Multiscale annealing combines a novel multilevel optimization scheme with the deterministic annealing continuation technique to a globally optimizing, yet highly efficient algorithm. Data resolution, group number and computational temperature of the annealing scheme are controlled by a single parameter using a criterion from statistical learning theory. Thus the optimization scale introduced by the temperature and the data scale are unified in a single framework.

Applications of the new algorithm are presented for unsupervised texture segmentation, spatial quantization of color images and structuring of image databases. We discuss new optimization criteria for histogram clustering and spatial quantization in detail. Results are presented for benchmark studies and real imagery.