Lightweight image super-resolution reconstruction method, system, storage medium and product

Abstract

The present invention relates to the field of image processing, and in particular to a lightweight image super-resolution reconstruction method, system, storage medium and product. The method comprises the following steps performed in sequence: S1: acquiring a low-resolution image data set, dividing the image data set into a training set and a verification set according to a preset ratio, and preprocessing the image data set; S2: constructing a lightweight image super-resolution reconstruction network based on CNN and Transformer fusion, inputting the training set processed in step S1 into the lightweight image super-resolution reconstruction network, adopting an L1 loss function to perform model training, and obtaining a trained lightweight image super-resolution reconstruction network; S3: inputting the verification set processed in step S1 into the trained lightweight image super-resolution reconstruction network, and obtaining a reconstructed super-resolution image; a dense feature fusion module is used to integrate local features and global features, and a cross-attention mechanism is used to enhance the correlation between features, so as to improve the effect of image super-resolution reconstruction.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a lightweight image super-resolution reconstruction method, system, storage medium and product. Background Art

[0002] Human perception of external information mainly depends on the visual system. As a carrier for recording and transmitting external information, images are widely used in various fields due to their richness and intuitiveness. However, due to the limitations and influence of imaging equipment, environmental noise, data transmission and other conditions, there are often image degradation processes such as noise, blur and downsampling in the imaging process, which reduces the resolution of the actual image and loses detail information. Single Image Super-Resolution (SISR) uses image processing algorithms to reconstruct a high-resolution (HR) image with rich details and clear texture from a low-resolution (LR) image. It is an effective and low-cost method to improve image quality. It has been widely used in urban monitoring management, multimedia consumption, medical imaging, remote sensing monitoring and other fields.

[0003] In the past decade, deep learning has developed rapidly. It has the ability to automatically learn features from data and handle complex tasks. It performs well in processing high-dimensional and complex data and is also widely used in image super-resolution reconstruction. Deep Convolutional Neural Networks (DCNN) is the most widely used model in the field of deep learning. Through the combination of multi-layer convolution, pooling and fully connected layers, it can automatically learn and extract image features and learn end-to-end feature mapping between low-resolution images and high-resolution images. The single image super-resolution reconstruction technology based on convolutional neural networks has improved from the initial three-layer convolution learning to a deeper network with more than 100 or even more than 400 layers, and its performance has also improved accordingly, but it requires greater computational costs and memory consumption, and cannot be deployed on embedded terminals such as mobile devices. In addition, due to the limitation of the kernel size of the convolution operation, DCNN can only process local features of the image. Transformer uses the self-attention mechanism to effectively capture the global information and long-distance dependencies in the data. It has made great progress in natural language processing and has also been borrowed and used in image processing. By effectively combining CNN and Transformer, we can complement each other's strengths and achieve better super-resolution reconstruction effects. However, the computational resources and memory overhead of these methods are too large to meet the needs of actual applications. Summary of the invention

[0004] The object of the present invention is to provide a lightweight image super-resolution reconstruction method, system, storage medium and product that reduce computing resources and memory overhead.

[0005] In order to achieve the above object, the present invention adopts such technical solution:

[0006] The lightweight image super-resolution reconstruction method comprises the following steps performed in sequence:

[0007] S1: Obtain a low-resolution image dataset, divide the image dataset into a training set and a validation set according to a preset ratio, use Bicubic downsampling to downsample the images in the image dataset to obtain corresponding LR images, and perform random data enhancement operations on each LR image to obtain an enhanced LR image, and randomly cut the enhanced LR image into 64 The image block;

[0008] S2: construct a lightweight image super-resolution reconstruction network based on the fusion of CNN and Transformer, input the training set processed in step S1 into the lightweight image super-resolution reconstruction network, use the L1 loss function to perform model training, and obtain a trained lightweight image super-resolution reconstruction network;

[0009] S3: Input the verification set processed in step S1 into the trained lightweight image super-resolution reconstruction network to obtain a reconstructed super-resolution image.

[0010] Preferably, the lightweight image super-resolution reconstruction network includes a shallow feature extraction module, a deep feature extraction module, a dense feature fusion module and an upsampling module.

