Satellite-ground network adaptive image semantic communication method, system and device based on NOMA technology, and storage medium

Abstract

The invention relates to a satellite-to-ground network adaptive image semantic communication method based on an NOMA technology, and the method comprises the following steps: S1, building a satellite-to-ground semantic communication frame integrated with the NOMA technology: carrying out the power superposition and serial interference elimination of different user transmission contents through the NOMA technology, a coding and decoding process is optimized through deep joint source channel coding based on hierarchical vision using a shift window; s2, establishing an adaptive semantic communication model: dynamically adjusting coding parameters based on a target bandwidth ratio and a signal-to-noise ratio of a satellite-ground channel; s3, proposing a dynamic weight adjustment algorithm: optimizing model convergence speed and transmission performance tradeoff under different channel conditions; and S4, designing a feature selection module: performing priority ranking on the image features, and selectively transmitting important features according to bandwidth resources. According to the method, the adaptive semantic communication model (ASeC-NOMA) which can be dynamically adjusted according to the target bandwidth ratio and the signal-to-noise ratio is constructed, and a large number of model storage parameters are saved while the transmission performance is ensured.

Description

Technical Field

[0001] The present invention belongs to the field of information and communication technology, and specifically relates to a satellite-to-ground network adaptive image semantic communication method, system, device and storage medium based on NOMA technology. Background Art

[0002] With the application of millimeter wave and massive multi-input multi-output technologies, 5G has entered the large-scale commercial stage. In order to realize the vision of intelligent connection of all things, industry and academia have begun to explore 6G. In the future 6G network, the satellite-ground fusion network has become an indispensable and important part of the 6G network with its full-range seamless coverage. However, the satellite-ground channel environment is extremely harsh, with multiple interference problems such as fading, noise, and data loss, which greatly affect the reliability of communication. At the same time, predictions show that the number of IoT devices in the world may reach tens of billions or even hundreds of billions in the future, which will lead to an exponential increase in the amount of data flowing in the combined network, and energy consumption will also increase. However, the spectrum resources in space are limited and expensive. Faced with the challenge of large-scale user access, how to improve spectrum utilization and achieve parallel access of multiple users without significantly reducing transmission quality is a key issue that needs to be considered in satellite-ground link transmission.

[0003] Due to the limited computing power of devices in traditional communication architectures and limited bandwidth resources, the data that can be transmitted by existing communication technologies has approached the capacity limit of the Shannon physical layer, which has become a bottleneck restricting the development of 6G technology. To this end, academia and industry have tried to use AI to empower networks and have carried out various studies on intelligent networks. Among them, the intelligent network semantic communication paradigm has become the key to breaking through the current difficulties faced by satellite-ground fusion networks. The encoding end of this paradigm can use deep neural networks to extract semantic information of the transmitted content, and the decoding end can restore semantic information and reconstruct the transmitted content with the help of the knowledge base built during the training process to achieve the purpose of efficient transmission. For image transmission tasks, the current mainstream technology on the encoding and decoding end is deep joint source channel coding based on CNN and its variants. However, the model capacity of CNN is limited. When the image resolution increases, CNN-based JSCC is difficult to effectively learn hierarchical semantic features and image details, which leads to a significant decrease in semantic transmission performance.

[0004] In addition, during the communication process, especially in the satellite-to-ground link, the channel state and transmission rate will change dynamically. Existing semantic communication models are usually trained for fixed signal-to-noise ratios and bandwidth ratios. In order to adapt to the channel environment, once the channel conditions change, the model needs to be retrained. This will cause the codec to store multiple model parameters under different channel conditions, which poses a major challenge to edge devices with limited storage resources. Summary of the invention

[0005] In view of the problems of low transmission quality caused by the harsh existing satellite-ground network communication environment, shortage of spectrum resources caused by large-scale terminal user access, and the inability of the existing semantic communication model to adapt to the channel environment, the present invention proposes a satellite-ground network adaptive image semantic communication method and system based on NOMA technology.

[0006] A satellite-ground network adaptive image semantic communication method based on NOMA technology of the present invention includes the following steps:

[0007] S1. Establish a satellite-ground semantic communication framework integrating NOMA technology: Use NOMA technology to perform power superposition and successive interference cancellation on the transmission content of different users, and optimize the encoding and decoding processes through deep joint source-channel coding based on the hierarchical vision Transformer (Swin Transformer) using a shifted window;

[0008] S2. Establish an adaptive semantic communication model: Dynamically adjust the encoding parameters based on the target bandwidth ratio and signal-to-noise ratio of the satellite-ground channel;

[0009] S3. Propose a dynamic weight adjustment algorithm: Optimize the trade-off between the model convergence speed and the transmission performance under different channel conditions;

[0010] S4. Design a feature selection module: Sort the image features by priority and selectively transmit important features according to the bandwidth resources.

[0011] Further, in S1, the satellite-ground semantic communication framework specifically includes:

[0012] S11. Optimize the spectral efficiency

[0013] Allow multiple users to access simultaneously in the same frequency band through NOMA technology, and use the power allocation strategy to achieve the superposition and separation of multi-user signals; Combine the power-domain non-orthogonal multiple access technology, and use the successive interference cancellation module to achieve the separation and decoding of multi-user signals;

[0014] S12. Source-channel joint coding based on Swin Transformer

[0015] The input image is first divided into non-overlapping image patches, and each image patch is projected into a feature space of a specific dimension through a linear embedding layer; By introducing the channel state information, this information is concatenated with the image features to form a joint representation, and is input into a multi-layer encoder based on Swin Transformer; During the encoding process, the Swin Transformer module uses the window-based multi-head self-attention mechanism and the shifted window multi-head self-attention mechanism to extract global and local features, so as to generate a feature matrix adapted to the channel conditions;

[0016] The encoded signal passes through a power superposition module to allocate power to multiple user signals and generate a transmission signal:

[0017]

[0018] Among them, y represents the signal to be sent by the sending end, and z j represents the signal formed by the j-th user after passing through the encoder, and P j represents the transmission power of the j-th user; after the signals are superposed, they are then transmitted to the receiving end through the satellite-ground channel;

[0019] S13. Transmission channel and noise processing:

[0020] The transmission signal y is transmitted through the analog channel, and the obtained signal is

[0021] S14. Receiving end signal processing:

[0022] The received signal passes through a successive interference cancellation module, and each user signal is separated in order from high to low according to power. The remaining signal is updated by iteratively canceling interference, and finally the decoded input signal of a single user is obtained; the separated signal is input into a decoder based on Swin Transformer, and the decoder reconstructs the image signal according to the channel state information and feature selection mask.

