Detail reconstruction image transmission method and system oriented to low compression ratio

Abstract

The invention discloses a detail reconstruction image transmission method and system oriented to a low compression ratio, and relates to the technical field of wireless communication, and the method comprises the steps: obtaining an input image, inputting the input image into a pre-established hybrid architecture model for semantic feature extraction, and outputting to obtain a hybrid feature map, local texture features and global semantic features are fused in the mixed feature map; performing dimensionality reduction on the mixed feature map through symbol compression to obtain a low-dimensional symbol representation sequence; inputting the low-dimensional symbol representation sequence into a pre-established preliminary reconstruction model for symbol decoding and global semantic feature recovery to generate a preliminarily reconstructed image, inputting the preliminarily reconstructed image into a pre-established detail enhancement network model, and gradually recovering high-frequency detail features of the image through a multi-layer dense residual structure to obtain a reconstructed image; and finally, a high-quality reconstructed image is output, so that the image transmission quality under the condition of a low compression ratio is improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a method and system for transmitting detailed reconstructed images with low compression ratios. Background Technology

[0002] While existing image transmission methods have made some progress, several challenges remain. First, image reconstruction quality often degrades significantly at low compression ratios. Current decoder designs, despite their good performance in information compression, often fail to retain sufficient detail during image reconstruction, leading to blurring or feature loss and impacting visual quality. Second, training deep learning models for image transmission still faces numerous challenges. As the complexity of model ensembles increases, the effects of channel embedding and error propagation become more pronounced. These issues can slow convergence during training, resulting in unsatisfactory training results and insufficient model robustness, ultimately affecting the overall performance of the semantic communication system. Furthermore, although existing image transmission systems have achieved initial validation in simulated channels, most studies have not yet fully validated them under real-world wireless channel conditions. Real-world channels may exhibit various complex factors such as multipath fading and frequency-selective fading, which can cause traditional systems to perform far worse in real-world environments than in simulated channels. Therefore, it is necessary to validate the performance of image transmission systems under real-world channel conditions to ensure image transmission quality under low compression ratios. Summary of the Invention

[0003] To address the shortcomings mentioned in the background art, the present invention aims to provide a method and system for transmitting detailed reconstructed images with low compression ratios.

[0004] Firstly, the objective of this invention can be achieved through the following technical solution: a method for transmitting detail-reconstructed images with low compression ratios, the method comprising the following steps: The input image is acquired and fed into a pre-established hybrid architecture model for semantic feature extraction, outputting a hybrid feature map, which integrates local texture features and global semantic features; the hybrid feature map is then reduced in dimensionality through symbolic compression to obtain a low-dimensional symbolic representation sequence. The low-dimensional symbolic representation sequence is input into a pre-established preliminary reconstruction model for symbolic decoding and global semantic feature recovery to generate a preliminary reconstructed image. The preliminary reconstructed image is then input into a pre-established detail enhancement network model, where high-frequency detail features of the image are gradually recovered through a multi-layer dense residual structure, ultimately outputting a high-quality reconstructed image.

[0005] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the semantic feature extraction process of the pre-established hybrid architecture model includes the following steps: The first stage of convolution operation is performed on the input image to obtain the first stage feature map; The feature map from the first stage is subjected to a deep convolution operation in the second stage to obtain a compact feature map in the second stage. The compact feature map in the second stage is processed by feature block segmentation and mapping to obtain the input sequence; The global semantic features are obtained by passing the input sequence through multiple Transformer encoders. The compact feature map of the second stage is reshaped to obtain local texture features; Global semantic features and local texture features are concatenated, and then a non-linear activation operation is performed on the concatenated features to finally obtain a hybrid feature map.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of performing the first-stage convolution operation on the input image, comprising: Perform the first stage of convolution on the input image, assuming the input is... , These represent the height, width, and number of RGB color channels of the image, respectively, using multiple convolutional kernels. Extract local feature information from the image, including edges, texture, and shape. ;in Represents the size of the convolution kernel, with a total of Each convolutional kernel is used, and then the features are filtered layer by layer through convolution operations. The calculation formula is: in, For ReLU function, This represents the convolution operation. This represents the bias corresponding to each convolutional kernel, generating the feature map for the first stage: , The process of performing a second-stage deep convolution operation on the feature map from the first stage includes: The feature map of the first stage Further input into deeper convolutional networks, through convolutional kernels For each position Convolution operation is performed on the feature maps. After the convolution operation, each convolution kernel Extract different feature maps to obtain Feature map And combine pooling operations to perform max pooling: Among them, step size Controlled downsampling ultimately generates a compact feature map for the second stage. ,in .

