Color image retrieval using the datasieve

With the increased use of on-line resources as well as the availability of digitized images, the need for efficient image retrieval algorithms is high. Research is focusing on systems that use the properties of the regions of an image in their retrieval process so that the search mechanism yields images that are more closely related to the query image. In this paper, we use the datasieve, a recursive nonlinear multiscale morphological operation, to give a scale-space decomposition of the input image. The datasieve uses efficient graph-based algorithms, and can typically compute an entire scale-space decomposition in less than half a second. We show that the 'seed nodes' obtained from the datasieve scale-tree provide a representative sample of the regions in the image. Image matching is then carried out by comparing the seed node features of the query image with the seed node features of the images in the database. The results presented show the method to be promising.