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

Abstract

The invention relates to a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture, and the method comprises the steps: collecting SAR image data, carrying out the preprocessing, generating a corresponding target recognition tag, and obtaining a data set with a tag; building a semi-decentralized federated learning framework, and initializing global model parameters; each client performs model training by using local private data, and obtains a local parameter consensus through communication between devices; the server randomly selects the clients for parameter transmission, aggregates model parameters, updates global model parameters, and broadcasts the updated global model parameters to all the clients; iterative training is carried out until global model parameters reach a convergence state, and a trained global model is obtained; and testing the trained global model, and selecting the tested global model to carry out SAR image target identification. Compared with the prior art, on the premise of ensuring model performance and data privacy protection, the communication overhead between the server and the client can be effectively reduced.

Description

Technical Field

[0001] The present invention relates to the application field of remote sensing synthetic aperture radar target recognition, and in particular to a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture. Background Art

[0002] Synthetic aperture radar (SAR) has the ability to work all day and all night, and has important applications in military surveillance, disaster monitoring, and environmental protection. However, SAR data collection is expensive and difficult to obtain, and data sharing involves privacy issues of sensitive information. The effective implementation of SAR automatic target recognition (ATR) faces major challenges of data scarcity and privacy protection. To address these issues, federated learning (FL) technology was proposed. By transferring model parameters instead of raw data between participants, multiple institutions can jointly train models without sharing data, ensuring data privacy.

[0003] However, the traditional FL architecture still has the problem of high communication overhead. The traditional FL architecture adopts a star topology, where the main server connects multiple clients, and each round of training includes local updates and global parameter aggregation. Such a centralized FL architecture has some limitations, such as long latency, large bandwidth utilization, and high power consumption, resulting in high communication overhead. Moreover, once the central server fails, the entire training process will be seriously affected.

[0004] With the advancement of communication technology, the cost of inter-satellite communication has been greatly reduced, and high-reliability, low-latency inter-satellite connections have been achieved. This makes it possible to replace as much communication between satellites and servers as possible with inter-satellite communication, thereby significantly reducing the overall communication cost of the system. Therefore, under the premise of ensuring data privacy, combining inter-satellite communication technology to reduce the communication cost between servers and satellites and realize joint training models has become a key challenge. Summary of the invention

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

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

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

[0008] S1: Collect SAR image data and preprocess it to generate corresponding target recognition labels to obtain labeled data sets;

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

[0010] S3: allocating local private data to each client according to the labeled data set, each client using the local private data to perform model training, obtaining a local parameter consensus through inter-device communication, and preparing to upload the updated local model parameters to the server;

[0011] S4: The server randomly selects clients for parameter transmission, aggregates model parameters and updates 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 a convergence state, and obtain a trained global model;

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

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

[0015] S101: Collect SAR image data through satellite mechanism;

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

[0017] S103: Generate corresponding target recognition labels according to 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; the clients communicate with each other, and the server communicates with each client via a 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 gradually extracts high-level features from the input SAR image 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 data set allocated based on the labeled data set. Each client performs local training based on the allocated global model parameters and optimizes the local model parameters to minimize the local cross entropy loss function. The optimization calculation expression of the local model parameters is:

[0027]

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

[0029] S302: Each client communicates with other clients in its cluster (k) Rounds of D2D communication are performed to iteratively update local parameters. In the kth iteration, client n exchanges parameters with neighbor client m. The update expression is:

[0030]

[0031] In the formula, is the local parameter of the client n in round t+1 after update, is the consensus parameter between client n and neighbor client m in the kth iteration, is the consensus parameter of client n itself in the kth iteration, is the local parameter of client n in round t, ζ (k) (n) is the neighbor set of client n in the kth iteration, is the local parameter of client m in round t;

[0032] S303: Each client aggregates its local consensus parameters Prepare to upload to the server.

