一种新的视频摘要可视化算法  

A Novel Approach for Abstractive Video Visualization

作　　者：彭帝超[1] 刘琳 陈广宇[1] 陈海东[1] 左伍衡[3] 陈为[1] 

机构地区：[1]CAD&CG国家重点实验(浙江大学)室,杭州310058 [2]邵逸夫医院影像科,杭州310001 [3]浙江工业大学之江学院信息工程分院,杭州310024

出　　处：《计算机研究与发展》2013年第2期371-378,共8页Journal of Computer Research and Development

基　　金：国家"八六三"高技术研究发展计划基金项目(2012AA120903);国家自然科学基金项目(61003193);浙江省科技厅公益基金项目(2011C21058)

摘　　要：提出一种摘要式浏览视频文件的可视化方法,将顺序视频转化成图像形式摘要,能够帮助读者快速有效地获得视频数据的结构信息.算法通过检测视频中每帧的尺度不变特征(SIFT),应用改进的词袋模型构建特征词库并统计词频,将整段视频映射为高维词库空间的一条曲线.通过多维尺度分析(MDS)方法对该曲线降维,生成反映视频语义信息的一条三维平滑曲线.实验结果表明,该曲线很好地体现视频中各帧之间的关联性和语义转折,可辅助读者快速理解视频情节结构.We present a novel approach for abstractive video visualization, which can help users understand the semantic information from the video in a fast and effective manner. We use the scale- invariant feature transform （SIFT） algorithm to detect features of each frame, together with a modified bag of words algorithm to construct a feature vocabulary in order to compute the feature frequencies. By mapping the video sequence onto a 3D curve in a high dimensional vocabulary space with the use of the multi-dimensional scaling （MDS） algorithm, the video is abstracted and embedded into a visually recognizable curve in 3D space. This generated visualization result can vividly illustrate the evolvement of the video contents, while well protecting and preserving the semantic meaning that are encoded within the video. Experimental results indicate that this curve-based visualization technique can uncover the semantic relationship between the frames, characterize the transition of video contents, and help the users understand the semantic structure of the underlying video sequence with a quick glance.

关 键 词：可视化 视频摘要 词袋 低维嵌入 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]