Deep semantic communication method based on mixed long-range and local modeling and fusion channel-space attention

Abstract

The invention relates to a deep semantic communication method based on mixed long-range and local modeling and fusion channel-space attention, and belongs to the technical field of communication and artificial intelligence, and the method comprises the following steps: S1, receiving a voice waveform signal; s2, performing multi-level semantic feature extraction on the voice waveform signal through a feature extraction network combining depth separable convolution and state space modeling; s3, fusing the multi-level semantic features by adopting a parallel channel-space attention fusion mechanism; s4, encoding the semantic features processed by the feature extraction network and the channel-space attention fusion mechanism; s5, transmitting the coded signal through a channel; s6, receiving the coded signal, and decoding the coded signal by adopting a decoder combined with a cross-layer feature fusion mechanism so as to reconstruct a voice waveform signal; and S7, keeping the robustness of voice transmission through the encoder and the decoder under the condition of low signal-to-noise ratio.

Description

Technical Field

[0001] This invention belongs to the field of communication and artificial intelligence technology, and relates to a deep semantic communication method based on hybrid long-range and local modeling and fusion channel-space attention. Background Technology

[0002] Voice communication, as an important component of modern wireless communication systems, continues to expand in data volume with the popularization of mobile networks, smart terminals, and IoT devices. However, in actual wireless networks, voice signals are often affected by factors such as multipath fading, noise interference, and channel estimation errors, gradually exposing the bottlenecks of traditional voice transmission schemes.

[0003] Current mainstream systems generally adopt a source coding and channel coding separation architecture. This paradigm compresses the speech signal into a bitstream and then uses channel coding to enhance transmission robustness. Although this architecture has advantages in engineering implementation, in scenarios where the channel degrades or the signal-to-noise ratio drops below 10dB, the error correction capability of channel coding deteriorates sharply, and the speech quality plummets to an unusable level, exhibiting a typical "cliff effect." This not only affects the continuity of communication but also makes it difficult to meet the extremely reliability requirements of applications such as real-time voice interaction.

[0004] To address this issue, joint source-channel coding methods have emerged. These methods utilize end-to-end neural networks to jointly optimize the source and channel coding processes, mitigating quality abrupt changes during speech transmission to some extent and achieving a smoother "graceful degradation." However, existing joint source-channel coding methods still face several technical challenges: long speech sequences are difficult to model effectively, especially under extremely low signal-to-noise ratio conditions, where global features are easily corrupted; deep feature representation is insufficient, making recovery difficult when key semantic cues are lost; high model complexity leads to high inference overhead and latency in real-time communication scenarios; and the lack of effective cross-layer semantic information fusion mechanisms prevents the full utilization of multi-scale speech features in the encoder.

[0005] Based on the above problems, existing technologies still cannot simultaneously achieve semantic robustness under low signal-to-noise ratio and the efficiency requirements of real-time transmission. There is an urgent need for a new semantic communication architecture that achieves comprehensive breakthroughs in structure, modeling capabilities, and feature fusion mechanisms. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a deep semantic communication method based on hybrid long-range and local modeling and fusion of channel-space attention. This invention achieves high semantic fidelity voice transmission even in harsh channel environments by constructing a novel neural network structure during semantic encoding and decoding, while significantly reducing computational complexity and inference latency. This overcomes the technical bottlenecks of traditional end-to-end semantic communication, such as susceptibility to noise interference, difficulty in modeling long sequences, and insufficient feature extraction.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep semantic communication method based on hybrid long-range and local modeling and fused channel-space attention includes the following steps: S1: Receives voice waveform signals; S2: Multi-level semantic feature extraction is performed on the speech waveform signal by a feature extraction network that combines deep separable convolution and state space modeling; S3: A parallel channel-space attention fusion mechanism is used to fuse the multi-level semantic features; S4: Encode the semantic features processed by the feature extraction network and the channel-space attention fusion mechanism; S5: Transmit the encoded signal through the channel; S6: Receive the encoded signal and use a decoder that incorporates a cross-layer feature fusion mechanism to decode the encoded signal in order to reconstruct the speech waveform signal; S7: Maintains robust voice transmission through encoders and decoders under low signal-to-noise ratio conditions.

