parameters for client n at iteration k. Here, η represents the local model parameters for client n at iteration k+1, and η is the learning rate. Let be the gradient of the local loss function in the k-th round;

[0053] S504: Inter-device communication between clients;

[0054] S505: The client obtains local parameter consensus through multiple rounds of inter-device communication iterations;

[0055] S506: The client obtains the final local parameter consensus in preparation for uploading to the server;

[0056] S507: The server receives local parameter consensus from the selected clients and updates the global model parameters:

[0057] S508: The server determines whether the change in global model parameters is less than the preset threshold or whether the number of iterations has reached the maximum number of iterations. If so, the global model parameters are considered to have reached a convergence state; otherwise, steps S501-S507 are repeated.

[0058] Furthermore, step S6 specifically includes the following steps:

[0059] S601: Select and prepare the test dataset;

[0060] S602: Load the trained global model parameters into the test environment;

[0061] S603: Perform inference on the global model in the test dataset to obtain the prediction result. The calculation expression for the prediction result is as follows:

[0062]

[0063] In the formula, For the predicted results, For global model parameters, x test For test data, f(·) is the model function;

[0064] S604: Compare the prediction results with the true labels y in the test dataset. test To compare the results, the validation accuracy Acc of the model is calculated. The expression for Acc is as follows:

[0065]

[0066] In the formula, D test For the test dataset, 1() is an indicator function that takes a value of 1 when the predicted result is the same as the true label, and 0 otherwise. For the i-th prediction result, y test,i For the i-th real label y test ;

[0067] S605: Evaluate model performance based on the model's validation accuracy (Acc);

[0068] S606: Optimize and improve the global model based on the model performance evaluation results.

[0069] Furthermore, step S605 also includes:

[0070] Record the training loss L for different numbers of clients train It is used to evaluate the learning process of the global model and the convergence speed under different client configurations;

[0071] Record the validation accuracy (Acc) of the global model under different communication overheads and in each communication round. round It is used to evaluate the changes and stability of model accuracy during training;

[0072] Record the global model validation accuracy (Acc) after completing the maximum number of communication rounds. final This is used to measure the overall effectiveness and performance of the training process with different numbers of clients.

[0073] This invention proposes a multi-client cooperative synthetic aperture radar (SAR) image target recognition method based on a semi-decentralized architecture, aiming to address the challenge of high communication overhead between the server and satellite in the traditional FL architecture. The traditional FL architecture employs a star topology, requiring all clients to communicate frequently with the central server, leading to high latency, high bandwidth utilization, and high power consumption. Furthermore, a failure of the central server severely impacts the entire training process.

[0074] Compared with the prior art, the present invention has the following advantages:

[0075] (1) This invention reduces the frequency of direct communication with the central server by introducing local aggregation and device-to-device (D2D) communication technologies, thereby reducing communication overhead. Specifically, satellites achieve local parameter consensus through D2D communication, and the master server only needs to transmit parameters with the selected satellites, significantly reducing the burden on the central server;

[0076] The innovation lies in the first application of a semi-decentralized architecture to multi-client collaborative SAR target recognition tasks, effectively solving the problem of high communication overhead in traditional FL architectures. By reducing the frequency of direct communication between the server and clients while ensuring data privacy, the robustness and communication efficiency of the overall system are improved.

[0077] (2) Experimental results show that the multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture proposed in this invention significantly reduces communication overhead while ensuring model performance and data privacy protection, outperforming the performance of the traditional centralized architecture. Specifically, in the semi-decentralized collaborative training tasks with 10 and 20 clients, the verification accuracy can reach as high as 97.69% and 96.07% respectively through iterative training.

[0078] Therefore, this invention is significant in addressing the high communication overhead problem in FL architecture. It effectively improves model performance and robustness while ensuring data privacy. This method has broad application prospects in military, civilian, and other fields. Attached Figure Description

[0079] Figure 1 This is a flowchart illustrating a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture provided in an embodiment of the present invention.

[0080] Figure 2 This is a schematic diagram of vehicle optical and SAR data in an MSTAR dataset provided in an embodiment of the present invention, wherein (a) ZSU234; (b) ZIL131; (c) T72; (d) T62; (e) D7; (f) BTR70; (g) BTR60; (h) BRDM2; (i) BMP2; (j) 2S1;

[0081] Figure 3 This is a schematic diagram of a semi-decentralized multi-client collaborative SAR image target recognition framework provided in an embodiment of the present invention;

[0082] Figure 4 This is a schematic diagram of a neural network structure provided in an embodiment of the present invention;

[0083] Figure 5 This is a schematic diagram illustrating the training loss results for collaborative tasks with different numbers of clients in a semi-decentralized and centralized framework provided in an embodiment of the present invention.

