基于时空信息融合的视频显著性区域检测算法  

Video saliency detection algorithm based on spatial-temporal information

作　　者：毕晓君[1] 陈晗 BI Xiaojun;CHEN Han(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2018年第11期1786-1792,共7页Journal of Harbin Engineering University

基　　金：科技部国际科技合作项目(2015DFG12150)

摘　　要：针对目前显著性区域检测算法处理存在背景扰动的视频时鲁棒性较差的问题,本文提出一种基于时空信息自适应融合的视频显著性区域检测算法。在时间上,利用光流法提取运动矢量场,综合像素点在连续多帧的运动熵以及与邻域点的方向一致性值表征时间显著性;在空间上,通过区域间的颜色对比度计算空间显著性值,以保证显著性区域的空间完整;最后,提出时、空显著图的自适应动态融合方法,以适应不同情形视频的显著性检测要求。实验结果表明:本文算法能够较为准确地检测出视频中的显著性区域,在精确度和鲁棒性等方面均具有较好的效果。Considering the poor robustness of the present video saliency detection algorithm when dealing with videos with background disturbance,a novel algorithm for detecting video salient region is proposed,which is based on adaptive fusion of spatial and temporal information.First,the motion vector field is extracted by optical flow method in time domain,which characterizes temporal saliency by combining the motion entropy of pixels in multiple continuous frames and the direction consistency value of the neighborhood points.In spatial domain,the spatial saliency is calculated by the color contrast between regions to ensure the spatial integrity of the salient area in each frame.Finally,an adaptive dynamic method based on spatial-temporal saliency fusion map is proposed,which can adapt to the saliency detection requirements of videos in different scenarios.The experimental results show that this algorithm can accurately detect the salient area in videos and has high accuracy and robustness.

关 键 词：视频显著性 时空信息 运动矢量场 颜色对比度 显著图 自适应融合 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]