[0033] Furthermore, in step S303, the optimized client model parameter aggregation expression is used to update the parameters, and the optimized client model parameter aggregation expression is:

[0034]

[0035] Where, d (k) The maximum number of client neighbors in the cluster at the kth 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 at the current iteration number k from the selected client m

[0039] S403: The server receives the local parameter consensus As a new global model parameter The global model parameters The calculation expression is:

[0040]

[0041] In the formula, is the global model parameter after the kth round update;

[0042] S404: The server updates the global model parameters Broadcast to all clients, the update expression of local parameters of each client is:

[0043]

[0044] In the formula, are the model parameters of client n at the beginning of the k+1th round.

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

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

[0047] S502: Each client uses a local data set to perform model training to minimize a local loss function, where the calculation expression of the local loss function is:

[0048]

[0049] Where, L n (θ) is the local loss function of client n, D n is the local data set, l is the cross entropy loss function, f(x i ; θ) is the predicted value of the model, y i Identify labels for targets;

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

[0051]

[0052] In the formula, are the local model parameters of client n at the kth iteration, is the local model parameter of client n at the k+1th iteration, η is the learning rate, is the gradient of the local loss function in the kth round;

[0053] S504: Clients communicate with each other.

[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 and prepares to upload it to the server;

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

[0057] S508: The server determines whether the change of the global model parameters is less than a preset threshold or whether the number of iterations reaches the maximum number of iterations. If so, it is considered that the global model parameters 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 a test data set;

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

[0061] S603: Infer the global model in the test data set to obtain a prediction result, the calculation expression of which is:

[0062]

[0063] In the formula, To predict the results, is the global model parameter, x test is the test data, f(·) is the model function;

[0064] S604: Compare the prediction result with the true label y of the test data set test For comparison, the verification accuracy Acc of the model is calculated. The calculation expression of the accuracy Acc is:

[0065]

[0066] Where D test is the test data set, 1() is the indicator function, which takes the value of 1 when the predicted result is the same as the true label, otherwise it takes the value of 0. is the i-th prediction result, y test,i is the i-th true label y test ;

[0067] S605: Evaluate model performance according to the verification accuracy Acc of the model;

[0068] S607: Optimize and improve the global model according to the model performance evaluation results.

[0069] Furthermore, step S605 further includes:

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

[0071] Record the verification accuracy Acc of the global model at different communication costs and in each communication round round , used to evaluate the change and stability of model accuracy during training;

[0072] Record the global model verification accuracy Acc after completing the maximum communication round final , used to measure the overall effectiveness and performance of the training process for different numbers of clients.

[0073] This paper proposes a multi-client collaborative synthetic aperture radar (SAR) image target recognition method based on a semi-decentralized architecture, aiming to solve the challenge of high communication overhead between servers and satellites in the FL architecture. The traditional FL architecture adopts a star topology, and all clients need to communicate frequently with the central server, resulting in high latency, high bandwidth utilization, and high power consumption. At the same time, once the central server fails, the entire training process will be seriously affected.

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

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

[0076] The innovation lies in the first application of semi-decentralized architecture to multi-client collaborative SAR target recognition tasks, which effectively solves the problem of high communication overhead in the traditional FL architecture. By reducing the frequency of direct communication between the server and the client 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 paper significantly reduces the communication overhead while ensuring model performance and data privacy protection, and is superior to the performance of the traditional centralized architecture. The specific performance is as follows: In the semi-decentralized 10 and 20 client collaborative training tasks, through iterative training, the verification accuracy can reach 97.69% and 96.07% respectively.

[0078] Therefore, the present invention is of great significance in solving the problem of high communication overhead in FL architecture. Under the premise of ensuring data privacy, the model performance and robustness are effectively improved. The method has broad application prospects in military, civilian and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 A schematic diagram of a flow chart of 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 Schematic diagram of vehicle optical and SAR data in an MSTAR data set 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 A schematic diagram of a semi-decentralized multi-client collaborative SAR image target recognition framework structure provided in an embodiment of the present invention;

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

[0083] Figure 5 A schematic diagram of training loss results of 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 A schematic diagram of verification accuracy of collaborative tasks with different numbers of clients as communication overhead varies in a semi-decentralized architecture (SDMC-SAR) and a traditional centralized architecture (star topology) provided in an embodiment of the present invention;

