双框架卷积神经网络用于运动模糊图像盲复原  

Two-Frame Convolutional Neural Network for Blind Motion Image Deblurring

作　　者：吴梦婷 李伟红 龚卫国[1,2] Wu Mengting;Li Weihong;Gong Weiguo(Key Laboratory of Optoelectronic Technology and System of Ministry of Education,Chongqing 400030;College of Optoelectronic Engineering,Chongqing University,Chongqing 400030)

机构地区：[1]光电技术及系统教育部重点实验室,重庆400030 [2]重庆大学光电工程学院,重庆400030

出　　处：《计算机辅助设计与图形学学报》2018年第12期2327-2334,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家科技惠民计划项目(2013GS500303)

摘　　要：为了实现全网络端对端的运动模糊图像盲复原,提出一种基于生成式对抗网络不估计模糊核的模糊图像盲复原方法.首先设计了双框架卷积神经网络——生成网络和判别网络,其中,生成网络是全卷积神经网络,实现模糊图像到复原图像的映射;判别网络是改进的VGG二分类网络,用于判定复原图像或原始清晰图像;然后采用最小均方差优化网络训练,添加图像保真项提高复原效果;最后通过生成网络和判别网络对抗训练获得清晰图像.在MS COCO数据集上的实验结果表明,该方法复原质量高、速度快.To achieve blind motion image deblurring by an end-to-end full network,a method without estimating blur kernel was proposed.Firstly,this paper designed a two-frame convolutional neural network which consists of a generative model G and a discriminative model D.The model G is a full-convolutional networks,which can map from blurred image to the deburred image.The model D is an improved VGG for classification,which was used for determining whether the input image is the deburred image or the original clear one.Then,the minimum mean square error was adopted for optimizing network by adding image fidelity items to improve the deblurring results.Finally,the deburred image can be obtained by executing adversarial training on the model G and the model D.Experiments were executed on MS COCO dataset,the results demonstrate that the proposed method can improve the quality of the blurred image and reduce the time-consuming.

关 键 词：运动模糊 图像盲复原 生成式对抗网络 残差网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]