A multi-domain adaptive semantic communication method and system

Abstract

This invention relates to the field of semantic communication technology, and more particularly to a multi-domain adaptive semantic communication method and system, comprising the following steps: dividing an input image into multiple image blocks using an image embedding module, extracting compressed semantic features, transmitting these features through an additive white Gaussian noise channel to obtain received semantic features, encoding source domain labels using a neural network to obtain domain embedding representations, establishing a set of domain prototypes, and at the receiving end, performing semantic reconstruction based on the received semantic features and target domain features, inputting the reconstructed semantic data into a downstream task model to obtain the task processing result. This multi-domain adaptive semantic communication method and system can improve semantic communication efficiency, enhance cross-domain semantic transmission performance, strengthen system robustness, achieve multi-domain adaptation capabilities, support multi-scenario applications, and significantly enhance the practical application value of semantic communication systems in multi-domain environments.

Description

Technical Field

[0001] This invention relates to the field of semantic communication technology, and in particular to a multi-domain adaptive semantic communication method and system. Background Technology

[0002] Traditional communication systems primarily focus on the accurate transmission of bit-level information, ensuring that the receiving end can fully reconstruct the data content from the sending end. However, in many intelligent application scenarios, such as autonomous driving, intelligent monitoring, drone vision tasks, and edge intelligent collaborative systems, downstream tasks truly require semantic information within the data, rather than all the raw data. Semantic communication systems extract key semantic features from the data using deep learning models and transmit only these features, thereby significantly improving communication efficiency under limited bandwidth conditions.

[0003] Currently, mainstream semantic communication methods typically employ deep neural networks for semantic encoding, with the Transformer architecture being the most common. However, the Transformer architecture suffers from high computational complexity when modeling long sequences, with its complexity increasing quadratically with sequence length. This can lead to significant latency and computational overhead in communication scenarios. In recent years, State Space Models (SSMs) have emerged as a novel approach for modeling long sequences. Among them, the Visual Mamba model achieves linear time complexity in two-dimensional image processing through a selective scanning mechanism, offering significant advantages in inference speed and memory usage.

[0004] Nevertheless, existing semantic communication research still faces the following technical challenges:

[0005] First, most studies only consider a single data domain environment, assuming that the data comes from the same distribution environment, and lack the ability to explicitly model the distribution of multi-domain data.

[0006] Second, existing systems lack explicit modeling mechanisms for domain characteristics, making it difficult to capture distributional differences between different domains;

[0007] Third, in cross-domain application scenarios, due to differences in the environment, sensor equipment, acquisition methods or application scenarios between the sending end and the receiving end, the input data usually has a significant distribution shift between different domains. This cross-domain difference will cause semantic features to degrade during the decoding stage, thereby affecting the performance of downstream tasks.

[0008] Fourth, existing technologies are insufficient in modeling the relationship between domain features and semantic content, making it difficult to achieve effective joint modeling of semantic content and domain information.

[0009] Therefore, there is an urgent need for a new semantic communication system architecture that can efficiently model semantic information and achieve multi-domain adaptive capabilities, thereby maintaining the robustness and generalization of semantic transmission in cross-domain scenarios. Summary of the Invention

[0010] The main objective of this invention is to overcome the shortcomings of the prior art and provide a multi-domain adaptive semantic communication method and system.

[0011] The technical solution adopted by this invention to achieve its technical objective is: a multi-domain adaptive semantic communication method, comprising the following steps:

[0012] S1, Semantic encoding step: The input image is divided into multiple image blocks by the image embedding module, mapped into a one-dimensional sequence representation, and then input into the state space model encoder based on the Visual Mamba structure to extract the compressed semantic features;

[0013] Among them, the state-space model encoder of the Visual Mamba structure maintains the latent state and models the semantics of the image through a selective scanning mechanism;

[0014] S2, Channel transmission step: After power normalization of the semantic features, the signals are transmitted through an additive white Gaussian noise channel to obtain the received semantic features.

