摘要
现有运动去模糊算法难以有效复原含有大尺度旋转的复合运动模糊,针对此问题提出了一种基于U-net模型的神经网络框架。该框架通过融合运动信息至网络输入,给定每一像素点不同的运动约束。经过网络的编码器与解码器结构,得到每一像素点的预测值,实现端对端的方式直接获得复原图像。实验在通用数据集上与当前先进去模糊算法进行比较,该方法相比性能最好的算法PSNR(peak signal-to-noise ratio)值提高了0.14 dB,相比实时性最好的算法运行时间减少了0.1 s;同时在含有旋转运动的测试集上进行验证,证明了该算法可获得较好的复原质量。
The existing motion deblurring algorithm is difficult to effectively recover the composite motion blur with large rotational motion.This paper proposed a neural network framework based on U-net model for this problem.By combining motion information to the input of network,this framework gave different motion constraint to each pixel.Through the structure of encoder and decoder,each pixel obtained the prediction value,thereby the blurry image could be recovered in an end-to-end manner.The experiment was compared with the current state-of-art deblurring algorithm on the universal data set.This method improved the PSNR value of the best algorithm by 0.14 dB,and reduced the running time of the best real-time algorithm by 0.1 s.At the same time,it was verified on the test dataset with rotational motion.This proves that the algorithm obtains better restoration quality.
作者
董星煜
刘传奇
赵健康
Dong Xingyu;Liu Chuanqi;Zhao Jiankang(School of Electronic Information&Electrical Engineering,Shanghai Jiao Tong University,Shanghai 201100,China)
出处
《计算机应用研究》
CSCD
北大核心
2021年第1期278-281,共4页
Application Research of Computers
基金
国家重点研发计划项目(2016YFC0200400)
国家自然科学基金资助项目(61673265)。
关键词
运动模糊
图像复原
卷积神经网络
运动约束
motion blur
image restoration
convolutional neural network
motion constraint