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of performing feature block division and mapping on the compact feature map of the second stage, including: The compact feature map generated in the second stage The input is fed into the Vision Transformer module, which divides the 2D feature map into several smaller blocks using feature partitioning. Specifically, it divides the feature map into smaller blocks. Divided into indivual Blocks, where each feature block contains partial local feature information, for each feature block Perform a linear mapping: in This means each small block Flattened into a vector, The three-dimensional block is transformed into a one-dimensional vector. For the mapping matrix, As a bias term, combined with the position encoding matrix The input sequence is: The input sequence is processed through multiple layers of a Transformer encoder to obtain global semantic features. ; The feature map of the second stage Perform a remodeling operation: feature map Divide into sizes Take small pieces, flatten each piece, and adjust the flattening according to the number of pieces. Forming a new feature matrix ; as a local texture feature; global semantic features Sum of local texture features The concatenation operation is performed, and the concatenated features will contain global semantic information from the Transformer encoder and local features extracted from the convolutional layers: in, It is a mapping matrix, responsible for mapping the concatenated feature representations to a new dimension. , For bias terms, Indicates a splicing operation; The concatenated features are processed using a non-linear activation operation: Finally, a hybrid feature map is generated. .

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of reducing the dimensionality of the hybrid feature map through symbolic compression, comprising: The channel dimension of the feature map is reduced by convolution operation to obtain the dimensionality-reduced feature map. The calculation formula is: in, For dimensionality reduction convolution kernels, For bias terms; For the dimensionality-reduced feature map The spatial resolution and channel dimension are compressed by using convolutional kernel operations and fixed-size pooling operations to obtain compressed feature maps. The process can be represented as: in, To compress the convolution kernel, For bias terms, This is a pooling operation; The compressed feature map is mapped to a symbolic representation space, where each symbol contains a corresponding I / Q dimension. The symbolic representation is defined as follows: , where each symbol The notation for the I / Q dimension is as follows: in, Here, is the quantization function, representing the sign quantization strategy; The mapped symbolic representation is serialized, and the spatial location and channel dimension of the feature map are rearranged into a one-dimensional symbolic sequence. The reconstructed symbolic sequence is stored and output in I / Q format as a low-dimensional symbolic representation sequence, as follows: in, .

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the low-dimensional symbol representation sequence into a pre-established preliminary reconstruction model for symbol decoding and global semantic feature recovery, comprising: Feature map of received recovery ,in , and These are the height, width, and number of channels, respectively. The feature map is divided into fixed-size blocks, each containing local semantic information. This division is achieved by segmenting the spatial dimensions of the feature map, as shown below: Each small block Each small block is then flattened into a one-dimensional vector, denoted as: The process of linearly projecting flattened small feature blocks onto a fixed feature vector space is represented as follows: in It is a linear projection matrix. As a bias term, during linear projection, a unified vector representation is generated by weighting and transforming small feature blocks, while also adding positional embedding information. The formula is as follows: The small patch features, after location embedding processing, are input into a multi-layer Transformer encoder. For the first... The layer adopts the following processing flow. The network structure is symmetrical to that of the sending end: in Presentation layer This indicates a multi-head attention mechanism. This represents a multilayer perceptron, with the initial input being... ; The final feature representation containing global dependencies is generated: The feature sequence generated by the encoder is restored to a two-dimensional feature map through linear mapping, and the features are rearranged according to the spatial position of the original small blocks to form a preliminary reconstructed image. The formula for the restoration process is as follows: in and The parameters are for restoring the mapping.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the initially reconstructed image into a pre-established detail enhancement network model, and progressively recovering the high-frequency detail features of the image through a multi-layer dense residual structure, including: The image is convolved using the first residual block to extract high-frequency detail features, including edge and texture information, as follows: in, Represents the first convolutional layer. It is a ReLU activation function; A short connection is established between the output features and input features in the first stage, directly passing the features of the original, initially reconstructed image to the next layer, as follows: The output features of the previous layer are sequentially input into the subsequent residual blocks, and each layer is responsible for extracting high-frequency information from different scales; Among them, the The formula for calculating each residual block is: in, Indicates the first Convolution operations of layers, A cascade representing all features from the preceding layers; Through a multi-layered dense residual structure, the output features of each layer are superimposed with the initial input features, forming a gradual accumulation of features, as follows: in, The number of layers in the residual block; The image after processing all residual blocks is then fused with features and upsampled to generate the final high-quality reconstructed image. .

[0011] Secondly, in order to achieve the above objectives, the present invention discloses an image transmission system for detail reconstruction with low compression ratios, comprising: The feature extraction module is used to acquire the input image, input the input image into a pre-established hybrid architecture model for semantic feature extraction, and output a hybrid feature map, wherein the hybrid feature map integrates local texture features and global semantic features; the hybrid feature map is reduced in dimensionality by symbol compression to obtain a low-dimensional symbol representation sequence; The image reconstruction module is used to input the low-dimensional symbolic representation sequence into a pre-established preliminary reconstruction model for symbolic decoding and global semantic feature recovery, generating a preliminary reconstructed image. The preliminary reconstructed image is then input into a pre-established detail enhancement network model, which gradually recovers the high-frequency detail features of the image through a multi-layer dense residual structure, and finally outputs a high-quality reconstructed image.

