Automatic Video Poster Generation

With the rapid growth of digital acquisition and internet techniques, the explosively increasing of digital video poses a great challenge to the management and maintenance of the digital video database. In this thesis, we propose a new video abstraction technique, called video poster, which expressively summarizes the video content and displays the summarization in an artistic manner. It not only provides an automatic tool for generating video posters, but also constructs an important possessing toolkit for the digital video database.There have been many works on video summarization and visualization. Neither one is satisfying in terms of an expressive, effective and attractive video abstraction. For the video content coverage, most video abstraction methods do not consider the structure information, such as the inter-shot relationship. Thus, the video summary is solely a collection of video clips that neglect the content coverage. In addition, there is a wide gap between the low level feature and video semantics for understanding the video content. The low level feature based approaches are inconsistent with human perception, and the semantics oriented methods are far from human expectation.We propose an automatic video poster generation framework. We present a set of rules of video poster page layouts, and introduce a new position importance driven automatic page layout algorithm. We adopt the a priori information of position importance and propose new global proposal and pair-wise proposal in the page layout framework. Our second contribution is a unified video abstraction framework that is capable of capturing both the video structure and attention model for video summarization. Additionally, we extend previous attention model by integrating a set of visual attention models extracted from video sequence. We also propose a new object motion attention model which is fast, accurate and format-free.