Image Super-resolution Reconstruction Based On Deep Learning

Spatial resolution is a kind of indication about image quality. Because of its high pixeldensity and good image quality, a high resolution image can provide much richer detailinformation about the corresponding imaging scene. Image super-resolution(SR) is the processof generating a high resolution image from one or several low resolution images about the sameimaging scene, and has become an active research direction in the field of computer vision andimage processing.This thesis surveys the state-of-art about the research of image super resolution andanalyzes several typical types of deep learning model. By combining the theoretical researchresults about deep learning and the need of static image super-resolution reconstruction, Thisthesis puts forward two new methods for image super resolution reconstruction.(1) Image super-resolution reconstruction based on RBM and sparse representation.Inspired by the generic structure characteristics on data representation based on RBM(Restricted Boltzmann Machine) and the process of joint dictionary learning based on sparsetheory, this thesis proposes a method for joint dictionary learning based on RBM andsuper-resolution image reconstruction based on sparse representation. Given the training setproduced from large amount of high-low resolution image patches, the proposed model cansimultaneously learn a joint dictionary composed of both high and low image patch dictionariesfrom a pair of feature space, and the learning process is unsupervised and automatic. With theabove joint dictionary by learning, high resolution image patch can be reconstructed basedon sparse representation. Then the reconstructed image patches are overlapped to form a largeimage, and a high resolution image can be got by means of iterative error compensation.Experimental results verify the effectiveness of the proposed method.(2)Image super-resolution reconstruction based on NSDAE and sparserepresentation.Inspired by the structure characteristics of auto-encoders and the de-blurring effect ofimage super resolution reconstruction, combined with non-negative image characteristics, thispaper presents a method of joint dictionary learning based onNSDAE(Non-negative Sparse Denoising Auto-Encoders) model. This model puts theconstraints of non-negativeness and sparseness on denoising auto-encoders. With the trainingset produced from large amount of high-low resolution image patches, a joint dictionary canbe learned based on the NSDAE model, Then the final high resolution image can bereconstructed via sparse representation. Experimental results demonstrate theeffectiveness of the proposed method.