[0015] S3, Domain Feature Encoding Step: The source domain label is encoded using a neural network to obtain the domain embedding representation;

[0016] S4. Domain Prototype Alignment Step: Construct a set containing K domain prototypes, where each domain corresponds to a feature prototype. Align the domain embedding representation with the domain prototype of its own domain, and maintain the separation of features between different domains through a comparative learning mechanism with other domain prototypes.

[0017] S5. Semantic decoding step: At the receiving end, semantic reconstruction is performed based on the received semantic features and target domain features.

[0018] S6. Downstream task processing step: Input the recovered semantic data into the downstream task model to obtain the task processing result.

[0019] Preferably, in the semantic encoding step, the input image, after embedding, obtains a sequence representation as follows: ;

[0020] in, For the encoded input image, a one-dimensional embedding sequence, For the encoding function of the image embedding module, The input image is H, where H and W represent the height and width of the image, respectively. , The side length of the image patch is used to determine the granularity of image segmentation;

[0021] The compressed semantic features are represented as follows: ;

[0022] in, For image semantic feature extraction function, Indicates semantic compression ratio. It is the amount of data in the original image. It is the amount of data for the encoded semantic features s.

[0023] Preferably, in the channel transmission step, the additive white Gaussian noise channel model is represented as: ;

[0024] in, Indicates sending a signal. This represents the received signal after transmission through the channel; This represents the additive white Gaussian noise component in the channel, which follows a mean of 0 and a variance of . The normal distribution, i.e. The variance Power intensity used to characterize noise.

[0025] Preferably, in the domain feature encoding step, the domain embedding is represented as: ;

[0026] in, This is a domain feature encoding function used to map the input to the domain feature embedding space. Domain labels are used to identify the distribution characteristics of target data domains. Input a tensor for an optional reference image.

[0027] Preferably, in the domain prototype alignment step, a global set containing K types of domain distribution prototypes is constructed: ;

[0028] in, A domain label used to uniquely identify the distribution characteristics of a specific data domain. For set P and field labels The corresponding feature prototype, and ;

[0029] The prototype alignment mechanism constrains the domain embeddings output by the encoder to aggregate toward the domain prototype to which they belong, and the prototype-based contrastive learning mechanism maintains the differences between different domain prototypes.

[0030] For target domain tags Its characteristic prototype This characterizes the statistical distribution center of the domain, while the domain embedding... Configured to be in the feature space with the feature prototype Perform distribution alignment to make the domain embedded While preserving individual differences, the feature prototypes are encompassed and characterized. The defined domain feature distribution pattern.

[0031] Preferably, in the semantic decoding step, the semantic reconstruction is represented as: ;

[0032] in, This is the decoding function of the image decoder. The communication received signal represents the semantic features of the source image. To embed a domain into a specified target domain, target domain features are introduced during the decoding phase, enabling the system to perform semantic transfer between different domains.

[0033] Preferably, during semantic decoding, the latent state changes of the state-space model are influenced by domain embedding, and the output matrix is ​​determined by semantic features and latent states. The changes in the state-space model are represented as follows:

[0034] ;

[0035] in, For the learnable basic state matrix, Both are learnable linear projection functions, where d is the domain embedding and s is the image semantic feature. , and Together, they constitute the core dynamic state-space parameters in the Mamba model.

[0036] Preferably, in the semantic decoding step, the state update and output of the state space model are represented as follows: , ;

[0037] in, Represents the discrete time step index in a one-dimensional embedded sequence. Indicates the first The field label for step input is eigenvectors, Indicates the first The hidden state vector of the step. , , For dynamic state-space parameters, Current number The final output of the step is represented;

[0038] As can be seen from the formula, when outputting y, elements from both the semantic content s and the domain feature d are obviously included in the decoding process.

[0039] Preferably, in the downstream task processing step, the downstream task model is an image classification model: ;

[0040] in, For downstream task adaptation modules, The image generated by the decoder, The predicted output for the corresponding downstream task is used to compare with the true label to evaluate the model performance.

[0041] The present invention also provides a multi-domain adaptive semantic communication system, comprising:

[0042] The semantic encoding module is used to divide the input image into multiple image blocks through the image embedding module, map them into a one-dimensional sequence representation, and then input them into the state space model encoder based on the Visual Mamba structure to extract the compressed semantic features.

