Abstract: The dramatic increase of video data makes it urgent to efficiently browse and understand video content. Video summarization is such a technology, which presents important and representative video content in a concise form, and makes it easy to browse and manage video. The ideal video summarization should meet the criteria of maximal coverage, important priority and minimal redundancy. However, the main technical challenge of the current video summarization is how to integrate the three standards into a framework to get the ideal video summarization. To overcome this problem, this paper proposes the maximal marginal importance and coverage algorithm (MMIC) for optimizing the redundancy, importance and coverage rate. In MMIC, K-RNN (K-regular nearest neighbor) graph based manifold ranking algorithm is used to find the important video frames. In addition, a summarization coverage criterion (SCC) is presented to intuitively guide the user to obtain appropriate summarization length. Extensive experiments are conducted on Open Video Project and YouTube datasets, and the results show the effectiveness and superiority of MMIC algorithm.

Key words: video summarization, video browsing, maximal marginal relevance, manifold ranking

摘要： 视频信息的剧增使得人们迫切需要一种能够高效浏览和理解视频内容的技术。视频摘要是其中一种有效的技术，它将重要的且具有代表性的视频内容以一种简洁的形式呈现出来，方便用户对视频的浏览和管理。理想的视频摘要需满足最大覆盖率、重要优先和最小冗余标准，而目前相关视频摘要的主要技术挑战是如何同时将这3个标准融入到一个框架中获取理想的视频摘要。针对这一问题提出了一种优化冗余性、重要性和覆盖率的最大边界重要和覆盖框架（maximal marginal importance and coverage，MMIC）。在MMIC中，利用基于K-RNN（K-regular nearest neighbor）图的流形排序算法计算视频帧的重要性，并且提出摘要覆盖率标准（summarization coverage criterion，SCC）用以直观指导用户获取合适的摘要长度。通过在Open Video Project和YouTube两个数据集上进行大量实验，验证了所提方法的有效性和先进性。

关键词: 视频摘要, 视频浏览, 最大边界相关, 流形排序