HIGH RESOLUTION DEMOIRE NETWORK

Photographing LCD screens for information recording has become a prevalent behavior, but moire artifact will appear in screen-shot pictures with different colors and shapes. In order to remove such artifacts, previous work mostly used multi-scale framework to identify the complex frequencies of moire, but the relationship between different scales used to be ignored. In this paper, we proposed a novel High-Resolution Demoire Network(HRDN) to fully explore the relationship between feature maps with different resolutions. It consists of three main components: parallel high resolution network, continuous information exchange modules and final feature fusion layer. Extensive experiments demonstrate our method outperforms the state-of-art both in quantity and quality, the PSNR of demoire images restored by HRDN reached 28.47, which is 0.72 higher than MopNet. To facilitate research in image demoireing we have published our code in github.