Frequency-space hybrid joint source channel coding method and wireless image transmission method

Through the frequency-space hybrid joint source-channel coding method, utilizing the channel adaptive attention mechanism and the frequency-space hybrid module, the problems of channel adaptation limitations and insufficient frequency modeling in wireless image transmission are solved, and higher quality image reconstruction is achieved.

CN120639906AActive Publication Date: 2025-09-12TONGJI UNIV

Patent Information

Application Number
CN202511064920.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-31
Publication Date
2025-09-12
Estimated Expiration
2045-07-31

AI Technical Summary

Technical Problem

Existing wireless image transmission methods have limitations in channel adaptation mechanisms, making it difficult to meet the differentiated requirements of features at different levels. They also lack frequency domain modeling, which limits their performance in frequency-sensitive wireless transmission scenarios.

Method used

A frequency-space hybrid joint source channel coding method is adopted, and attention is dynamically adjusted at the encoder level through the channel adaptive attention mechanism. Combined with frequency and spatial feature modeling, adaptive processing of different channel signal-to-noise ratios is achieved.

Benefits of technology

It significantly reduces the distortion and bit error rate during wireless image transmission and improves the image reconstruction quality, especially in high-resolution and fast channel fading scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120639906A_ABST
    Figure CN120639906A_ABST
Patent Text Reader

Abstract

The invention relates to a frequency-space hybrid joint source channel coding method and an image transmission method. The method comprises the following steps: inputting an image to be transmitted into a joint source channel coder, and mapping the image into a channel input vector; an image block embedding module and a frequency space mixing module are sequentially arranged in each encoder layer of the joint source channel encoder, the frequency space mixing module independently divides a high-frequency feature and a low-frequency feature into two branches for parallel processing, and outputs of the high-frequency and low-frequency branches are spliced along a channel dimension and then are subjected to spatial feature modeling. The output of the frequency space mixing module is input into a next encoder layer; a channel adaptive attention mechanism is introduced into an encoder layer, real-time channel SNR is embedded into attention calculation of each encoder layer, and attention focuses are dynamically adjusted according to the real-time channel SNR. Compared with the prior art, the self-adaption to different channel signal-to-noise ratios is realized, and the reconstruction quality of two types of information of high-frequency texture and global semantics is effectively considered under the same code length.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of wireless image transmission, and in particular to a source-channel joint coding method for wireless image transmission and a wireless image transmission method. Background Art

[0002] In recent years, the rapid development of deep learning has significantly improved the performance of joint source-channel coding (JSCC) in wireless image transmission. Traditional communication systems separate source coding and channel coding. This separation design theoretically achieves optimal performance only when the codeword length is infinite, but in practice may be suboptimal with finite codeword lengths. To achieve optimal performance with limited codeword lengths, researchers are increasingly exploring end-to-end joint source-channel coding (JSCC) methods.

[0003] Deep learning-based JSCC methods, such as DeepJSCC and its variants, build models based on convolutional neural networks (CNNs) and perform better than traditional methods under low channel signal-to-noise ratio (SNR) and limited bandwidth conditions. ADJSCC introduces a channel attention mechanism and embeds the channel SNR into the model as global context information, enabling the model to adaptively adjust the coding strategy according to the real-time channel SNR, while significantly reducing computational complexity and storage overhead. DeepJSCC-1 and DeepJSCC-1++ adapt to different bandwidths and channel SNRs by slicing the latent code or utilizing the Swin Transformer. SwinJSCC also uses the Swin Transformer as the backbone network and adjusts the output of the Transformer stage according to the real-time channel SNR through a channel modulation network to achieve higher quality transmission and reconstructed images. However, existing methods still face the following two major limitations:

[0004] First, there are limitations to channel adaptation mechanisms. Most existing methods employ a one-size-fits-all strategy for channel adaptation across different layers of features, making it difficult to meet the differentiated requirements for low-level details and high-level semantics under different channel conditions. This hinders smooth performance transitions and weakens the robustness of features to noise.