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

[0086] Figure 8 A schematic diagram of the final verification accuracy of collaborative tasks with different numbers of customers in a semi-decentralized framework provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0088] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0089] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[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 is used to realize the joint training of models by multiple satellite organizations without sharing SAR data, and reduce the communication overhead between the server and the client under the premise of ensuring model performance and data privacy protection, including the following steps:

[0092] S1: SAR image data acquisition and preprocessing: SAR image data is acquired and preprocessed to generate corresponding target recognition labels to obtain a labeled data set. This embodiment further divides the labeled data set 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 multiple clients of the server and cluster, and the clients in the cluster communicate with each other; build a neural network model for target recognition and initialize global model parameters for collaborative training;

[0094] S3: Local model training and updating: local private data is allocated to each client according to the labeled data set, each client uses the local private data for model training, 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 for parameter transmission, 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 converged state, and obtain a trained global model;

[0097] S6: Model testing and performance evaluation: Use the test data set to test the trained global model, obtain the verification accuracy, and evaluate the overall performance of the model. The global model that passes the test is selected for SAR image target recognition.

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

[0099] SAR image data acquisition and preprocessing in step S1: SAR image data is acquired, preprocessed, labels are generated and the data set is divided into a training set and a test set.

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

[0101] Step S101: Data collection: The satellite mechanism collects SAR image data.

[0102] Step S102: Data preprocessing: Perform preprocessing operations such as denoising, correction, and enhancement on the collected SAR image data to improve data quality.

[0103] Step S103: Generate labels. Generate corresponding target identification labels based on SAR image data. Each satellite structure n forms a local data set Where |D (n) | is the size of the dataset, and Represent the input SAR image and the corresponding label respectively.

[0104] Step S104: Divide 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 that contains SAR images of 10 types of military vehicles. Figure 2 The schematic diagram of optical and SAR images corresponding to various types of vehicles in the MSTAR dataset is shown. Table 1 shows the number of samples corresponding to training and testing for the 10 types of targets in the MSTAR dataset.

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

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

[0107] Table 1 Number of training and test samples corresponding to 10 types of targets 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, it includes: determining the client U participating in the collaborative training fl , the number of clients N and the central server S flThe example of multi-client collaborative SAR target recognition in a semi-decentralized architecture simulates configuration scenarios of 2, 4, 6, 8, 10, 15 and 20 clients (i.e., N = 2, 4, 6, 8, 10, 15, 20), and compares the semi-decentralized distributed training results with the centralized distributed training.

[0112] Figure 3 Schematic diagram of a semi-decentralized multi-client collaborative SAR image target recognition framework. Specifically, the system model is described as follows: The semi-decentralized architecture consists of multiple satellites forming a cluster, each of which is capable of device-to-device (D2D) communication. The master server is located on the ground, communicates with the cluster via satellite links, and is 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 path between the satellite and the server.

[0113] In this semi-decentralized framework, SAR data is not shared between satellites, and satellites use D2D links to reach local parameter consensus. The master server communicates with randomly selected satellites through parameter transmission to implement a joint training model. This semi-decentralized architecture significantly reduces the communication cost between satellites and servers while ensuring data privacy.

[0114] Step S202: Initialize global model parameters, set model parameters such as initial weights and learning rates. Specifically, it includes: defining the architecture of the model; initializing global model parameters θ according to the preset neural network structure; glb , the global model parameters after initialization are Set the learning rate η and other related hyperparameters, such as batch size B, number of training rounds E, etc.

[0115] In order to effectively perform SAR target recognition, a neural network model is designed in this example. Figure 4 As shown in Figure 2, 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 batch normalization, ReLU activation, and pooling layers. This layered approach ensures that the model can capture the spatial and hierarchical characteristics of the input data. Specifically, the backbone network consists of four convolutional modules. After the feature extraction is completed, the fully connected layer acts as a classifier. The fully connected layer maps the extracted features to the final output category. Specifically, the fully connected layer consists of two fully connected 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 logits into probabilities for multi-class classification.

