Towards intelligent retrieval in image databases

Research by content in image databases should benefit from the database approach as much as from the image processing and analysis field. Previously, we emphasized the use of a query language in an object-oriented database, namely OQL (Object Query Language) and O/sub 2/. This offers greater expressive power, flexibility, direct mixing with structured data, and the possibility to store the queries for reuse. In this paper, we extend our approach by proposing a genetic algorithm that can construct automatically these queries from a set of example and counter-example images. Genetic algorithms are used partly because this problem is proved to be NP-complete, and partly because there is no clear specific property that can be exploited. The main advantage is that user's selection of good and bad images eliminates the need for him or her to tune unintuitive knobs. Also, during an interactive session, the first presented images can be exploited to rapidly discard irrelevant pictures. This work is promising and offers the opportunity to be extended in several combined directions, including previous proposals.