[0005] Second, there is a lack of explicit frequency modeling. Most existing methods only extract spatial features, ignoring frequency domain information, which is crucial for wireless image transmission. CNN-based JSCC only extracts local spatial features with a limited receptive field, while SwinJSCC increases the spatial receptive field. However, for wireless image transmission, preserving both spatial and frequency features is crucial. Frequency features reflect the energy distribution in the frequency domain. The lack of frequency-aware processing ultimately limits their performance in frequency-sensitive wireless transmission scenarios. Summary of the Invention

[0006] The purpose of the present invention is to provide a source-channel joint coding method and a wireless image transmission method for end-to-end frequency-space mixing of wireless image transmission.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] As a first aspect of the present invention, a frequency-space hybrid joint source-channel coding method is provided, wherein the method maps an image to be transmitted into a channel input vector from a joint source-channel encoder;

[0009] The joint source-channel encoder includes multiple encoder layers, each of which is sequentially provided with an image embedding module and a frequency-space mixing module. The frequency-space mixing module independently divides high-frequency features and low-frequency features into two branches for parallel processing, concatenates the outputs of the high-frequency and low-frequency branches along the channel dimension, and performs spatial feature modeling. The output of the frequency-space mixing module is input into the next encoder layer.

[0010] The encoder layer also introduces a channel-adaptive attention mechanism, which embeds the real-time channel signal-to-noise ratio into the attention calculation of each encoder layer and dynamically adjusts the attention focus according to the real-time channel signal-to-noise ratio.

[0011] As a preferred technical solution, the channel adaptive attention mechanism is specifically as follows:

[0012] The signal-to-noise ratio of the channel is encoded into a high-dimensional embedding vector μ using the SNR modulator;

[0013] Project the input feature map X into query Q, key K and value V;

[0014] The high-dimensional embedding vector μ is additively fused into the query Q and key K to obtain the modulated query Q′ and key K′;

[0015] The dot product of the modulated query Q′ and key K′ is softmaxed and then multiplied by the value V to obtain the output of the adaptive attention mechanism.

[0016] As a preferred technical solution, the frequency-space hybrid module divides the input features into non-overlapping windows for the high-frequency feature branch and maps them into high-frequency query Q h , high frequency key K h and high frequency value V h ;

[0017] Using the adaptive attention mechanism, a high-dimensional embedding vector of the channel signal-to-noise ratio is obtained through the SNR modulator, and the high-frequency query and high-frequency key are modulated. The dot product of the modulated high-frequency query and high-frequency key is softmaxed and then multiplied by the high-frequency value to obtain the output of the high-frequency branch.

[0018] As a preferred technical solution, the frequency-space hybrid module reshapes the input features into a two-dimensional feature map for the low-frequency feature branch and applies average pooling for downsampling; uses the original input feature map to obtain the low-frequency query Q l , the low-frequency key K is obtained by mapping the two-dimensional feature map after pooling and downsampling l Sum V l ;

[0019] Using the adaptive attention mechanism, a high-dimensional embedding vector of the channel signal-to-noise ratio is obtained through the SNR modulator, and the low-frequency query and low-frequency key are modulated. The dot product of the modulated high-frequency query and high-frequency key is softmaxed and then multiplied by the high-frequency value to obtain the output of the low-frequency branch.

[0020] As a preferred technical solution, the SNR modulator projects the signal-to-noise ratio of the channel into the latent space through two fully connected layers; then uses 1×1 convolution to perform channel projection and map it to the attention dimension d to obtain a high-dimensional embedding vector.

[0021] As a preferred technical solution, the output of the channel adaptive attention mechanism is expressed as follows:

[0022]

[0023] Where Q′ and K′ represent the modulated query and key; softmax is the normalized exponential function; V represents the value obtained by projecting the feature map; and d represents the attention dimension.

[0024] As a preferred technical solution, the outputs of the high-frequency and low-frequency branches are spliced ​​along the channel dimension and then sent to the SwinTransformer block for spatial feature modeling.

[0025] As a preferred technical solution, the Swin Transformer block adopts an adaptive attention mechanism to replace the standard attention mechanism.

