Sparse Support Regression for Image Super-Resolution

Abstract

In most optical imaging systems and applications, images with high resolution (HR) are desired and often required. However, charged coupled device (CCD) and complementary metal-oxide semiconductor (CMOS) sensors may be not suitable for some imaging applications due to the current resolution level and consumer price. To transcend these limitations, in this paper, we present a novel single image super-resolution method. To simultaneously improve the resolution and perceptual image quality, we present a practical solution that combines manifold learning and sparse representation theory. The main contributions of this paper are twofold. First, a mapping function from low-resolution (LR) patches to HR patches will be learned by a local regression algorithm called sparse support regression, which can be constructed from the support bases of LR-HR dictionary. Second, we propose to preserve the geometrical structure of image patch dictionary, which is critical for reducing artifacts and obtaining better visual quality. Experimental results demonstrate that the proposed method produces high-quality results, both quantitatively and perceptually.