Learning-Based Single Image Super Resolution

Abstract

Recent advancements in signal processing techniques have led to obtain more high resolution images. A high resolution image refers to an image with high density of pixels. The importance and desire of high resolution images are obvious in the field of electronic and digital imaging applications.The quality of an image can be improved either by hardware or software approaches. Hardware approaches are straightforward solutions to enhance the quality of a given image, but some constraints, such as chip size increment, making them expensive to some extend. Therefore, most of the researchers are focused on software methods. Super resolution is one of the software image processing approaches where a high resolution image can be recovered from low resolution one(s). The main goal of super resolution is the resolution enhancement. This topic has been widely brought into attention in image processing society due to the current and future application demands especially in the field of medical applications. Super resolving a high resolution image can be performed from either a single low resolution or many low resolution images. This thesis is completely concentrated on Single Image Super Resolution (SISR) where a single low resolution image is the candidate to be exploited as the input image. There are several classes of methods to obtain SISR where three important ones, i.e., the Example-based, Regression-based and Self-similarity-based are investigated within this thesis. This thesis evaluates the performance of the above-mentioned methods. Based on achieved results, the Regression method shows better performance compared to other approaches. Furthermore, we utilize parameters, such as patch size, to improve the numerical and virtual results in term of PSNR and resolution, respectively. These modifications are applied to the Regression-based and Self-similarity-based methods. The modified algorithms in both methods lead to improve results and obtain the best ones.