Takashi Endo

Mutual Spotting Retrieval between Speech and Video Image Using Self-Organized Network Databases

Video codec technology like MPEG and improved performance of microprocessors enable environments to be setup in which large volumes of video images can be stored. The ability to perform search and retrieve operations on stored video is therefore becoming more important. This paper proposes a technique for performing mutual spotting retrieval between speech and video images in which either speech or video is used as a query to retrieve the other. This technique makes use of a network that self organizes itself incrementally and represents redundant structures in degenerate form, which makes for efficient searches. As a result, the capacity of a database can be decreased by about one half for speech and by about three fourths for video when expressed in network form. Applying this technique to a database consisting of six-hours worth of speech and video, it was found that a search from video to speech could be performed in 0.5 seconds per frame.