[0026] As a second aspect of the present invention, a wireless image transmission method is provided, the method steps comprising:

[0027] For the image to be transmitted, the joint source-channel encoder as described above is used to map it into a channel input vector;

[0028] Transmitting a channel input vector through a wireless channel;

[0029] The joint source-channel decoder receives the channel output vector and the channel signal-to-noise ratio and decodes to obtain a reconstructed image.

[0030] As a preferred technical solution, the structure of the joint source channel decoder is the same as that of the joint source channel encoder, with the opposite data processing direction, including a multi-level frequency-space mixing module and a channel adaptive attention mechanism; the joint source channel decoder takes the received channel output vector as input, and then uses the frequency-space mixing module to reconstruct the features in the spatial domain and frequency domain layer by layer; in each layer, the channel adaptive attention mechanism is used to dynamically adjust the attention distribution according to the signal-to-noise ratio of the current channel; finally, the decoder outputs a reconstructed image with the same size as the original image.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1) The frequency-space hybrid joint source channel coding method proposed in this invention, by introducing a channel-adaptive attention mechanism at each level of the encoder and superimposing a frequency-space hybrid module, achieves adaptation to different channel signal-to-noise ratios and effectively balances the reconstruction quality of both high-frequency texture and global semantic information under the same code length conditions, significantly reducing the distortion and bit error rate during wireless image transmission.

[0033] 2) This paper implements hierarchical channel adaptation. The channel-adaptive attention mechanism embeds the real-time channel SNR into the attention calculation of each encoder layer, thereby enabling dynamic adjustment of feature extraction throughout the encoding process. At low channel SNRs, the model focuses more on low-level local features; at high channel SNRs, the model emphasizes high-level semantic features to achieve more accurate and detailed image reconstruction.

[0034] 3) This invention introduces frequency domain modeling for frequency-space mixing to improve the quality of reconstructed images in frequency-sensitive wireless transmission scenarios. The frequency attention module is able to decouple high-frequency and low-frequency components in parallel and then connect their respective outputs. This enables the model to simultaneously capture fine-grained local details (high-frequency features) and global structural information (low-frequency features). Through the frequency-space mixing module, collaborative feature extraction in the frequency and spatial domains is achieved, achieving higher-quality wireless image transmission and reconstructed images from a dual-dimensional perspective. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The figure is a schematic diagram of a wireless image transmission process based on source-channel joint coding according to the present invention.

[0036] Figure 2 Schematic diagram of the channel adaptive attention mechanism of the present invention.

[0037] Figure 3This is a schematic diagram of the frequency modeling module of the present invention.

[0038] Figure 4 Schematic diagram of the overall architecture of the wireless image transmission method of the present invention.

[0039] Figure 5 The experimental results of the proposed solution on different datasets and indicators are shown below. a) PSNR indicator comparison on the CIFAR10 dataset, b) PSNR indicator comparison on the CelebA-HQ dataset, c) MS-SSIM indicator comparison on the CIFAR10 dataset, d) MS-SSIM indicator comparison on the CelebA-HQ dataset. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0041] Example 1

[0042] The present invention discloses a frequency-space hybrid joint source channel coding method for wireless image transmission, the specific implementation of which includes the following technical solutions: The overall architecture of the method is as follows: Figure 1 shown.

[0043] The image to be transmitted x∈R H×w×3 Input the joint source channel encoder E designed by the present invention φ , and maps it to the channel input vector z∈C k The encoding process can be expressed as: z = E φ (x,γ)∈C k .

[0044] The joint source-channel encoder consists of multiple layers, each integrating a frequency-space hybrid module and a channel-adaptive attention mechanism (CAAM). First, the image input is split into small blocks and feature embedded in the image embedding module. Then, within each encoder layer, the feature embedding is separated into high-frequency and low-frequency branches by a frequency modeling module, extracting detailed texture and overall structural information, respectively. The modeling is then performed using the channel-adaptive attention mechanism. The outputs of the high-frequency and low-frequency branches are concatenated and fused along the channel dimension to generate frequency-domain features. The fused frequency-domain features are then fed into a Swin Transformer block for spatial modeling to capture local and global spatial relationships in the image. Furthermore, each attention module embeds the current channel signal-to-noise ratio. CAAM dynamically adjusts the attention distribution across different layers, enabling the model to flexibly allocate information between low-level details and high-level semantics based on channel conditions. Finally, the features output by the multi-layer encoder are integrated into a single channel input vector that satisfies power constraints, achieving efficient mapping from image to channel representation.