[0023] Furthermore, in S2, the adaptive semantic communication model includes the following steps:

[0024] S21. Target bandwidth ratio and signal-to-noise ratio modeling

[0025] At the input end, the target bandwidth ratio ρ and signal-to-noise ratio SNR of the satellite-ground channel are input as link state information, and they are mapped into an auxiliary coding feature tensor u = MLP([SNR, ρ]) through a fully connected layer; this auxiliary information is concatenated with the image features and then input into the encoder to dynamically adjust the coding process and adapt to different channel conditions;

[0026] S22. Adaptive feature extraction at the encoding end

[0027] Perform image block division and embedding: The input image S j is divided into non-overlapping blocks and projected into a feature space of a specific dimension through a linear embedding layer; the generated tensor X j is concatenated with the link state information u to guide the subsequent feature extraction process;

[0028] Pass through a multi - layer encoder based on Swin Transformer: The encoder uses Swin Transformer modules to extract global and local features of the image layer by layer; the module combines link state information through window multi - head self - attention and shifted window multi - head self - attention mechanisms to dynamically optimize feature extraction, making the encoding result adapt to the current channel conditions;

[0029] Output features and transmission: The final encoding result is adjusted to a feature matrix Y that adapts to the transmission bandwidth j , and important features are selected through the feature selection mask M to generate the transmission signal z j = f θ (ρ, SNR, S j );

[0030] S23. Non - orthogonal multiple access mechanism

[0031] Adopt power - domain NOMA technology to achieve non - orthogonal transmission of multiple user signals; at the encoding end, according to the power allocation coefficient P of the user j superimpose the signals in power to generate the total transmission signal: This signal is transmitted through the satellite - to - ground channel and is affected by channel noise interference.

[0032] Furthermore, in S2, the adaptive semantic communication model further includes the following steps:

[0033] S24. Adaptive reconstruction at the decoding end

[0034] Successive interference cancellation (SIC) separates signals: The receiving end separates each user signal from high to low power through a successive interference cancellation module; for the user signal with the highest current power, the following steps are taken:

[0035] 1) Decode the user signal:

[0036] 2) Update the remaining signals:

[0037] 3) Repeat the above steps until all user signals are separated;

[0038] Decoder based on Swin Transformer: The separated user signals are input into the decoder, and the decoder reconstructs the target image layer by layer by splicing the link state information u and the transmission features

[0039] S25. Adaptive transmission optimization mechanism

[0040] Bandwidth adaptation: By dynamically adjusting the bandwidth ratio ρ, limit the number of features N actually transmitted F, only transmit the most important features;

[0041] Signal-to-noise ratio adaptation: Under low signal-to-noise ratio conditions, improve transmission reliability by increasing redundant information; under high signal-to-noise ratio conditions, reduce redundant information to enhance transmission efficiency;

[0042] Dynamic warrant adjustment algorithm: Introduce a dynamic weight allocation mechanism during training, and dynamically adjust the weights according to the image reconstruction loss under different bandwidth ratios or signal-to-noise ratio conditions to balance the performance of the model under different channel conditions.

[0043] Furthermore, in S3, the dynamic weight adjustment algorithm includes the following steps:

[0044] S31. Core objective of dynamic weight adjustment

[0045] To balance the performance of the model under different channel conditions, NDWA dynamically allocates training weights to avoid the model from sacrificing performance under high bandwidth ratio or high signal-to-noise ratio conditions due to optimizing low bandwidth ratio or low signal-to-noise ratio conditions; improve the convergence speed of the model through the weight dynamic adjustment mechanism to make it adapt to various channel environments;

[0046] S32. Initialization of dynamic weights

[0047] Allocate initial weights for each channel condition where w represents the set of discrete values of the bandwidth ratio or signal-to-noise ratio; the initial weights are equal to ensure that the model evenly optimizes all channel conditions at the initial stage of training;

[0048] S33. Weight update mechanism during the training phase

[0049] In each training cycle, for each bandwidth ratio ρ w or signal-to-noise ratio SNR, calculate the image reconstruction quality index; by comparing with the optimal model reconstruction performance under fixed conditions, calculate finally, the performance gap of the current model: where is the optimal PSNR under fixed conditions, is the PSNR of the current model at the t-th round of training;

[0050] According to the performance gap dynamically adjust the weights, and the weight update formula is:

[0051]

[0052] where α is the scaling coefficient, k is the balance parameter,; the weights are restricted to the range [0, 10], that is, when the weight is less than 0, it is set to 0, and when it is greater than 10, it is set to 10 to avoid unstable training;

[0053] Normalize the weights for all channel conditions to ensure that the sum of the total weights is 1:

[0054]

[0055] where W is the total number of bandwidth ratios or signal-to-noise ratios.

[0056] Furthermore, in S3, the dynamic weight adjustment algorithm further includes the following steps:

[0057] S34. Sampling strategy for bandwidth and signal-to-noise ratio

[0058] During the training process, the bandwidth ratio ρ w and the signal-to-noise ratio SNR are uniformly sampled from a fixed range; the sampled ρ w and SNR are input into the adaptive semantic communication model together with the input image, and the reconstructed image is obtained through model calculation;

[0059] S35. Weighted optimization of the loss function

[0060] In each training, the loss values under different conditions are weighted and summed, and the optimization objective is the weighted mean square error:

[0061]

[0062] where is the dynamic weight in the t-th round of training, and K represents the number of image blocks and also the number of users;

[0063] S36. Model performance evaluation and weight update

[0064] In the validation stage of training, calculate the performance difference according to the reconstruction performance of the test set images under different conditions and update the weights according to the weight adjustment rule to ensure that the model is evenly optimized under all conditions;

[0065] S37. Weight constraint and convergence control

[0066] By imposing upper and lower bounds on the weights, ensure that the weights do not decrease or increase infinitely under specific conditions;

[0067] S38. Flow of the weight adjustment algorithm

[0068] First, initialize the initial weights corresponding to different bandwidth ratios. During the training process, the algorithm iterates through multiple training rounds. In each round of training, the data is divided into several batches for processing; in each batch, first randomly sample the bandwidth ratio and signal-to-noise ratio to simulate the fluctuations in the actual communication environment; subsequently, input the image data and the sampled communication conditions into the model, generate the corresponding reconstructed image, and calculate the loss function of the current batch to measure the error of the reconstruction quality.

[0069] After each round of training, the algorithm enters the verification phase, evaluates the reconstruction performance differences under different bandwidth conditions, and dynamically adjusts the weights based on this feedback, enabling the model to better adapt to different bandwidth environments; after multiple rounds of training and weight optimization, the trained model parameters and the optimized dynamic weight allocation strategy are finally obtained.

[0070] Further, in S4, the feature selection module includes the following parts:

[0071] S41. Importance evaluation and ranking

[0072] The FSM evaluates the importance of the feature tensors extracted at the encoding end through global average pooling, sorts them in descending order according to the importance scores of the feature channels, and determines the feature priorities; the feature channels with high scores are preferentially selected for transmission to ensure the transmission efficiency of semantic information.

[0073] S42. Dynamic feature selection

[0074] According to the target bandwidth ratio and signal-to-noise ratio, the FSM dynamically adjusts the number of features to be transmitted, generates a feature selection mask, and only transmits the top-ranked important feature channels. The unselected channels are masked to reduce the transmission of invalid data.

