| With the popularization of intelligent terminals,the number of images stored in the network has grown rapidly.Image resolution determines the visual effect,but high resolution means more storage space.To reduce the bandwidth during network transmission and the hard disk space during storage,the image super-resolution algorithm allows for down-sample images to reduce the resolution,and then restore the original image before using the up-sampling operation.The challenge of the image superresolution task is to generate real,clear,good visual effect images and reduce the algorithm complexity.In recent years,the technology of image super-resolution has been continuously updated,but the reconstructed results still have problems.Aiming at resolving the challenges and problems in image super-resolution tasks,the main research results of this paper are as follows:(1)Aiming at solving the problem that the existing algorithms cannot effectively extract the key features that appear repeatedly in the image itself,this paper designs and implements a residual fractal convolutional block based on the self-attention mechanism.This block extracts multi-scale features through a multi-branch convolutional structure with exponentially increased receptive fields and realizes staged feature fusion at the end of the branch.It enables the network to gain the ability to focus on the key features in the input image itself,to realize the self-attention mechanism.In addition,this paper further deepens and widens the residual fractal convolutional block to obtain a deep residual fractal network and a wide residual fractal network respectively,improving the ability of feature extraction by introducing more convolutional modules.The experimental results on multiple datasets show that the proposed method can obtain better reconstruction results under fewer levels of parameters.(2)Aiming at solving the problem that the attention mechanism cannot be effectively applied to existing lightweight networks due to the insufficient ability of semantic feature extraction,this paper designs and implements an edge attention model based on extracting high-frequency features.Through down-sampling and up-sampling operations,the structure can actively extract high-frequency features existing in the image.The edge attention module can reduce the learning difficulty of key features and the dependence of the attention mechanism on semantic features.The experimental results show that the edge attention mechanism can reduce the dependence on semantic features compared with other attention mechanisms,and can be directly applied to low-level features,which is more suitable for lightweight networks.(3)Aiming at solving the problem that the existing lightweight networks ignore the blurred area that affects the visual effect in the reconstruction results,this paper designs and implements a refinement module based on U-Net.This module refers to the U-Net structure in the field of image segmentation to use its capability of image processing at the pixel level.The module also effectively combines the task characteristics of super-resolution by introducing adaptive modifications such as skip connections and interpolation,making the module more suitable for superresolution tasks.The module is used to further refine the existing blurred texture in the reconstruction results and has a plug-and-play property for networks with the same module in the reconstruction stage.The experimental results on multiple datasets show that the module can effectively refine the blurred area and obtain more realistic and clearer reconstruction results.Furthermore,the worse the initial effect,the more obvious the refined effect. |
