Method and apparatus for indexing and retrieving images using visual keywords

摘要

A method, an apparatus and a computer program product for indexing and retrieving image data using visual keywords (108) is disclosed. Visual keywords (108) are prototypical, visual tokens (104) and are extracted from samples of visual documents (100) in a visual-content domain via supervised and/or unsupervised learning processes. An image or a video-shot key frame is described and indexed by a signature (112) that registers the spatial distribution of the visual keywords (108) present in its visual content. Visual documents (100) are retrieved for a sample query (120) by comparing the similarities between the signature (112) of the query (120) and those of visual documents (100) in the database. The signatures (112) of visual documents (100) are generated based on spatial distributions of the visual keywords (108). Singular-value decomposition (114) is applied to the signatures (112) to obtain a coded description (116).