|
Image Deraining Based on Adaptive Perceptual Pyramid Network
YANG Ai-ping， WANG Chao-chen， WANG Jian， ZHANG Teng-fei
2022, 43 (4):
470-479.
DOI: 10.12068/j.issn.1005-3026.2022.04.003
Although deep learning-based methods have made great progresses in the field of single image deraining， existing methods still have some problems such as loss of detail and residue of rain streaks. To overcome these short-comings， this paper proposed an adaptive perceptual pyramid network to remove dense rain streaks while revising image details at the same time and thus significantly improve the visual quality. Firstly， a multi-scale pyramid network with progressive connected sub-networks is constructed based on wavelet transform to extract and remove rain streaks iteratively. Each sub-network with the adaptive rain streak perceptual block as its core module is constructed through symmetric skip-connection to extract the shallow layer features to deeper layers to be reused. The adaptive rain streak perceptual block uses non-local operation and shared dilated convolution to expand the receptive fields， which can effectively perceive and remove rain streak adaptively. In order to remove rain streaks with different scales and impose some constrains of the network training， a multi-scale loss function is designed so as to gradually remove rain streaks from coarse to fine， which can effectively prevent artifacts. Extensive experimental results on synthetic and real datasets demonstrate the superiority of the proposed method over the state-of-the-art methods， which can remove rain streaks effectively with preserving vivid image details.
References |
Related Articles |
Metrics
|
