图像盲去模糊综述:从变分方法到深度模型以及延伸讨论  

Blind image deblurring:an overview from variational approaches to deep representation models and beyond

作　　者：李海波[1,2] 邵文泽[2,3] LI Haibo;SHAO Wenze(School of Computer Science and Communication,KTH Royal Institute of Technology,Stockholm 10044,Sweden;College of Telecommunications&Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]瑞典皇家理工学院计算机科学与通信学院,斯德哥尔摩10044 [2]南京邮电大学通信与信息工程学院,江苏南京210003 [3]南京理工大学高维信息智能感知与系统教育部重点实验室,江苏南京2100094

出　　处：《南京邮电大学学报（自然科学版）》2020年第5期84-94,共11页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61771250,61972213,11901299);中央高校基本科研业务费专项资金(30918014108)资助项目。

摘　　要：图像盲去模糊不仅是低层视觉领域的基础性问题,同时也是计算成像领域的前沿性课题。由于实际模糊成像过程的模糊核往往复杂多变、不易参数化,探讨不受限图像非参盲去模糊具有显而易见的现实意义。然而时至今日,在恢复模型的直观性以及估计算法的精确性、鲁棒性和时效性均衡方面,现有方法依然未能给出令人信服的答案。为了进一步推动该领域的深入研究,文中从空间不变非参盲去模糊这个根本性问题出发,对当前基于变分贝叶斯(VariationalBayes)、最大后验估计(MaximumaPosterior)以及深度表示学习(DeepRepresentationLearning)的代表性方法作了简要、清晰的回顾。同时,对于该领域今后值得重点解决的关键科学问题进行了相关展望。最后,结合图像超分辨这个与非参盲去模糊密切相关的前沿问题进行了延伸讨论。The blind image deblurring,i.e.,blind image deconvolution,is a fundamental issue in Iowr-Iev-el vision and a cutting-edge problem in computational imaging.Since complex and varying blur kernels are hardly parameterized in the realistic imaging,it makes great sense to study unrestricted nonparametric blind image deblurring(UN-BID)in terms of real applications.So up till now;existing blind deblurring methods are not satisfactory at all,either in the intuition of restoration models or in the balance of estima­tion accuracy,robustness and efficiency of numerical algorithms.Towards this situation,this paper focuses on the spatially-invariant UN-BID,making a concise yet clear review^on various kinds of approaches for the fundamental problem,including variational Bayes,maximum a posterior and deep representation learn­ing.Based on the overview;perspectives on the deserved efforts for the field in the near future are pointed out.Finally,an extending discussion on the hot topic super-resolution closely relative to the nonparametric blind image deblurring is carried out.

关 键 词：盲去模糊 反卷积 变分贝叶斯 最大后验 深度神经网络 超分辨率 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]