Design of Parallelized Training System of Single Class Cascade Classifier

This paper proposes a new training method of a cascade classifier in order to implement on Hadoop MapReduce platform. Learning process of cascade classifier requires many computations whereas the serialized algorithm does not fit to a parallel platform well. The parallelization is achieved by dividing the training into two parts. Before starting learning for adaptation to required false positive rate, the unit classifiers are trained independently using positive examples and small set of negative examples. To make a chain of classifiers, the latter part performs training using only negative examples.