[0117] In addition, in the example, the specific hyperparameters are set as follows: batch size is 100, maximum communication rounds is 1500, and learning rate is 3×10 -4 , and the optimizer uses AdamW. All experiments are performed on a single RTX 3090 GPU. The global model parameters are initialized through the above steps to ensure the basic performance of the model and provide a basis for subsequent local training and global aggregation.

[0118] Step S203: The server distributes initial global model parameters. fl The global model parameters that will be initialized Distribute to each client U fl .

[0119] Step S3, local model training and updating: Each client uses local private data for model training, 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 local private data to train the model. Each client (n) holds a local data set D (n) , including the input SAR image and the corresponding labels In each iteration k, each client uses 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 the kth round of D2D communication, the local parameters are iteratively updated. In the kth round of iteration, client n exchanges parameters with its neighbor client m through the following formula To update:

[0124]

[0125] in, is the consensus parameter between client n and neighbor client m in the kth iteration, ζ (k) (n) is the neighbor set of client n in the kth iteration. In order to better converge, we use the following formula

[0126] Aggregate local model parameters:

[0127]

[0128] Among them, 0 <d (k) <1 / D (k) , D (k) The maximum number of client neighbors in the cluster at the kth iteration.

[0129] Step S303: Prepare to upload the updated model parameters to the server. (k) After the D2D communication is completed, the selected client is ready to aggregate its local consensus parameters Upload 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: receiving 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 updates the received local model parameters. As a new global model parameter The update formula is as follows:

[0135]

[0136] in, Represents the global model parameters after the kth round of update.

[0137] Step S404: Distribute global model parameters. The server distributes the updated global model parameters to Broadcast to all clients to prepare for the next round of local model training, specifically:

[0138]

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

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

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

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

[0143]

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

[0145] Step S502: Local model training. Client n uses local private data to perform model training and minimize the local loss function:

[0146]

[0147] Among them, l represents the cross entropy loss function, f(x i ; θ) represents the predicted value of the model, y i represents the true label.

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

[0149]

[0150] Among them, η represents the learning rate, Represents the gradient of the local loss function at the kth round.

[0151] Step S504: Device-to-device (D2D) communication. The clients perform several rounds of iterations through D2D communication to reach a local parameter consensus:

[0152]

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

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

[0155]

[0156] Among them, t (k) Indicates the number of D2D communication iterations in the kth round.

[0157] Step S506: Upload the local parameter consensus. The client uploads the updated local parameter consensus Upload 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 state. If the global model parameter θ glb If the change in is less than a preset threshold or reaches the maximum number of iterations, the model is considered to have converged, otherwise steps S5-1 to S5-7 are repeated.

[0162] Step S6, model testing and performance evaluation: Use the test data set to verify the trained global model, obtain the verification accuracy, and evaluate the overall performance of the model.

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

[0164] Step S601: Prepare a test data set. Select and prepare an independent test data set D separated from the training data set. test to ensure the objectivity of the test results.

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

[0166] Step S603: Perform model inference on the test data set. te Infer the global model and obtain the model's prediction results

[0167]

[0168] Among them, x test is the test data, and f(·) is the model function.

[0169] Step S604: Calculate the verification accuracy. and the true label y of the test dataset test For comparison, calculate the model's verification accuracy Acc to evaluate the model's recognition performance:

[0170]

[0171] Among them, 1() is an indicator function, which takes the value of 1 when the predicted result is the same as the true label, otherwise it is 0.

[0172] Step S605: Evaluate model performance. In addition to verifying the accuracy, the model can also be comprehensively evaluated by calculating other performance indicators to analyze the convergence speed and stability of the model in practical applications. The details are as follows:

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

[0174]

[0175] Figure 5 The training loss of different numbers of client collaborative tasks in semi-decentralized and centralized frameworks is shown. The horizontal axis represents the communication rounds (com.rnds.), and the vertical axis represents the training loss (training loss). The blue curve represents the traditional star topology federated learning (Star topology), and the orange curve represents the semi-decentralized architecture (SDMC-SAR) proposed in this invention.

