Survey on rain removal from videos or a single image

作     者：Hong WANG Yichen WU Minghan LI Qian ZHAO Deyu MENG Hong WANG;Yichen WU;Minghan LI;Qian ZHAO;Deyu MENG

作者机构：School of Mathematics and Statistics Ministry of Education Key Lab of Intelligent Networks and Network SecurityXi'an Jiaotong University Pazhou Lab 

出 版 物：《Science China(Information Sciences)》 (中国科学:信息科学(英文版))

年 卷 期：2022年第65卷第1期

页      面：72-94页

核心收录：

学科分类：08[工学] 080203[工学-机械设计及理论] 0802[工学-机械工程] 

基　　金：supported by the National Key R&D Program of China (Grant No. 2020YFA0713900) National Natural Science Foundation of China (Grant Nos. 11690011, 61721002, U1811461) 

主　　题：rain removal maximum a posterior estimation deep learning generalization performance comprehensive repository 

摘      要：Rain can cause performance degradation of outdoor computer vision tasks. Thus, the exploration of rain removal from videos or a single image has drawn considerable attention in the field of image processing. Recently, various deraining methodologies have been proposed. However, no comprehensive survey work has yet been conducted to summarize existing deraining algorithms and quantitatively compare their generalization ability, and especially, no off-the-shelf toolkit exists for accumulating and categorizing recent representative methods for easy performance reproduction and deraining capability evaluation. In this regard, herein, we present a comprehensive overview of existing video and single image deraining methods as well as reproduce and evaluate current state-of-the-art deraining methods. In particular, these approaches are mainly classified into model-and deep-learning-based methods, and more elaborate branches of each method are presented. Inherent abilities, especially generalization performance, of the state-of-the-art methods have been both quantitatively and visually analyzed through thorough experiments conducted on synthetic and real benchmark datasets. Moreover, to facilitate the reproduction of existing deraining methods for general users, we present a comprehensive repository with detailed classification, including direct links to 85 deraining papers, 24 relevant project pages, source codes of 12 and 25 algorithms for video and single image deraining, respectively, 5 and 10 real and synthesized datasets, respectively, and 7 frequently used image quality evaluation metrics, along with the corresponding computation codes. Research limitations worthy of further exploration have also been discussed for future research along this direction.