[0084] Figure 6 This is a schematic diagram illustrating the verification accuracy of collaborative tasks with varying communication overhead in a semi-decentralized architecture (SDMC-SAR) and a traditional centralized architecture (star topology) with different numbers of clients, as provided in an embodiment of the present invention.

[0085] Figure 7 This is a schematic diagram illustrating the verification accuracy of collaborative tasks with different numbers of clients in a semi-decentralized architecture (SDMC-SAR) and a traditional centralized architecture (star topology) as the number of communication rounds varies in an embodiment of the present invention.

[0086] Figure 8 This is a schematic diagram illustrating the final verification accuracy of a collaborative task with different numbers of clients in a semi-decentralized framework provided in an embodiment of the present invention. Detailed Implementation

[0087] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0088] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0089] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0090] Example 1

[0091] like Figure 1 As shown, this embodiment provides a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture, which enables multiple satellite agencies to jointly train models without sharing SAR data. This method reduces communication overhead between the server and client while ensuring model performance and data privacy protection. The method includes the following steps:

[0092] S1: SAR image data acquisition and preprocessing: Acquire SAR image data and preprocess it to generate corresponding target identification labels and obtain a labeled dataset. In this embodiment, the labeled dataset is further divided into a training set and a test set.

[0093] S2: Build a semi-decentralized federated learning framework and initialize global model parameters: Build a semi-decentralized federated learning framework, which includes a server and multiple clients in the cluster. Clients within the cluster communicate with each other. Construct a neural network model for target recognition and initialize global model parameters for collaborative training.

[0094] S3: Local model training and update: Based on the labeled dataset, local private data is allocated to each client. Each client uses the local private data to train the model, obtains local parameter consensus through device-to-device communication (D2D), and prepares to upload the updated local model parameters to the server.

[0095] S4: Global model aggregation and update: The server randomly selects clients to transmit parameters, aggregates model parameters and updates global model parameters, and broadcasts the updated global model parameters to all clients;

[0096] S5: Iterative training until the global model converges: Repeat steps S3 and S4 until the global model parameters reach a convergent state, and obtain the trained global model.

[0097] S6: Model Testing and Performance Evaluation: Test the trained global model using the test dataset, obtain the validation accuracy, evaluate the overall performance of the model, and select the global model that passes the test for SAR image target recognition.

[0098] The following combination Figure 1 The flowchart of the multi-client cooperative synthetic aperture radar (SAR) image target recognition method based on a semi-decentralized architecture in this embodiment is described.

[0099] Step S1 involves SAR image data acquisition and preprocessing: acquiring SAR image data, preprocessing the data, generating labels, and dividing the dataset into training and test sets.

[0100] Step S1 includes the following sub-steps:

[0101] Step S101: Data Acquisition. The satellite agency acquires SAR image data.

[0102] Step S102: Data preprocessing. The acquired SAR image data undergoes preprocessing operations such as denoising, correction, and enhancement to improve data quality.

[0103] Step S103: Generate tags. Generate corresponding target identification tags based on SAR image data. Each satellite agency n forms a local dataset. Where |D (n) | represents the size of the dataset. and These represent the input SAR image and its corresponding label, respectively.

[0104] Step S104: Split the dataset. In the example of collaborative multi-client SAR target recognition, we used the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset. The MSTAR dataset is a widely used SAR target recognition dataset containing SAR images of 10 types of military vehicles. Figure 2 This section presents schematic diagrams of optical and SAR images corresponding to various vehicle types in the MSTAR dataset. Table 1 shows the number of training and testing samples for the 10 target classes in the MSTAR dataset.

[0105] Each satellite agency will process its own labeled dataset D (n) Divided into training set and test set In this example, the training and test sets consist of target data with pitch angles of 17° and 15°, respectively.

[0106] Specifically, in the collaborative multi-client SAR target recognition example, we assume that samples of each class are randomly and uniformly distributed among all clients, ensuring that there is no overlap or intersection of data between clients. The example simulates Independent and Identically Distributed (IID) scenarios, uniformly and randomly dividing the dataset among the clients participating in collaborative training. To verify the effectiveness and generalization of the proposed method, the example simulates configurations of 2, 4, 6, 8, 10, 15, and 20 clients, and compares the semi-decentralized distributed training results with the centralized distributed training results.

