VIDEO INDEXING BASED ON IMAGE AND SOUND

摘要

Video indexing is a major challenge for both scientific and economic reasons. Information extraction can sometimes be easier from sound channel than from image channel, We first present a multi-channel and multi-modal query interface, to query sound, image and script through "pull" and "push" queries. We then summarize the segmentation phase, which needs information from the image channel. Detection of critical segments is proposed. It should speed-up both automatic and manual indexing. We then present an overview of the information extraction phase. Information can be extracted from the sound channel, through speaker recognition, vocal dictation with unconstrained vocabularies, and script alignment with speech (or "script warping"). We present experiment results for these various techniques. Speaker recognition methods were tested on the TIMIT and NTIMIT database. Vocal dictation was experimented on newspaper sentences spoken by several speakers. Script alignment was tested on part of a cartoon movie, "Ivanhoe". For good quality sound segments, error rates are low enough for use in indexing applications. Major issues are the processing of sound segments with noise or music, and performance improvement through the use of appropriate, low-cost parallel architectures or networks of workstations.