Lightweight image super-resolution method based on feature distillation and inter-layer information interaction

Abstract

The present invention relates to the field of image processing technology, and discloses a lightweight image super-resolution method based on feature distillation and inter-layer information interaction, comprising the following steps: constructing a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module; constructing a lightweight image super-resolution network based on the dynamic asymmetric distillation module and the inter-layer full-dimensional information interaction module; and realizing image super-resolution using the lightweight image super-resolution network; wherein the lightweight image super-resolution network uses a convolutional layer to extract shallow features from an input low-resolution image, uses a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module to extract deep features from shallow features, and uses a convolutional layer and upsampling to reconstruct images from deep features to obtain high-resolution images. The present invention extracts key features in a lighter and more efficient manner, thereby minimizing computational overhead and the amount of parameters while improving the reconstruction effect.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a lightweight image super-resolution method based on feature distillation and inter-layer information interaction. Background Art

[0002] High-definition images are widely used in medical diagnosis, remote sensing technology, industrial inspection, autonomous driving and other fields. However, due to the performance limitations of acquisition equipment and the scarcity of high-definition resources, it is necessary to introduce image super-resolution (SR) technology to improve the details and clarity of low-resolution images and retain the original pixel information as much as possible to reconstruct high-resolution (HR) images. SR methods based on deep convolutional neural networks (DCNNs) have powerful feature extraction capabilities and have therefore received widespread attention and research. SRCNN introduced CNN into SR in an end-to-end manner for the first time. VDSR significantly improved the convergence efficiency of the network by introducing residual learning technology. EDSR further improved VDSR and designed a deep network with a parameter count of 43M, achieving excellent super-resolution effects. SRGAN, based on the generative adversarial network architecture, improves the quality and visual experience of high-resolution images. Methods based on DCNNs tend to use deeper networks and a large number of parameters, which limits their practical applications. Therefore, it is particularly important to develop lightweight and efficient super-resolution models.

[0003] Recently, scholars have conducted in-depth research on lightweight SR networks. In order to solve the problem that the performance of the model decreases when processing complex image features due to parameter reduction, a popular method is to establish a multi-scale feature extraction network. IDN fuses multi-scale features through the information distillation mechanism to extract key information. Furthermore, RFDN adopts a progressive refinement module and introduces channel segmentation technology to shorten the running time of the network. RLFN introduces contrast loss and achieves significant SR performance. However, the above SR methods based on the information distillation mechanism still have three problems: First, due to the limitations of traditional convolution, the feature distillation block cannot fully utilize the directionality of spatial information, and there is unnecessary computational overhead. Second, in the distillation process, the above method uses a longer connection path, which leads to inefficient feature representation and affects the rapid transmission of feature information. Third, there is information loss in the distillation process, and the information integration at different stages is insufficient, and there is a lack of effective use of feature information.

[0004] In addition, lightweight SR networks usually use residual modules or densely connected modules to capture information from low-level features to high-level features. Different hierarchical features can complement each other in the fusion process to enhance the feature representation ability of the model. However, this method cannot fully explore the potential correlation between the internal space and channels of the hierarchical features, and lacks effective use of the interactive information between different levels. In addition, the subsequent residual blocks only receive the fused complex features, resulting in the useful original residual features and intermediate features being masked, thereby hindering the performance of SR. Summary of the invention

[0005] The purpose of the present invention is to solve the problems in the prior art.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide a lightweight image super-resolution method based on feature distillation and inter-layer information interaction, comprising the following steps:

[0007] Construct a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module;

[0008] A lightweight image super-resolution network is constructed based on a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module;

[0009] Implementing image super-resolution using the lightweight image super-resolution network;

[0010] Among them, the lightweight image super-resolution network uses convolutional layers to extract shallow features of the input low-resolution image, uses dynamic asymmetric distillation modules and inter-layer full-dimensional information interaction modules to extract deep features of the shallow features, and uses convolutional layers and upsampling to reconstruct the deep features to obtain a high-resolution image.

