基于多尺度残差生成对抗网络的单图像盲去运动模糊方法  

Method of single image blind deblurring based on multi scale residual generative adversarial networks

作　　者：陈乔松[1] 隋晓旭 官旸珺 范金松 邓欣[1] 王进[1] Chen Qiaosong;Sui Xiaoxu;Guan Yangjun;Fan Jinsong;Deng Xin;Wang Jin(Chongqing Key Laboratory of Data Engineering&Visual Computing,Chongqing University of Posts&Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学数据工程与可视计算重点实验室,重庆400065

出　　处：《计算机应用研究》2021年第3期919-922,共4页Application Research of Computers

摘　　要：图像盲去运动模糊一直是计算机视觉领域的一个经典问题,它的目的是在模糊核未知的情况下恢复清晰图像。考虑到更大的感受野以及多尺度信息对恢复清晰图像中的全局信息以及局部细节信息具有重要作用,因此提出的方法对DeblurGAN方法进行改进,提出一种基于条件生成对抗网络的GR-DeblurGAN(granular resi-dual DeblurGAN)的单图像盲去运动模糊方法,采用细粒度残差模块(granular residual block)作为骨干网络,以此在不增加参数量的情况下,扩大感受野,获得多尺度信息。最后在两个广泛使用的数据集:GoPro数据集以及Kohler数据集上进行算法性能评估,并与代表性算法进行对比。从实验结果可以看出,提出的方法改进效果明显,并且在计算开销上面优于其他算法。Blind image deblurring is a classical problem in computer vision field.The goal of BID is to restore clear image with unknown fuzzy core.Considering that the larger receptive field and multi-scale information play an important role in recovering the global information and partial detail information.This paper improved the DeblurGAN(deblur generative adversarial networks)method and proposed a single image blind deblurring method that GR-DeblurGAN(granular residual deblur generative adversarial networks)based on conditional generative adversarial networks.This paper used granular residual block as the backbone so that it will expand the receptive field and acquire the multi-scale information.Finally,it evaluated the algorithm on two widely used data sets:GoPro and Kohler and compared with representative algorithms.It can be seen from the experimental results that the proposed algorithm has obvious improvement effects.

关 键 词：图像盲去模糊 细粒度残差模块 生成对抗网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]