Image demosaicking based on moving least square

Abstract

We present a demosaicking method based on moving least square to estimate to-be-interpolated color pixels. Demosaicking is a severely ill-posed problem and we need to reconstruct two color channels for each pixel. To predict the missing color pixels using the given color pixels, we rely on a generic local approximation function using moving least square. We assume that image details have similar geometric structure in a local region, and we utilize moving least square to represent this geometric feature. Experimental results indicate that the proposed method outperforms existing approaches in both objective and subjective performance.