An integrated theory of image database modeling, indexing, and content-based retrieval.

摘要

In the area of content-based image retrieval, image databases are automatically indexed and accessed with image lower-level analysis. Database level manipulations, including database organization, partition or clustering, and management, are also based on image-level analysis. In this dissertation, the philosophy of viewing an image database as a whole is emphasized. A theory of modeling and analyzing image databases is proposed to provide a quantitative measure opt the complexity of an image database with respect to a feature representation. The measure, called complexities of image databases (CID), can be used to compare (1). The different difficulties of retrieving images from image databases based on the same feature representation; and (2). The performance difference of applying various features to retrieve images from the same image database. The concept is a natural generalization of the perplexity in language modeling and speech recognition. After quantizing images using vector quantization, statistical models, such as N-block models and trigger models with respect to a given geometric configuration, are proposed for the probability estimation either in a single image or in a database as a whole. Based on the image-level estimations, several new techniques are developed for extracting comprehensive image features, which are experimentally demonstrated superior to existing techniques. Based on the database-level estimations, the cross-entropy of the adopted mode is used to define CID of the database. Extensive experiments on both synthetic and real image data demonstrate that the proposed measure of image database complexity is related to the empirical performance obtained from the standard query set and is in consistent with the perception intuition: a simpler image database has a lower complexity. Applications include the benchmarking of content-based image retrieval and the combination of image search engines.