[0043] The channel transmission module is used to normalize the power of the semantic features and then transmit them through an additive white Gaussian noise channel to obtain the received semantic features.

[0044] The domain feature encoding module is used to encode the source domain label through a neural network to obtain the domain embedding representation;

[0045] The domain prototype alignment module is used to establish a set of domain prototypes containing K domains, where each domain corresponds to a feature prototype. The domain embedding representation is aligned with the domain prototype of its own domain, and the features between different domains are kept separate through a comparative learning mechanism with other domain prototypes.

[0046] The semantic decoding module is used to perform semantic reconstruction at the receiving end based on the received semantic features and target domain features. The potential state changes of the state space model are affected by the domain features, and the output matrix is ​​determined by the semantic features.

[0047] The downstream task processing module is used to input the recovered semantic data into the downstream task model to obtain the task processing results.

[0048] The working principle of this multi-domain adaptive semantic communication method and system is as follows: At the sending end, the semantic encoding module divides the input image into multiple image blocks and maps them into a one-dimensional sequence. The compressed semantic features are extracted by a state-space model encoder based on Visual Mamba. At the same time, the domain feature encoding module encodes the source domain labels to obtain the domain embedding representation. The domain embedding is aligned with the domain prototype of its own domain by the domain prototype alignment module. Then, the contrastive learning mechanism is used to keep the prototypes of different domains separate. The compressed semantic features are transmitted to the receiving end through an additive white Gaussian noise channel after power normalization. At the receiving end, the semantic decoding module performs semantic reconstruction based on the received noisy semantic features and target domain features. The potential state changes of the state-space model are affected by the domain features, and the output matrix is ​​determined by the semantic features, thereby realizing the joint modeling of semantic content and domain information. Finally, the reconstructed semantic data is input into the downstream task model to obtain the processing results of tasks such as image classification, object detection or segmentation, thus completing end-to-end semantic communication in a cross-domain environment.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] This multi-domain adaptive semantic communication method and system can improve the efficiency of semantic communication: the present invention adopts the VisualMamba state space model as the core architecture of semantic encoding and decoding, and realizes linear complexity modeling through selective scanning mechanism, avoiding the high latency problem caused by the quadratic computational complexity of the Transformer structure, significantly reducing the computational overhead of the communication system, and is suitable for communication scenarios with high real-time requirements.

[0051] This multi-domain adaptive semantic communication method and system has the ability to achieve multi-domain adaptation: by introducing a domain feature encoding module 3 and a domain prototype alignment module 4, by establishing a set of domain prototypes containing K domains, and by adopting a prototype alignment mechanism and a prototype-based contrastive learning mechanism, the domain features output by the encoder gradually approach the domain prototypes of their respective domains, while maintaining a large difference between different domain prototypes, thereby achieving adaptive modeling of multi-domain data distribution, enabling the system to adapt to different data distribution environments.

[0052] This multi-domain adaptive semantic communication method and system improves the performance of cross-domain semantic transmission: by introducing target domain features into the semantic decoding module and setting the potential state changes of the state space model to depend on the domain features and the output matrix to depend on the semantic features, the joint modeling of semantic content and domain information is realized, effectively reducing the semantic degradation problem caused by domain offset.

[0053] This multi-domain adaptive semantic communication method and system enhances system robustness: by using the prototype alignment mechanism and contrastive learning mechanism in the domain prototype alignment module 4, the distinguishability between features of different domains is maintained, the representation within the same domain remains highly consistent, and the separation structure between different domains is maintained clearly, thereby reducing the performance degradation caused by domain offset and improving the robustness and generalization ability of the semantic communication system in cross-domain scenarios.

