出版社：The Institute of Image Information and Television Engineers
摘要：We propose a query-by-example method that can retrieve a variety of shots relevant to a query, but these shots contain significantly different features due to varied shooting techniques and settings. Thus, we use rough set theory to extract multiple classification rules that characterize different subsets of example shots. We elaborate on how to extract useful rules from only a small number of example shots provided by the user. We incorporate bagging and the random subspace method into rough set theory. The former is useful to extract rules that cover a variety of shots, and the latter is useful to avoid extracting rules that overfit the example shots. Finally, although our method needs counter example shots, they are not provided by the user. Therefore, we use partially supervised learning to collect counter example shots from shots other than example shots. Experimental results on TRECVID 2009 video data validate the effectiveness of our method.
