Performance-optimized feature ordering in content-based image retrieval

We present a method to improve the performance of content-based image retrieval (CBIR) systems. The idea is based on the concept of query models [1], which generalizes the notion of similarity in multi-feature queries. In a query model features are organized in layers. Each succeeding layer has to investigate only a subset of the image set the preceding layer had to examine. For the purpose of performance acceleration we group features into two types: features for quick elimination of rather not similar images and features for the detailed analysis of result set candidates. Performance optimization is based on a model for predicting the number of images to be retrieved and on a model describing relationships between features. Results in our test environment show significant reduction of query execution time.