[0176] from Figure 5 It can be seen that although the traditional star topology has a faster decline in training loss in the early stage, the semi-decentralized architecture (SDMC-SAR) proposed in the present invention can also reach a similar convergence level in the end. This shows that although the traditional method converges faster in the early stage, both methods can converge to a lower training loss in the end.

[0177] At the same time, the semi-decentralized architecture still maintains a relatively stable training loss in most cases, which shows its stability under large-scale client configuration. Specifically, in the collaborative training tasks of different numbers of clients, the final convergence effect of the semi-decentralized architecture is close to that of the traditional star topology, which verifies its ability to maintain model performance.

[0178] 2. Record the verification accuracy Acc of the global model in different communication overheads and each communication round round , helps understand the changes and stability of model accuracy during training;

[0179] Figure 6 The verification accuracy of the semi-decentralized architecture (SDMC-SAR) and the traditional centralized architecture (star topology) under different client number configurations is shown. The figure shows that under the same overhead conditions, the SDMC-SAR architecture has higher verification accuracy under all client number configurations. This shows that SDMC-SAR can use communication resources more efficiently and improve the overall performance of the model while reducing communication overhead.

[0180] Figure 7 The verification accuracy of the semi-decentralized architecture (SDMC-SAR) and the traditional centralized architecture (star topology) under different client number configurations is shown. It can be seen that although in the early stage of the change of communication rounds, the verification accuracy of the star topology increases rapidly due to the centralized parameter aggregation method. However, as the communication rounds increase, the SDMC-SAR architecture gradually improves the verification accuracy through effective inter-device communication and local parameter consensus, and gradually achieves performance that is basically equivalent to that of the star topology.

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

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

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

[0184]

[0185] Step S606: Optimize the model. According to the performance evaluation results, optimize and improve the model, such as adjusting hyperparameters, improving model structure, and optimizing training strategies. Analyze the effectiveness and reliability of the model, so as to further improve its performance and generalization ability, and determine its applicability in practical applications.

[0186] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture, characterized in that: The following steps are involved: S1: Collect SAR image data and preprocess it to generate corresponding target recognition labels to obtain labeled data sets; S2: Build a semi-decentralized federated learning framework, which includes multiple clients in the server and cluster, and the clients in the cluster communicate with each other; build a neural network model for target recognition, initialize and distribute global model parameters to each client; S3: allocating local private data to each client according to the labeled data set, each client using the local private data to perform model training, obtaining a local parameter consensus through inter-device communication, and preparing to upload the updated local model parameters to the server; S4: The server randomly selects clients for parameter transmission, aggregates model parameters and updates 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 a convergence state, and obtain a trained global model; S6: Test the trained global model and select the global model that has passed the test for SAR image target recognition.

2. According to the method of claim 1, a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture is characterized in that: Step S1 specifically includes the following steps: S101: Collect SAR image data through satellite mechanism; S102: preprocessing the collected SAR image data, the preprocessing including denoising, correction and enhancement; S103: Generate corresponding target recognition labels according to the preprocessed SAR image data.

3. According to the method of multi-client collaborative SAR image target recognition based on semi-decentralized architecture according to claim 1, it is characterized in that: 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; the clients communicate with each other, and the server communicates with each client via a 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.

4. According to a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture as described in claim 1, it is characterized in that: The neural network model for target recognition includes a backbone network for feature extraction and a classifier composed of fully connected layers for classification; The backbone network gradually extracts high-level features from the input SAR image through a series of convolutional layers; The classifier maps the high-level features extracted by the backbone network to the final output category.