[0054] This multi-domain adaptive semantic communication method and system supports multiple application scenarios: It is applicable to various visual tasks, and the downstream task processing module 6 can adapt to various task models such as image classification, object detection, or image segmentation. Specific application scenarios include: in UAV visual communication scenarios, it can transmit only key semantic features under limited communication bandwidth conditions, reducing communication resource consumption, and adapting to different UAV devices and shooting environments through a domain feature alignment mechanism; in autonomous driving vehicle-road cooperative systems, it can solve the problem of data distribution differences between different vehicle sensors, improving the stability of semantic recognition in cross-vehicle environments; in smart city video surveillance systems, it can adapt to data domain differences caused by different camera devices and lighting conditions, achieving efficient semantic transmission of large-scale video surveillance data. Furthermore, this invention can also be applied to edge intelligent collaborative systems and remote medical image analysis, significantly enhancing the practical application value of semantic communication systems in multi-domain environments. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the steps of a multi-domain adaptive semantic communication method.

[0056] Figure 2 This is a system framework diagram of a multi-domain adaptive semantic communication system.

[0057] The module consists of: 1. Semantic coding module; 2. Channel transmission module; 3. Domain feature coding module; 4. Domain prototype alignment module; 5. Semantic decoding module; and 6. Downstream task processing module. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0059] In the description of this invention, it should be noted that when an element is referred to as being "fixed to" or "set on" another element, it can be directly on or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to or indirectly connected to the other element.

[0060] In the description of this invention, it should be noted that the terms "center," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. "Several" means one or more, unless otherwise explicitly specified.

[0061] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0062] Example 1:

[0063] Please see Figures 1-2 A multi-domain adaptive semantic communication method includes the following steps:

[0064] S1. Semantic encoding step: The input image is divided into multiple image blocks by the image embedding module, mapped into a one-dimensional sequence representation, and then input into the state space model encoder based on the Visual Mamba structure to extract compressed semantic features; wherein, the state space model encoder based on the Visual Mamba structure maintains the latent state and models the image semantics through a selective scanning mechanism.

[0065] The input image, after embedding, yields a sequence representation as follows: ;

[0066] in, For the encoded input image, a one-dimensional embedding sequence, For the encoding function of the image embedding module, The input image is H, where H and W represent the height and width of the image, respectively. , The side length of the image patch is used to determine the granularity of image segmentation.

[0067] The compressed semantic features are represented as follows: ;

[0068] in, For image semantic feature extraction function, Indicates semantic compression ratio. It is the amount of data in the original image. It is the amount of data for the encoded semantic features s.

[0069] S2, Channel transmission step: After power normalization of the semantic features, the signals are transmitted through an additive white Gaussian noise channel to obtain the received semantic features.

[0070] The additive white Gaussian noise channel model is expressed as: ;

[0071] in, Indicates sending a signal. This represents the received signal after transmission through the channel; This represents the additive white Gaussian noise component in the channel, which follows a mean of 0 and a variance of . The normal distribution, i.e. The variance Power intensity used to characterize noise.

[0072] S3, Domain Feature Encoding Step: The source domain label is encoded using a neural network to obtain the domain embedding representation.

[0073] Domain embedding is represented as: ;

[0074] in, This is a domain feature encoding function used to map the input to the domain feature embedding space. Domain labels are used to identify the distribution characteristics of target data domains. Input a tensor for an optional reference image.

[0075] S4. Domain Prototype Alignment Step: Establish a domain prototype set containing K domains, where each domain corresponds to a feature prototype. Align the domain embedding representation with the domain prototype of its parent domain, and maintain feature separation between different domains through a contrastive learning mechanism with other domain prototypes. The domain prototype set is as follows: ;

[0076] in, A domain label used to uniquely identify the distribution characteristics of a specific data domain. For set P and field labels The corresponding feature prototype, and .

[0077] The prototype alignment mechanism makes the domain features output by the encoder close to the prototype of its domain, and the prototype-based contrastive learning mechanism maintains the difference between prototypes of different domains.

[0078] For the target domain label α, its feature prototype This characterizes the statistical distribution center of the domain, while the domain embedding... Configured to be in the feature space with the feature prototype Perform distribution alignment to make the domain embedded While preserving individual differences, the feature prototypes are encompassed and characterized. The defined domain feature distribution pattern.