[0107] Table 1 shows the number of training and testing samples for the 10 target classes in the MSTAR dataset.

[0108]

[0109]

[0110] Step S2, Build a semi-decentralized federated learning framework and initialize global model parameters: Build a semi-decentralized federated learning framework and initialize global model parameters for collaborative training.

[0111] Step S201: Establish a semi-decentralized federated learning framework and determine the participating clients and servers. Specifically, this includes determining the clients U participating in collaborative training. fl The number of clients N and the central server S flAn example of multi-client collaborative SAR target recognition with a semi-decentralized architecture was simulated, with configurations of 2, 4, 6, 8, 10, 15, and 20 clients (i.e., N = 2, 4, 6, 8, 10, 15, 20). The results of semi-decentralized distributed training were compared with those of centralized distributed training.

[0112] Figure 3 This is a schematic diagram of a semi-decentralized multi-client collaborative SAR image target recognition framework. Specifically, the system model is as follows: This semi-decentralized architecture consists of multiple satellites forming a cluster, each capable of device-to-device (D2D) communication. The master server is located on the ground, communicating with the cluster via satellite links and responsible for aggregating global model parameters. The D2D communication links within the cluster are represented by white arrows, and the yellow arrows represent the parameter transmission paths between the satellites and the server.

[0113] In this semi-decentralized framework, SAR data is not shared between satellites; instead, satellites reach local parameter consensus via D2D links. The master server communicates with randomly selected satellites through parameter transmission to jointly train the model. This semi-decentralized architecture significantly reduces communication costs between satellites and servers while ensuring data privacy.

[0114] Step S202: Initialize global model parameters, setting initial weights and learning rate, etc. Specifically, this includes: defining the model architecture; initializing global model parameters θ according to the preset neural network structure. glb The initialized global model parameters are Set the learning rate η and other relevant hyperparameters, such as batch size B and number of training epochs E.

[0115] To effectively perform SAR target identification, a neural network model was designed in the example, such as... Figure 4 As shown, the model mainly consists of two parts: a backbone network for feature extraction and a classifier composed of fully connected layers for classification.

[0116] Figure 4The overall architecture of the network is shown. The backbone network progressively extracts high-level features from the input SAR image through a series of convolutional layers, each followed by a batch normalization, ReLU activation, and pooling layer. This hierarchical approach ensures that the model can capture the spatial and hierarchical features of the input data. Specifically, the backbone network consists of four convolutional modules. After feature extraction, fully connected layers act as classifiers. The fully connected layers map the extracted features to the final output categories. Specifically, the fully connected layers consist of two layers. The first fully connected layer has 256 units, followed by a ReLU activation function and a Dropout layer to prevent overfitting. The second fully connected layer maps the 256 units to the number of output categories, and finally uses a Softmax activation function to convert the logits into multi-class classification probabilities.

[0117] Furthermore, in this example, the specific hyperparameter settings are as follows: Batch Size is 100, Maximum Communication Rounds is 1500, and Learning Rate is 3×10⁻⁶. -4 The optimizer used was AdamW. All experiments were performed on a single RTX 3090 GPU. The steps outlined above initialized the global model parameters, ensuring the model's fundamental performance and laying the foundation for subsequent local training and global aggregation.

[0118] Step S203: The server distributes the initial global model parameters. Server S fl initialize global model parameters Distributed to various client U fl .

[0119] Step S3, Local Model Training and Update: Each client uses local private data to train the model, obtains local parameter consensus through device-to-device (D2D) communication, and prepares to upload the updated model parameters to the server.

[0120] Step S3 includes the following sub-steps:

[0121] Step S301: Each client uses its local private data for model training. Each client (n) holds a local dataset D. (n) , including input SAR images and corresponding tags In each iteration k, each client base is based on the global model parameters. Perform local training and optimize local model parameters. To minimize the local cross-entropy loss function:

[0122]

[0123] Step S302: Obtain local parameter consensus through device-to-device (D2D) communication. Each client communicates with other clients in its cluster. (k) In each round of D2D communication, local parameters are iteratively updated. During the k-th iteration, client n exchanges parameters with its neighbor client m using the following formula. Update:

[0124]

[0125] in, Let ζ be the consensus parameter between client n and neighboring client m during the k-th iteration. (k) (n) represents the set of neighbors of client n at the k-th iteration. For better convergence, we use the following formula...

