Exploring hypergraph-based semi-supervised ranking for query-oriented summarization
Information Sciences: an International Journal
| Hi-index | 0.00 |
We define a user-specific video summarization problem. With a comprehensive understanding of the challenges ( semantic and intention gaps), we simplify the problem with some assumptions, and formalize it into an optimization problem, which is intractable to solve directly. We propose a Bayesian and video hypergraph framework for video summarization. The Bayesian framework aims at analyzing users' browsing logs and inferring users' preference towards videos. There are two distinct features involved this framework, continuous streaming collaborative filtering via random walks, and modeling users' preference by Latent Dirichlet Allocation ( LDA ). LDA is also used as a dimension reduction method. With users' preference probabilities calculated in Bayesian framework, the optimization can be integrated into a video hypergraph framework in order to guarantee the coherence of video summary. We finally approximately solve the optimization problem by choosing representative clips within sub-graph partitioned by spectral clustering. Experiments show that our solution supports user-specific video summarization.