[0075] S43. Channel adaptation and decoding reconstruction

[0076] At the receiving end, the feature tensors are restored using the same feature selection mask as at the sending end, and the important features are reconstructed through a decoder to finally generate a high-quality image reconstruction result. Under limited bandwidth and signal-to-noise ratio conditions, the FSM preferentially transmits features with a higher semantic concentration.

[0077] The present invention also relates to a satellite-ground network adaptive image semantic communication system based on NOMA technology, and the system includes a computer module that utilizes the above satellite-ground fusion network multi-objective spatio-temporal switching decision method.

[0078] The present invention also relates to a computer device, including a memory and a processor. When the processor executes the computer program, the steps of the satellite-ground network adaptive image semantic communication method based on NOMA technology are implemented.

[0079] The present invention also relates to a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the satellite-ground network adaptive image semantic communication method based on NOMA technology are implemented.

[0080] Beneficial effects

[0081] The present invention proposes a space-ground network semantic communication framework (SeC-NOMA) integrating NOMA technology. Its encoding and decoding ends perform joint source-channel coding based on the Swin Transformer model, solving the problems of high model complexity and poor image reconstruction effect existing in deep JSCC based on CNN. At the same time, the NOMA technology applied in SeC-NOMA improves the spectrum efficiency utilization rate, providing a solution idea for the scenario of large-scale access of users in the space-ground integrated network. The present invention proposes an adaptive scheme and constructs an adaptive semantic communication model (ASeC-NOMA) that can be dynamically adjusted according to the target bandwidth ratio and signal-to-noise ratio, saving a large number of model storage parameters while ensuring transmission performance. At the same time, in order to balance the reconstruction quality of different bandwidth ratios or signal-to-noise ratios and solve the problem of too slow convergence speed caused by taking the bandwidth ratio and signal-to-noise ratio as training parameters together, the present invention also proposes a dynamic weight training strategy (NDWA), ensuring the balance of transmission performance under different channel conditions while accelerating the training speed.

[0082] In addition, aiming at the problem that the space-ground link has poor performance compared with other channels due to complex interference, the present invention proposes a feature selection module (FSM) to screen transmission features according to feature importance to improve the image reconstruction effect. Through a large number of simulation experiments, the present invention first verifies that the model complexity based on the Swin Transfomer is significantly lower than that based on CNN. Secondly, it verifies that the proposed adaptive model under the space-ground link and other various channels can be comparable to the model trained under specific conditions, and in some cases, the PSNR and SSIM can be 2dB and 0.2 higher respectively. Thirdly, it verifies the significant improvement of the feature ranking module (FSM) on the image reconstruction performance, and can even improve the PSNR and SSIM by more than 5dB and 0.1. Finally, it verifies that the ASeC-NOMA model has a higher spectrum efficiency utilization rate than the model using OMA technology in communication, and is a feasible solution for the scenario of large-scale user access. Brief Description of the Drawings

[0083] Figure 1 It is the architecture diagram of SeC-NOMA with multi-user simultaneous access in the space-ground network of the present invention.

[0084] Figure 2 It is the detailed framework diagram of the semantic communication system ASeC-NOMA of the present invention.

[0085] Figure 3 It is the design diagram of the adaptive scheme of the present invention.

[0086] Figure 4 It is the viewable diagram of the self-attention mechanism complexity of the present invention.

[0087] Figure 5This is a comparison chart of the performance of ASeC-NOMA and other models in the Gaussian channel of the present invention.

[0088] Figure 6 This is a schematic diagram of the performance of ASeC-NOMA (BW-NDWA) and ASeC-NOMA (BW-NDWA-FSM) in the satellite-ground channel of the present invention under different bandwidth ratios.

[0089] Figure 7 This is a schematic diagram of the performance of ASeC-NOMA (SNR-NDWA) and ASeC-NOMA (SNR-NDWA-FSM) in the satellite-ground channel of the present invention under different signal-to-noise ratios.

[0090] Figure 8 This is a viewable diagram before and after image transmission under different conditions of the satellite-ground channel of the present invention.

[0091] Figure 9 This is a schematic diagram of the performance when three users in the satellite-ground channel use ASeC-NOMA (BW-NDWA) and ASeC-OMA (BW-NDWA) for transmission in the present invention.

[0092] Figure 10 This is a schematic diagram of the performance when three users in the satellite-ground channel use ASeC-NOMA (SNR-NDWA) and ASeC-OMA (SNR-NDWA) for transmission in the present invention.

[0093] Figure 11 This is a schematic diagram of the comparison of the spectral efficiency of OMA and NOMA in ASeC-NOMA (BW-NDWA) and ASeC-NOMA (SNR-NDWA) of the present invention. Detailed implementation manners

[0094] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but it is not limited to the present invention.

[0095] The adaptive image semantic communication method integrating NOMA technology in the satellite-ground network of the present invention includes the following steps:

[0096] S1. Establish a satellite-ground semantic communication framework (SeC-NOMA) integrating NOMA technology: Use NOMA technology to improve spectral efficiency, optimize the encoding and decoding process through deep joint source-channel coding (JSCC) based on Swin Transformer, and at the same time reduce the model complexity and improve the image reconstruction quality;

[0097] S11. Optimize spectral efficiency

[0098] The NOMA technology allows multiple users to access simultaneously in the same frequency band, and a power allocation strategy is used to achieve the superposition and separation of multi-user signals; combined with the power-domain non-orthogonal multiple access technology, a serial interference cancellation module is used to separate and decode multi-user signals;

[0099] S12. Source-channel joint coding based on Swin Transformer

[0100] The input image is first divided into non-overlapping image patches, and each image patch is projected into a feature space of a specific dimension through a linear embedding layer; by introducing channel state information, this information is concatenated with the image features to form a joint representation and input into a multi-layer encoder based on Swin Transformer; during the encoding process, the Swin Transformer module uses window-based multi-head self-attention mechanism and shifted window multi-head self-attention mechanism to extract global and local features, thus generating a feature matrix adapted to the channel conditions;

[0101] The encoded signal passes through a power superposition module to allocate power to multiple user signals and generate a transmission signal:

[0102]

[0103] where y represents the signal that the transmitting end needs to send, and z j represents the signal formed by the j-th user after passing through the encoder, and P j represents the transmission power of the j-th user; after the signals are superimposed, they are then transmitted to the receiving end through the satellite-ground channel;

[0104] S13. Transmission channel and noise processing:

[0105] The transmission signal y is transmitted according to the analog channel, and the obtained signal is

[0106] S14. Receiving-end signal processing:

[0107] The received signal passes through a serial interference cancellation module, and the user signals are separated in order from high to low according to power, and the remaining signals are updated by iteratively canceling interference, and finally the decoded input signal of a single user is obtained; the separated signal is input into a decoder based on Swin Transformer, and the decoder reconstructs the image signal according to the channel state information and feature selection mask.

