Single-Image Super-Resolution Based On Semi-Supervised Learning

High-resolution (HR) images are always welcome in kinds of applications, such as medical diagnosis, biometrics identification, remote surveillance. Unfortunately, it is extremely expensive to collect HR images. Thus, the problem of generating high-resolution images from the given low-resolution (LR) images becomes more and more important now. To deal with the problem, super-resolution (SR) which enhances the resolution of an image with the help of one or more LR images is proposed.Generally, SR algorithms can be categorized into the multi-frame based algorithms and single-image based algorithms. Based on the assumption of sub-pixel of shifting, multi-frame SR algorithms generate a HR image with the help of a series of LR im-ages in common scenes. For single-image SR, a set of LR and HR pairs are used to reconstruct a HR image from a given LR image. Different from multi-frame based al-gorithms, it is unnecessary to ensure the common scenes shared by the LR test image and training images.In this paper,we propose a local semi-supervised learning-based algorithm for single-image SR. Different from most of example-based algorithms, the information of test patches is considered during learning local regression functions which map a LR patch to a HR patch. Localization strategy is generally adopted in single-image SR with nearest neighbor-based algorithms. However, the poor generalization of the nearest neighbor estimation decreases the performance of such algorithms. Though the problem can be fixed by local regression algorithms, the sizes of local training sets are always too small to improve the performance of nearest neighbor-based algorithms sig-nificantly. To overcome the difficulty, the semi-supervised regression algorithm is used here. Unlike supervised regression, the information about test samples is considered in semi-supervised regression algorithms, which makes the semi-supervised regression more powerful. Experiments verify the effectiveness of our algorithm.We organize the thesis as follows:First of all, we introduced the development of SR techniques, and present the current research status around the world; Secondly, the problem of SR is introduce; Thirdly, we make analysis of the existing algorithms. Fourthly, we briefly review the semi-supervised learning. Finally, those existing algo-rithms which have been proposed for SR are based on supervised learning, which share the merits of attractive results and excellent generalization.