[0008] Furthermore, in step S2, the feature extraction network uses the depthwise separable convolutional-selective state-space module DWmamba to model long-sequence speech. Specifically, it models local temporal dependencies through a depthwise separable convolutional layer and captures global long-range dependencies through a selective state-space model. The steps are as follows: S21: Let the input feature sequence be... , For batch size, For sequence length, To increase the feature dimension, the input signal is first normalized, and then the feature dimension is increased by passing it through a linear projection layer.

[0009] in Represents an up-dimensional mapping matrix. Indicates the dimensionality increase factor; S22: Feature Sequence Input depthwise separable convolutions can be used to model local temporal dependencies.

[0010] S23: Local features Input into the selective state-space model Mamba:

[0011] S24: Global features Projecting back to the original feature dimension:

[0012] in Represents the dimension reduction mapping matrix. ; The entire process is simplified to: .

[0013] Furthermore, in step S3, the channel-space attention fusion mechanism fully utilizes multi-dimensional features by performing channel attention modeling and spatial attention modeling on the multi-level semantic features respectively. Specifically, it includes the following steps: S31: Input is a feature sequence ,in For frame number, For frame length, Given the number of channels; first, apply global max pooling and global average pooling along the spatial dimensions respectively to obtain two channel description vectors:

[0014] The two description vectors are used to model the inter-channel dependencies through a multilayer perceptron with shared weights to obtain the channel attention weights, which are expressed as follows:

[0015] in , express function; S32: Apply global max pooling and global average pooling along the channel dimension to obtain two spatial description vectors:

[0016] in ; S33: Obtained by concatenating the channel attention weights and spatial attention weights:

[0017] S34: Contextual information is captured using a multilayer perceptron; S35: Segment the fused features to restore the original channel attention weight shape. Spatial attention weight shape ; and will and The final attention weights are obtained by activating the functions separately:

[0018] S36: Integrating attention by weighted modulation of input features: .

[0019] Furthermore, the cross-layer feature fusion mechanism described in step S6 concatenates features from different stages of the encoder with features from corresponding stages of the decoder along the channel dimension to preserve the detailed information of the original signal. Specifically, it includes: The feature map from the current stage of the decoder is concatenated with the cross-layer feature map of the corresponding level in the encoder along the channel dimension; Use a 1x1 convolutional layer to compress the channels; The output of the fused convolution is passed through an activation function; An additional convolutional layer is used to further reduce the number of channels in the feature map; For the decoder side stage1 Output feature sequence Corresponding encoder stage3 The output hop connection sequence through the channel is ,in For frame number, For frame length, The number of channels is described as follows: .

[0020] The beneficial effects of this invention are as follows: By combining deep separable convolution with a selective state-space architecture and a channel-space attention fusion mechanism, this invention effectively solves the problem of poor voice transmission performance under low signal-to-noise ratio conditions in existing technologies while ensuring channel robustness. This invention provides a high-performance, highly robust deep semantic communication method that meets the real-time voice communication requirements in complex communication scenarios. This invention utilizes deep neural networks to model continuous speech waveforms, achieving efficient semantic feature extraction and robust coding, and employs a joint source-channel coding mechanism to ensure robust transmission under low signal-to-noise ratio conditions. The system is trained and tested using the LibriSpeech and Edinburgh datasets, and the results show that this invention effectively overcomes the "cliff effect" of traditional communication.

[0021] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 System architecture diagram; Figure 2 Here is a structural diagram of DWMamba; Figure 3 This is a structural diagram of the channel-space fusion attention module; Figure 4 This is a structural diagram of cross-layer feature fusion; Figure 5 Training flowchart; Figure 6 This paper compares the performance of the present invention with that of existing technologies through a Gaussian white noise channel, where (a) is the perceived evaluation of speech quality, (b) is the signal distortion ratio, and (c) is the root mean square error. Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0025] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0026] Example 1: This invention provides a deep semantic communication method based on hybrid long-range and local modeling and fusion of channel-space attention for end-to-end voice transmission over a wireless channel, comprising the following steps: S1: Receives voice waveform signals; S2: A lightweight feature extraction network combining deep separable convolution and state space modeling is introduced for multi-level semantic feature extraction of the speech waveform signal. The feature extraction network employs the Deep Separable Convolution-Selective State Space Model (DWmamba) module, which combines deep separable convolution with a state space model for efficient modeling of long speech sequences. Specifically, the DWmamba module models local temporal dependencies through a deep separable convolutional layer and captures global long-range dependencies through a selective state space model module.