[0108] S2. Propose an adaptive semantic communication model (ASeC-NOMA): Based on the target bandwidth ratio and signal-to-noise ratio (SNR) of the satellite-ground channel, dynamically adjust the coding parameters, use auxiliary information to enhance the coding robustness, reduce the model memory occupancy, and achieve efficient adaptation to various channel environments;

[0109] S21. Target Bandwidth Ratio and Signal-to-Noise Ratio Modeling

[0110] At the input end, the target bandwidth ratio ρ and signal-to-noise ratio SNR of the satellite-ground channel are input as link state information, and through the fully connected layer, they are mapped to the auxiliary coding feature tensor u = MLP([SNR, ρ]); after splicing this auxiliary information with the image features, it is input into the encoder to dynamically adjust the coding process and adapt to different channel conditions;

[0111] S22. Adaptive Feature Extraction at the Encoding End

[0112] Image Blocking and Embedding: The input image S j is segmented into non-overlapping blocks and projected into the feature space of a specific dimension through the linear embedding layer; the generated tensor X j is spliced with the link state information u to guide the subsequent feature extraction process;

[0113] Pass through the multi-layer encoder based on Swin Transformer: The encoder adopts the Swin Transformer module to extract the global and local features of the image layer by layer; the module combines the link state information through the window multi-head self-attention and shifted window multi-head self-attention mechanisms to dynamically optimize the feature extraction and make the coding result adapt to the current channel conditions;

[0114] Output Feature and Transmission: The final coding result is adjusted to the feature matrix Y that adapts to the transmission bandwidth j , and the important features are selected through the feature selection mask M to generate the transmission signal z j = f θ (ρ, SNR, S j );

[0115] S23. Non-Orthogonal Multiple Access Mechanism

[0116] Adopt the power domain NOMA technology to achieve non-orthogonal transmission of multiple user signals; at the encoding end, according to the power allocation coefficient P of the user j superimpose the signals in power to generate the total transmission signal: This signal is transmitted through the satellite-ground channel and is interfered by channel noise.

[0117] S24. Adaptive Reconstruction at the Decoding End

[0118] Successive Interference Cancellation SIC to Separate Signals: At the receiving end, through the successive interference cancellation module, each user signal is separated from high to low in power; for the user signal with the highest current power, the following steps are taken:

[0119] 4) Decode the user signal:

[0120] 5) Update the remaining signals:

[0121] 6) Repeat the above steps until all user signals are separated;

[0122] Decoder based on Swin Transformer: The separated user signals are input into the decoder, and the decoder reconstructs the target image layer by layer by splicing the link state information u and the transmission features

[0123] S25. Adaptive transmission optimization mechanism

[0124] Bandwidth adaptability: By dynamically adjusting the bandwidth ratio ρ, limit the number of features N actually transmitted F , and only transmit the most important features;

[0125] Signal-to-noise ratio adaptability: Under low signal-to-noise ratio conditions, improve transmission reliability by increasing redundant information; under high signal-to-noise ratio conditions, reduce redundant information to improve transmission efficiency;

[0126] Dynamic weight adjustment algorithm: Introduce a dynamic weight allocation mechanism during training, and dynamically adjust the weights according to the image reconstruction loss under different bandwidth ratios or signal-to-noise ratio conditions to balance the performance of the model under different channel conditions.

[0127] S3. Propose a dynamic weight adjustment algorithm (NDWA): By optimizing the trade-off between the model convergence speed and the transmission performance under different channel conditions, improve the training efficiency and training performance of the adaptive model;

[0128] S31. Core objective of dynamic weight adjustment

[0129] To balance the performance of the model under different channel conditions, NDWA dynamically allocates training weights to avoid the model damaging the performance under high bandwidth ratio or high signal-to-noise ratio conditions due to optimizing low bandwidth ratio or low signal-to-noise ratio conditions; improve the convergence speed of the model through the weight dynamic adjustment mechanism to make it adapt to various channel environments;

[0130] S32. Initialization of dynamic weights

[0131] Allocate initial weights for each channel condition where w represents the set of discrete values of the bandwidth ratio or signal-to-noise ratio; the initial weights are equal to ensure that the model evenly optimizes all channel conditions at the initial stage of training;

[0132] S33. Weight update mechanism during the training phase

[0133] In each training cycle, for each bandwidth ratio ρ wOr the signal-to-noise ratio SNR, calculate the image reconstruction quality metric; finally, calculate the performance gap of the current model by comparing it with the reconstruction performance of the optimal model under fixed conditions: where is the optimal PSNR under fixed conditions, is the PSNR of the current model during the t-th round of training;

[0134] According to the performance gap dynamically adjust the weights, and the weight update formula is:

[0135]

[0136] where α is the scaling coefficient, k is the balance parameter,; the weights are restricted to the range [0, 10], that is, when the weight is less than 0, it is set to 0, and when it is greater than 10, it is set to 10 to avoid unstable training;

[0137] Normalize the weights for all channel conditions to ensure that the total weight sum is 1:

[0138]

[0139] where W is the total number of bandwidth ratios or signal-to-noise ratios.

[0140] S34. Sampling Strategies for Bandwidth and Signal-to-Noise Ratio

[0141] During training, the bandwidth ratio ρ w and the signal-to-noise ratio SNR are uniformly sampled from a fixed range; the sampled ρ w and SNR are input into the adaptive semantic communication model together with the input image, and the reconstructed image is obtained through model calculation;

[0142] S35. Weighted Optimization of the Loss Function

[0143] In each training, the loss values under different conditions are weighted and summed, and the optimization objective is the weighted mean square error:

[0144]

[0145] where is the dynamic weight during the t-th round of training, and K represents the number of image blocks and also the number of users;

[0146] S36. Model Performance Evaluation and Weight Update

[0147] In the validation stage of training, calculate the performance difference according to the reconstruction performance of the test set images under different conditions and update the weights according to the weight adjustment rule to ensure that the model is balanced and optimized under all conditions;

[0148] S37. Weight Constraint and Convergence Control

[0149] By imposing upper and lower bounds on the weights, it is ensured that the weights do not decrease or increase infinitely under specific conditions;

[0150] S38. Weight Adjustment Algorithm Process

[0151] First, initialize the initial weights corresponding to different bandwidth ratios. During the training process, the algorithm traverses multiple training rounds. In each round of training, the data is divided into several batches for processing; in each batch, first randomly sample the bandwidth ratio and signal-to-noise ratio to simulate the fluctuations in the actual communication environment; subsequently, input the image data and the sampled communication conditions into the model to generate the corresponding reconstructed image, and calculate the loss function of the current batch to measure the error of the reconstruction quality.

[0152] After each round of training, the algorithm enters the verification stage, evaluates the differences in reconstruction performance under different bandwidth conditions, and dynamically adjusts the weights based on this feedback, enabling the model to better adapt to different bandwidth environments; after multiple rounds of training and weight optimization, finally obtain the trained model parameters and the optimized dynamic weight allocation strategy.

