映射结合聚类的视频关键帧提取

发布时间：2018-03-17 09:10

本文选题：映射　切入点：聚类　出处：《中国图象图形学报》2016年12期 　论文类型：期刊论文

【摘要】：技术在多媒体数据处理和计算机视觉中都扮演着重要的角色。基于聚类的摘要方法多结合图像全局或局部特征,对视频帧进行集群分类操作,再从各类中获取具有代表性的关键帧。然而这些方法多需要提前确定集群的数目,自适应的方法也不能高效的获取聚类的中心。为此,提出一种基于映射和聚类的图像密度值分析的关键帧选取方法。方法首先利用各图像间存在的差异,提出将其映射至2维空间对应点的度量方法,再依据点对间的相对位置和邻域密度值进行集群的聚类,提出根据聚类的结果从视频中获取具有代表性的关键帧的提取方法。结果分别使用提出的度量方法对Olivetti人脸库内图像和使用关键帧提取方法对Open Video库进行测试,本文关键帧提取方法的平均查准率达到66%、查全率达到74%,且F值较其他方法高出11%左右达到了69%。结论本文提出的图像映射后聚类的方法可有效进行图像类别的识别,并可有效地获取视频中的关键帧,进而构成视频的摘要内容。
[Abstract]:Technology plays an important role in multimedia data processing and computer vision. However, most of these methods need to determine the number of clusters in advance, and adaptive methods can not efficiently get the center of clustering. A key frame selection method for image density value analysis based on mapping and clustering is proposed. Then the cluster is clustered according to the relative position of point pairs and the value of neighborhood density. This paper presents a method of extracting representative key frames from video based on the clustering results. Results the proposed method is used to test the image in Olivetti face database and the key frame extraction method to test the Open Video library. The average precision of the key frame extraction method in this paper is 66, the recall rate is 74, and the F value is about 11% higher than other methods. Conclusion the post-clustering method of image mapping proposed in this paper can effectively identify the image categories. The key frames in the video can be obtained effectively, and then the summary content of the video can be formed.
【作者单位】：合肥工业大学计算机与信息学院;
【基金】：安徽省自然科学基金项目(J2014AKZR0055) 中国博士后基金项目(2014M561817)~~
【分类号】：TP391.41

【参考文献】