| The advent of computer,information processing,and visual communication technology has ushered in a new age of information technology.The amount of data accessible to people is on the rise,thus necessitating the continual advancement and growth of information processing technology to offer more convenient,expeditious,and varied services to individuals.The utilization of digital image and its related processing technology has become a major component of information processing technology,which is increasingly being employed in numerous areas.For digital images in some cases generally require high-resolution images,such as medical images require to be able to show those subtle lesions that can not be distinguished by the human eye;satellite ground requires images to be able to distinguish at least the face of a person,and even expect the effect to be comparable to ID photos;some detection and identification control devices require images of high enough resolution to ensure the accuracy of measurement and control.Therefore,improving image resolution has been a goal pursued in the field of image acquisition.The field of computer vision has seen a surge in interest in the use of image superresolution reconstruction algorithms to analyze the necessary image data and thereby enhance image quality,due to the advancement of science and technology,as well as the growth of deep learning.However,there are still some problems with the current methods.First,some of the current methods tend to be single-scale,which means that a network model can only achieve one scale of image super-resolution,which seriously affects the realistic application of the model.Secondly,it has been a constant goal to further improve the effectiveness of the model.Finally,due to the high equipment requirements of deep learning,some of the more effective networks are often difficult to deploy to the industry due to the excessive parameters and operations.In this thesis,we focus on the existing problems of image super-resolution,from image feature extraction and supplementary location information to image representation at arbitrary scales,etc.The main work and contributions of this thesis contain the following points.(1)This thesis employs a Cross Transformer to capture the features within the image and utilizes the self-attention mechanism to identify the essential visual features of the image,thus aiding the network in recovering the image with enhanced visual effects and a greater measurement index.Additionally,we analyze the significance of location information for image super-resolution and propose a dynamic location coding module,then fuse and learn location information with image visual features,and finally complete image super-resolution at any scale by implicit image representation.Our approach yields superior outcomes in both objective evaluation index and visual performance to the more sophisticated techniques.(2)In order to further realize the lightweight of the model,this thesis adopts a sparse self-attentive mechanism to construct a lightweight and efficient transformer model for extracting features that are conducive to the recovery of image texture details.This thesis also proposes a dual-stream interactive transformer for location information supplementation,which can alleviate the discomfort of image oversampling to a certain extent and make image recovery more accurate through the accurate use of location information.Meanwhile,in order to further refine the texture details,this thesis proposes a texture estimation module and a texture enhancement loss function,which can obtain more realistic and natural images by further constraining the texture details.Finally,validation of the proposed method in this thesis on multiple super-resolution datasets for experiments was conducted to assess the algorithm’s effect.The experimental results demonstrate that our method achieves better results on several benchmark datasets. |
