Piece-Wise Kernel Regression for Example-Based Super-Resolution

Abstract

Many existed example-based super-resolution algorithms focus on learning a regression function which maps a low-resolution image patch to a high-resolution image patch. Even though this strategy often works well, the speciality of image patches is omitted during learning process. To this end, a novel piece-wised kernel regression algorithm for super-resolution is proposed for benefitting from the speciality of training image patches. The kernel-based regression function learned by this algorithm will vary along with the changing of low-resolution image patches, which enable the learned regression function more flexible to match the speciality of low-resolution image patches. Experimental results show the convinced performance of the reported algorithm.