MMDM: Multi-Frame and Multi-Scale for Image Demoireing

Shuai Liu, Chenghua Li, Nan Nan, Ziyao Zong, Ruixia Song; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 434-435

The imaging characteristics of digital sensors often lead to the moire patterns, which are widely distributed over the frequency domain and have irregular colors and shapes. The images with moire patterns could lead to a serious decline in the visual quality. The difficulty of demoireing lies in that the moire patterns mix both low and high frequency information to be processed. In this paper, we propose MMDM, an effective image demoireing network, which uses multiple images as inputs and multi-scale feature encoding module as low-frequency information enhancement. Our MMDM has three key modules: the newly designed multi-frame spatial transformer networks (M-STN), the multi-scale feature encoding module (MSFE), and the enhanced asymmetric convolution block (EACB). Especially, the M-STN aims to align the multiple input images simultaneously. The MSFE is for multiple frequency information encoding, which is built on the efficient EACB module. Experiments prove the effectiveness of MMDM. Also, our model achieves the 2nd place on both demoiring track and denoising track in the NTIRE2020 Challenge. Code is avaliable at: https://github.com/q935970314/MMDM