[0153] S4. Design Feature Selection Module (FSM): Prioritize image features and selectively transmit important features according to bandwidth resources to maximize the image reconstruction effect.

[0154] S41. Importance Evaluation and Ranking

[0155] FSM evaluates the importance of the feature tensors extracted at the encoding end through global average pooling, sorts them in descending order according to the importance scores of the feature channels to determine the feature priorities; the feature channels with high scores are preferentially selected for transmission to ensure the transmission efficiency of semantic information;

[0156] S42. Dynamic Feature Selection

[0157] According to the target bandwidth ratio and signal-to-noise ratio, dynamically adjust the number of features to be transmitted. FSM generates a feature selection mask, only transmits the top-ranked important feature channels, and the unselected channels are masked to reduce the transmission of invalid data;

[0158] S43. Channel Adaptation and Decoding Reconstruction

[0159] At the receiving end, use the same feature selection mask as the sending end to restore the feature tensor, and reconstruct the important features through the decoder, finally generating a high-quality image reconstruction result. Under the limited bandwidth and signal-to-noise ratio conditions, FSM preferentially transmits the features with higher semantic concentration.

[0160] Such as Figure 1As shown, the present invention proposes a semantic communication architecture SeC-NOMA under the satellite-ground network, and designs an adaptive semantic communication model ASeC-NOMA on this basis, such as Figure 2 shown to solve the problem of changing channel conditions in the satellite-ground network.

[0161] The basic end-to-end model of semantic communication usually includes an encoder, a channel, a decoder, and a semantic knowledge base. Due to the complexity of the information layer and the adverse conditions of the satellite-ground communication environment, the present invention adopts a deep joint source-channel coding (JSCC) method to combine source compression and channel coding in order to better cope with the transmission challenges of the satellite-ground network.

[0162] SeC-NOMA System Architecture

[0163] To support large-scale user access and improve spectrum utilization, the present invention proposes a satellite-ground semantic communication framework SeC-NOMA that combines non-orthogonal multiple access (NOMA) technology. This architecture uses a deep JSCC method based on SwinTransformer in the encoder and decoder to reduce the model complexity and enhance the image transmission quality.

[0164] The input image S j After being block-processed, it is converted into a feature tensor through a linear embedding layer, and the obtained tensor X j contains the channel state information u = MLP([SBR, ρ]), which helps to adjust the encoding process to adapt to different channel conditions. The encoder uses a window-based multi-head self-attention (MSA) mechanism and a shifted window self-attention (SW-MSA) mechanism to extract the local and global features of the image, and then generates the encoded image feature Y j . The computational complexity of the two is as Figure 4 shown. Then, important features are selected through the feature selection mask M to generate the transmission signal z j = f θ (ρ, SNR, S j ).

[0165] The encoded image feature z j will be superimposed through a power allocation module to generate the total transmission signal Under NOMA technology, the signals of multiple users will share the same spectrum resources. The transmission signal is sent through the satellite-ground channel and received at the receiving end.

[0166] The receiving end separates the signals of each user from high to low power through a successive interference cancellation (SIC) module.

[0167] The separated user signals It is input into the decoder, and the decoder reconstructs the target image layer by layer by splicing the link state information u and the transmission features.

[0168] ASeC-NOMA Adaptive Model

[0169] To cope with the changes in different bandwidth and signal-to-noise ratio (SNR) conditions, the ASeC-NOMA model introduces an adaptive mechanism that can adjust the transmission features according to different channel states. By taking the target bandwidth ratio and signal-to-noise ratio as inputs, the encoder can dynamically adjust the quantity of transmitted information, thereby enhancing the robustness and quality of image transmission.

[0170] The ASeC-NOMA model dynamically adjusts the encoding method of the input image according to the current bandwidth ratio ρ and signal-to-noise ratio SNR during the encoding process, ensuring that the key features of the image are preferentially transmitted when the bandwidth is low or the signal-to-noise ratio is poor. In this way, the model can adapt to changing channel conditions.

[0171] During the training process, in order to balance the reconstruction quality under different channel conditions, a dynamic weight adjustment algorithm (NDWA) is proposed. By assigning dynamic weights to different channel conditions, it ensures that the image transmission performance under low bandwidth ratio or low SNR conditions will not be overly sacrificed.

[0172] To further improve the image reconstruction quality, the feature selection module (FSM) prioritizes the extracted image features according to the transmitted bandwidth and signal-to-noise ratio, and selects the most important features for transmission. This method effectively reduces the bandwidth consumption and optimizes the image reconstruction effect when the transmission bandwidth is limited.

[0173] This invention mainly focuses on the downlink scenario of NOMA, and the model details are as Figure 2 shown.

[0174] Neural Network Architecture of ASeC-NOMA

[0175] Figure 2 It is the neural network for the encoding and decoding parts of the ASeC-NOMA system. Each image S j is initially divided into non-overlapping blocks, and then each block is projected into a feature space with dimension c through a linear embedding layer. In this invention, the size of the block is set to 2×2. This process of block division and linear embedding converts the original image from dimension C×H×W to the feature tensor X j , whose dimension is c×H / 2×W / 2. Before passing X j to the subsequent transformation layer, the link information is appended to each block, obtaining a tensor X u with dimension (c + n jAmong them, the vector u is obtained by inputting the signal-to-noise ratio and the ρ value into a fully connected layer u = MLP([SNR, ρ]).

[0176] Then X j is fed into the Swin Transformer block in the encoder. After being processed through all I stages, the output dimension is c×h I ×w I . After that, this output tensor is reshaped and linearly projected to form a matrix Y T ×N F , where N j represents the number of transmissible features, satisfying N F N F N T = 2ρN, and N T represents the number of feature maps. Subsequently, the matrix Y j is processed by the mask M generated by the adaptive module. This mask selectively retains only the features intended to be transmitted through the channel, thereby obtaining the actually transmitted codeword z j .

[0177] After obtaining the signal z j through the serial interference cancellation process, it is first converted into a real-valued tensor, and then the first dimension is set to zero according to the mask M of the FSM, obtaining As Figure 2 shown, the link state information u in the encoder is copied and assigned to U d , and then concatenated with each block of Y j . This operation generates a tensor forming The resulting concatenated tensor is then projected into a vector of dimension c. The decoder consists of I stages, and each stage includes a block splitting mechanism and a Swin Transformer block. The block splitting mechanism uses a pixel shuffle operation to increase the spatial resolution of the input tensor. The upsampled tensor is then processed by a Swin Transformer block, whose structure is the same as that in the encoder. After completing all I stages, the latent tensor is converted into a reconstructed image

