Deep Demosaicking with Luminance and Chrominance Estimations

Abstract

A digital camera acquires images using a single electronic sensor with a color filter array (CFA). The raw image contains luminance, defined as a spatial map of intensity, and chrominance, defined as a spatial map of each color information. Since the luminance and chrominance components have different demosaicking complexities, they should be modeled separately. In this paper, we propose a novel convolutional neural network (CNN)-based demosaicking method that separately estimates the luminance and chrominance components. Specifically, we apply two-stage CNNs consisting of a luminance component estimation network and a chrominance component estimation network. The proposed method suppresses artifacts such as false colors and reduces the computational complexity. Experimental results on several benchmark datasets demonstrate that the proposed method provides results that are better or competitive with conventional demosaicking algorithms while reducing the computational complexity.