[0011] Preferably, the shallow feature extraction module adopts Convolution transforms the low-resolution images in the input training set into the feature space to generate shallow features. , the operation process of the shallow feature extraction module is as follows:

[0012] ;

[0013] In the formula, represents the input low-resolution image, represents a standard 3×3 convolution, Represents the extracted shallow features;

[0014] The deep feature extraction module includes a trunk branch and a bypass branch. The trunk branch uses a multi-layer stacked feature mixing module to capture the global information in the image, and the bypass branch uses a hierarchical edge focusing module to improve the network's ability to obtain edge information and local details. The processing process of the deep feature extraction module is expressed as follows:

[0015] ;

[0016] ;

[0017] In the formula, Indicates n feature mixing modules, Indicates that n The features extracted by the feature mixing module are represents the hierarchical edge focusing module, Represents the features extracted by the hierarchical edge-focusing module.

[0018] Preferably, the feature mixing module adopts a Transformer structure, which is composed of a spatial adaptive feature modulation module, a layer normalization module, and a convolution channel mixing module. The structure of the feature mixing module is similar to the Vit network structure, and the layer normalization replaces the batch normalization to standardize each channel in the feature map. The overall operation of the feature mixing module is expressed as:

[0019] ;

[0020] In the formula, F represents the input feature map, is the layer normalization module operation, represents the spatial adaptive feature modulation module, It is a convolutional channel mixing module. The feature mixing module includes residual connections to preserve the original feature information while converting the features.

[0021] The spatial adaptive feature modulation module is used to effectively obtain the spatial features of the image at multiple scales. Feature modulation is performed at different spatial levels, and channel segmentation is performed on the normalized input features to generate four components. The first component is processed by a 3×3 depth-separable convolution, and the remaining components are sent to the multi-scale feature generation unit. The specific operation is expressed as follows:

[0022] ;

[0023] in, Represents the input graph X Perform channel splitting operation; represents a depth-wise separable convolution with a kernel size of 3×3, Indicates that the features of a specific level are upsampled to the original resolution through nearest neighbor interpolation p , Indicates that the input feature is downsampled to 2 i The size of Represents a splicing operation, Represents 1×1 convolution, which connects these multi-scale features through 1×1 convolution and outputs the final feature ;

[0024] The convolutional channel mixing module is used to enhance the local spatial modeling capability of the trunk branch and perform channel mixing. The convolutional channel mixing module contains a 3×3 convolution and a 1×1 convolution. The first 3×3 convolution encodes the spatial local context and doubles the number of channels of the input features of the mixed channel; the 1×1 convolution reduces the channel to the original number of input bits and uses the GELU activation function to apply to the hidden layer for nonlinear mapping;

[0025] The hierarchical edge focusing module is used to enhance the image edges and local details. The hierarchical edge focusing module consists of multiple diversion attention modules and hierarchical fusion submodules to extract features at different levels. The hierarchical edge focusing module processes the original feature map in the deep feature extraction stage and combines the enhanced features with the original features through residual connections to maintain the richness of texture information. The process is expressed as follows:

[0026] ;

[0027] in, represents the shallow features generated by the shallow feature extraction module, Indicates the first diversion attention module for shallow features The subsequent split attention module further processes the features output by the previous split attention module. represents the hierarchical fusion submodule,

[0028] The split attention module uses dynamic channel segmentation technology to carefully process image features, effectively controlling the number of feature map channels while retaining key information to the greatest extent. The split attention module first divides the input feature map into X Split into two branches X 1 and X 2 , the first branch X 1 After a series of convolution and ReLU activation layers to refine the features, the second branch X 2 remain unchanged so that the original information is preserved;

[0029] First of all, X 1 Perform continuous convolution and ReLU activation operations to extract key feature information Then, the channel attention module dynamically adjusts the response strength of the feature channel to filter out the key information and suppress the unimportant information to improve the representativeness of the feature. Multiply element-wise with the output of the channel attention module to obtain a weighted feature map Finally, the two branch features are concatenated and a convolution layer is used to integrate the final features and output the final result. The above process is expressed as:

[0030] ; In the formula, Represents a split operation; Indicates a convolution with a kernel size of 3×3;

[0031] represents the activation function; represents the operation through the channel attention module; Represents element-wise multiplication; Represents a convolution with a kernel size of 1×1; Represents a splicing operation;

[0032] is the final output of the split attention module;

[0033] The hierarchical fusion submodule is used to perform feature enhancement and feature fusion. The hierarchical fusion submodule first uses two independent channel attention CAs to respectively x 1 and x 2 Adaptively recalibrate them, then concatenate and fuse them to produce an enhanced feature representation, and finally refine the features through a convolutional layer. The process is expressed as:

[0034] .

