A compression framework for content analysis

Presents a statistical coding framework that supports content analysis and retrieval in the compressed domain. An unsupervised learning approach based upon latent variable modeling is adopted to learn a collection, or mixture, of local linear subspaces that are designed for compression, while providing a probabilistic model of the source which is useful for inferring image content. The compressed bitstream is organized to enable the progressive decoding of the compressed data, such that the bitstream is only decompressed up to the level necessary to satisfy the query. We describe methods of extracting relevant features from the compressed representation that support querying based on single and multiple example images, high-level class categories such as people, and low-level features like particular colors and textures. Retrieval experiments have shown that this representation provides good inferencing with very little decompression.