[0012] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs a detail reconstruction image transmission method for low compression ratio as described above.

[0013] In another aspect of the present invention, in order to achieve the above-mentioned objective, a computer-readable storage medium is disclosed, wherein a computer program is stored in the computer program, and when the computer program is loaded and executed by a processor, it employs a detail reconstruction image transmission method for low compression ratios as described above.

[0014] The beneficial effects of this invention are: This invention demonstrates superior image reconstruction quality. By optimizing semantic feature extraction and detail enhancement, it maintains high-precision image restoration even under low compression ratios, achieving clearer edges, textures, and overall structure. Simultaneously, in terms of bandwidth utilization, this scheme efficiently compresses and transmits critical information, achieving high visual fidelity with lower data transmission volume, significantly reducing bandwidth consumption and improving communication efficiency. Finally, regarding the training strategy, a two-stage training method with channel enhancement is proposed. This allows the model to learn global semantic features under ideal channel conditions, while further optimizing detail recovery capabilities in real channel environments affected by noise and fading, improving the model's adaptability and robustness. This ensures the stability and reliability of the semantic communication system under complex channel conditions and enhances image transmission quality under low compression ratios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram illustrating the estimation performance under different partitioning schemes in an embodiment of the present invention; Figure 3This is a comparative diagram of the estimation performance of various methods under different signal-to-noise ratios in the embodiments of the present invention; Figure 4 This is a schematic diagram illustrating the estimated performance when different subcarrier frequencies are selected as references in an embodiment of the present invention; Figure 5 This is a comparative diagram of the performance estimation of various methods under different bandwidths in the embodiments of the present invention; Figure 6 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: like Figure 1 As shown, a method for transmitting detail-reconstructed images with low compression ratios includes the following steps: S101: Obtain the input image, input the input image into the pre-established hybrid architecture model for semantic feature extraction, and output a hybrid feature map, wherein the hybrid feature map integrates local texture features and global semantic features; reduce the dimensionality of the hybrid feature map through symbol compression to obtain a low-dimensional symbol representation sequence; The semantic feature extraction process of a pre-established hybrid architecture model includes the following steps: The first stage of convolution operation is performed on the input image to obtain the first stage feature map; The feature map from the first stage is subjected to a deep convolution operation in the second stage to obtain a compact feature map in the second stage. The compact feature map in the second stage is processed by feature block segmentation and mapping to obtain the input sequence; The global semantic features are obtained by passing the input sequence through multiple Transformer encoders. The compact feature map of the second stage is reshaped to obtain local texture features; Global semantic features and local texture features are concatenated, and then a non-linear activation operation is performed on the concatenated features to finally obtain a hybrid feature map.

[0018] The specific process is as follows: First, the input image undergoes a first-stage convolution operation, with the input being... , These represent the image's height, width, and the number of RGB color channels, respectively. Multiple convolutional kernels are used. Extracting local feature information from an image, including basic features such as edges, texture, and shape. This includes... This represents the size of the convolution kernel, and there are a total of There are several such convolutional kernels. Then, the features are filtered layer by layer through convolution operations, calculated as follows: in, For ReLU function, This represents the convolution operation. This represents the bias corresponding to each convolutional kernel. In this way, the ReLU activation function suppresses negative values, highlighting local edges, textures, and shape features, making high-contrast information more apparent. Furthermore, this invention incorporates a normalization operation after the convolutional layers to ensure stable data distribution across different batches, reduce internal covariate shift issues, and improve model training efficiency and convergence speed. Simultaneously, during feature extraction, to reduce information redundancy and enhance the expression of key features, this invention employs a channel attention mechanism, enabling the network to adaptively focus on important channels, improving the quality of feature representation. Finally, through multi-layer convolutional processing, the first-stage feature map is generated: , This provides initial input for subsequent feature compression and deep semantic modeling.

[0019] The feature map of the first stage Further input into deeper convolutional networks, using larger convolutional kernels. For each position The feature maps are convolutionally processed to extract more complex patterns and local detail features. After the convolution operation, each convolution kernel... Extract different feature maps to obtain Feature map This process incorporates pooling operations (max pooling): Among them, step size By controlling downsampling, a compact feature map is ultimately generated. ,in This feature map is presented at a lower resolution while preserving the main semantic information and local details of the image.

