Indexing and content-based image retrieval

Based on the use of KD-tree, we present a method of indexing of the color images. This method combines color´s information of the pixels and their positions in the images. It also aims at reducing the size of signatures of each image, and consequently reducing the computation time between two signatures. The method of similarity measure used is the EMD distance. The latter based on the minimization of cost required to transform one distribution to another, and is one of the only distances which can work on signatures which are not necessarily the same number of bins. The performances of this method are compared to those obtained by using color histograms in terms of time and accuracy.