| Image super-resolution is the task of restoring a low-resolution image to a high-resolution image with higher visual quality.With super-resolution techniques,images can be rendered in finer detail,thus allowing researchers to observe more complex patterns and structures.However,traditional methods often suffer from problems in establishing spatial mappings between low-resolution and high-resolution images,such as lack of clear definitions and inefficient algorithms.In contrast,deep learningbased methods can directly learn the mapping relationship between lowresolution and high-resolution images through end-to-end model training,enabling high-quality feature learning and extraction with significant improvements in performance and various evaluation criteria.Currently,deep learning-based super-resolution algorithms still face some challenging problems,including difficulties in recovering fine details and high computational complexity.To address these problems and effectively improve the performance of image super-resolution reconstruction,the main research and innovations of this thesis are as follows.(1)Aiming at the problem of high image computational complexity and poor performance for different image sizes of existing super-resolution multi-headed self-attention calculus,an image super-resolution algorithm based on multi-dimensional multi-headed self-attention is proposed.The algorithm obtains two forms of local and global self-attentive operations by decomposing the attention of the whole planar space in spatial dimensions,which reduces the computational complexity while effectively capturing global long-range features and better adapting to image inputs of different sizes.In addition,the algorithm embeds an implicit position encoding module for super-resolution,which encodes the position information of each pixel into a fixed-length vector in order to allow the model to better learn the position relationship between pixels and thus effectively reconstruct the true texture in the image.The algorithm excels in reducing computational complexity and adapting to different size image inputs,allowing better reconstruction of images with realistic textures.(2)Aiming at the problem of difficult reconstruction of real texture details caused by complex textures of images,a novel image superresolution algorithm based on full-region multiscale features(Hybrid CNN-Transformer networks,HCT)is proposed.The algorithm achieves full-region multiscale feature extraction by feeding images into a network architecture consisting of multi-headed self-attentive branches and convolutional branches and complementing the network learning results at each stage by fusing the two branches.In addition,the algorithm uses a dynamic interaction module(Dual-Global-Local,DGL)to realize the feature information interaction between the local features extracted by CNN and the global features extracted by multi-headed self-attention to achieve multi-scale feature fusion.The experimental results show that the algorithm can reconstruct higher quality high-resolution images so that the output image presents clearer and more realistic detailed features and texture details.(3)The image super-resolution reconstruction system is designed and implemented to provide users with a convenient and independent service for image uploading and image reconstruction.The system is designed to have a scientific division of functions,including a visualization interface module,a user information management module,and an image superresolution reconstruction module.The visualization interface module is responsible for providing user-friendly interaction to enable users to use the system more conveniently;the user information management module is responsible for managing user accounts and message center information;the image super-resolution reconstruction module is responsible for performing image super-resolution reconstruction and managing the history of super-resolution operations.The system uses lightweight web architecture to facilitate installation and deployment and is divided into front-end,back-end,and data storage to improve user experience. |
