Automatic design of cascaded classifiers

Abstract

Cascades of boosted classifiers have become increasingly popular in machine vision and have generated a lot of recent research. Most of it has focused on modifying the underlying Adaboost method and far less attention has been given to the problem of dimensioning the cascade, i.e. determining the number and the characteristics of the boosted classifiers. To a large extent, the designer of a cascade must set the parameters in the cascade using ad-hoc methods. We propose to automatically build a cascade of classifiers, given just a family of weak classifiers a desired performance level and little more. First, a boosted classifier with the desired performance is built using any boosting method. This classifier is then "sliced" using dynamic programming into a cascade of classifiers in a nearly computation-cost-optimal fashion. © Springer-Verlag 2004 References.