基于经验模式分解和多种评价准则的电子稳像  

Video stabilization based on empirical mode decomposition and several evaluation criterions

作　　者：於俊[1,2] 汪增福[1,2,3] 

机构地区：[1]中国科学技术大学语音及语言信息处理国家工程实验室,安徽合肥230027 [2]中国科学技术大学自动化系,安徽合肥230027 [3]中国科学院合肥智能机械研究所,安徽合肥230031

出　　处：《浙江大学学报（工学版）》2014年第3期423-429,共7页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(61303150);安徽自主创新专项资金智能语音技术研发和产业化专项基金资助项目(13z02008);中央高校基本科研业务费专项资金青年创新基金资助项目(WK2100100020);中国博士后科学基金资助项目(2012M521248)

摘　　要：针对摄影中易产生视频抖动的问题,提出一个实时鲁棒的视频去抖动系统.该系统具有如下特性:1)提取参考帧和当前帧的Sift特征点,并对它们进行匹配,通过随机采样一致(RANSAC)算法来得到全局运动参数;2)基于经验模式分解以及多种评价准则(全局运动参数和特征点对位置误差)来确定抖动参数以实现对当前帧的运动补偿;3)结合图像纹理合成算法来修复运动补偿后的视频帧,从而得到稳定和完整的输出视频.通过比较抖动视频和去抖动后的视频结果表明:该系统能够在保持实时性的同时提高视频的平均信噪比约7.2dB,大大提高人对视频中内容的辨识度和观察舒适感.In view of the probtem of video jitter in photograph,a real-time and robust video stabilization system was proposed.It has following characteristics：1）feature points are matched between reference frame and current frame with Sift feature extraction,and global motion parameters are obtained with random sample consensus（RANSAC）algorithm;2）motion compensation is applied to current frame with jittered parameters obtained by empirical mode decomposition and several evaluation criterions（global motion parameters and location errors of corresponding feature points）;3）stable and complete video is obtained after each frame is repaired with texture synthesis.Comparing between jitter video and stabilized video,the experimental results confirm that the system can increase the average peak signal-to-noise ratio around 7.2dB in real-time,thus can increase the ability of identification and perceptive comfort on video content.

关 键 词：图像匹配 运动估计 运动决定 运动补偿 视频修复 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]