[0045] To ensure robustness of each layer of encoded features, this paper designs a channel-adaptive attention mechanism that embeds the real-time channel signal-to-noise ratio (SNR) into the attention calculation, achieving hierarchical adaptive weighting of features at different encoder layers. This enables the model to dynamically adjust its attention focus based on the instantaneous channel SNR, thereby adapting to the changing channel SNR in the early stages of feature modeling, enhancing its robustness and adaptability in JSCC scenarios.

[0046] like Figure 1 As shown in Figure 1, the channel adaptive attention mechanism CAAM contains an SNR modulator that encodes the signal-to-noise ratio (SNR) γ of the channel into a high-dimensional embedding vector μ. Specifically, the SNR modulator projects the SNR into the latent space through two fully connected layers with GELU activations and uses 1×1 convolution for channel projection, which is directly mapped to the attention dimension d. The channel adaptive attention mechanism converts the input feature map X∈R N×d The projection is into query Q, key K, and value V. The high-dimensional embedding vector μ obtained by the SNR modulator is additively fused to the query Q and key K to obtain modulated Q′ and K′. The dot product of modulated Q′ and K′ is softmaxed and then multiplied by the value V to obtain the output of CAAM. The overall process can be expressed as follows:

[0047] Q=XW Q ,K=XW K ,V=XW V ,

[0048] Q′=Q+μ,

[0049] K′=K+μ,

[0050]

[0051] The present invention also designs a frequency-space hybrid module for modeling in the frequency domain, dividing high-frequency and low-frequency features into two branches for independent processing. This module uses a two-stage process: frequency processing first, followed by spatial processing. This module can simultaneously consider both feature and spatial characteristics, enabling the model to utilize both frequency and spatial information for transmission, thereby improving the transmission quality of datasets of different types and resolutions.

[0052] like Figure 3 As shown in Figure 2, the high-frequency branch divides the input features into non-overlapping windows and maps them to high-frequency queries Q h , key K h Sum V h Combined with the SNR modulator, the high-dimensional embedding vector μ of the signal-to-noise ratio is used to modulate the high-frequency query Q′ h and key K′ h , and then calculate the output Z of the high-frequency branch through CAAM h For the low-frequency branch, the input features are first reshaped into a two-dimensional feature map and average pooling is applied for downsampling. The mapping of the original input features is used as the low-frequency query Q l , the mapping of the two-dimensional feature map after pooling downsampling is used as the low-frequency key K l and low frequency value V l Similarly, the low-frequency query Q′ is modulated using the SNR modulator l and low-frequency key K′ l , and then calculate the output Z of the low-frequency branch through CAAM l . Preferably, the outputs of the high-frequency and low-frequency branches are concatenated along the channel dimension and fused through a 1×1 convolution before outputting them.

[0053] The output of the frequency domain modeling is fed into the Swin Transformer block (SwinT) for spatial feature modeling. The standard attention mechanism in SwinT is replaced by CAAM, which enhances the model's adaptability to different channel SNRs.

[0054] Example 2

[0055] As another embodiment of the present invention, this embodiment also provides a wireless image transmission method based on the source-channel joint coding method described in the above embodiment 1. Figure 3 The specific steps are as follows:

[0056] Step 1: Encoding of input image and generation of channel input vector.

[0057] The image to be transmitted x∈R H×W×3 After the joint source-channel encoder E φ , is mapped to the channel input vector z∈C k The encoding process can be expressed as: z = E φ (x,γ)∈C k A channel-adaptive attention mechanism is introduced at each encoder layer to enhance the model's adaptability to different channel conditions, and a frequency-space mixing module is used to simultaneously capture features in both the frequency and spatial domains.

[0058] Step 2: Channel transmission.

