An Optimal Greedy Approximate Nearest Neighbor Method in Statistical Pattern Recognition
The insufficient performance of statistical recognition of composite objects (images, speech signals) is explored in case of medium-sized database (thousands of classes). In contrast to heuristic approximate nearest-neighbor methods we propose a statistically optimal greedy algorithm. The decision is made based on the Kullback-Leibler minimum information discrimination principle. The model object to be checked at the next step is selected from the class with the maximal likelihood (joint density) of distances to previously checked models. Experimental study results in face recognition task with FERET dataset are presented. It is shown that the proposed method is much more effective than the brute force and fast approximate nearest neighbor algorithms, such as randomized kd-tree, perm-sort, directed enumeration method.