[0126] Aggregate local model parameters:

[0127]

[0128] Among them, 0 <d (k) <1 / D (k) D (k) This represents the maximum number of client neighbors in the cluster during the k-th iteration.

[0129] Step S303: Prepare to upload the updated model parameters to the server. (In t) (k) After the round of D2D communication is completed, the selected client prepares to aggregate its consensus parameters locally. Uploaded to the main server.

[0130] Step S4: Global model aggregation and update: The server randomly selects a client, transmits parameters with the randomly selected client, aggregates model parameters, and updates the global model.

[0131] Step S4 includes the following sub-steps:

[0132] Step S401: Select a client. The server randomly selects a client m for parameter transmission.

[0133] Step S402: Receive the updated local model parameters. The server receives the local parameter consensus obtained through D2D communication from the selected client m.

[0134] Step S403: Update global model parameters. The server will receive the local model parameters. As new global model parameters The updated formula is as follows:

[0135]

[0136] in, This represents the global model parameters after the k-th round of updates.

[0137] Step S404: Distribute global model parameters. The server will distribute the updated global model parameters. The broadcast is sent to all clients, preparing for the next round of local model training, specifically:

[0138]

[0139] in, This represents the model parameters of client n at the start of round k+1.

[0140] Step S5: Iterative training until the global model converges: Repeat steps S3 and S4 until the global model parameters reach a convergent state.

[0141] Step S5 includes the following sub-steps:

[0142] Step S501: Distribute global model parameters. The server will distribute the updated global model parameters. Broadcast to all clients:

[0143]

[0144] in, This represents the model parameters of client n at the start of round k.

[0145] Step S502: Local Model Training. Client n trains the model using local private data, minimizing the local loss function:

[0146]

[0147] Where l represents the cross-entropy loss function, f(x) i ;θ) represents the model's predicted value, y i This indicates the actual label.

[0148] Step S503: Local Model Update. Each client n updates its local model parameters using the gradient descent algorithm:

[0149]

[0150] Where η represents the learning rate. This represents the gradient of the local loss function in the k-th round.

[0151] Step S504: Device-to-Device (D2D) Communication. Clients communicate with each other through D2D communication for several rounds of iteration to reach a consensus on local parameters.

[0152]

[0153] in, This represents the local parameters of client n after the t-th iteration. This represents the communication weight between clients n and m in the k-th round.

[0154] Step S505: Achieve local parameter consensus. Client n achieves local parameter consensus after multiple rounds of D2D communication iterations:

[0155]

[0156] Among them, t (k) This represents the number of D2D communication iterations in the k-th round.

[0157] Step S506: Upload local parameter consensus. The client will upload the updated local parameter consensus. Uploaded to the server.

[0158] Step S507: Global Model Update. The server receives local parameter consensus from the selected client and updates the global model parameters:

[0159]

[0160] Where m represents a randomly selected client.

[0161] Step S508: Check the convergence status. If the global model parameters θ glb If the change is less than the preset threshold or the maximum number of iterations is reached, the model is considered to have converged; otherwise, repeat steps S5-1 to S5-7.

[0162] Step S6, Model Testing and Performance Evaluation: Validate the trained global model using the test dataset, obtain the validation accuracy, and evaluate the overall performance of the model.

[0163] Step S6 includes the following sub-steps:

[0164] Step S601: Prepare the test dataset. Select and prepare an independent test dataset D, separate from the training dataset. test This is to ensure the objectivity of the test results.

[0165] Step S602: Load global model parameters. Load the finally converged global model parameters. Load it into the test environment.

[0166] Step S603: Perform model inference on the test dataset. Use the test dataset D. te Perform inference on the global model to obtain the model's prediction results.

[0167]

[0168] Where, x test For the test data, f(·) is the model function.

[0169] Step S604: Calculate the validation accuracy. Use the model's prediction results... Compared with the true label y of the test dataset test To compare the models, the validation accuracy (Acc) is calculated to evaluate the model's recognition performance.

[0170]

[0171] Here, 1() is an indicator function, which takes the value 1 when the prediction result is the same as the true label, and 0 otherwise.

[0172] Step S605: Evaluate model performance. In addition to verifying accuracy, other performance metrics can be calculated to comprehensively evaluate the model, analyzing its convergence speed and stability in practical applications. Specifically:

[0173] 1. Record the training loss L for different numbers of clients. train This metric provides information about the model's learning process and convergence speed under different client configurations:

[0174]

[0175] Figure 5 This paper illustrates the training loss for collaborative tasks with different numbers of clients in both semi-decentralized and centralized frameworks. The horizontal axis represents the number of communication rounds (com.rnds.), and the vertical axis represents the training loss. The blue curve represents federated learning in a traditional star topology, while the orange curve represents the semi-decentralized architecture (SDMC-SAR) proposed in this invention.

