Content-based video similarity computation and indexing

While automatic management of video sources is increasingly important, conventional techniques are insufficient for effective and efficient video access because of video data's special features. In this thesis, we develop content-based methods for automatic video structuring, indexing, and retrieval. A block-based image similarity measure and a Step-variable algorithm are proposed for scene change detection. Our method out-performs existing approaches in both speed and accuracy and detects both camera breaks and gradual transitions. While the block-based measure emphasizes overall image features, the wavelet-based measure captures intra-shot features because wavelet coefficients reflect detailed image content. We derive an image similarity measure from wavelet coefficients and use it in designing a Seek and Spread algorithm to extract key frames for video browsing, indexing, and retrieval. The region-based measurement supports more accurate and specific image retrieval. We apply a recursive split and merge algorithm for image segmentation using Luv color information. Since video sequences are often incrementally and dynamically identified, video objects require flexible, dynamic indexing. We implement a Conceptual Clustering Mechanism supporting object-oriented techniques. By incorporating these techniques with the classified video features, the video information management system supports, for example, dynamic creation of video programs from existing objects based on semantic features.