[0079] S5. Semantic decoding step: At the receiving end, semantic reconstruction is performed based on the received semantic features and target domain features. Semantic reconstruction is represented as: ;

[0080] in, This is the decoding function of the image decoder. The communication received signal represents the semantic features of the source image. To embed a domain into a specified target domain, target domain features are introduced during the decoding phase, enabling the system to perform semantic transfer between different domains.

[0081] During semantic decoding, the latent state changes of the state-space model are influenced by domain features, and the output matrix is ​​determined by semantic features. The changes in the state-space model are represented as follows:

[0082] ;

[0083] in, For the learnable basic state matrix, Both are learnable linear projection functions, where d is the domain embedding and s is the image semantic feature. , and Together, they constitute the core dynamic state-space parameters in the Mamba model.

[0084] The state update and output of the state-space model are represented as follows: , ;

[0085] in, Represents the discrete time step index in a one-dimensional embedded sequence. Indicates the first The field label for step input is eigenvectors, Indicates the first The hidden state vector of the step. , , For dynamic state-space parameters, Current number The final output of the step is represented;

[0086] As can be seen from the formula, when outputting y, elements from both the semantic content and the domain feature sequence are obviously included in the decoding process.

[0087] S6. Downstream task processing step: Input the recovered semantic data into the downstream task model to obtain the task processing result. The downstream task model is an image classification model. ;

[0088] in, For downstream task adaptation modules, The image generated by the decoder, The predicted output for the corresponding downstream task is used to compare with the true label to evaluate the model performance.

[0089] Specifically, in use, the semantic encoding module 1 performs block embedding processing on the input image, dividing the image into multiple image blocks of size p, which are then mapped to a one-dimensional sequence representation. .

[0090] Subsequently, the sequence is input into a state-space model encoder based on the Visual Mamba architecture. This encoder maintains the latent state with linear complexity through a selective scanning mechanism, models the semantics of the image, and finally outputs compressed semantic features. ;

[0091] The compressed semantic features are power normalized by channel transmission module 2 and then transmitted through an additive white Gaussian noise channel to obtain the received semantic features. .

[0092] Meanwhile, the domain feature encoding module 3 encodes the source domain label using a neural network to obtain the domain embedding representation. Alternatively, richer domain feature representations can be extracted by inputting a reference image.

[0093] The domain embedding representation d is fed into the domain prototype alignment module 4, which establishes a set of domain prototypes containing K domains. ,in This represents the feature prototype of the k-th domain; through a prototype alignment mechanism, the model guides the domain features output by the encoder to gradually approach the domain prototype of their respective domain. This enhances the consistency of representations within the same domain, while a prototype-based contrastive learning mechanism ensures that domain prototypes from different domains remain significantly different, thus achieving effective separation of domain features.

[0094] At the receiving end, the semantic decoding module 5 performs semantic reconstruction based on the received semantic features and target domain features, represented as follows: During semantic decoding, the latent state changes of the state-space model are influenced by domain features, and the output matrix is ​​determined by semantic features, specifically set as follows: The corresponding state update and output are represented as follows: , This allows semantic content and elements in the domain feature sequence to be integrated into the decoding process during the output stage.

[0095] Finally, the restored semantic data is input into the downstream task processing module 6, which can be adapted to the image classification model. It can perform specific tasks such as object detection models or image segmentation models, and output the processing results of downstream tasks, thereby completing the end-to-end semantic communication process from multi-domain image input to task output.

[0096] Example 2:

[0097] Please see Figures 1-2 Based on the above embodiments, a multi-domain adaptive semantic communication system includes a semantic encoding module 1, a channel transmission module 2, a domain feature encoding module 3, a domain prototype alignment module 4, a semantic decoding module 5, and a downstream task processing module 6.

[0098] In this embodiment, the semantic encoding module 1 is used to divide the input image into multiple image blocks through the image embedding module, map them into a one-dimensional sequence representation, and then input them into the state space model encoder based on the Visual Mamba structure to extract the compressed semantic features.

[0099] In this embodiment, the channel transmission module 2 is used to normalize the power of the semantic features and then transmit them through an additive white Gaussian noise channel to obtain the received semantic features.