[0035] Preferably, the dense feature fusion module is a key component designed to combine the different levels of features extracted by the Transformer and CNN in the network. The dense feature fusion module first processes the input features through two 1×1 convolutional layers. and , respectively and The purpose is to retain the spatial information of the features while reducing the computational complexity of subsequent operations. The processed features are adaptively averaged pooled to obtain global information, and then transformed through the fully connected layer (FC) to obtain their respective channel attention vectors, as shown below:

[0036] ;

[0037] in, and Respectively represent the feature maps and The calculated channel attention module weights are used to evaluate and The importance of each channel in for , for ;

[0038] represents a 1×1 convolution operation, represents adaptive average pooling, represents the fully connected layer, Represents the Sigmoid function, which is used to compress the output of the fully connected layer to between 0 and 1 and generate a weighting factor;

[0039] Then, the original features and By weighting these weighting factors, the features most relevant to the reconstruction task are strengthened, and other irrelevant information is suppressed. Through the cross-attention module, the weighted features are cross-combined to strengthen the complementarity between features, as shown below:

[0040] ;

[0041] in, Indicates that the application After channel attention weight Feature map, Indicates that the application After the channel attention weight Feature map, the design of the cross attention mechanism enables the two sets of feature maps extracted by Transformer and CNN to learn each other's importance and make full use of all available feature information;

[0042] Finally, the cross-weighted features are concatenated and integrated through another 1×1 convolutional layer to output the final fused features:

[0043] ;

[0044] In the formula, It means concatenating two feature maps in the channel dimension. Through the processing of the dense feature fusion module, the network can establish effective connections between features at different levels, reduce the amount of calculation of each module, and improve the quality of the reconstructed high-resolution image;

[0045] The upsampling module maps the features processed by the dense feature fusion module to a high-resolution space to generate the final super-resolution image:

[0046] ;

[0047] in, represents the final super-resolution reconstructed image, Represents the upsampling operation. The upsampling module includes a 3×3 convolution layer and a Pixel Shuffle layer.

[0048] A super-resolution reconstruction system comprises a processor and a memory, wherein a computer program is stored in the memory, and the computer program is executed by the processor to implement any of the methods described above.

[0049] A computer storage medium stores a computer program, which implements any of the above methods when executed by a processor.

[0050] A computer product comprises a computer program, wherein when the computer program is executed by a processor, the method described in any one of the above is implemented.

[0051] By adopting the aforementioned design scheme, the beneficial effects of the present invention are as follows: in terms of feature extraction, the present invention adopts a method of fusing global features and local features, designs a hierarchical edge focusing module to extract local features and edge detail information of the image, and improves the problem of the Transformer-based lightweight image super-resolution reconstruction network ignoring local information of the image; in terms of feature fusion, a dense feature fusion module is used to integrate local features and global features, and a cross-attention mechanism is used to enhance the correlation between features, making the features richer and improving the effect of image super-resolution reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the structure of the lightweight image super-resolution reconstruction network of the present invention;

[0053] Figure 2 It is a schematic diagram of the structure of the spatial adaptive modulation module and the convolution channel mixing module of the present invention;

[0054] Figure 3 It is a structural schematic diagram of the attention diversion module of the present invention;

[0055] Figure 4 It is a structural schematic diagram of the hierarchical fusion module of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The lightweight image super-resolution reconstruction method comprises the following steps performed in sequence:

[0058] S1: Obtain a low-resolution image dataset, divide the image dataset into a training set and a validation set according to a preset ratio, use Bicubic downsampling to downsample the images in the image dataset to obtain the corresponding LR images, and perform random data enhancement operations on each LR image to obtain an enhanced LR image, and randomly cut the enhanced LR image into 64 The image block;

[0059] In this embodiment, the DIV2K dataset is used as the image dataset, and the training set and the validation set are divided according to a preset ratio of 8:1. The random data enhancement operations include left-right flipping and 90-degree rotation.

