Image Super-Resolution Based on Data-Driven Gaussian Process Regression

Abstract

In this paper, we aim at producing the super-resolution image from a single low-resolution image based on Gaussian Process regression. Gaussian Processes provide a framework for deriving regression techniques with explicit uncertainty models. Super resolution can be transformed into a regression problem. We show how Gaussian Processes with covariance functions can be used for image super-resolution. Furthermore, considering that the training data have greatly effect on the super-resolution performance and the unsuitable training data would result in unexpected details, we adopt a data-driven scheme to learn a regression map for each query patch. There are two advantages of our approach: 1) we establish a map between the low-resolution space and the high-resolution space independent of a specified regression function; 2) the data-driven learning scheme improves the super-resolution performance. We estimate our approach on the popular testing images which are used in other super-resolution literatures, and the results demonstrate that our approach is efficient and it manifests a high-quality performance compared with several popular super-resolution methods.