[0011] Preferably, the method of realizing image super-resolution by using the lightweight image super-resolution network comprises the following steps:

[0012] The 3×3 convolutional layer is used to extract shallow features of the input low-resolution image, which can be expressed as:

[0013] ;

[0014] in, represents the input low-resolution image, is the output of the 3×3 convolution, represents a 3×3 convolutional layer;

[0015] The dynamic asymmetric distillation module and the inter-layer full-dimensional information interaction module are used to extract deep features from shallow features, which can be expressed as:

[0016] ;

[0017] ;

[0018] in, is the output of the ith dynamic asymmetric distillation module, is the output of the jth inter-layer full-dimensional information interaction module; the output of the 4th inter-layer full-dimensional information interaction module is the final deep feature;

[0019] The deep features are reconstructed using convolutional layers and upsampling to obtain a high-resolution image, which is expressed as:

[0020] ;

[0021] in, represents the output high-resolution image, Indicates a rebuild operation.

[0022] Preferably, the internal execution process of the dynamic asymmetric distillation module comprises the following steps:

[0023] The input features are distilled using asymmetric separable convolution and additional convolutional layers, expressed as:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] in, represents the input features of the dynamic asymmetric distillation module; , and They represent the first group of distillation characteristics, the second group of distillation characteristics, and the third group of distillation characteristics, respectively; , and They represent the first group of coarse features, the second group of coarse features, and the third group of coarse features respectively; , and Represent the first convolution layer, the second convolution layer, and the third convolution layer respectively; represents asymmetric separable convolution;

[0030] The distilled features are fused and enhanced, expressed as:

[0031] ;

[0032] in, This is the final output of the dynamic asymmetric distillation module. ; To enhance the spatial attention block, is a 1 × 1 convolution, Represents a join operation.

[0033] Preferably, asymmetric separable convolution It is expressed as:

[0034] ;

[0035] ;

[0036] ;

[0037] in, and They represent the outputs of the two branches in the asymmetric separable convolution, Asymmetric separable convolution The output, represents a local separable convolution with a convolution kernel of 1×3. Indicates that the convolution kernel is a local separable convolution of 3×1; Represents the activation function.

[0038] Preferably, the locally separable convolution includes partial convolution and point-by-point convolution, which is expressed as:

[0039] ;

[0040] ;

[0041] in, represents the output of partial convolution, After locally separable convolution Output: Indicates the specified channel subset, represents the expansion rate; represents a partial convolution operation, Represents a point-wise convolution operation.

[0042] Preferably, the internal execution process of the inter-layer full-dimensional information interaction module includes the following steps:

[0043] Extract spatial and channel correlation information at different levels from the input features to obtain a feature matrix with global attention weights;

[0044] Dynamically weight the feature matrix with global attention weights to obtain a reshaped feature matrix;

[0045] The integrated and reshaped feature matrix is ​​used as the output of the inter-layer full-dimensional information interaction module.

[0046] Preferably, extracting spatial and channel association information at different levels from the input features comprises the following steps:

[0047] Input features based on inter-layer full-dimensional information interaction module , introduce local separable convolution to generate query matrix , key matrix , value matrix , respectively expressed as:

[0048] ;

[0049] in, , and They represent the linear transformation matrix corresponding to the query vector, the linear transformation matrix corresponding to the key vector, and the linear transformation matrix corresponding to the value vector, respectively. k represents the kth input of the inter-layer information interaction module.

[0050] Calculate the first intermediate result , expressed as:

[0051] ;

[0052] in, represents the normalized exponential function;

[0053] After rotation , expressed as:

[0054] ;

[0055] Calculate the second intermediate result , expressed as:

[0056] ;

[0057] Second intermediate result After matrix transposition, the feature matrix with global attention weights is output , expressed as:

[0058] ;

[0059] in, represents the rotation operation along the spatial axis, Indicates reverse rotation operation.

[0060] Preferably, the dynamically weighting the feature matrix with global attention weights comprises the following steps:

[0061] Calculate the weight matrix , expressed as:

[0062] ;

[0063] ;

[0064] in, represents the matrix reshaping operation, and n represents the number of inputs of the inter-layer information interaction module;

[0065] Combine the weight matrix with the feature maps of different layers Multiplication realizes dynamic weighting and obtains a weighted feature group, which is expressed as:

[0066] ;

[0067] in, represents 1 × 1 convolution;

[0068] Reshape the weighted feature groups into , , expressed as:

[0069] ;

[0070] in, Represents the reshaped feature matrix.

[0071] Preferably, the integrated reshaping feature matrix is ​​expressed as:

[0072] ;

[0073] in, represents the output of the inter-layer full-dimensional information interaction module, represents the splicing operation along the channel, represents a 1 × 1 convolution.

