New classifier with reduced computational complexity

Proposes a new classifier architecture that can reduce the computational complexity substantially. In the proposed classifier, the authors store the distance between any pair of the classes and select some of the classes as a reference set. Then, the classifier calculates the distance of the input to a class as usual if the class is in the reference set; otherwise, it estimates the distance with the stored class distances and the distances to the reference classes. In the proposed classifier computational complexity of the classifier is reduced if the number of the reference classes is small and the distance estimation procedure is simple. The authors explain how to estimate the distances and how to select the reference set with the minimization of the misclassification risk. The authors designed a classifier for digit recognition based on the proposed method. The simulation result shows usefulness of the proposed design procedure for the classifier with reduced computational complexity.