Self-refining deep symmetry enhanced network for rain removal

• Main Authors: LIU, Hong, YE, Hanrong, LI, Xia, SHI, Wei, LIU, Mengyuan, SUN, Qianru
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2019
• Subjects: Image Restoration; Rotation Equivariance; CNN; Numerical Analysis and Scientific Computing
• Online Access: https://ink.library.smu.edu.sg/sis_research/4449; https://ink.library.smu.edu.sg/context/sis_research/article/5452/viewcontent/ICIP2019_pre.pdf
• Institution: Singapore Management University

Rain removal aims to remove the rain streaks on rain images. Traditional methods based on convolutional neural network (CNN) have achieved impressive results. However, these methods are under-performed when dealing with tilted rain streaks, because CNN is not equivariant to object rotations. To tackle this problem, we propose the Deep Symmetry Enhanced Network (DSEN) that explicitly extracts and learns from rotation-equivariant features from rain images. In addition, we design a self-refining strategy to remove rain streaks in a coarse-to-fine manner. The key idea is to reuse DSEN with an information link which passes the gradient flow to the finer stage. Extensive experimental results on both synthetic and real-world rain images show that our method of self-refined DSEN yields top performance for rain removal.