[0060] S2: Construct a lightweight image super-resolution reconstruction network based on the fusion of CNN and Transformer, input the training set processed in step S1 into the lightweight image super-resolution reconstruction network, use the L1 loss function to train the model, and obtain a trained lightweight image super-resolution reconstruction network;

[0061] The lightweight image super-resolution reconstruction network includes a shallow feature extraction module, a deep feature extraction module, a dense feature fusion module and an upsampling module;

[0062] The shallow feature extraction module adopts Convolution transforms the input training set into feature space to generate shallow features ,The operation process of the shallow feature extraction module is as follows:

[0063] ;

[0064] In the formula, represents the input low-resolution image, represents a standard 3×3 convolution, Represents the extracted shallow features;

[0065] The deep feature extraction module includes a trunk branch and a bypass branch. The trunk branch uses a multi-layer stacked feature mixing module (FMM) to capture the global information in the image, and the bypass branch uses a hierarchical edge focus module (HEFM) to improve the network's ability to obtain edge information and local details. The processing process of the deep feature extraction module is expressed as follows:

[0066] ;

[0067] ;

[0068] In the formula, Indicates n feature mixing modules, Indicates that n The features extracted by the feature mixing module are represents the hierarchical edge focusing module, represents the features extracted by the hierarchical edge focusing module;

[0069] The feature mixing module adopts the Transformer structure, such as Figure 1 As shown in part (a), it consists of a spatially adaptive feature modulation module (SAFM), a layer normalization module (LayerNorm, LN) and a convolutional channel mixing module (Convolutional Channel Mixing). The structure of the feature mixing module is similar to the Vit network structure, and the layer normalization (LayerNorm, LN) replaces the batch normalization (BatchNorm) to standardize each channel in the feature map to improve the stability and convergence speed of the training process. The overall operation of the feature mixing module is expressed as:

[0070] ;

[0071] In the formula, F Represents the input feature map. There are many feature mixing modules. The first feature mixing module inputs the shallow features extracted by 3×3 convolution. , the input of the subsequent feature mitigation module is the output of the previous feature mixing module, so here F It is just a general term;

[0072] is the layer normalization module operation, represents the spatial adaptive feature modulation module, It is a convolutional channel mixing module. The feature mixing module includes residual connections to preserve the original feature information while converting the features.

[0073] The spatial adaptive feature modulation module is to effectively obtain the spatial features of the image at multiple scales, such as Figure 2 The spatial adaptive feature modulation module is used to adjust the input image Feature modulation at different spatial levels enhances the adaptability of the network. First, the channel segmentation operation is performed on the normalized input features to generate four components. Secondly, the first component is processed by a 3×3 depth-separable convolution, and the remaining components are sent to the multi-scale feature generation unit. The specific operation is expressed as:

[0074] ;

[0075] in, Represents the input graph X Perform channel splitting operation; represents a depth-wise separable convolution with a kernel size of 3×3, Indicates that the features of a specific level are upsampled to the original resolution through nearest neighbor interpolation p , Indicates that the input feature is downsampled to 2 i The size of Represents a splicing operation, Represents 1×1 convolution, which connects these multi-scale features through 1×1 convolution and outputs the final feature .

[0076] The Convolutional Channel Mixer (CCM) module is used to enhance the local spatial modeling capability of the trunk branches and perform channel mixing, such as Figure 2 As shown in part (b) of the module. This module contains a 3×3 convolution and a 1×1 convolution. The first 3×3 convolution encodes the spatial local context and doubles the number of channels of the input features of the mixed channels; the subsequent 1×1 convolution reduces the channel to the original number of input bits. The GELU activation function is applied to the hidden layer for nonlinear mapping.

[0077] The hierarchical edge focusing module is inspired by the feature pyramid, such as Figure 1As shown in part (b) of , it aims to enhance the edges and local details of the image. The hierarchical edge focusing module consists of multiple Pay Attention to Shunting Modules (PASM) and Hierarchical Fusion Modules (HF) to extract features at different levels. The hierarchical edge focusing module processes the original feature map in the deep feature extraction stage and combines the enhanced features with the original features through residual connections to maintain the richness of texture information. The process is expressed as follows:

[0078] ;

[0079] in, represents the shallow features generated by the shallow feature extraction module, Indicates the first diversion attention module for shallow features The subsequent split attention module further processes the features output by the previous split attention module. Represents the hierarchical fusion submodule.