[0059] The channel input vector z is transmitted through a wireless channel, including but not limited to an AWGN channel. In this embodiment, the AWGN channel is taken as an example. The output vector z can be expressed as:

[0060] Step 3: Image decoding and reconstruction.

[0061] Joint source-channel decoder D ψ Receive channel output vector and channel SNRγ, reconstructed image The decoding process can be expressed as: Finally, by minimizing the original image x and the reconstructed image The distortion is measured by the mean square error (MSE) between them.

[0062] The overall structure of the joint source-channel decoder is similar to the joint source-channel encoder described above, both employing a multi-layer frequency-space hybrid module and a channel-adaptive attention mechanism (CAAM), but differ in their data processing direction and functional objectives. The decoder takes the received channel vector as input and then reconstructs features in the spatial and frequency domains layer by layer, gradually improving the spatial resolution and semantic information of the features. Within each layer, the CAAM dynamically adjusts the attention distribution based on the signal-to-noise ratio of the current channel, enabling the model to effectively focus on reliable features under different channel conditions, enhancing the robustness and quality of image reconstruction. Ultimately, the decoder outputs a reconstructed image with the same size as the original image.

[0063] The overall architecture of the wireless image transmission method is as follows Figure 4 shown.

[0064] To verify the effectiveness of this method, the proposed model is trained and evaluated on the CIFAR10 and CelebA-HQ datasets. The CIFAR10 dataset contains 50,000 RGB color images of 32x32 pixels. The CelebA-HQ dataset contains 30,000 face images, which are randomly cropped to 256x256 pixels during training. The model uses the Adam optimizer with a learning rate of 1×10 -4 The batch size for training the CIFAR10 dataset is 128, and the batch size for training the CelebA-HQ dataset is 8. The model is trained under a channel SNR of [1, 4, 7, 10, 13] dB. The method of this patent is compared with DeepJSCC and its adaptive variant ADJSCC, as well as the Transformer-based SwinJSCC (especially the channel-adaptive version SwinJSCC_w / _SA). The performance of this patent method is evaluated using the peak signal-to-noise ratio (PSNR) and the perceptual multi-scale structural similarity index (MS-SSIM). The experimental results are shown in Figure 2. Figure 5 shown.

[0065] The present invention introduces a channel-adaptive attention mechanism at each level of the encoder and superimposes a frequency-space hybrid module, thereby achieving adaptation to different channel signal-to-noise ratios and effectively taking into account the reconstruction quality of both high-frequency texture and global semantic information under the same code length conditions, significantly reducing the distortion and bit error rate during wireless image transmission. Experimental results show that in a 1dB-13dB wide dynamic channel environment, this method outperforms existing comparison schemes such as DeepJSCC, ADJSCC, and SwinJSCC_w / _SA in comprehensive indicators such as PSNR and MS-SSIM, especially in high-resolution and fast channel fading scenarios. It can be seen that this method not only verifies the performance upper limit of deep learning joint source channel coding under limited code length, but also provides an industrially applicable technical solution for mobile terminals, satellite / UAV links, and other image transmission applications with limited bandwidth or poor channel conditions.

[0066] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0067] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A frequency-space hybrid joint source channel coding method, characterized in that: The method maps the image to be transmitted into a joint source-channel encoder into a channel input vector; The joint source-channel encoder includes multiple encoder layers, each of which is sequentially provided with an image embedding module and a frequency-space mixing module. The frequency-space mixing module independently divides high-frequency features and low-frequency features into two branches for parallel processing, concatenates the outputs of the high-frequency and low-frequency branches along the channel dimension, and performs spatial feature modeling. The output of the frequency-space mixing module is input into the next encoder layer. The encoder layer also introduces a channel-adaptive attention mechanism, which embeds the real-time channel signal-to-noise ratio into the attention calculation of each encoder layer and dynamically adjusts the attention focus according to the real-time channel signal-to-noise ratio.

2. The frequency-space hybrid joint source channel coding method according to claim 1, characterized in that: The channel adaptive attention mechanism is specifically as follows: The signal-to-noise ratio of the channel is encoded into a high-dimensional embedding vector μ using the SNR modulator; Project the input feature map X into query Q, key K and value V; The high-dimensional embedding vector μ is additively fused into the query Q and key K to obtain the modulated query Q ′ and key K ′ ; The modulated query Q ′ and key K ′ The dot product of is softmaxed and then multiplied by the value V to get the output of the adaptive attention mechanism.

