摘要
为了提高视频关键帧提取的质量和效率,提出一种基于最优距离聚类和特征融合表达的视频关键帧提取算法。在视频帧间差异性分析基础上,寻找并确定最优帧间距离阈值,采用无监督聚类算法对帧间距离进行聚类,获得类别数目最优的类图像集;计算图像的颜色复杂度和信息熵并融合,按照类中图像特征值"平均"的思想提取类代表帧,组成视频关键帧。对4个监测视频进行实验,结果显示:该算法提取关键帧的平均保真度为96.72%、平均压缩率为96.42%,运行时间也较短,与两种典型的基于聚类的关键帧提取方法相比,在相同的压缩率情况下,算法保真度大幅度提高,而运行时间较小或相当。该算法解决了无监督聚类对阈值的依赖性问题,兼顾了视频中运动目标变化和环境异常两种情况,具有良好的性能和适应性。
In order to extract key frames from the monitoring video accurately and efficiently,a key frame extracting algorithm which is based on optimal distance threshold clustering and feature fusion expression is proposed.In order to obtain the frame class image set with optimal clustering number,we analyze the differences between frames of the video,and determine the optimal distance threshold which is used for unsupervised clustering of inter-frame distances.In order to extract the representative frame of each cluster,we calculate and merge color complexity and information entropy,and extracte representative frame based on‘cluster average’concept.Representative frame extracted from each cluster is assigned to the key frame image set.Test results of the four monitoring videos show that,the average fidelity and average compression ratio are 96.72%and 96.42%,and the running time is shorter.Compared with the two typical algorithms based on clustering,the fidelity of the proposed algorithm is greatly improved,while the running time is smaller or equivalent,when the compression rate is the same.This algorithm solves the problem of the dependency of unsupervised clustering on the threshold and takes moving target changes and environment anomaly into account,having good performance and adaptability.
作者
孙云云
江朝晖
单桂朋
刘海秋
饶元
Sun Yunyun;Jiang Zhaohui;Shan Guipeng;Liu Haiqiu;Rao Yuan(School of Information and Computer Science,Anhui Agricultural University,Hefei 230036,China;Key Laboratory of Technology Integration and Application in Agricultural Internet of Things,Ministry of Agriculture,P.R.China,Hefei 230036,China)
出处
《南京理工大学学报》
EI
CAS
CSCD
北大核心
2018年第4期416-423,共8页
Journal of Nanjing University of Science and Technology
基金
安徽省科技攻关项目(1501031102)
农业部农业物联网技术集成与应用重点实验室开放基金(2016KL01)
安徽农业大学2018年度研究生创新基金(2018yjs-63)
关键词
监测视频
关键帧提取
最优距离阈值
无监督聚类
特征融合
monitoring video
key frame extraction
optimal distance threshold
unsupervised clustering
feature fusion