[0020] The compact feature map generated in the second stage The input is fed into the Vision Transformer module, which divides the 2D feature map into several smaller blocks using feature partitioning. Divided into indivual The image is divided into blocks, each containing a portion of local feature information. This block-based approach helps capture local features from different regions while maintaining the integrity of spatial information, allowing subsequent Transformer processing to fully utilize the image's structural information. Then, each feature block... Perform a linear mapping: in This means each small block Flattened into a vector, The three-dimensional blocks are transformed into a one-dimensional vector. This step ensures that the features can be processed by the Transformer while reducing computational complexity. This is a mapping matrix, responsible for mapping the flattened feature blocks to a... - A 3D embedding space to accommodate the input requirements of the Transformer. This is a bias term. To enhance the model's ability to understand spatial information, this invention incorporates location embedding information (location encoding matrix). This provides location information for each feature block, enabling the Transformer to preserve spatial structure information and prevent the loss of feature order when learning the relationships between feature blocks. The input sequence is: The input sequence is modeled using a multi-layer Transformer encoder. This encoder consists of six identical stacked coding layers, each employing a standard two-sublayer structure. The multi-head self-attention sublayer uses eight attention heads, which uniformly distribute the input features across these attention heads via linear projection, computing self-attention in parallel. The feedforward neural network sublayer uses a two-layer fully connected structure, with its hidden layer dimension set to four times the input dimension, and the GeLU function chosen as the activation function. Residual connections and layer normalization mechanisms are applied after each sublayer to stabilize the training process.

[0021] Through multi-layer processing by the Transformer encoder, we obtain global semantic features. This feature encodes global information for the entire input image, enabling the model to understand the main semantic structure and object relationships of the image at a holistic level. However, relying solely on global features can lead to the loss of detailed information. Therefore, we will use global semantic features... Local feature maps extracted by the second-stage convolution The fusion process begins with the feature maps from the second stage. We will perform a reshaping operation. Specifically, we will reshape the feature map. Divide into sizes The reshaping operation involves flattening each small piece and adjusting the size based on the number of pieces. Form a new feature matrix Then, we will use global semantic features. and reshaped local features Perform a concatenation operation. The concatenated features will contain global semantic information from the Transformer encoder and local features extracted from the convolutional layers: in, It is a mapping matrix, responsible for mapping the concatenated feature representations to a new dimension. , For bias terms, This indicates a concatenation operation, followed by processing the fused features using a non-linear activation operation: Finally, a hybrid feature map is generated. .

[0022] The process of reducing the dimensionality of the hybrid feature map through symbolic compression includes: The semantic features are subjected to dimensionality reduction and pooling operations to compress spatial resolution and channel dimension, and the feature maps are converted into I / Q symbols, which are then mapped to the symbolic representation space through quantization. The specific content is as follows: Next, a series of convolution operations are performed on the blended feature maps. The dimensionality is reduced by the channel dimension while maintaining the spatial resolution. Dimensionality reduction preserves the core semantic features of the image and reduces feature redundancy to accommodate subsequent symbolic mapping computations. The dimensionality-reduced feature map is... The calculation formula is: in, It is a dimensionality-reducing convolution kernel, specifically designed to reduce the number of channels and highlight the core semantic features of the image during compression. This is a bias term used to adjust the feature value distribution, making the model more stable. Dimensionality reduction compresses the channel dimension of the image while preserving core semantic feature information.

[0023] The reduced-dimensionality feature map is further compressed in terms of spatial resolution and channel dimension through convolutional kernel operations and fixed-size pooling operations to reduce data dimensionality and enhance information compactness. In this process, pooling operations serve to remove redundancy and enhance feature aggregation capabilities, helping to improve the model's adaptability under different channel conditions. Finally, this process maps the spatial dimension of the feature map to a lower fixed resolution and converts it into an I / Q symbol dimension representation to adapt to the requirements of wireless channel transmission. The compressed feature map is... This process can be represented as: in, To compress the convolution kernel, specifically designed to further reduce data dimensionality and enhance the expression of key information. This is a bias term to ensure the stability of the data after nonlinear transformation. Pooling is a pooling operation that spatially compresses the feature map, reducing its spatial resolution.

[0024] The compressed low-dimensional feature map is mapped to a symbol representation space, where each symbol contains a corresponding I / Q dimension for subsequent channel transmission. The symbol representation is defined as follows: , where each symbol Symbols representing the I / Q dimensions. The mapping process, combined with a symbol quantization strategy, further compresses the symbols, ensuring compact representation and efficient transmission performance. in, Let be the quantization function, representing the symbolic quantization strategy.

[0025] The mapped symbolic representation is serialized, rearranging the spatial locations and channel dimensions of the feature map into a one-dimensional symbolic sequence. The reconstructed symbolic sequence is stored and output in I / Q format, as follows: in, Finally, the symbol sequence is stored and output in I / Q format.