[0080] The structure of the split attention module is as follows Figure 3 As shown in the figure, the split attention module uses dynamic channel segmentation technology to carefully process image features, effectively controlling the number of feature map channels while retaining key information to the greatest extent. The split attention module first splits the input feature map into X Split into two branches X 1 and X 2 , the first branch X 1 After a series of convolution and ReLU activation layers to refine the features, the second branch X 2 Leave it unchanged so that the original information is preserved.

[0081] First of all, X 1 Perform continuous convolution and ReLU activation operations to extract key feature information Then, the channel attention module (CA) dynamically adjusts the response strength of the feature channel to filter out key information and suppress unimportant information to improve the representativeness of the feature. Multiply element-wise with the output of the channel attention module to obtain a weighted feature map Finally, the two branch features are concatenated and passed through a convolutional layer to integrate the final features and output the final result. The above process is expressed as:

[0082] ; In the formula, Represents a split operation; Indicates a convolution with a kernel size of 3×3;

[0083] represents the activation function; represents the operation through the channel attention module; Represents element-wise multiplication; Represents a convolution with a kernel size of 1×1; Represents a splicing operation;

[0084] is the final output of the split attention module.

[0085] The hierarchical fusion submodule is used for feature enhancement and feature fusion, such as Figure 4 The hierarchical fusion submodule first uses two independent channel attention CAs to respectively x 1 and x 2 Adaptively recalibrate them, then concatenate and fuse them to produce an enhanced feature representation, and finally refine the features through a convolutional layer. The process is expressed as:

[0086] ;

[0087] The dense feature fusion module is a key component designed to integrate the different levels of features extracted by Transformer and CNN in the network, such as Figure 1 The dense feature fusion module first processes the input features through two 1×1 convolutional layers. and , respectively and The purpose is to preserve the spatial information of the feature while reducing the computational complexity of subsequent operations. The processed features are respectively subjected to adaptive average pooling (AAP) to obtain global information, and then feature transformation is performed through the fully connected layer (FC) to obtain the respective channel attention vectors, as shown below:

[0088] ;

[0089] in, and Respectively represent the feature maps and The calculated channel attention module weights are used to evaluate and The importance of each channel in for , for .

[0090] represents a 1×1 convolution operation, represents adaptive average pooling, represents the fully connected layer, Represents the Sigmoid function, which is used to compress the output of the fully connected layer to between 0 and 1 and generate a weighting factor.

[0091] Then, the original features and By weighting these weighting factors, the features most relevant to the reconstruction task are strengthened, and other irrelevant information is suppressed. Through the cross-attention module, the weighted features are cross-combined to strengthen the complementarity between features, as shown below:

[0092] ;

[0093] in, Indicates that the application After channel attention weight Feature map, Indicates that the application After the channel attention weight Feature map, the design of the cross-attention mechanism enables the two sets of feature maps extracted by Transformer and CNN to learn each other's importance and make full use of all available feature information.

[0094] Finally, the cross-weighted features are concatenated and integrated through another 1×1 convolutional layer to output the final fused features:

[0095] ;

[0096] In the formula, It means splicing two feature maps in the channel dimension. Through the processing of the dense feature fusion module, the network can establish effective connections between features at different levels, reduce the computational complexity of each module, and improve the quality of the reconstructed high-resolution image.

[0097] The upsampling module maps the features processed by the dense feature fusion module to a high-resolution space to generate the final super-resolution image:

[0098] ;

[0099] in, represents the final super-resolution reconstructed image, Represents the upsampling operation. The upsampling module includes a 3×3 convolution layer and a Pixel Shuffle layer.

[0100] S3: Input the verification set processed in step S1 into the trained lightweight image super-resolution reconstruction network to obtain a reconstructed super-resolution image.

[0101] This embodiment also provides a system for implementing the above method.

[0102] A super-resolution reconstruction system includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is executed by the processor to implement any of the methods described above.

