Design of robust super resolution algorithms for deteriorated natural images

Abstract

Spatial resolution of an image is limited by the size and density of image sensors. newlineResolution of an image can be increased either by employing high density sensors newlineor by using signal processing techniques. But, the quality of acquired images can newlinebe deteriorated by abnormalities in acquisition medium, damages in sensors, noise newlineduring acquisition and further processing stages, etc. Super resolution refers to a newlinecategory of signal processing techniques that are used to obtain a high resolution (HR) newlineimage from one or more low resolution (LR) images. It has been an attractive topic newlineof research since the last three decades. Applications of super resolution include newlinemedical images, satellite images, face images, surveillance images, text images, newlinefingerprints, microscopic images, etc. Each domain of applications demands specific newlinerequirements and hence poses unique challenges. newlineThe presence of noise in the LR observation severely degrades the performance newlineof a majority of the existing super resolution algorithms. This thesis mainly attempts newlineto develop robust super resolution algorithms, which can reconstruct clean HR newlineimages even from noisy LR observations. Super resolution algorithms can be broadly newlineclassified into learning based methods and reconstruction based methods. In this newlinethesis, three learning based algorithms are proposed for single image super resolution newline(SISR) and face hallucination. These proposed methods require a set of example newlineimages for training. Moreover, two reconstruction based methods are proposed for newlinemulti-frame image super resolution (MFSR) and deteriorated color images. newline