Content-based video abstraction

This paper addresses automatic scene change detection, key-frame selection, and similarity ranking which constitute the main steps of a content-based video abstraction system. Unlike other methods, the proposed algorithm performs scene change detection and key-frame selection in one step. We treat scene change detection as a two-class classification problem and employ automatic threshold selection techniques originally developed for image binarization. A quantitative measure for retrieval of similar scenes according to their color content is also defined. The described scheme can be applied to both uncompressed and MPEG compressed video, and can be implemented in real-time. Performance of the algorithm has been analyzed on real TV sequences, and comparison with some previously introduced techniques are provided.