[0103] This embodiment also provides a computer storage medium storing a computer program that can implement the above method.

[0104] A computer storage medium stores a computer program, which implements any of the above methods when executed by a processor.

[0105] This embodiment also provides a computer product for implementing the above method.

[0106] A computer product comprises a computer program, and when the computer program is executed by a processor, the method described in any one of the above items is implemented.

[0107] In summary, in terms of feature extraction, the present invention adopts a method of fusing global features and local features, designs a hierarchical edge focusing module to extract local features and edge detail information of the image, and improves the problem of the Transformer-based lightweight image super-resolution reconstruction network ignoring local information of the image; in terms of feature fusion, instead of the previous image fusion method that only adopts a simple pixel addition method, a dense feature fusion module is used to integrate local features and global features, and the cross-attention mechanism is used to enhance the correlation between features, so that the features are richer, which can improve the effect of image super-resolution reconstruction.

[0108] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A lightweight image super-resolution reconstruction method, characterized by: The steps include the following steps: S1: Obtain a low-resolution image dataset, divide the image dataset into a training set and a validation set according to a preset ratio, use Bicubic downsampling to downsample the images in the image dataset to obtain corresponding LR images, and perform random data enhancement operations on each LR image to obtain an enhanced LR image, and randomly cut the enhanced LR image into 64 The image block; S2: construct a lightweight image super-resolution reconstruction network based on the fusion of CNN and Transformer, input the training set processed in step S1 into the lightweight image super-resolution reconstruction network, use the L1 loss function to perform model training, and obtain a trained lightweight image super-resolution reconstruction network; The lightweight image super-resolution reconstruction network includes a shallow feature extraction module, a deep feature extraction module, a dense feature fusion module and an upsampling module; The shallow feature extraction module adopts Convolution transforms the low-resolution images in the input training set into the feature space to generate shallow features. , the operation process of the shallow feature extraction module is as follows: ； In the formula, represents the input low-resolution image, represents a standard 3×3 convolution, Represents the extracted shallow features; The deep feature extraction module includes a trunk branch and a bypass branch. The trunk branch uses a multi-layer stacked feature mixing module to capture the global information in the image, and the bypass branch uses a hierarchical edge focusing module to improve the network's ability to obtain edge information and local details. The processing process of the deep feature extraction module is expressed as follows: ； ； In the formula, Indicates n feature mixing modules, Indicates that n The features extracted by the feature mixing module are represents the hierarchical edge focusing module, represents the features extracted by the hierarchical edge focusing module; The feature mixing module adopts the Transformer structure, which is composed of a spatial adaptive feature modulation module, a layer normalization module and a convolution channel mixing module. The structure of the feature mixing module is similar to the Vit network structure, and the layer normalization replaces the batch normalization to standardize each channel in the feature map. The overall operation of the feature mixing module is expressed as: ； In the formula, F represents the input feature map, is the layer normalization module operation, represents the spatial adaptive feature modulation module, It is a convolutional channel mixing module. The feature mixing module includes residual connections to preserve the original feature information while converting the features. The spatial adaptive feature modulation module is used to effectively obtain the spatial features of the image at multiple scales. Feature modulation is performed at different spatial levels, and channel segmentation is performed on the normalized input features to generate four components. The first component is processed by a 3×3 depth-separable convolution, and the remaining components are sent to the multi-scale feature generation unit. The specific operation is expressed as follows: ； in, Represents a channel split operation; represents a depth-wise separable convolution with a kernel size of 3×3, Indicates that the features of a specific level are upsampled to the original resolution through nearest neighbor interpolation p , Indicates that the input feature is downsampled to 2 i The size of Represents a splicing operation, Represents 1×1 convolution, which connects these multi-scale features through 1×1 convolution and outputs the final feature ; The convolutional channel mixing module is used to enhance the local spatial modeling capability of the trunk branch and perform channel mixing. The convolutional channel mixing module contains a 3×3 convolution and a 1×1 convolution. The first 3×3 convolution encodes the spatial local context and doubles the number of channels of the input features of the mixed channel; the 1×1 convolution reduces the channel to the original number of input bits and uses the GELU activation function to apply to the hidden layer for nonlinear mapping; The hierarchical edge focusing module is used to enhance the image edges and local details. The hierarchical edge focusing module consists of multiple diversion attention modules and hierarchical fusion submodules to extract features at different levels. The hierarchical edge focusing module processes the original feature map in the deep feature extraction stage and combines the enhanced features with the original features through residual connections to maintain the richness of texture information. The process is expressed as follows: ； in, represents the shallow features generated by the shallow feature extraction module, Indicates the first diversion attention module for shallow features The subsequent split attention module further processes the features output by the previous split attention module. Represents the hierarchical fusion submodule; The split attention module uses dynamic channel segmentation technology to carefully process image features, effectively controlling the number of feature map channels while retaining key information to the greatest extent. The split attention module first divides the input feature map into X Split into two branches X 1 and X 2. The first branch X 1 After a series of convolution and ReLU activation layers to refine the features, the second branch X 2 remain unchanged so that the original information is preserved; First of all, X 1 Perform continuous convolution and ReLU activation operations to extract key feature information Then, the channel attention module dynamically adjusts the response strength of the feature channel to filter out the key information and suppress the unimportant information to improve the representativeness of the feature. Multiply element-wise with the output of the channel attention module to obtain a weighted feature map Finally, the two branch features are concatenated and a convolution layer is used to integrate the final features and output the final result. The above process is expressed as: ; In the formula, Represents a channel split operation; Indicates a convolution with a kernel size of 3×3; represents the activation function; represents the operation through the channel attention module; Represents element-wise multiplication; Represents a convolution with a kernel size of 1×1; Represents a splicing operation; is the final output of the split attention module; The hierarchical fusion submodule is used to perform feature enhancement and feature fusion. The hierarchical fusion submodule first uses two independent channel attention CAs to respectively x 1 and x 2 performs adaptive recalibration, then concatenates and fuses them to produce an enhanced feature representation, and finally refines the features through a convolutional layer. The process is expressed as: ； S3: Input the verification set processed in step S1 into the trained lightweight image super-resolution reconstruction network to obtain a reconstructed super-resolution image.

