Improved multi-scale dynamic feature encoding network for image demoiréing

Highlights：

• We further propose a dynamic feature encoding (DFE) module that can encode the variations of moire patterns in images so that our model can better cope with the dynamic texture of moire.

• We propose a novel L1 wavelet loss function which calculates the L1 distance of the decomposed high and low elements between clean and demoired images. Our proposed network can gain 0.74 dB in PSNR after adding this wavelet loss.

• Our image demoireing method outperforms the state-of-the-arts on two demoireing benchmarks. The proposed improved Multi-scale convolutional network with Dynamic feature encoding for image DeMoireing (MDDM+) is an enhanced version of our preliminary work, which achieved the 2nd place winner in the ICCV2019 AIM Demoireing Challenge.

摘要

•We propose a novel image demoireing method to remove the moire patterns in images through a progressive and multi-scale residual network, which enables our model to learn the representations on multiple frequency bands.•We further propose a dynamic feature encoding (DFE) module that can encode the variations of moire patterns in images so that our model can better cope with the dynamic texture of moire.•We propose a novel L1 wavelet loss function which calculates the L1 distance of the decomposed high and low elements between clean and demoired images. Our proposed network can gain 0.74 dB in PSNR after adding this wavelet loss.•Our image demoireing method outperforms the state-of-the-arts on two demoireing benchmarks. The proposed improved Multi-scale convolutional network with Dynamic feature encoding for image DeMoireing (MDDM+) is an enhanced version of our preliminary work, which achieved the 2nd place winner in the ICCV2019 AIM Demoireing Challenge.