[0178] Design the adaptive module

[0179] The deep JSCC model is relatively large, has a high memory complexity, and processes image data, making retraining time-consuming. For each IoT device, it is unrealistic to store multiple sets of model parameters under different channel conditions. To enable the model to adapt to different channel conditions, the present invention proposes an adaptive training model called ASeC-NOMA, which incorporates the link bandwidth, signal-to-noise ratio, and input image into the model. To better reflect real-world channel conditions, in addition to introducing the signal-to-noise ratio as a parameter, the link quality is also considered. Based on this link quality, the original signal-to-noise ratio parameter is perturbed, and the resulting signal-to-noise ratio and link bandwidth ratio are mapped to a tensor The size of this tensor matches the dimension of the tensor generated by the embedding layer in SwinTransformer, as Figure 3 shown. In addition, the present invention aims to dynamically adjust the bandwidth required for each signal transmission according to the corresponding bandwidth ratio. As Figure 3 shown, the encoder maps the input image to dimension ρ and transmits the corresponding first ρ elements through the mask M, converting Y j to z j . At the receiving end, the mask in the FSM is also used to convert to the corresponding converted to the corresponding

[0180] Non-Dynamic Weight Adjustment Algorithm (NDWA)

[0181] 1. Core objective of dynamic weight adjustment:

[0182] To balance the performance of the model under different channel conditions (bandwidth ratio and signal-to-noise ratio), NDWA dynamically allocates training weights to prevent the model from sacrificing performance under high bandwidth ratio or high signal-to-noise ratio conditions due to optimizing low bandwidth ratio or low signal-to-noise ratio conditions. In addition, the convergence speed of the model is improved through the weight dynamic adjustment mechanism to make it adapt to various channel environments.

[0183] 2. Initialization of dynamic weights:

[0184] Assign initial weights to each channel condition where w represents the set of discrete values of the bandwidth ratio or signal-to-noise ratio. The initial weights are equal to ensure that the model evenly optimizes all channel conditions in the initial stage of training.

[0185] 3. Weight update mechanism during the training phase:

[0186] First, in each training cycle, for each bandwidth ratio ρ w or signal-to-noise ratio SNR, calculate the image reconstruction quality metric (such as peak signal-to-noise ratio PSNR or structural similarity SSIM). By comparing with the optimal model reconstruction performance under fixed conditions, finally, calculate the performance gap of the current model: Among them is the optimal PSNR under fixed conditions, and is the PSNR of the current model during the t-th round of training.

[0187] Then, according to the performance gap dynamically adjust the weights, and the weight update formula is:

[0188]

[0189] where α is the scaling coefficient and k is the balancing parameter to ensure that the weight adjustment does not overly favor certain conditions; the weights are restricted to the range [0, 10], that is, when the weight is less than 0, it is set to 0, and when it is greater than 10, it is set to 10, to avoid unstable training.

[0190] Finally, normalize the weights for all channel conditions to ensure that the total weight sum is 1:

[0191]

[0192] where W is the total number of bandwidth ratios or signal-to-noise ratios.

[0193] 1. Sampling strategies for bandwidth and signal-to-noise ratio:

[0194] During the training process, the bandwidth ratio ρ w and the signal-to-noise ratio SNR are uniformly sampled from a fixed range. The sampled ρ w and SNR are input into the adaptive semantic communication model together with the input image, and the reconstructed image is obtained through model calculation.

[0195] 2. Weighted optimization of the loss function:

[0196] In each training, the loss values under different conditions are weighted and summed, and the optimization objective is the weighted mean square error (MSE):

[0197]

[0198] where is the dynamic weight during the t-th round of training, and K represents the number of image blocks and also the number of users.

[0199] 3. Model performance evaluation and weight update:

[0200] In the validation stage of training, according to the reconstruction performance of the test set images under different conditions, calculate the performance difference and update the weights according to the weight adjustment rule to ensure that the model is evenly optimized under all conditions.

[0201] 4. Weight constraint and convergence control:

[0202] By imposing upper and lower bounds on the weights, it is ensured that the weights do not decrease or increase infinitely under specific conditions (such as high bandwidth or high signal-to-noise ratio). Experiments show that weight constraints can effectively stabilize the convergence of the model.

[0203] Feature Sorting Module (FSM)

[0204] 1. Importance Evaluation and Sorting:

[0205] The FSM evaluates the importance of the feature tensors extracted at the encoding end through global average pooling, sorts them in descending order according to the importance scores of the feature channels, and determines the feature priorities. The feature channels with high scores are preferentially selected for transmission to ensure the transmission efficiency of semantic information.

[0206] 2. Dynamic Feature Selection:

[0207] According to the target bandwidth ratio and signal-to-noise ratio, the number of transmitted features is dynamically adjusted. The FSM generates a feature selection mask and only transmits the top-ranked important feature channels, and the unselected channels will be masked to reduce the transmission of invalid data.

[0208] 3. Channel Adaptation and Decoding Reconstruction:

[0209] At the receiving end, the feature tensor is restored using the same feature selection mask as the sending end, and the important features are reconstructed by the decoder to finally generate a high-quality image reconstruction result. In the case of limited bandwidth and signal-to-noise ratio conditions, the FSM can preferentially transmit features with higher semantic concentration, thereby significantly improving the image reconstruction quality.

[0210] 3. System Training and Performance Simulation Analysis

[0211] In an embodiment of the present invention, the relevant parameters are shown in Table 1.

[0212] Table 1 Simulation Parameters of the ASeC-NOMA Semantic Communication System

[0213]

[0214] Such as Figure 5As shown, in one embodiment of the present invention, under the Gaussian channel (AWGN), the performance of the ASeC-NOMA model is significantly better than that of the traditional separate coding model BPG LDPC and the CNN-based joint source-channel coding model DeepJSCC CNN. Specifically, ASeC-NOMA exhibits higher peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) under all signal-to-noise ratio (SNR) conditions. For example, at an SNR of 10 dB, the PSNR of ASeC-NOMA is at least 7 dB higher than that of BPG LDPC and at least 8 dB higher than that of DeepJSCC CNN. Under high SNR conditions, the PSNR and SSIM values of ASeC-NOMA are close to 30 dB and 0.95, indicating that the reconstructed image is almost indistinguishable from the original image, which demonstrates the superior performance of ASeC-NOMA under the Gaussian channel.

[0215] As Figure 6 shown, in one embodiment of the present invention, under the satellite-ground channel, the performances of ASeC-NOMA (BW-NDWA) and ASeC-NOMA (BW-NDWA-FSM) are good under different bandwidth ratio conditions. The PSNR and SSIM values of ASeC-NOMA (BW-NDWA) increase with the increase of the bandwidth ratio. For example, when the bandwidth ratio is 1 / 4, its PSNR and SSIM values are approximately 2 dB and 0.02 higher than those of ASeC-NOMA (fixed) under fixed conditions. After introducing the FSM module, the PSNR and SSIM values of ASeC-NOMA (BW-NDWA-FSM) are further improved, with the maximum improvements exceeding 5 dB and 0.1 respectively, which indicates that the FSM module significantly improves the image transmission quality under the satellite-ground channel.

