Abstract
We propose a new compressed-domain image retrieval scheme exploiting the similarity of PIFS codes. In PIFS encoding, the compressed code contains contractive mapping information between similar regions in an image. These mapping information can be represented as vectors carrying structural feature of the image. We introduce a new simialrity measure between two vector sets, and exploit them in image categorization and retrieval. In this report, we explain this scheme and demonstrate its usefulness through experiments.