Open AccessMSFA-Net: a Network for Single Image Deraining
Author(s) -
Shun Wu,
Jun Zhou
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1584/1/012047
Subject(s) - block (permutation group theory) , feature (linguistics) , computer science , residual , image (mathematics) , artificial intelligence , net (polyhedron) , pattern recognition (psychology) , scale (ratio) , exploit , feature learning , computer vision , algorithm , mathematics , geography , cartography , philosophy , linguistics , geometry , computer security
Rain streaks degrade the quality of image. Many methods have been proposed to solve single image rain streaks removal recently. However, some methods over-smooth the recovered image. A deep network architecture called Multi-Scale Feature Attention Network (MSFA-Net) is proposed in this paper. We propose a novel basic block structure to exploit the image features, which consists of multi-scale residual learning block and feature attention block. Several basic block structures with a local residual learning compose a group architecture. The outputs of each group architecture are concatenated for final multi-scale feature fusion. Then the features are fed into feature attention block and reconstruction module. Finally, a global residual learning module restore the clean image. Besides, the feature attention block combines channel attention with spatial attention. The proposed MSFA-Net removes the rain streaks which study a non-linear mapping relationship between the rainy and clean image from synthesized dataset. Through comparing with other state-of-the-art algorithms, our algorithm performs better for both synthesized rainy image data and real rainy image data.