[0216] As Figure 7 shown, in one embodiment of the present invention, under the satellite-ground channel, the performances of ASeC-NOMA (SNR-NDWA) and ASeC-NOMA (SNR-NDWA-FSM) are good under different signal-to-noise ratio conditions. The PSNR and SSIM values of ASeC-NOMA (SNR-NDWA) increase with the increase of the signal-to-noise ratio. For example, when the signal-to-noise ratio is 7 dB, its PSNR and SSIM values are approximately 3 dB and 0.02 higher than those of ASeC-NOMA (fixed) under fixed conditions. After introducing the FSM module, the PSNR and SSIM values of ASeC-NOMA (SNR-NDWA-FSM) are further improved, with the maximum improvements exceeding 0.5 dB and 0.01 respectively, which indicates that the FSM module significantly improves the image transmission quality under the satellite-ground channel.

[0217] As Figure 8As shown, in an embodiment of the present invention, under the satellite-ground channel, the visible views before and after image transmission under different conditions demonstrate the effectiveness of the ASeC-NOMA model. The comparison before and after image transmission shows that the ASeC-NOMA model can effectively reconstruct images under different conditions. For example, under low bandwidth ratio and low signal-to-noise ratio conditions, the image reconstruction quality is still good, and the PSNR and SSIM values are high. After introducing the FSM module, the image reconstruction quality is further improved, and the details are clearer, indicating that the ASeC-NOMA model has good adaptability and robustness under the satellite-ground channel.

[0218] As Figure 9 shown, in an embodiment of the present invention, under the satellite-ground channel, the performance of three users when transmitting using ASeC-NOMA (BW-NDWA) and ASeC-OMA (BW-NDWA) is good. ASeC-NOMA (BW-NDWA) performs well under different bandwidth ratio conditions, and the PSNR and SSIM values are comparable to those of ASeC-NOMA (fixed) under fixed conditions. For example, when the bandwidth ratio is 1 / 4, the PSNR and SSIM values of ASeC-NOMA (BW-NDWA) are approximately 3 dB and 0.05 higher than those of ASeC-OMA (BW-NDWA) respectively. This indicates that ASeC-NOMA has better spectrum utilization and transmission performance in multi-user scenarios and can effectively support the simultaneous transmission of multiple users.

[0219] As Figure 10 shown, in an embodiment of the present invention, under the satellite-ground channel, the performance of three users when transmitting using ASeC-NOMA (SNR-NDWA) and ASeC-OMA (SNR-NDWA) is good. ASeC-NOMA (SNR-NDWA) performs well under different signal-to-noise ratio conditions, and the PSNR and SSIM values are comparable to those of ASeC-NOMA (fixed) under fixed conditions. For example, when the signal-to-noise ratio is 7 dB, the PSNR and SSIM values of ASeC-NOMA (SNR-NDWA) are approximately 3 dB and 0.05 higher than those of ASeC-OMA (SNR-NDWA) respectively. This indicates that ASeC-NOMA has better spectrum utilization and transmission performance in multi-user scenarios and can effectively support the simultaneous transmission of multiple users.

[0220] As Figure 11As shown, in an embodiment of the present invention, under the satellite-ground channel, the comparison of the spectral efficiency of OMA and NOMA in ASeC-NOMA (BW-NDWA) and ASeC-NOMA (SNR-NDWA) shows that the spectral efficiency of NOMA is significantly higher than that of OMA. For example, when the bandwidth ratio is 1 / 4, the spectral efficiency of NOMA is three times that of OMA. This indicates that NOMA has higher spectral efficiency in multi-user scenarios, can support more users to transmit data simultaneously, and thus provides superior performance under the satellite-ground channel.

[0221] Although the present invention has been described with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed, as long as they do not deviate from the spirit and scope of the present invention defined by the appended claims. It should be understood that different dependent claims and the features described in the present invention can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. A satellite - to - ground network adaptive image semantic communication method based on NOMA technology, comprising the following steps: S1. Establish a satellite - to - ground semantic communication framework integrating NOMA technology: Use NOMA technology to perform power superposition and successive interference cancellation on the transmission content of different users, and optimize the encoding and decoding processes through deep joint source - channel coding based on Swin Transformer; S2. Establish an adaptive semantic communication model: Dynamically adjust the encoding parameters based on the target bandwidth ratio and signal - to - noise ratio of the satellite - to - ground channel; S3. Propose a dynamic weight adjustment algorithm: Optimize the model convergence speed and the trade - off of transmission performance under different channel conditions; S4. Design a feature selection module: Rank the priorities of image features and selectively transmit important features according to bandwidth resources.

2. The method for satellite-ground network adaptive image semantic communication based on NOMA technology according to claim 1, wherein In S1, the satellite - to - ground semantic communication framework specifically includes: S11. Optimize the spectral efficiency Allow multiple users to access simultaneously in the same frequency band through NOMA technology, and use the power allocation strategy to achieve the superposition and separation of multi - user signals; Combine the power - domain non - orthogonal multiple access technology, and use the successive interference cancellation module to achieve the separation and decoding of multi - user signals; S12. Source - channel joint coding based on Swin Transformer The input image is first divided into non - overlapping image patches, and each image patch is projected into a feature space of a specific dimension through a linear embedding layer; By introducing channel state information, this information is concatenated with the image features to form a joint representation and input into a multi - layer encoder based on Swin Transformer; During the encoding process, the Swin Transformer module uses window - based multi - head self - attention mechanism and shifted window multi - head self - attention mechanism to extract global and local features, thereby generating a feature matrix adapted to the channel conditions; The encoded signal passes through the power superposition module to allocate power to multiple user signals and generate a transmission signal: Among them, y represents the signal to be sent by the transmitting end, and z j represents the signal formed by the j-th user after passing through the encoder, and P j represents the transmission power of the j-th user; after the signals are superimposed, they are then transmitted to the receiving end through the satellite-ground channel; S13. Transmission channel and noise processing: The transmitted signal y is transmitted through an analog channel, and the resulting signal is S14. Receiver signal processing: The received signal passes through the successive interference cancellation module, and the user signals are separated in order from high to low power. The remaining signals are updated by iteratively eliminating interference, and finally the decoded input signal of a single user is obtained; The separated signal is input into the decoder based on Swin Transformer, and the decoder reconstructs the image signal according to the channel state information and the feature selection mask.