[0074] Preferably, the lightweight image super-resolution network is used to realize image super-resolution. The low-resolution image is used as input, and the output is a Dimensional high-resolution output, where scale is the required magnification.

[0075] The present invention has the following beneficial effects:

[0076] (1) The present invention improves the reconstruction effect while minimizing the computational overhead and the number of parameters;

[0077] (2) The asymmetric separable convolution module of the present invention extracts horizontal and vertical features separately and integrates local separable convolution to minimize computational overhead. The local separable convolution introduces an adjustable dilation rate to extract key features in a lighter and more efficient way. The dynamic asymmetric distillation module of the present invention combines the information distillation mechanism and asymmetric separable convolution to capture key features, while using a shorter connection path to achieve equivalent feature representation, and multiplexes the distillation information at different stages to compensate for the information loss in the distillation process, thereby achieving efficient multi-scale feature extraction.

[0078] (3) The inter-layer full-dimensional information interaction module of the present invention realizes the interaction of inter-layer information by focusing the spatial and channel correlation information to different levels and dynamically assigning weights according to the inter-layer similarity, making full use of the hierarchical features to provide details and semantic information for SR.

[0079] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 A method step diagram of an embodiment of the present invention;

[0081] Figure 2 Schematic diagram of a lightweight image super-resolution network framework according to an embodiment of the present invention;

[0082] Figure 3 Schematic diagram of local separable convolution according to an embodiment of the present invention;

[0083] Figure 4 Schematic diagram of asymmetric separable convolution according to an embodiment of the present invention;

[0084] Figure 5 A schematic diagram of a dynamic asymmetric distillation module according to an embodiment of the present invention;

[0085] Figure 6 Schematic diagram of an inter-layer full-dimensional information interaction module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0086] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0087] S101, construct dynamic asymmetric distillation module and inter-layer full-dimensional information interaction module;

[0088] S102, build a lightweight image super-resolution network based on the dynamic asymmetric distillation module and the inter-layer full-dimensional information interaction module;

[0089] S103, realizing image super-resolution using the lightweight image super-resolution network;

[0090] Among them, the lightweight image super-resolution network uses convolutional layers to extract shallow features of the input low-resolution image, uses dynamic asymmetric distillation modules and inter-layer full-dimensional information interaction modules to extract deep features of the shallow features, and uses convolutional layers and upsampling to reconstruct the deep features to obtain a high-resolution image.

[0091] See also Figure 2 As shown in the figure, it is a schematic diagram of the lightweight image super-resolution network framework of an embodiment of the present invention, which involves 1 3×3 convolution, 6 dynamic asymmetric distillation modules, 4 inter-layer full-dimensional information interaction modules and 1 image reconstruction module. A hierarchical feature progressive aggregation structure is designed, which integrates complementary information across layers in groups and progressively, enhances the intermediate residual features, and the image reconstruction module uses PixelShuffle for upsampling. The above process can be defined as:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] in, and Represent the input low-resolution image and the output high-resolution image respectively, is the output of the 3×3 convolution, is the output of the ith dynamic asymmetric distillation module, is the output of the j-th inter-layer full-dimensional information interaction module; represents a 3×3 convolutional layer, is the function corresponding to the inter-layer full-dimensional information interaction module, To rebuild the module.

[0097] See also Figure 3 FIG. 1 is a schematic diagram of a local separable convolution according to an embodiment of the present invention. The local separable convolution module includes a partial convolution and a point-by-point convolution. , Selectively perform convolution operations on some input channels, and introduce a dilation rate to expand the receptive field of the convolution kernel; then, use point-by-point convolution to integrate cross-channel information to obtain the final output. The operation is defined as:

[0098] ;

[0099] ;

[0100] in represents the output after partial convolution, represents the output of the locally separable convolution; Indicates the specified channel subset, represents the expansion rate; represents partial convolution, represents point-by-point convolution; in the embodiment of the present invention, C' is set to one third of the number of channels (the number of channels is 48 by default), and the expansion rate is set to 2.