[0027] S3: A parallel channel-space attention fusion mechanism is employed to fuse the multi-level semantic features. This mechanism performs channel attention modeling and spatial attention modeling on the multi-level semantic features respectively, thereby fully utilizing the multi-dimensional features. Specifically, the channel-space attention fusion mechanism further fuses the channel attention weights and spatial attention weights through a multilayer perceptron to enhance contextual information.

[0028] S4: Encode the semantic features processed by the feature extraction network and the channel-space attention fusion mechanism; S5: Transmit the encoded signal through the channel; S6: Receive the encoded signal and decode it using a decoder incorporating a cross-layer feature fusion mechanism to reconstruct the speech waveform signal. The cross-layer feature fusion mechanism concatenates features from different stages of the encoder with features from corresponding stages of the decoder along the channel dimension to preserve the detailed information of the original signal. Specifically, the cross-layer feature fusion mechanism further includes a fusion convolutional layer for channel compression and information integration of the concatenated features.

[0029] S7: Under low signal-to-noise ratio conditions, the codec maintains robustness of voice transmission.

[0030] Example 2: The network architecture used in this method is as follows: Figure 1 As shown, this architecture utilizes deep neural networks to model continuous speech waveforms, achieving efficient semantic feature extraction and robust coding. It also employs a joint source-channel coding mechanism to ensure robust transmission under low signal-to-noise ratio conditions. The system is trained and tested using the LibriSpeech and Edinburgh datasets, and the results demonstrate that the system effectively overcomes the "cliff effect" of traditional communication. In its specific implementation, this invention is described in detail from four parts: system architecture, the DWmamba module, the channel-space fusion attention module, and the cross-layer feature fusion module.

[0031] S1: At the transmitting end, the system input is the raw speech signal, which is sampled to obtain a discrete-time sequence, which can be represented as follows: ,in Indicates the number of sampling points.

[0032] S2: Voice signal After semantic encoder Mapped to a sequence of semantic symbols

[0033] S3: Symbol sequence before transmission Normalized to satisfy the average transmit power constraint:

[0034] S4: The signal passes through the physical channel After transmission, the observed signal at the receiving end can be modeled as

[0035] The fading factor is expressed as: , , representing additive Gaussian noise, with a noise power of , It is an identity matrix. " indicates a convolution operation.

[0036] S5: Receiver semantic decoder According to the received signal Reconstructing the speech:

[0037] Among them, Represents the trainable parameters of the semantic encoder. This is the reconstructed signal.

[0038] S6: Unlike traditional communication systems, semantic communication systems consider the existence of a semantic gap between the original audio and the reconstructed audio in the semantic domain. This patent treats the speech recovery process as a signal reconstruction task, utilizing the characteristics of speech signals to minimize the error between signal values. The weight parameters of the neural network are represented as semantic encoder and semantic encoder, respectively. The system's loss function can be expressed as:

[0039] in These represent the number of input speech channels, the number of frames, and the frame length, respectively. This represents the sample point at each time step in the sample. This is to maintain the stability of the loss function. Set to 0.0001.

[0040] like Figure 2 As shown, the DWmamba module used in this invention combines the advantages of depthwise separable convolution and selective state-space models. This module is mainly used for efficiently modeling long speech sequences and capturing local and long-range dependency features. Its inference process is as follows: S21: Let the input feature sequence be... , For batch size, For sequence length, The feature dimension is defined as follows. To improve the model's expressive power, the input is first normalized to eliminate scale differences between different samples and feature dimensions, and then the feature dimension is increased through a linear projection layer.

[0041] in Represents an up-dimensional mapping matrix. This indicates the dimensionality increase factor. This dimensionality increase operation increases the feature representation capability, enabling subsequent sequence modeling to be performed in a higher-dimensional space.

[0042] S22: Feature Sequence Input depthwise separable convolutions can be used to model local temporal dependencies.