[0026] After symbolic data is generated, to ensure stable transmission in a wireless environment, this invention utilizes the GNU Radio software platform combined with the USRP wireless device to construct a real-time semantic wireless communication system, supporting data transmission and channel simulation. The complexity of wireless channels, such as multipath effects, fading, and noise interference, can affect signal transmission stability. Therefore, this invention performs a series of preprocessing steps before signal transmission to enhance anti-interference capabilities. First, the fused semantic feature data is modulated and encapsulated, with semantic frame header information added to provide key information such as frame synchronization and length indication, ensuring the receiver can correctly parse the data structure. Subsequently, the data is segmented into fixed-size semantic units and a Cyclic Redundancy Check (CRC) is added to enhance error detection capabilities and improve the system's anti-interference performance. The system performs power adjustment and preprocessing on the signal to ensure it meets the transmission requirements of the wireless channel and completes radio frequency transmission through the USRP device. At the receiver, after acquiring the signal, the USRP device performs frame boundary detection and timing synchronization to ensure accurate data decoding. After channel equalization and error correction, the received signal parses the semantic frame header, extracts the payload data, and performs denoising and signal compensation to finally recover the semantic feature data. This wireless transmission scheme ensures efficient transmission and robust decoding of semantic information, maintaining high-quality data recovery capabilities even in complex channel environments. After being decoded at the receiving end, the wirelessly transmitted symbolic data is remapped back to the feature map for subsequent feature reconstruction and detail restoration.

[0027] S102: Input the low-dimensional symbolic representation sequence into the pre-established preliminary reconstruction model for symbolic decoding and global semantic feature recovery to generate a preliminary reconstructed image. Input the preliminary reconstructed image into the pre-established detail enhancement network model, and gradually recover the high-frequency detail features of the image through a multi-layer dense residual structure, finally outputting a high-quality reconstructed image.

[0028] The process of inputting the low-dimensional symbolic representation sequence into a pre-established preliminary reconstruction model for symbolic decoding and global semantic feature recovery includes: Feature map of receiver recovery in , and These are the height, width, and number of channels, respectively. Next, we will... The feature map is divided into fixed-size blocks, each containing local semantic information. This partitioning method ensures that the network preserves spatial relationships during processing, allowing local features to be better integrated with global information. The block partitioning, achieved by segmenting the spatial dimensions of the feature map, can be represented as: Each small block Representing local information of a fixed region in the feature map, this block-segmentation process is similar to the sliding window approach in convolution, but here it focuses more on the discretization and independent processing of local features to enable flexible semantic feature mapping later. Each small block is then flattened into a one-dimensional vector, denoted as: This provides input for subsequent linear projection. The flattened vector represents the feature information of all channels within the small block, converting it into a linear spatial representation, which facilitates subsequent feature transformation and processing. Simultaneously, this operation reduces data dimensionality and computational complexity, allowing the network to learn in a more compact representation space and improving data processing efficiency.

[0029] Then, the flattened small feature blocks are linearly projected onto a fixed feature vector space. This process can be represented as: in It is a linear projection matrix. This is the bias term. During linear projection, a unified vector representation is generated by weighting and transforming the small feature blocks, while simultaneously adding positional embedding information. The positional embedding information is generated based on the spatial location of the small blocks in the original feature map, used to preserve the spatial relationship of feature blocks, thereby ensuring that spatial semantics are not lost during global feature modeling. Its formula is: The small patch features, after location embedding, are input into a multi-layer Transformer encoder. Specifically, for the... layer We adopt the following processing flow, which is symmetrical to the network structure of the sending end: in Presentation layer This indicates a multi-head attention mechanism. This represents a multilayer perceptron, with the initial input being... Multi-head attention mechanisms extract global semantic information from images by calculating the correlations between small feature patches. Simultaneously, they enhance feature representation by performing non-linear transformations on the semantic features using an MLP. In this process, the output of each encoder layer serves as the input to the next layer, progressively extracting higher-order semantic information and ultimately generating a feature representation that includes global dependencies. Finally, the feature sequence generated by the encoder is restored to a two-dimensional feature map through linear mapping, and the features are rearranged according to the spatial location of the original small blocks to form a preliminary reconstructed image. The formula for the restoration process is: in and These are the parameters for restoring the mapping. Initial image reconstruction. The main semantic information of the input image is recovered, including most of the global structure and semantic consistency. However, due to the lack of high-frequency detail information, there is blurring in details such as texture and edges. Therefore, it is necessary to enhance the details by using a multi-layer dense residual structure.

