Abstract
Lots of mobile devices such as digital camera, mobile phones, and
personal digital assistants (PDAs) adopt single image sensors to acquire
scene images for real-time processing. Convolutional neural networks
(CNN) have been shown good performance in lots of area about image
processing. In this paper, we develop a demosaicking method based on
deep residual CNN (DRCNN). Firstly, we initialize the mosaicked images
using conventional demosaicking algorithms and learn the DRCNN for three
color channels. In DRCNN, we adopt binary convolution unit (BCU),
efficient layer aggregation networks (ELAN), and dense residual blocks
(DRB). Experimental results show that our proposed algorithms
outperforms existing approaches in PSNR, computational complexity, and
visual quality.