부산대학교 도서관

In a space of the tree-structured object representations, we study a classification fidelity in terms of an error probability depending on a processed information amount. Using an information-theoretic model, a lower bound to the average error probability as a function of the average mutual information between the objects and their class-label decisions is given. For any collection of the discriminant functions defined at the successive representation levels, an algorithm of a guided search for the decision on a submitted object is proposed. Also, a redundancy of the average error probability relative to the lower bound is defined. Given datasets of face and signature images, the evaluations of the above characteristics show a possibility of a trade-off between the average error probability and a computational complexity of the decision algorithm.