[0043] Depthwise separable convolution divides convolution into two steps: channel-wise convolution and point-wise convolution. This can capture the temporal dependencies between adjacent features while maintaining computational efficiency, thus enabling efficient capture of local patterns.

[0044] S23: To further model long-range dependencies, local features Input into the selective state-space model Mamba module.

[0045]

[0046] Selective state-space models achieve global sequence modeling through an efficient state-space mechanism, thereby overcoming the limitations of convolution.

[0047] S24: Finally, global features Projecting back to the original feature dimension

[0048] in Represents the dimension reduction mapping matrix. This design ensures that no original information is lost while enhancing feature representation capabilities, thereby improving the model's robustness and generalization ability. The entire process can be simplified to...

[0049] like Figure 3 As shown, in the channel-spatial fusion attention module of the present invention, the input is a feature sequence. ,in For frame number, For frame length, The number of channels. Channel-spatial fusion mainly includes three steps: channel attention modeling, spatial attention modeling, and feature fusion.

[0050] S31: Channel Attention: First, apply global max pooling and global average pooling along the spatial dimension to obtain two channel description vectors:

[0051] Subsequently, these two descriptive vectors are passed through a multilayer perceptron with shared weights to model inter-channel dependencies and obtain channel attention weights. The above process can be represented as:

[0052] in , express function.

[0053] S32: Spatial Attention: Similarly, global max pooling and global average pooling are applied to the channel dimension to obtain two spatial description vectors.

[0054]

[0055] in .

[0056] S33: Feature Fusion: In order to fully integrate channel and spatial attention, the channel attention weights and spatial attention weights are first concatenated.

[0057]

[0058] Subsequently, A multilayer perceptron is used to capture richer contextual information.

[0059] Next, the fused features are segmented to restore the original channel attention weight shape. Spatial attention weight shape And will and The final attention weights are obtained by activating the functions respectively.

[0060]

[0061] Finally, attention is fused by weighted modulation of input features:

[0062] The structure diagram of the cross-layer feature fusion module is as follows: Figure 4 As shown, the model adopts a three-level connection mechanism, which directly transmits the feature maps of each stage of the encoder to the corresponding upsampling stage in the decoder to achieve effective information fusion.

[0063] S41: The feature map from the current stage of the decoder is concatenated with the cross-layer feature map from the corresponding level in the encoder along the channel dimension. This operation combines deep semantic information with shallow spatial detail information to form a more comprehensive feature representation.

[0064] S42: The number of channels in the concatenated feature map is doubled. To restore the original size and further fuse information, we use a 1x1 convolutional layer to compress the channels. This not only reduces computational complexity but also allows the model to learn how to optimally integrate features from two different levels.

[0065] S43: The output of the fused convolution is then passed through an activation function to improve the training stability of the model.

[0066] S44: Finally, we use an additional convolutional layer to further reduce the number of channels in the feature map to match the input requirements of the next stage of the decoder. This step ensures the dimensionality consistency of the features during the transfer process, preparing for subsequent upsampling operations. exist Figure 1 Decoder end stage1 Output feature sequence Corresponding encoder stage3 The output hop connection sequence through the channel is ,in For frame number, For frame length, If the number of channels is given, the above process can be described as follows:

[0067] Figure 5 The training process of this model is illustrated in the diagram. To verify the technical advantages of this invention, we compared the audio transmission performance of the trained model with that of several representative traditional source-channel separation coding models. The performance metrics for comparison included: perceptual evaluation of speech quality, with a score ranging from -0.5 to 4.5, where a higher score indicates better speech quality; signal-to-distortion ratio, which measures the energy ratio between the original signal and the distorted signal, with a higher value being better; and root mean square error, with a lower value being better.

[0068] Traditional Model 1: Uses 8-bit PCM source coding, Turbo code channel coding (coding rate of 1 / 2), and 16QAM modulation; Traditional Model 2: Uses 8-bit PCM source coding, LDPC code channel coding (coding rate of 3 / 4), and 64QAM modulation; Traditional Model 3: Uses 8-bit PCM source coding, Turbo code channel coding (coding rate of 1 / 4), and QPSK modulation.

[0069] The final experimental results are shown in the figure below. Figure 6 As shown in (a)-(c), experiments have verified that this system can maintain stable voice quality under adverse channel conditions, effectively overcoming the "cliff effect" in traditional communication.