2. The lightweight image super-resolution reconstruction method according to claim 1, characterized in that: The dense feature fusion module is a key component designed to combine the different levels of features extracted by Transformer and CNN in the network. The dense feature fusion module first processes the input features through two 1×1 convolutional layers. and , respectively and The purpose is to preserve the spatial information of the features and reduce the computational complexity of subsequent operations. The processed features are respectively subjected to adaptive average pooling to obtain global information, and then feature transformation is performed through the fully connected layer to obtain their respective channel attention vectors, as shown below: ； in, and Respectively represent the feature maps and The calculated channel attention module weights are used to evaluate and The importance of each channel in for , for ; represents a 1×1 convolution operation, represents adaptive average pooling, represents the fully connected layer, Represents the Sigmoid function, which is used to compress the output of the fully connected layer to between 0 and 1 and generate a weighting factor; Then, the original features and By weighting these weighting factors, the features most relevant to the reconstruction task are strengthened, and other irrelevant information is suppressed. Through the cross-attention module, the weighted features are cross-combined to strengthen the complementarity between features, as shown below: ； in, Indicates that the application After channel attention weight Feature map, Indicates that the application After the channel attention weight Feature map, the design of the cross attention mechanism enables the two sets of feature maps extracted by Transformer and CNN to learn each other's importance and make full use of all available feature information; Finally, the cross-weighted features are concatenated and integrated through another 1×1 convolutional layer to output the final fused features: ； In the formula, It means concatenating two feature maps in the channel dimension. Through the processing of the dense feature fusion module, the network can establish effective connections between features at different levels, reduce the amount of calculation of each module, and improve the quality of the reconstructed high-resolution image; The upsampling module maps the features processed by the dense feature fusion module to a high-resolution space to generate the final super-resolution image: ； in, represents the final super-resolution reconstructed image, Represents the upsampling operation. The upsampling module includes a 3×3 convolution layer and a Pixel Shuffle layer.

3. A super-resolution reconstruction system, comprising a processor and a memory, wherein a computer program is stored in the memory, characterized in that: The computer program is executed by the processor to implement the method according to any one of claims 1 to 2.

4. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

5. A computer product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

Images