[0176] from Figure 5 As can be seen, although the traditional star topology reduces training loss faster in the early stages, the semi-decentralized architecture (SDMC-SAR) proposed in this invention can eventually achieve a similar level of convergence. This indicates that although the traditional method converges faster in the initial stages, both methods can eventually converge to a lower training loss.

[0177] Meanwhile, the semi-decentralized architecture maintained a relatively stable training loss in most cases, demonstrating its stability under large-scale client configurations. Specifically, in collaborative training tasks with varying numbers of clients, the semi-decentralized architecture achieved convergence results close to those of the traditional star topology, validating its ability to maintain model performance.

[0178] 2. Record the global model's validation accuracy (Acc) across different communication overheads and communication rounds. round This helps to understand the changes and stability of model accuracy during the training process;

[0179] Figure 6 The figure illustrates the validation accuracy of the semi-decentralized architecture (SDMC-SAR) and the traditional centralized architecture (star topology) under varying communication overhead configurations with different numbers of clients. The figure shows that, under the same overhead conditions, the SDMC-SAR architecture exhibits higher validation accuracy across all client number configurations. This indicates that SDMC-SAR can reduce communication overhead while more efficiently utilizing communication resources, thereby improving the overall performance of the model.

[0180] Figure 7 The demonstration shows the verification accuracy of a semi-decentralized architecture (SDMC-SAR) and a traditional centralized architecture (star topology) under varying communication rounds with different client counts. It can be seen that although the star topology shows a faster increase in verification accuracy in the initial stages of communication round changes due to its centralized parameter aggregation, as the number of communication rounds increases, the SDMC-SAR architecture gradually improves its verification accuracy through effective inter-device communication and local parameter consensus, eventually reaching performance levels comparable to the star topology.

[0181] 3. Record the global model validation accuracy (Acc) after completing the maximum number of communication rounds (e.g., 1500 rounds). final The overall effectiveness and performance of the training process are measured for different numbers of clients (e.g., 1, 2, 4, 6, 8, 10, 15, and 20).

[0182] Figure 8 The global model validation accuracy is shown after completing the maximum number of communication rounds (e.g., 1500 rounds). Table 2 shows the final validation accuracy for collaborative tasks with different numbers of clients in the semi-decentralized and centralized frameworks. It can be seen that the semi-decentralized framework, while reducing communication overhead, still maintains performance roughly equivalent to the centralized architecture.

[0183] Table 2. Final validation accuracy of collaborative tasks with different numbers of clients in semi-decentralized and centralized frameworks.

[0184]

[0185] Step S606: Optimize the model. Based on the performance evaluation results, optimize and improve the model. This may include adjusting hyperparameters, improving the model structure, and optimizing the training strategy. Analyze the effectiveness and reliability of the model to further improve its performance and generalization ability, and determine its applicability in practical applications.

[0186] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture, characterized in that, Includes the following steps: S1: Acquire SAR image data and preprocess it to generate corresponding target recognition labels, thus obtaining a labeled dataset; S2: Build a semi-decentralized federated learning framework, which includes a server and multiple clients in the cluster. Clients within the cluster communicate with each other. Construct a neural network model for target recognition, initialize it, and distribute global model parameters to each client. S3: Allocate local private data to each client based on the labeled dataset. Each client uses the local private data to train the model. The local parameters are agreed upon through inter-device communication. The updated local model parameters are then prepared to be uploaded to the server. S4: The server randomly selects a client to transmit parameters, aggregates the model parameters and updates the global model parameters, and broadcasts the updated global model parameters to all clients; S5: Repeat steps S3 and S4 until the global model parameters reach convergence and the trained global model is obtained. S6: Test the trained global model and select the global model that passes the test for SAR image target recognition; Step S2 specifically includes the following steps: S201: Establish a semi-decentralized federated learning framework, determine the participating clients and servers, and the number of clients; clients communicate with each other via devices, and the server communicates with each client via satellite link and is responsible for aggregating global model parameters; S202: Pre-build a neural network model for target recognition and initialize global model parameters; S203: The server distributes the initialized global model parameters to each client; The neural network model for target recognition includes a backbone network for feature extraction and a classifier consisting of fully connected layers for classification. The backbone network extracts high-level features from the input SAR image step by step through a series of convolutional layers; The classifier maps the high-level features extracted by the backbone network to the final output category. Step S3 specifically includes the following steps: S301: Each client holds a local dataset allocated based on the labeled dataset. Each client performs local training based on the allocated global model parameters, optimizing the local model parameters to minimize the local cross-entropy loss function. The optimization calculation expression for the local model parameters is as follows: In the formula, These are the local model parameters for client n at iteration k. For the i-th SAR image of client n, To identify the label for the i-th target of client n, For local datasets; S302: Each client communicates with other clients in its cluster. In round-by-round D2D communication, local parameters are iteratively updated. In the k-th iteration, the expression for client n exchanging parameters with neighboring client m to update local parameters is: In the formula, For the local parameters of client n after round t+1 update, For the first The consensus parameters between client n and neighboring client m during round iteration. For the first The consensus parameter of client n during round iteration, For client n, the local parameters in round t. For the first The set of neighbors of client n during round iteration, For client m, the local parameters in round t; S303: Each client reaches a consensus on its locally aggregated parameters. Ready to upload to the server; In step S303, the optimized client-side model parameter aggregation expression is used for parameter update. The optimized client-side model parameter aggregation expression is: In the formula, For the first The maximum number of client neighbors in the cluster during round iteration.

2. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1, characterized in that, Step S1 specifically includes the following steps: S101: Acquire SAR image data via satellite; S102: Preprocess the acquired SAR image data, including denoising, correction and enhancement; S103: Generate corresponding target identification labels based on the preprocessed SAR image data.

3. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1, characterized in that, Step S4 specifically includes the following steps: S401: The server randomly selects a client m for parameter transmission; S402: The server obtains the local parameter consensus for the current iteration number k from the selected client m. ; S403: The server will receive local parameters for consensus. As new global model parameters The global model parameters The calculation expression is: In the formula, These are the global model parameters after the k-th round update; S404: The server will update the global model parameters. The broadcast is sent to all clients, and the update expression for each client's local parameters is: In the formula, These are the model parameters for client n at the start of round k+1.

4. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1, characterized in that, Step S5 specifically includes the following steps: S501: Each client receives the global model parameters sent by the server and updates its local model parameters. S502: Each client uses a local dataset for model training to minimize the local loss function, the expression for which the local loss function is calculated is: In the formula, Let n be the local loss function for client n. For local datasets, Let cross-entropy be the loss function. The predicted value of the model. Identify labels for the target; S503: Each client uses the gradient descent algorithm to update the local model parameters. The corresponding update expression is: In the formula, These are the local model parameters for client n at iteration k. These are the local model parameters for client n at iteration k+1. For learning rate, Let be the gradient of the local loss function in the k-th round; S504: Inter-device communication between clients; S505: The client obtains local parameter consensus through multiple rounds of inter-device communication iterations; S506: The client obtains the final local parameter consensus in preparation for uploading to the server; S507: The server receives local parameter consensus from the selected clients and updates the global model parameters: S508: The server determines whether the change in global model parameters is less than the preset threshold or whether the number of iterations has reached the maximum number of iterations. If so, the global model parameters are considered to have reached a convergence state; otherwise, steps S501-S507 are repeated.

5. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1, characterized in that, Step S6 specifically includes the following steps: S601: Select and prepare the test dataset; S602: Load the trained global model parameters into the test environment; S603: Perform inference on the global model in the test dataset to obtain the prediction result. The calculation expression for the prediction result is as follows: In the formula, For the predicted results, These are global model parameters. For test data, For model functions; S604: Compare the prediction results with the true labels of the test dataset. Compare the results and calculate the validation accuracy of the models. The accuracy The calculation expression is: In the formula, For the test dataset, This is an indicator function; it takes a value of 1 when the predicted result matches the true label, and 0 otherwise. For the i-th prediction result, For the i-th real label ; S605: Based on the model's validation accuracy Evaluate model performance; S606: Optimize and improve the global model based on the model performance evaluation results.

6. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 5, characterized in that, Step S605 also includes: Record the training loss for different numbers of clients It is used to evaluate the learning process of the global model and the convergence speed under different client configurations; Record the validation accuracy of the global model under different communication overheads and in each communication round. It is used to evaluate the changes and stability of model accuracy during training; Record the global model validation accuracy after completing the maximum number of communication rounds. This is used to measure the overall effectiveness and performance of the training process with different numbers of clients.

Images