Embedded object dictionaries for image database browsing and searching

We describe a technique to encode images for content-based access and retrieval. Potential query terms are first extracted from an image and represented in terms of multiresolution subbands. A vector quantizer structure then maps the subbands of each image object onto a set of embedded dictionaries. An algorithm is used to exploit the occurrence and query probabilities of the objects for efficient coding and retrieval. Furthermore, a new browsing tool based on multiresolution prototypes is proposed. A prototype object is associated with each dictionary entry. Prototype objects may be substituted for subband data for high quality image browsing during retrieval.