An index structure for content-based retrieval from a video database

We present a technique for clustering an index database in order to build an index structure for content based retrieval of video to reduce the query time and make it easier for users to browse the database. The video indexes are classified by an image search engine, and the key index used in the clustering process is chosen as a representative index. When querying a large scale video database, the user can understand the contents of the video database by browsing the representative indexes and can also search for a target video scene efficiently by searching through the representative indexes. We developed a video retrieval system and conducted retrieval experiments in order to compare the retrieval accuracy of three types of index structure. The index structure constructed by k-nearest neighbor clustering achieved higher retrieval accuracy than those constructed by overlapping clustering or serially partitioned clustering