5. According to a method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture as described in claim 1, it is characterized in that: Step S3 specifically includes the following steps: S301: Each client holds a local data set allocated based on the labeled data set. Each client performs local training based on the allocated global model parameters and optimizes the local model parameters to minimize the local cross entropy loss function. The optimization calculation expression of the local model parameters is: In the formula, are the local model parameters of client n at the kth iteration, is the i-th SAR image of client n, Identify the label of the i-th target of client n, D (n) For local datasets; S302: Each client communicates with other clients in its cluster (k) Rounds of D2D communication are performed to iteratively update local parameters. In the kth iteration, client n exchanges parameters with neighbor client m. The update expression is: In the formula, is the local parameter of the client n in round t+1 after update, is the consensus parameter between client n and neighbor client m in the kth iteration, is the consensus parameter of client n itself in the kth iteration, is the local parameter of client n in round t, ζ (k) (n) is the neighbor set of client n in the kth iteration, is the local parameter of client m in round t; S303: Each client aggregates its local consensus parameters Prepare to upload to the server.

6. According to claim 5, a multi-client collaborative SAR image target recognition method based on a semi-decentralized architecture is characterized in that: In step S303, the optimized client model parameter aggregation expression is used to update the parameters. The optimized client model parameter aggregation expression is: Where, d (k) The maximum number of client neighbors in the cluster at the kth iteration.

7. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1 is 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 at the current iteration number k from the selected client m S403: The server receives the local parameter consensus As a new global model parameter The global model parameters The calculation expression is: In the formula, is the global model parameter after the k-th round update; S404: The server updates the global model parameters Broadcast to all clients, the update expression of local parameters of each client is: In the formula, are the model parameters of client n at the beginning of the k+1th round.

8. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1 is characterized in that: Step S5 specifically includes the following steps: S501: Each client receives the global model parameters sent by the server and updates the local model parameters; S502: Each client uses a local data set to perform model training to minimize a local loss function, where the calculation expression of the local loss function is: Where, L n (θ) is the local loss function of client n, D n is the local data set, l is the cross entropy loss function, f(x i ; θ) is the predicted value of the model, y i Identify labels for targets; S503: Each client updates the local model parameters using the gradient descent algorithm. The corresponding update expression is: In the formula, are the local model parameters of client n at the kth iteration, is the local model parameter of client n at the k+1th iteration, η is the learning rate, is the gradient of the local loss function in the kth round; S504: Clients communicate with each other. S505: The client obtains local parameter consensus through multiple rounds of inter-device communication iterations; S506: The client obtains the final local parameter consensus and prepares to upload it to the server; S507: The server receives local parameter consensus from the selected client and updates the global model parameters: S508: The server determines whether the change of the global model parameters is less than a preset threshold or whether the number of iterations reaches the maximum number of iterations. If so, it is considered that the global model parameters have reached a convergence state, otherwise, steps S501-S507 are repeated.

9. The method for multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 1 is characterized in that: Step S6 specifically includes the following steps: S601: Select and prepare a test data set; S602: Load the trained global model parameters into the test environment; S603: Infer the global model in the test data set to obtain a prediction result, the calculation expression of which is: In the formula, To predict the results, is the global model parameter, x test is the test data, f(·) is the model function; S604: Compare the prediction result with the true label y of the test data set test For comparison, the verification accuracy Acc of the model is calculated. The calculation expression of the accuracy Acc is: Where D test is the test data set, 1() is the indicator function, which takes the value of 1 when the predicted result is the same as the true label, otherwise it takes the value of 0. is the i-th prediction result, y test,i is the i-th true label y test ; S605: Evaluate model performance according to the verification accuracy Acc of the model; S607: Optimize and improve the global model according to the model performance evaluation results.

10. The method of multi-client collaborative SAR image target recognition based on a semi-decentralized architecture according to claim 9 is characterized in that: Step S605 also includes: Record the training loss L under different number of clients train , used to evaluate the learning process of the global model and the convergence speed under different client configurations; Record the verification accuracy Acc of the global model at different communication costs and in each communication round rpund , used to evaluate the change and stability of model accuracy during training; Record the global model verification accuracy Acc after completing the maximum communication round final , used to measure the overall effectiveness and performance of the training process for different numbers of clients.

Images