Super-resolution based on fast linear kernel regression

Jianmin Li, Yanyun Qu, Ying Gu, Tian-Zhu Fang, Cuihua Li

Abstract: This paper aims at how to reconstruct a super-resolution image in a fast speed based on linear kernel regression. We use linear kernel regression to learn a map between the space of high-resolution images and the space of blurred high-resolution images that are the interpolation results obtained from the corresponding low-resolution images, because linear kernel regression is simple and has low computational complexity compared with nonlinear kernel regression. In a computational viewpoint, the super-resolution image reconstruction can be transformed to solve linear equations whose size depends on the number of the training data. When the amount of the training data is large, it is time-consuming to solve the regression problem. In order to solve this problem quickly, we select part of the pairwise patches from the training dataset by K-means method instead of all the training data and use these patches to construct a moderate scale regression problem. Furthermore, we implement orthogonal matching pursuit to improve the speed of solving linear equations obtaining from the moderate linear regression problem. The experimental results show that it achieves good performance and it is superior to other super-resolution methods in terms of PSNR.