[0100] In this embodiment, the domain feature encoding module 3 is used to encode the source domain label through a neural network to obtain the domain embedding representation.

[0101] In this embodiment, the domain prototype alignment module 4 is used to establish a set of domain prototypes containing K domains, where each domain corresponds to a feature prototype. The domain embedding representation is aligned with the domain prototype of its own domain, and the features between different domains are kept separate through a comparative learning mechanism with other domain prototypes.

[0102] The domain prototype alignment module 4 firstly uses a prototype alignment mechanism to ensure that the input domain features have a high similarity to the corresponding domain prototype. The model is guided to make the style features output by the encoder gradually approach the prototype of its domain, thereby enhancing the consistency of representation within the same domain and providing a stable reference representation for domain features. Secondly, a prototype-based contrastive learning mechanism is further introduced to maintain a large difference from other domain prototypes and to maintain a clear separation structure between different domains, thereby achieving effective separation of domain features. This structured representation is particularly important for multi-domain style transfer tasks and can improve the system's ability to model multi-domain data distributions.

[0103] In this embodiment, the semantic decoding module 5 is used to perform semantic reconstruction at the receiving end based on the received semantic features and target domain features, wherein the potential state changes of the state space model are affected by the domain features, and the output matrix is ​​determined by the semantic features;

[0104] In this embodiment, the downstream task processing module 6 is used to input the recovered semantic data into the downstream task model to obtain the task processing result.

[0105] Specifically, in use, the semantic encoding module 1 first divides the input image into multiple image blocks and maps them into a one-dimensional sequence through the image embedding module, and then uses the state space model encoder based on the Visual Mamba structure to extract compressed semantic features.

[0106] The feature is normalized by the power of the channel transmission module 2 and transmitted to the receiver through the additive white Gaussian noise (AWGN) channel to obtain the received semantic feature. At the same time, the domain feature encoding module 3 encodes the source domain label to obtain the domain embedding representation, and aligns it with the domain prototype of the domain through the domain prototype alignment module 4. Meanwhile, a contrastive learning mechanism is used to maintain the difference between different domain prototypes.

[0107] The receiving semantic decoding module 5 performs semantic reconstruction based on the received semantic features and the target domain features. The potential state changes of the state space model are affected by the domain features, and the output matrix is ​​determined by the semantic features.

[0108] Finally, the restored semantic data is input to the downstream task processing module 6 to complete specific tasks such as image classification, object detection, or image segmentation.

[0109] The solution in this embodiment can be selectively combined with solutions in other embodiments.

[0110] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of this invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of this invention, or equivalent structural, procedural, or functional transformations made using the description and drawings of this invention, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of protection of this invention.

Claims

1. A multi-domain adaptive semantic communication method, characterized in that, Includes the following steps: S1, Semantic encoding step: The input image is divided into multiple image blocks by the image embedding module, mapped into a one-dimensional sequence representation, and then input into the state space model encoder based on the Visual Mamba structure to extract the compressed semantic features; S2, Channel transmission step: After power normalization of the semantic features, the signals are transmitted through an additive white Gaussian noise channel to obtain the received semantic features. S3, Domain Feature Encoding Step: The source domain label is encoded using a neural network to obtain the domain embedding representation; S4. Domain Prototype Alignment Step: Construct a set of domain prototypes, where each domain corresponds to a feature prototype. Align the domain embedding representation with the domain prototype of its own domain, and maintain the separation of features between different domains through a comparative learning mechanism with other domain prototypes. S5. Semantic decoding step: At the receiving end, semantic reconstruction is performed based on the received semantic features and target domain features. S6. Downstream task processing step: Input the recovered semantic data into the downstream task model to obtain the task processing result.

2. The multi-domain adaptive semantic communication method according to claim 1, characterized in that, In the semantic encoding step, the input image, after embedding, obtains a sequence representation as follows: ; in, For the encoded input image, a one-dimensional embedding sequence, For the encoding function of the image embedding module, The input image is H, where H and W represent the height and width of the image, respectively. , The side length of the image patch is used to determine the granularity of image segmentation; The compressed semantic features are represented as follows: ; in, For image semantic feature extraction function, Indicates semantic compression ratio. It is the amount of data in the original image. It is the amount of data for the encoded semantic features s.