[0101] See also Figure 4 As shown, it is a schematic diagram of an asymmetric separable convolution according to an embodiment of the present invention. The asymmetric separable convolution module includes two 3×1 local separable convolutions and two 1×3 local separable convolutions, which constitute two opposite asymmetric branches. The operation is defined as:

[0102]

[0103]

[0104]

[0105] in, , Represent the outputs of the two asymmetric branches, is the final output of the asymmetric separable convolutional block, represents the local separable convolution operation under different convolution kernels, yes function.

[0106] See also Figure 5 , which is a schematic diagram of a dynamic asymmetric distillation module according to an embodiment of the present invention. The dynamic asymmetric distillation module includes three steps of feature distillation, feature fusion and feature enhancement using an asymmetric separable convolution module and an additional convolution layer.

[0107] Specifically, the feature distillation using asymmetric separable convolutional modules and additional convolutional layers is as follows: The distillation process is divided into three stages. In the first stage, the input feature Divided into two paths, one path obtains the first set of distilled features through 1 × 1 convolution The other way is to enhance the feature representation through an asymmetric separable convolution module to obtain coarse features ;Will As the input of the second stage, the distilled features are reused , repeat the above operation to obtain the second set of distillation characteristics and coarse features ; Finally, As the input of the third stage, the distilled features are reused and , the third set of distilled features is extracted using 3× 3 convolution The above operation can be defined as:

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] in, , and They represent the first group of distillation characteristics, the second group of distillation characteristics, and the third group of distillation characteristics, respectively; , and They represent the first group of coarse features, the second group of coarse features, and the third group of coarse features respectively; , and Represent the first convolution layer, the second convolution layer, and the third convolution layer respectively; Represents an asymmetric separable convolution.

[0114] Specifically, in the feature fusion and feature enhancement stage, the distilled features are concatenated through 1 × 1 convolution, and then the ESA layer is introduced to amplify the spatial details in the fused features. The operation can be defined as:

[0115] ;

[0116] in, It is the final output of the dynamic asymmetric distillation module; To enhance the spatial attention block, It is a 1 × 1 convolution.

[0117] See also Figure 6As shown, it is a schematic diagram of the inter-layer full-dimensional information interaction module of an embodiment of the present invention. The inter-layer full-dimensional information interaction module includes three steps of extracting spatial and channel association information of different levels, calculating the similarity between layers for dynamic weighting, and integrating hierarchical features.

[0118] Specifically, the extraction of spatial and channel association information at different levels is as follows: for the intermediate feature map , , n represents the number of inputs of the inter-layer full-dimensional information interaction module. In the embodiment of the present invention, n=2, and n=3 can also be taken. First, introduce local separable convolution to generate query, key, and value matrices , and then , The matrix multiplication is passed through the activation function to obtain a size of Spatial attention map. Afterwards, the spatial attention map is combined with Multiply to get the intermediate result The next stage is to copy After rotation , used as the subsequent channel attention. Then, the size is Channel attention map and Multiply to get the intermediate result , after matrix transposition, the feature matrix with global attention weights is output , the above process can be expressed as:

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] in, is the input embedding matrix of channel attention, represents the rotation operation along the spatial axis, Indicates reverse rotation operation.

[0125] Specifically, the similarity between the calculation layers is dynamically weighted as follows: the weight matrix of the hierarchical features is obtained by matrix transposition and multiplication: , and then through a 1 × 1 convolution layer to enhance the feature representation, and then combine it with the feature maps of different layers Multiplication realizes dynamic weighting, and finally reshapes the intermediate feature group into , , the above process can be expressed as:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] in, represents a matrix reshaping operation, represents a 1 × 1 convolution.

[0131] Specifically, the integration of hierarchical features is as follows: first, the hierarchical features are concatenated along the channel, and then the dimension is reduced through a 1×1 convolution layer. The operation can be defined as:

[0132] ;

[0133] in, It represents the full-dimensional information interaction module between layers. represents a 1 × 1 convolution.

[0134] Specifically, in S103, the lightweight image super-resolution network is used to achieve image super-resolution, including training and testing the lightweight image super-resolution network; the embodiment of the present invention uses the DIV2K dataset for training, which consists of 800 high-resolution training images; in addition, five SR benchmark datasets are selected: Set5, Set14, B100 and Urban100, Manga109 as test sets to evaluate the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) on the Y channel of the image. In the training stage, the LR image is cut into many image blocks of size 64×64, and the initial learning rate is set to 5×10-4; the Adam optimizer with default parameters is used to perform 1000 rounds of iterative training on the model with a batch size of 64; the training platform is built on the PyTorch1.7.0 platform, and the NVIDIA GeForce TITAN RTX 24GB GPU is used for training and testing.

