Research On Gradient Prior And Sparse Representation Based Image Super-Resolution Algorithm

The vision is an important way for people to get outside information.The most essential object observed by the vision is the image with a variety of information,so the image becomes the main carrier of human access to information.In real life,due to the limitation of hardware imaging equipment,manufacturing process and information transmission conditions,people usually only get the image with low resolution.However the low resolution image cannot satisfy the applications in some areas such as video surveillance and medical detection etc..Image super-resolution technology can solve the limitation mentioned above and improve image resolution by software,which help people achieve more information.At present this technology has become the hot research topic.The idea of image super-resolution reconstruction is to use one or more low resolution images from the same scene as input and reconstruct higher resolution image.In this dissertation,image super-resolution reconstruction algorithms are studied.Firstly,the research background and significance of super-resolution reconstruction technology are introduced.Then,three super-resolution methods are described,and analyze the sparse representation model.Finally,the dissertation proposes the image super-resolution algorithm based on gradient prior and sparse representation to overcome the shortcomings of the dictionary learning in the one class.The main contributions of this dissertation are as follows:(1)The gradient prior information of image patches is introduced into the classification model of image patches.The gradient prior information of the image patches is used to classify the image patches,and similar image patches are divided into the same class.(2)The process of dictionary training is optimized.This dissertation improves the dictionary learning method based on sparse representation,that is,high resolutiondictionaries are learned and low resolution dictionaries are obtained by extracting features of high resolution dictionaries.Thus this optimization greatly reduces the cost in dictionary training process.The proposed algorithm is used to do experiments on several images.The experimental results show that the proposed algorithm can improve the visual quality,and the PSNR and SSIM of the image are also raised.