The Algorithms Of Image Super Resolution Recovery Based On Structural Similarity

This thesis primarily studies the image super-resolution algorithms based on structural similarity and its application.Image super-resolution reconstruction is a significant research subject in computer vision and has attracted much interest.The primary objective of it is to recover the missing information and generate high-quality image from an observed degraded image.Digital image,which carries critical details and information,plays an important role in the age of highly-developed computer technology and is required for many research fields such as medical diagnostic,remote sensing,intelligent surveillance.This thesis gives a brief summary of the existing methods and algorithms for super-resolution image reconstruction firstly.Then based on these algorithms,we propose some new methods via structural similarity and carry out a set of experimentations.The primary coverage of the work is depicted as follows:1.Based on spares representation super-resolution method,the work employs the superior advantages of clustering and regression.A collection of images often share same local geometrical structure,on this basis,the training samples are divided into numerous overlapped sub-patches,then the sub-features extracting from these sub-patches are grouped into different clusters.According to the class labels,original patches are divided into different subspaces.Anchored centres for different subspaces are defined by training joint dictionaries,the hierarchical similarity can be learned.Variable mappings between low resolution and high resolution are then obtained.Finally,the degraded image can be reconstructed.2.This work extends the super-resolution method in luminance channel to multiple channels.For the sake of collecting chrominance information,the training patches with multi-channel constraints are extracted from RGB channels directly at the same time as grouping the statistical priors leaned by clustering,which can reconstruct most of the structures and textures with less blurring and richer details.In addition,this approach integrates clustering,collaborative representation,and regression mapping relationships to reconstruct high-resolution image.Also,a continuous reconstructive structure which take example by neural network is established to improve the reconstruction efficiency.3.This part introduces an image super-resolution via deep undesigned feature learning and designed feature tuning with self-similarity.Two independent feature extracting convolutional networks are trained to learn the undesigned feature maps of low resolution(LR)and high resolution(HR)images.Then the residual network acts as an enhancement for reconstruction to fill the gap caused by networks connection procedure.Finally,designed features containing self-similar samples are extracted using differential operators to tune the ultimate reconstructed image.At the end,we analyse the challenges and opportunities of image super-resolution.