[0030] The process of inputting the initially reconstructed image into a pre-established detail enhancement network model, and gradually restoring the high-frequency detail features of the image through a multi-layered dense residual structure, includes: Through the first layer of residual blocks Convolution operations are performed to extract high-frequency detail features from the image, including edge and texture information. Since high-frequency features typically contain detailed information about the image, such as object outlines and texture variations, convolution operations capture local high-frequency features using multiple small convolution kernels and enhance these features with non-linear activation functions, ensuring clearer edge information. in, Represents the first convolutional layer. This is the ReLU activation function.

[0031] Short connections, or skip connections, are established between the output and input features in the first stage to directly pass the features of the original image to the next layer. This fusion operation preserves the original semantic information of the image and provides a reference benchmark for subsequent enhancement, preventing the loss of detailed information. The output features of the previous layer are sequentially input into subsequent residual blocks, with each layer responsible for extracting high-frequency information from different scales. In the early residual blocks, the focus is on enhancing the image's edge contours to ensure clear and continuous edges. In the middle residual blocks, the texture features and local contrast are further enhanced, resulting in richer image details. In the later residual blocks, the details of the entire image are integrated, optimizing the smoothness and consistency of overall details. The formula for calculating each residual block is: in, Indicates the first Convolution operations of layers, This represents a cascade of features from all previous layers. The purpose of cascading is to fully integrate information from different layers, improve the model's feature representation ability, and avoid information loss due to the increase in the number of deep network layers.

[0032] By employing a dense residual structure, it is ensured that each layer of output features is superimposed on the initial input features, resulting in a gradual accumulation of features. This mechanism guarantees that high-frequency features are continuously enhanced in each residual block, while avoiding information loss or feature degradation caused by multiple superpositions. in, This represents the number of layers in the residual block.

[0033] The image after processing all residual blocks is then fused and upsampled to generate the final image. , in, and These represent the height and width after upsampling, respectively. Upsampling operations can employ bilinear interpolation or transposed convolution to ensure the image size is restored to the target resolution and to reduce artifacts.

[0034] A two-stage sequential training strategy is employed. First, the semantic encoder and preliminary reconstruction module are trained. Then, the detail augmentation module is trained using a dataset with channel noise and fading characteristics, improving the model's adaptability and reconstruction quality in real wireless environments. The specific content is as follows: First, in the first phase of training, a specific dataset was used (trained using the CIFAR-10 dataset and validated using the CIFAR-100 dataset), with an initial learning rate of 0.0001 and The optimizer samples 64 images from the dataset in each training cycle. After data preprocessing such as normalization and random horizontal flipping, the images are input into the semantic encoder and preliminary reconstruction module. Under ideal channel conditions, the global semantic features of the images are extracted and initially reconstructed. At the same time, the model checkpoints are saved periodically for training monitoring.

[0035] Subsequently, in the second phase of training, to simulate the effects of channel noise and fading in real-world communication environments, channel characteristic data was generated using a specific dataset (high- and low-resolution pair datasets). Specifically, for each low-resolution image... First, it is converted into a symbolic representation using a trained encoder. Then, by simulating the actual communication channel... Adding noise and fading effects yields a noisy symbolic representation. Next, the preliminary reconstruction module was used to... Decode the image to generate a low-resolution reconstructed image. ; and compare it with the corresponding high-resolution image Pairing to form data pairs This is to accurately reflect the distortion that may occur during image transmission under channel conditions.

[0036] Next, batch data is sampled from the aforementioned channel-assisted dataset. Extract low-resolution images respectively and high-resolution images The data is then fed into the detail enhancement module for targeted training. An initial learning rate is used in this stage. (Set to 0.0001 here) and The optimizer updates the parameters of the detail enhancement module and adjusts the learning rate every 200,000 training steps. Simultaneously, the model state is saved periodically. The two sequential training stages described above: the first stage ensures the model accurately recovers the global semantic information of the image under ideal channel conditions; the second stage first introduces interference characteristics from the actual communication environment using channel-aided data, and then specifically optimizes the detail enhancement module. This achieves high quality and high detail fidelity in image reconstruction under actual channel conditions.

[0037] Specifically, the present invention will be further illustrated below through embodiments: The estimation performance of the method proposed in this invention in a real system is as follows: Figures 3-5 As shown. DRISCom is the method proposed in this invention.

[0038] like Figure 3 As shown, at a compression ratio of 1 / 6, the PSNR value of the present invention is higher than that of the comparative methods, regardless of whether the signal-to-noise ratio is high or low. This indicates that the present invention performs better at low compression ratios.

[0039] Comparison of the performance of this invention with various methods at different compression ratios Figure 4 As shown, the proposed solution achieves a higher PSNR value than all other methods when the compression ratio is below 0.32. Although it is surpassed by BPG + Capacity as the compression ratio continues to increase, it still verifies that this solution has good performance at low compression ratios.

