Video retrieval using an adaptive video indexing technique and automatic relevance feedback

Abstract

This work demonstrates content-based retrieval techniques for video databases using an adaptive video indexing (AVI) and a neural network model. The AVI utilizes a "template frequency model" for embedding spatial-temporal contents which are a key in characterizing the time-varying nature of video. This model can naturally be adopted to characterize video at various levels from shot, group, and story levels, in order to facilitate a multiple-level access video database. The AVI retrieval system achieves excellent retrieval accuracy, substantially higher than that of the key-frame based video indexing (KFVI), a popular benchmark for video retrieval. Furthermore, AVI structure can be integrated to a specialized neural network model to perform automatic relevance feedback retrieval. This offers advantages both in minimizing human-user involvement, and in considerably enhancing retrieval accuracy in the context of adaptive systems.