3. The method for satellite-ground network adaptive image semantic communication based on NOMA technology according to claim 1, characterized in that In S2, the adaptive semantic communication model includes the following steps: S21. Model the target bandwidth ratio and signal - to - noise ratio At the input end, the target bandwidth ratio ρ and signal - to - noise ratio SNR of the satellite - to - ground channel are input as link state information, and are mapped into an auxiliary coding feature tensor u = MLP([SBR, ρ]) through a fully - connected layer; This auxiliary information is concatenated with the image features and then input into the encoder to dynamically adjust the encoding process and adapt to different channel conditions; S22. Adaptive feature extraction at the encoding end Perform image chunking and embedding: Input image S j is divided into non-overlapping chunks and projected into a feature space of a specific dimension through a linear embedding layer; the generated tensor X j concatenates the link state information u to guide the subsequent feature extraction process; Pass through a multi - layer encoder based on Swin Transformer: The encoder uses Swin Transformer modules to extract global and local features of the image layer by layer; the module combines link state information through window multi - head self - attention and shifted window multi - head self - attention mechanisms to dynamically optimize feature extraction, making the encoding result adapt to the current channel conditions; Output Features and Transmission: The final encoded result is adjusted to the feature matrix Y that adapts to the transmission bandwidth j , and the important features are selected through the feature selection mask M to generate the transmission signal z j = f θ (ρ, SNR, S j ); S23, Non - orthogonal multiple access mechanism The non - orthogonal transmission of multiple user signals is achieved by using power - domain NOMA technology; at the encoding end, according to the power allocation coefficient P of the user j the signals are power - superposed to generate the total transmission signal: This signal is transmitted through the satellite - to - ground channel and is interfered by channel noise.

4. The method for satellite-ground network adaptive image semantic communication based on NOMA technology according to claim 3, wherein In S2, the adaptive semantic communication model further includes the following steps: S24, Adaptive reconstruction at the decoding end Successive interference cancellation (SIC) to separate signals: At the receiving end, through the successive interference cancellation module, each user signal is separated according to the power from high to low; for the user signal with the highest current power, the following steps are taken: 1) Decode the user signal: 2) Update the remaining signal: 3) Repeat the above steps until all user signals are separated; Decoder Based on Swin Transformer: Isolated User Signal Input into the decoder, the decoder reconstructs the target image layer by layer by splicing the link state information u and the transmission features S25, Adaptive transmission optimization mechanism Bandwidth Adaptation: By dynamically adjusting the bandwidth ratio ρ, the number of features N actually transmitted is restricted F , and only the most important features are transmitted; Signal - to - noise ratio (SNR) adaptation: Under low SNR conditions, increase redundant information to improve transmission reliability; under high SNR conditions, reduce redundant information to improve transmission efficiency; Dynamic weight adjustment algorithm: Introduce a dynamic weight allocation mechanism during training, and dynamically adjust the weights according to the image reconstruction loss under different bandwidth ratios or SNR conditions to balance the performance of the model under different channel conditions.

5. The method for satellite-ground network adaptive image semantic communication based on NOMA technology according to claim 1, wherein In S3, the dynamic weight adjustment algorithm includes the following steps: S31, The core objective of dynamic weight adjustment To balance the performance of the model under different channel conditions, NDWA dynamically allocates training weights to avoid the model damaging its performance under high - bandwidth - ratio or high - SNR conditions due to optimizing low - bandwidth - ratio or low - SNR conditions; improve the convergence speed of the model through the weight dynamic adjustment mechanism to make it adapt to various channel environments; S32, Initialization of dynamic weights Assign initial weights to each channel condition where ω represents a set of discrete values of bandwidth ratio or signal-to-noise ratio; the initial weights are equal to ensure that the model optimizes all channel conditions evenly at the initial stage of training; S33, Weight update mechanism during the training phase In each training cycle, for each bandwidth ratio ρ w or signal-to-noise ratio SNR, calculate the image reconstruction quality metric; finally, calculate the performance gap of the current model by comparing it with the reconstruction performance of the optimal model under fixed conditions: where is the optimal PSNR under fixed conditions, is the PSNR of the current model at the t-th round of training; According to the performance gap Dynamically adjust the weights. The weight update formula is as follows: Where α is the scaling coefficient, k is the balance parameter,; the weights are restricted to the range [0, 10], that is, when the weight is less than 0, it is set to 0, and when it is greater than 10, it is set to 10 to avoid unstable training; Normalize the weights for all channel conditions to ensure that the sum of the total weights is 1: Where W is the total number of bandwidth ratios or SNR.

6. The method for satellite-ground network adaptive image semantic communication based on NOMA technology according to claim 5, wherein In S3, the dynamic weight adjustment algorithm further includes the following steps: S34, Sampling strategy for bandwidth and SNR During the training process, the bandwidth ratio ρ w and the signal-to-noise ratio SNR are uniformly sampled from a fixed range; the sampled ρ w and SNR are input into the adaptive semantic communication model together with the input image, and the reconstructed image is obtained through model calculation; S35, Weighted optimization of the loss function In each training, perform a weighted sum of the loss values under different conditions, and the optimization objective is the weighted mean square error: Among them is the dynamic weight of the t-th round of training, and K represents the number of image patches and also the number of users; S36, Model performance evaluation and weight update During the validation phase of training, calculate the performance difference based on the reconstruction performance of the test set images under different conditions And update the weights according to the weight adjustment rules to ensure balanced optimization of the model under all conditions; S37, Weight constraint and convergence control By imposing upper and lower bounds on the weights, ensure that the weights do not decrease or increase infinitely under specific conditions; S38, Flow of the weight adjustment algorithm First, initialize the initial weights corresponding to different bandwidth ratios. During training, the algorithm traverses multiple training epochs. In each epoch, the data is divided into several batches for processing; in each batch, first randomly sample the bandwidth ratio and SNR to simulate the fluctuations of the actual communication environment; then, input the image data and the sampled communication conditions into the model to generate the corresponding reconstructed image, and calculate the loss function of the current batch to measure the error of the reconstruction quality. After each round of training, the algorithm enters the verification phase, evaluates the reconstruction performance differences under different bandwidth conditions, and dynamically adjusts the weights based on this feedback, enabling the model to better adapt to different bandwidth environments; after multiple rounds of training and weight optimization, the trained model parameters and the optimized dynamic weight allocation strategy are finally obtained.

7. The method for satellite-ground network adaptive image semantic communication based on NOMA technology according to claim 1, wherein In S4, the feature selection module includes the following parts: S41. Importance evaluation and sorting The FSM evaluates the importance of the feature tensors extracted at the encoding end through global average pooling, sorts them in descending order according to the importance scores of the feature channels, and determines the feature priorities; the feature channels with high scores are preferentially selected for transmission to ensure the transmission efficiency of semantic information. S42. Dynamic feature selection According to the target bandwidth ratio and signal-to-noise ratio, the FSM dynamically adjusts the number of features to be transmitted, generates a feature selection mask, and only transmits the top-ranked important feature channels. The unselected channels are masked to reduce the transmission of invalid data. S43. Channel adaptation and decoding reconstruction At the receiving end, the feature selection mask consistent with the sending end is used to restore the feature tensors, and the decoder reconstructs the important features to finally generate a high-quality image reconstruction result. Under limited bandwidth and signal-to-noise ratio conditions, the FSM preferentially transmits the features with higher semantic concentration.

8. An adaptive image semantic communication system for satellite-ground network based on NOMA technology, characterized in that, The system includes a computer module that utilizes the above multi-objective space-ground fusion network spatio-temporal switching decision method.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the space-ground network adaptive image semantic communication method based on the NOMA technology.

10. A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the space-ground network adaptive image semantic communication method based on the NOMA technology.

Images