[0040] Figure 5 By comparing the models trained with and without channel assistance, it is evident that the model trained with channel assistance outperforms the model trained without channel assistance at all signal-to-noise ratios, thus verifying the effectiveness of the proposed two-stage serial training with channel assistance.

[0041] Example 2: To achieve the above objective, such as Figure 6As shown, based on Embodiment 1, this invention discloses a detail reconstruction image transmission system for low compression ratios, comprising: The feature extraction module 11 is used to acquire the input image, input the input image into a pre-established hybrid architecture model for semantic feature extraction, and output a hybrid feature map, wherein the hybrid feature map integrates local texture features and global semantic features; the hybrid feature map is reduced in dimensionality by symbol compression to obtain a low-dimensional symbol representation sequence; The image reconstruction module 12 is used to input the low-dimensional symbol representation sequence into a pre-established preliminary reconstruction model for symbol decoding and global semantic feature recovery, generate a preliminary reconstructed image, input the preliminary reconstructed image into a pre-established detail enhancement network model, and gradually recover the high-frequency detail features of the image through a multi-layer dense residual structure, and finally output a high-quality reconstructed image.

[0042] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, 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, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0043] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0044] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for transmitting detail-reconstructed images with low compression ratios, characterized in that, The method includes the following steps: The input image is acquired and fed into a pre-established hybrid architecture model for semantic feature extraction. The output is a hybrid feature map, which integrates local texture features and global semantic features. The hybrid feature map is reduced in dimensionality by symbolic compression to obtain a low-dimensional symbolic representation sequence. The low-dimensional symbolic representation sequence is input into a pre-established preliminary reconstruction model for symbolic decoding and global semantic feature recovery to generate a preliminary reconstructed image. The preliminary reconstructed image is then input into a pre-established detail enhancement network model, where high-frequency detail features of the image are gradually recovered through a multi-layer dense residual structure, ultimately outputting a high-quality reconstructed image.

2. The method for transmitting detailed reconstructed images with low compression ratios according to claim 1, characterized in that, The semantic feature extraction process of the pre-established hybrid architecture model includes the following steps: The first stage of convolution operation is performed on the input image to obtain the first stage feature map; The feature map from the first stage is subjected to a deep convolution operation in the second stage to obtain a compact feature map in the second stage. The compact feature map in the second stage is processed by feature block segmentation and mapping to obtain the input sequence; The global semantic features are obtained by passing the input sequence through multiple Transformer encoders. The compact feature map of the second stage is reshaped to obtain local texture features; Global semantic features and local texture features are concatenated, and then a non-linear activation operation is performed on the concatenated features to finally obtain a hybrid feature map.

3. The method for transmitting detailed reconstructed images with low compression ratios according to claim 2, characterized in that, The process of performing the first-stage convolution operation on the input image includes: Perform the first stage of convolution on the input image, assuming the input is... , These represent the height, width, and number of RGB color channels of the image, respectively, using multiple convolutional kernels. Extract local feature information from an image, including edges, texture, and shape. ,;in Represents the size of the convolution kernel, with a total of Each convolutional kernel is used, and then the features are filtered layer by layer through convolution operations. The calculation formula is: in, For ReLU function, This represents the convolution operation. This represents the bias corresponding to each convolutional kernel, generating the feature map for the first stage: ， The process of performing a second-stage deep convolution operation on the feature map from the first stage includes: The feature map of the first stage Further input into deeper convolutional networks, through convolutional kernels For each position Convolution operation is performed on the feature maps. After the convolution operation, each convolution kernel Extract different feature maps to obtain Feature map And combine pooling operations to perform max pooling: Among them, step size Controlled downsampling ultimately generates a compact feature map for the second stage. ,in .

4. The method for transmitting detailed reconstructed images with low compression ratios according to claim 3, characterized in that, The process of performing feature block division and mapping on the compact feature map of the second stage includes: The compact feature map generated in the second stage The input is fed into the Vision Transformer module, which divides the 2D feature map into several smaller blocks using feature partitioning. Specifically, it divides the feature map into smaller blocks. Divided into indivual Blocks, where each feature block contains partial local feature information, for each feature block Perform a linear mapping: in This means each small block Flattened into a vector, The three-dimensional block is transformed into a one-dimensional vector. For the mapping matrix, As a bias term, combined with the position encoding matrix The input sequence is: The input sequence is processed through multiple layers of a Transformer encoder to obtain global semantic features. ; The feature map of the second stage Perform a remodeling operation: feature map Divide into sizes Take small pieces, flatten each piece, and adjust the flattening according to the number of pieces. Forming a new feature matrix ; as a local texture feature; global semantic features Sum of local texture features The concatenation operation will produce features that include global semantic information from the Transformer encoder and local features extracted from the convolutional layers. in, It is a mapping matrix, responsible for mapping the concatenated feature representations to a new dimension. , For bias terms, Indicates a splicing operation; The concatenated features are processed using a non-linear activation operation: Finally, a hybrid feature map is generated. .