[0135] Specifically, The low-resolution image is input into the lightweight image super-resolution network trained in the embodiment of the present invention, and the output obtained is a dimensional high-resolution output, where scale is the required magnification factor, for example, scale=2, we get High resolution.

[0136] It can be seen that the present invention minimizes the computational overhead and the number of parameters while improving the reconstruction effect; the asymmetric separable convolution module extracts horizontal and vertical features separately and integrates local separable convolutions to minimize the computational overhead, thereby extracting key features in a lighter and more efficient way; the dynamic asymmetric distillation module refines features step by step by combining the information distillation mechanism and asymmetric separable convolution, while using a shorter connection path to achieve equivalent feature representation, and multiplexing the distillation information at different stages to make up for the information loss in the distillation process, thereby achieving efficient multi-scale feature extraction; the inter-layer full-dimensional information interaction module focuses the spatial and channel correlation information on different levels, and dynamically allocates weights according to the inter-layer similarity, to achieve the interaction of inter-layer information, and make full use of the hierarchical features to provide details and semantic information for SR.

[0137] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A lightweight image super-resolution method based on feature distillation and inter-layer information interaction, characterized in that: The following steps are involved: Construct a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module; A lightweight image super-resolution network is constructed based on a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module; Implementing image super-resolution using the lightweight image super-resolution network; The lightweight image super-resolution network uses a convolutional layer to extract shallow features from an input low-resolution image, uses a dynamic asymmetric distillation module and an inter-layer full-dimensional information interaction module to extract deep features from shallow features, and uses a convolutional layer and upsampling to reconstruct deep features to obtain a high-resolution image. The method of realizing image super-resolution by using the lightweight image super-resolution network comprises the following steps: The 3×3 convolutional layer is used to extract shallow features of the input low-resolution image, which can be expressed as: I0=f SF (I LR ); Among them, I LR represents the input low-resolution image, I0 is the output of the 3×3 convolutional layer, and f SF (·) indicates a 3×3 convolutional layer; The dynamic asymmetric distillation module and the inter-layer full-dimensional information interaction module are used to extract deep features from shallow features, which can be expressed as: in, is the output of the ith dynamic asymmetric distillation module, is the output of the jth inter-layer full-dimensional information interaction module; the output of the 4th inter-layer full-dimensional information interaction module is the final deep feature; The deep features are reconstructed using convolutional layers and upsampling to obtain a high-resolution image, which is expressed as: Among them, I HR represents the output high-resolution image, f Rec (·) indicates a reconstruction operation; The internal execution process of the dynamic asymmetric distillation module includes the following steps: The input features are distilled using asymmetric separable convolution and additional convolutional layers, expressed as: I d1 =f D1 (I in ); I a1 =f ASCB (I in ); I d2 =f D2 (I a1 +I d1 ); I a2 =f ASCB (I a1 +I d1 ); I d3 =f D3 (I a2 +I d1 +I d2 ); Among them, I in I represents the input features of the dynamic asymmetric distillation module; d1 ,I d2 and I d3 I represent the first group of distillation characteristics, the second group of distillation characteristics and the third group of distillation characteristics respectively; a1 and I a2 Respectively represent the first group of coarse features and the second group of coarse features; f D1 (·),f D2 (·) and f D3 (·) denotes the first, second and third convolution layers respectively; f ASCB (·) represents asymmetric separable convolution; The distilled features are fused and enhanced, expressed as: I out =f ESA (f Conv1 (Concat(I d1 ,I d2 ,I d3 ))); Among them, I out This is the final output of the dynamic asymmetric distillation module. f ESA (·) is the enhanced spatial attention block, f Conv1 (·) is a 1×1 convolution, and Concat(·) represents a concatenation operation; The internal execution process of the inter-layer full-dimensional information interaction module includes the following steps: Extract spatial and channel correlation information at different levels from the input features to obtain a feature matrix with global attention weights; Dynamically weight the feature matrix with global attention weights to obtain a reshaped feature matrix; Integrate and reshape the feature matrix as the output of the inter-layer full-dimensional information interaction module; The step of extracting spatial and channel correlation information at different levels from the input features to obtain a feature matrix with global attention weights includes the following steps: Input feature H based on the inter-layer full-dimensional information interaction module k ∈R (H,W,C) , introduce local separable convolution to generate query matrix Q s , key matrix K s , value matrix V s , respectively expressed as: Q s =H k ·W q ,K s =H k ·W k ,V s =H k ·W v ; Among them, W q , W k and W v They represent the linear transformation matrix corresponding to the query vector, the linear transformation matrix corresponding to the key vector, and the linear transformation matrix corresponding to the value vector, respectively. k represents the kth input of the inter-layer full-dimensional information interaction module. Calculate the first intermediate result A s , expressed as: A s =SoftMax(Q s K sT )·V s ; Among them, SoftMax(·) represents the normalized exponential function; Q s , K s 、V s After rotation, we get Q c , K c 、V c , expressed as: Q c =f R (Q s ),K c =f R (K s ),V c =f R (V s ); Calculate the second intermediate result A c , expressed as: A c =SoftMax(Q c K cT )·V c ; Second intermediate result A c After matrix transposition, the feature matrix H′ with global attention weights is output. k ∈R (H,W,C) , expressed as: Among them, f R (·) represents the rotation operation along the spatial axis, Indicates a reverse rotation operation.

