| Most of the information obtained by people comes from images.The amount of information and the authenticity of information largely depend on the quality of images.Resolution is used to measure the quality of image,which reflects the clarity of the image detail texture.More information can be obtained by improving the resolution.Image super-resolution(SR)reconstruction aims to recover the ground-truth images from the degenerative images,which can effectively improve the clarity of image and recover the high-frequency information of the image.Therefore,the reconstruction algorithms have important research value.In recent years,deep learning has made excellent achievements,and the SR reconstruction networks have also made a breakthrough.In this paper,super-resolution reconstruction networks with powerful learning and representation ability are constructed based on neural models to reconstruct images with high spatial resolution and high fidelity.The specific research contents are as follows:Firstly,a super-resolution model based on second-order progressive feature fusion network is constructed,and its core part is the stacked progressive feature fusion blocks.In order to make full use of the features in the network,the progressive feature fusion structure is proposed,and the progressive feature fusion block is constructed by using this structure.In addition,the second-order feature fusion mechanism is proposed,which uses the same structure to fuse features at the local and global levels to make the best use of the learned feature information.Experiments show that the performance of the second-order progressive feature fusion network is better than most of classical models on several benchmark sets.Secondly,a new dual-branch feature learning network for more accurate super-resolution reconstruction is constructed,and its core part is the stacked dual-branch feature learning blocks.In order to obtain as many effective image features as possible,a dual-branch feature learning block is proposed,which uses a multi-level dual-branch structure to extract image information.At the same time,some components to improve feature utilization are introduced into block to maximize the learning ability of the dual-branch feature learning block.Experiments show that the performance of dual-branch feature learning network is better than most models including second-order progressive feature fusion network on several benchmark sets,which indicates that dual-branch feature learning network can learn more representative features and generate more accurate super-resolution reconstruction results. |
