Retrieval of videos by fuzzy clustering

Abstract

Fuzzy clustering has been developed as an effective way of searching for a video by using a query or by browsing a video database. Each video in the database is segmented into shots by fuzzy clustering on the basis of the coarse color composition of frames. The type of shot boundaries, i.e., whether the change is discontinuous (called a cut boundary) or continuous (such as fade and wipe), can be detected by variation in the membership values at the boundary. Every video is summarized by a set of representative images of its shots. Fuzzy clustering is also used for searching for a target video by browsing a video database. In this search, representative images of shots are displayed and one of them is selected if it is the target video. The display region is narrowed successively by tracing hierarchical fuzzy clusters of shots. Additionally, the color correlogram of a video retrieved by querying shot images is shown to be superior to that retrieved according to a histogram.