Adaptive Multispectral Encoding Network for Image Demoiréing

Qiang Dai, Xi Cheng, Li Zhang, Le-Kang Sun

Abstract: Moiré often appears when photographing textured objects, which can seriously degrade the quality of captured photographs. Due to the wide distribution of moiré and the dynamic properties of moiré, it is a challenge to effectively remove moiré patterns. For this purpose, we present an adaptive multispectral encoding network (AMSDM) for image demoiréing. In AMSDM, we leverage a multiscale network structure to process moiré images at different spatial resolutions, which can relieve the issue of moiré with distributed frequency spectrum. To solve the issue of dynamic properties of moiré, we design an adaptive multispectral encoding (AMSE) module to encode moiré patterns adaptively, which helps AMSDM restore moiré images clearly. Besides, a demoiréing convolutional network block (DMCNB) in the AMSE module makes AMSDM have the adaptability and the long-range correlation; thus, it can learn both global and local information about moiré images. Extensive experimental results indicate that our proposed AMSDM significantly outperforms state-of-the-art (SOTA) methods and achieves a great balance between performance and efficiency.