3. The frequency-space hybrid joint source channel coding method according to claim 2, characterized in that: The frequency-space hybrid module divides the input features into non-overlapping windows for the high-frequency feature branch and maps them into high-frequency queries Q h , high frequency key K h and high frequency value V h ; Using the adaptive attention mechanism, a high-dimensional embedding vector of the channel signal-to-noise ratio is obtained through the SNR modulator, and the high-frequency query and high-frequency key are modulated. The dot product of the modulated high-frequency query and high-frequency key is softmaxed and then multiplied by the high-frequency value to obtain the output of the high-frequency branch.

4. The frequency-space hybrid joint source channel coding method according to claim 2, characterized in that: The frequency-space hybrid module reshapes the input features into a two-dimensional feature map and applies average pooling for downsampling for the low-frequency feature branch; uses the original input feature map to obtain the low-frequency query Q l , the low-frequency key K is obtained by mapping the two-dimensional feature map after pooling and downsampling l Sum V l ; Using the adaptive attention mechanism, a high-dimensional embedding vector of the channel signal-to-noise ratio is obtained through the SNR modulator, and the low-frequency query and low-frequency key are modulated. The dot product of the modulated high-frequency query and high-frequency key is softmaxed and then multiplied by the high-frequency value to obtain the output of the low-frequency branch.

5. A frequency-space hybrid joint source channel coding method according to any one of claims 2 to 4, characterized in that: The SNR modulator projects the signal-to-noise ratio of the channel into the latent space through two fully connected layers; then uses 1×1 convolution to perform channel projection and map it to the attention dimension d to obtain a high-dimensional embedding vector.

6. The frequency-space hybrid joint source channel coding method according to claim 5, characterized in that: The output of the channel-adaptive attention mechanism is expressed as follows: Where Q ′ , K ′ Represents the modulated query and key; softmax is the normalized exponential function; V represents the value obtained by projecting the feature map; d represents the attention dimension.

7. The frequency-space hybrid joint source channel coding method according to claim 1, characterized in that: The outputs of the high-frequency and low-frequency branches are concatenated along the channel dimension and fed into the Swin Transformer block for spatial feature modeling.

8. The frequency-space hybrid joint source channel coding method according to claim 7, characterized in that: The Swin Transformer block adopts an adaptive attention mechanism to replace the standard attention mechanism.

9. A wireless image transmission method, characterized in that: The method steps include: For an image to be transmitted, mapping it into a channel input vector using the joint source-channel encoder according to any one of claims 1 to 8; Transmitting a channel input vector through a wireless channel; The joint source-channel decoder receives the channel output vector and the channel signal-to-noise ratio and decodes to obtain a reconstructed image.

10. The wireless image transmission method according to claim 9, characterized in that: The joint source channel decoder has the same structure as the joint source channel encoder, but with the opposite data processing direction. It includes a multi-level frequency-space hybrid module and a channel-adaptive attention mechanism. The joint source channel decoder takes the received channel output vector as input and then uses the frequency-space hybrid module to reconstruct features in the spatial and frequency domains layer by layer. In each layer, the channel-adaptive attention mechanism is used to dynamically adjust the attention distribution according to the signal-to-noise ratio of the current channel; Finally, the decoder outputs a reconstructed image with the same size as the original image.

Citation Information

Patent Citations

  • Image semantic end-to-end transmission method based on double attention mechanism

    CN117746211A

  • Semantic communication method based on multi-scale feature adaptive fusion

    CN118585949A

  • Cooperative non-orthogonal image semantic communication method and system based on information deentanglement

    CN118803277A

  • 3D wavelet based video codec with human perceptual model

    WO2000074385A2

  • Adaptive line equalizer for improving data communication over less than perfect power lines or transmission lines

    WO2014145044A2

Cited By

  • Adaptive rate wireless image communication method based on channel attention and mask auto-encoder

    CN121334375A