5. The method for transmitting detailed reconstructed images with low compression ratios according to claim 1, characterized in that, The process of reducing the dimensionality of the hybrid feature map through symbolic compression includes: The channel dimension of the feature map is reduced by convolution operation to obtain the dimensionality-reduced feature map. The calculation formula is: in, For dimensionality reduction convolution kernels, For bias terms; For the dimensionality-reduced feature map The spatial resolution and channel dimension are compressed by using convolutional kernel operations and fixed-size pooling operations to obtain compressed feature maps. The process can be represented as: in, To compress the convolution kernel, For bias terms, This is a pooling operation; The compressed feature map is mapped to a symbolic representation space, where each symbol contains a corresponding I / Q dimension. The symbolic representation is defined as follows: , where each symbol The notation for the I / Q dimension is as follows: in, Here, is the quantization function, representing the sign quantization strategy; The mapped symbolic representation is serialized, and the spatial location and channel dimension of the feature map are rearranged into a one-dimensional symbolic sequence. The reconstructed symbolic sequence is stored and output in I / Q format as a low-dimensional symbolic representation sequence, as follows: in, .

6. The method for transmitting detailed reconstructed images with low compression ratios according to claim 1, characterized in that, The process of inputting the low-dimensional symbol representation sequence into a pre-established preliminary reconstruction model for symbol decoding and global semantic feature recovery includes: Feature map of received recovery ,in , and These are the height, width, and number of channels, respectively. The feature map is divided into fixed-size blocks, each containing local semantic information. This division is achieved by segmenting the spatial dimensions of the feature map, as shown below: Each small block Each small block is then flattened into a one-dimensional vector, denoted as: The process of linearly projecting flattened small feature blocks onto a fixed feature vector space e is represented as follows: in It is a linear projection matrix. As a bias term, during linear projection, a unified vector representation is generated by weighting and transforming small feature blocks, while also adding positional embedding information. The formula is as follows: The small patch features, after location embedding processing, are input into a multi-layer Transformer encoder. For the first... The layer adopts the following processing flow. The network structure is symmetrical to that of the sending end: in Presentation layer This indicates a multi-head attention mechanism. This represents a multilayer perceptron, with the initial input being... ; The final feature representation containing global dependencies is generated: The feature sequence generated by the encoder is restored to a two-dimensional feature map through linear mapping, and the features are rearranged according to the spatial position of the original small blocks to form a preliminary reconstructed image. The formula for the restoration process is as follows: in and The parameters are for restoring the mapping.

7. The method for transmitting detailed reconstructed images with low compression ratios according to claim 1, characterized in that, The process of inputting the initially reconstructed image into a pre-established detail enhancement network model, and gradually restoring the high-frequency detail features of the image through a multi-layered dense residual structure, includes: The image is convolved using the first residual block to extract high-frequency detail features, including edge and texture information, as follows: in, Represents the first convolutional layer. It is a ReLU activation function; A short connection is established between the output features and input features in the first stage, directly passing the features of the original, initially reconstructed image to the next layer, as follows: The output features of the previous layer are sequentially input into the subsequent residual blocks, and each layer is responsible for extracting high-frequency information from different scales; Among them, the The formula for calculating each residual block is: in, Indicates the first Convolution operations of layers, A cascade representing all features from the preceding layers; Through a multi-layered dense residual structure, the output features of each layer are superimposed with the initial input features, forming a gradual accumulation of features, as follows: in, The number of layers in the residual block; The image after processing all residual blocks is then fused with features and upsampled to generate the final high-quality reconstructed image. 。 8. A detail reconstruction image transmission system for low compression ratios, employing the detail reconstruction image transmission method for low compression ratios according to any one of claims 1 to 7, characterized in that, include: The feature extraction module is used to acquire the input image, input the input image into a pre-established hybrid architecture model for semantic feature extraction, and output a hybrid feature map, wherein the hybrid feature map integrates local texture features and global semantic features; The hybrid feature map is reduced in dimensionality by symbolic compression to obtain a low-dimensional symbolic representation sequence. The image reconstruction module is used to input the low-dimensional symbolic representation sequence into a pre-established preliminary reconstruction model for symbolic decoding and global semantic feature recovery, generating a preliminary reconstructed image. The preliminary reconstructed image is then input into a pre-established detail enhancement network model, which gradually recovers the high-frequency detail features of the image through a multi-layer dense residual structure, and finally outputs a high-quality reconstructed image.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs a detail reconstruction image transmission method for low compression ratios, as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs a detail reconstruction image transmission method for low compression ratios as described in any one of claims 1 to 7.

Images