[0070] Example 3: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0071] Example 4: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0072] Example 5: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0073] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0074] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0075] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0076] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0077] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0078] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0079] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0080] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep semantic communication method based on hybrid long-range and local modeling and fusion of channel-space attention, characterized in that: Includes the following steps: S1: Receives voice waveform signals; S2: Multi-level semantic feature extraction is performed on the speech waveform signal by a feature extraction network that combines deep separable convolution and state space modeling; S3: A parallel channel-space attention fusion mechanism is used to fuse the multi-level semantic features; S4: Encode the semantic features processed by the feature extraction network and the channel-space attention fusion mechanism; S5: Transmit the encoded signal through the channel; S6: Receive the encoded signal and use a decoder that incorporates a cross-layer feature fusion mechanism to decode the encoded signal in order to reconstruct the speech waveform signal; S7: Maintains robust voice transmission through encoders and decoders under low signal-to-noise ratio conditions.

2. The deep semantic communication method based on hybrid long-range and local modeling and fusion of channel-space attention as described in claim 1, characterized in that: In step S2, the feature extraction network uses the depthwise separable convolutional-selective state-space module DWmamba to model long-sequence speech. Specifically, it models local temporal dependencies through a depthwise separable convolutional layer and captures global long-range dependencies through a selective state-space model. The steps are as follows: S21: Let the input feature sequence be... , For batch size, For sequence length, To increase the feature dimension, the input signal is first normalized, and then the feature dimension is increased by passing it through a linear projection layer. in Represents an up-dimensional mapping matrix. Indicates the dimensionality increase factor; S22: Feature Sequence Input depthwise separable convolutions can be used to model local temporal dependencies. S23: Local features Input into the selective state-space model Mamba: S24: Global features Projecting back to the original feature dimension: in Represents the dimension reduction mapping matrix. ; The entire process is simplified to: 。 3. The deep semantic communication method based on hybrid long-range and local modeling and fusion of channel-space attention as described in claim 1, characterized in that: In step S3, the channel-space attention fusion mechanism fully utilizes multi-dimensional features by performing channel attention modeling and spatial attention modeling on the multi-level semantic features respectively. Specifically, it includes the following steps: S31: Input is a feature sequence ,in For frame number, For frame length, Given the number of channels; first, apply global max pooling and global average pooling along the spatial dimensions respectively to obtain two channel description vectors: The two description vectors are used to model the inter-channel dependencies through a multilayer perceptron with shared weights to obtain the channel attention weights, which are expressed as follows: in , express function; S32: Apply global max pooling and global average pooling along the channel dimension to obtain two spatial description vectors: in ; S33: Obtained by concatenating the channel attention weights and spatial attention weights: S34: Contextual information is captured using a multilayer perceptron; S35: Segment the fused features to restore the original channel attention weight shape. Spatial attention weight shape ; and will and The final attention weights are obtained by activating the functions separately: S36: Integrating attention by weighted modulation of input features: 。 4. The deep semantic communication method based on hybrid long-range and local modeling and fusion of channel-space attention as described in claim 1, characterized in that: The cross-layer feature fusion mechanism described in step S6 concatenates features from different stages of the encoder with features from corresponding stages of the decoder along the channel dimension to preserve the detailed information of the original signal. Specifically, it includes: The feature map from the current stage of the decoder is concatenated with the cross-layer feature map of the corresponding level in the encoder along the channel dimension; Use a 1x1 convolutional layer to compress the channels; The output of the fused convolution is passed through an activation function; An additional convolutional layer is used to further reduce the number of channels in the feature map; For the decoder side stage1 Output feature sequence Corresponding encoder stage3 The output hop connection sequence through the channel is ,in For frame number, For frame length, The number of channels is described as follows: 。 5. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the deep semantic communication method based on hybrid long-range and local modeling and fused channel-space attention as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the deep semantic communication method based on hybrid long-range and local modeling and fused channel-space attention as described in any one of claims 1-4.

7. A computer program product, characterized in that: Includes a computer program that, when executed by a processor, implements the deep semantic communication method based on hybrid long-range and local modeling and fused channel-space attention as described in any one of claims 1-4.

Images