Learning visual keywords for content-based retrieval

Today keyword-based teat retrieval systems have enjoyed reasonable success in real usage. Despite the simplicity of the keyword metaphor, practical test search engines are able to handle huge volumes of free text documents. In image and video retrieval, search mainly relies on pre-annotated words or/and primitive visual features. We propose the notion of visual keywords for content-based retrieval. The visual keywords of a given visual content domain are typical visual entities that are extracted from statistical learning. A visual content is spatially described in terms of the extracted visual keywords and coded via singular value decomposition for similarity matching. We demonstrate our framework in retrieval of natural scene images