Novel quality-efficient universal demosaicing for arbitrary color filter array images

Abstract

To save cost, nowadays most digital cameras equip with a single sensor covered with a specific color filter array (CFA) and each pixel in the captured CFA image is composed of only one color. Universal demosaicing is the process to reconstruct the full-RGB (red-green-blue) color image through an input arbitrary CFA image. In this paper, we propose a novel quality-efficient universal demosaicing method which exploits both the edge information and the color correlation during demosaicing and can work for arbitrary RGB and non-RGB CFA images. In the proposed method, two main contributions are made and they are (1) an effective universal edge-directed color difference based interpolation is designed to specially demosaic the edge regions which usually appear severe artifacts by using the existing universal demosaicing methods and (2) a new Fourier transform-based technique combined with the Sobel edge detection is developed to derive the masks for extracting gradient information from arbitrary CFA images so as to provide better weights in the proposed universal edge-directed interpolation, leading to better quality of the demosaiced images. Based on the 24 typical testing images from Kodak collection, experimental results showed that the proposed universal demosaicing method possesses the quality advantage in terms of average color peak signal-to-noise ratio (CPSNR) and S-CIELAB and demonstrates less color artifacts in terms of visual perception when compared with the three state-of-the-art universal demosaicing methods and some existing demosaicing methods designed for non-RGB CFA images.