Single Image Super Resolution Method Research Based On Learning

High-resolution image has great value in multimedia and research field because of having much richer image detail information.Image Super-resolution(SR)is an software processing algorithm that aims to reconstruct low-resolution image(LR)to corresponding high resolution image(HR)in the same image scene.This method decreases the hardware cost on enhancing image super-resolution in some degree and has extensive application value.Meanwhile,it has been becoming one of the significant study field of computer vision.In this thesis,we survey and summarize the state-of-art research of image super resolution.Based on sparse representation and deep learning theory,we analysis the present technologies and propose corresponding improvement algorithm.(1)Single Image Super-Resolution Based on Sparse Representation in Common SpaceThere are many similar structural information in high and low resolution image.In order to enhance their correlation coefficients,this thesis presents a single image super-resolution algorithm based on sparse representation in common space.We apply canonical correlation analysis to find the relationship between high resolution and low resolution image pairs.Meanwhile,the solution of sparse representation can be solved in common space transformed by canonical correlation analysis.By using a different training approach called online dictionary learning,the training speed is improved considerably.The principal component analysis method is used to reduce the image feature in order to improve the speed of reconstructing high resolution image.In addition,a post-processing method based image prior is used to eliminate blurring and ringing artifacts around major edges and further improve the image quality.Experimental results demonstrate that our proposed algorithms obtain improved quality and excellent running performance.(2)Single Image Super-Resolution Using Deep Convolutional NetworksIn order to simplify tedious image reconstruction process in traditional methodand break the limit of improving image reconstruction quality,this thesis proposes a single image super-resolution method using the combination of multiple deep convolutional neural networks.We aim to learn an end to end image super resolution algorithm frame.By combining the convolutional networks with receptive field of different sizes,our proposed method takes full advantage of diversity filters to extract image features.Compared with other excellent algorithms,experimental results demonstrate our network presents good reconstruction ability for single image and accelerates convergence rate on training stage.