Optimizing progressive query-by-example over pre-clustered large image databases

The typical mode for querying in an image content-based information system is query-by-example, which allows the user to provide an image as a query and to search for similar images (i.e., the nearest neighbors) based on one or a combination of low-level multidimensional features of the query example. Off-lime, this requires the time-consuming pre-computing of the whole set of visual descriptors over the image database. On-line, one major drawback is that multidimensional sequential NN-search is usually exhaustive over the whole image set face to the user who has a very limited patience. In this paper, we propose a technique for improving the performance of image query-by-example execution strategies over multiple visual features. This includes first, the pre-clustering of the large image database and then, the scheduling of the processing of the feature clusters before providing progressively the query results (i.e., intermediate results are sent continuously before the end of the exhaustive scan over the whole database). A cluster eligibility criterion and two filtering rules are proposed to select the most relevant clusters to a query-by-example. Experiments over more than 110 000 images and five MPEG-7 global features show that our approach significantly reduces the query time in two experimental cases: the query time is divided by 4.8 for 100 clusters per descriptor type and by 7 for 200 clusters per descriptor type compared to a "blind" sequential NN-search with keeping the same final query result. This constitutes a promising perspective for optimizing image query-by-example execution.