Multi-stage network for single image Demoireing
Ru Li, Junwei Xie, Yuhui Lin, Tong Tong, Weixin Zheng, Qinquan Gao · 2022
The moiré pattern refers to the interference fringes generated by two equal amplitude sinusoids with close frequencies. In digital imaging, images collected in some scenes are vulnerable to moiré, such as images taken from knitted fabrics and images taken on LED screens, the visual quality of which is always damaged seriously. The difficulty of demoiréing lies in the moiré patterns distributed through different bands of frame frequencies and vary in colors and shapes. To fully learn the global information of the moiré images and remove the moiré patterns in a wide range of frequency bands, we proposed a multi-stage and multi-patch network, which can recover non-homogeneous moire images by aggregating the features of different spatial regions of patches in different stages. To increase the receptive field, we also introduce a novel Atrous Fusion Module in different atrous rates to learn multi-scale information. By taking advantage of these improvements, our proposed network can achieve superior accuracy than state-of-the-art approaches on the public dataset in the NTIRE2020 Single Image Demoiréing Challenge.