3. The multi-domain adaptive semantic communication method according to claim 2, characterized in that, In the channel transmission step, the additive white Gaussian noise channel model is represented as: ; in, Indicates sending a signal. This represents the received signal after transmission through the channel; This represents the additive white Gaussian noise component in the channel, which follows a mean of 0 and a variance of . The normal distribution, i.e. The variance Power intensity used to characterize noise.

4. The multi-domain adaptive semantic communication method according to claim 1, characterized in that, In the domain feature encoding step, the domain embedding is represented as: ; in, This is a domain feature encoding function used to map the input to the domain feature embedding space. Domain labels are used to identify the distribution characteristics of target data domains. Input a tensor for an optional reference image.

5. The multi-domain adaptive semantic communication method according to claim 1, characterized in that, In the domain prototype alignment step, a global set containing K types of domain distribution prototypes is constructed: ; in, A domain label used to uniquely identify the distribution characteristics of a specific data domain. For set P and field labels The corresponding feature prototype, and ; The prototype alignment mechanism constrains the domain embeddings output by the encoder to aggregate toward the domain prototype to which they belong, and the prototype-based contrastive learning mechanism maintains the differences between different domain prototypes.

6. The multi-domain adaptive semantic communication method according to claim 1, characterized in that, In the semantic decoding step, semantic reconstruction is represented as: ; in, This is the decoding function of the image decoder. The communication received signal represents the semantic features of the source image. Domain embedding for a specified target domain; By introducing target domain features during the decoding stage, the system can perform semantic transfer between different domains.

7. The multi-domain adaptive semantic communication method according to claim 6, characterized in that, During semantic decoding, the latent state changes of the state-space model are influenced by domain embedding, and the output matrix is ​​determined by semantic features and latent states. The changes in the state-space model are represented as follows: ； in, For the learnable basic state matrix, Both are learnable linear projection functions, where d is the domain embedding and s is the image semantic feature. , and Together, they constitute the core dynamic state-space parameters in the Mamba model.

8. The multi-domain adaptive semantic communication method according to claim 7, characterized in that, In the semantic decoding step, the state update and output of the state space model are represented as follows: , ; in, Represents the discrete time step index in a one-dimensional embedded sequence. Indicates the first The field label for step input is eigenvectors, Indicates the first The hidden state vector of the step. , , For dynamic state-space parameters, Current number The final output of the step is represented; As can be seen from the formula, when outputting y, elements from both the semantic content s and the domain feature d are obviously included in the decoding process.

9. The multi-domain adaptive semantic communication method according to claim 1, characterized in that, In the downstream task processing step, the downstream task model is an image classification model: ; in, For downstream task adaptation modules, The image generated by the decoder, The predicted output for the corresponding downstream task is used to compare with the true label to evaluate the model performance.

10. A multi-domain adaptive semantic communication system, applied in the multi-domain adaptive semantic communication method according to any one of claims 1-9, characterized in that, include: The semantic encoding module (1) is used to divide the input image into multiple image blocks through the image embedding module, map them into a one-dimensional sequence representation, and then input them into the state space model encoder based on the Visual Mamba structure to extract the compressed semantic features. The channel transmission module (2) is used to normalize the power of the semantic features and then transmit them through an additive white Gaussian noise channel to obtain the received semantic features. The domain feature encoding module (3) is used to encode the source domain label through a neural network to obtain the domain embedding representation; Domain prototype alignment module (4) is used to establish a set of domain prototypes, where each domain corresponds to a feature prototype. The domain embedding representation is aligned with the domain prototype of its own domain, and the features between different domains are kept separate through a comparative learning mechanism with other domain prototypes. The semantic decoding module (5) is used to perform semantic reconstruction at the receiving end based on the received semantic features and target domain features, wherein the potential state changes of the state space model are affected by the domain features and the output matrix is ​​determined by the semantic features; The downstream task processing module (6) is used to input the recovered semantic data into the downstream task model to obtain the task processing results.

Images