2. The lightweight image super-resolution method based on feature distillation and inter-layer information interaction according to claim 1, characterized in that: Asymmetric separable convolution f ASCB (·) is expressed as: I1=f LSConv1×3 (σ(f LSConv3×1 (I in ))); I2=f LSConv3×1 (σ(f LSConv1×3 (I in ))); I' out =σ(I1+I2); Among them, I1 and I2 represent the outputs of the two branches in the asymmetric separable convolution, and I′ out is an asymmetric separable convolution f ASCB The output of (·), f LSConv1×3 represents a local separable convolution with a convolution kernel of 1×3, f LSConv3×1 represents a locally separable convolution with a convolution kernel of 3×1; σ(·) represents an activation function.

3. The lightweight image super-resolution method based on feature distillation and inter-layer information interaction according to claim 2 is characterized in that: The locally separable convolution includes partial convolution and point-by-point convolution, which can be expressed as: I′ in =f PConv (I in ,C′,d); I” out =f PWConv (I′ in ); Among them, I′ in Represents the output of the partial convolution, I' out is a locally separable convolution f LSConv The output of; C ′ represents the specified channel subset, d represents the expansion rate; f PConv (·) represents a partial convolution operation, f PWConv (·) represents a point-wise convolution operation.

4. The lightweight image super-resolution method based on feature distillation and inter-layer information interaction according to claim 1, characterized in that: The step of dynamically weighting the feature matrix with global attention weights to obtain a reshaped feature matrix includes the following steps: Calculate the weight matrix A l , expressed as: H′ group =f reshape (H′1,…H′ k ,…H′ n ); A l =SoftMax(f reshape (H′ group )·H′ group ); Among them, f reshape (·) represents the matrix reshaping operation, and n represents the number of inputs of the inter-layer full-dimensional information interaction module; The weight matrix is ​​combined with the feature map H′ of different layers group Multiplication realizes dynamic weighting and obtains a weighted feature group, which is expressed as: H″ group =f conv1 (A l )·H′ group ; Among them, f conv1 (·) represents 1×1 convolution; Reshape the weighted feature group into H″ k ∈R (H,W,C) , k={1,…,n}, expressed as: (H″1,…H″ k ,…H″ n )=f reshape (H″ group ); Among them, H k Represents the reshaped feature matrix.

5. The lightweight image super-resolution method based on feature distillation and inter-layer information interaction according to claim 1, characterized in that: The integration reshapes the feature matrix, expressed as: I HGIB =f conv1 ([H″1,…H″ k ,…H″ n ]); Among them, I HGIB represents the output of the inter-layer full-dimensional information interaction module, [·] represents the splicing operation along the channel, and f conv1 (·) denotes a 1×1 convolution.

6. The lightweight image super-resolution method based on feature distillation and inter-layer information interaction according to claim 1, characterized in that: The lightweight image super-resolution network is used to realize image super-resolution, and an H×W×C low-resolution image is used as input, and the output obtained is a high-resolution output of (H*scale)×(W*scale)×C dimensions, where scale is the required magnification.

Images