| The overwhelming volume of Web videos returned from search engines makes effective browsing and searching a challenging task. It becomes necessary to summarize the major event highlights with a concise structure showing the flow of events, together with the representative text keywords and visual shots. This thesis explores the issues of event discovery and structure construction from the Web video search results. Both text and visual features are explored to discover the events. Visual near-duplicate feature trajectory, combining with visual co-occurrence, is proposed to extract visual features. Text and visual features is then joined together by bipartite graph, and event structure is constructed by linking and aligning events along the timeline on the basis of event similarity, which demonstrates the major highlights by concisely depicting the evolution of events. Effective visualization and browsing of Web videos is enhanced by further associating representative text keywords and visual shots to every event. Experiments on a large-scale Web video data set collected from YouTube are conducted for performance evaluation. Our proposed methods show better results in precision, and the mined event structure is meaningful.The major contributions of this thesis are:First, a new framework, which combines the relatively coarse granularity text semantics and the relatively fine granularity visual features, is proposed for web news videos event mining.Second, in order to improve the event mining performance, visual feature trajectory is proposed by involving linked near-duplicate keyframes.Third, visual near-duplicate co-occurrence and feature trajectory are further explored, and we use the K-means with constraints to cluster the trajectories while taking advantage of the co-occurrence.Fourth, an event structure construction algorithm is proposed to visualize the detected events and their evolution, which may prompt a brand-new style of interaction in video search. |
PDF Full Text Request