Equivalent relationship of feedforward neural networks and real-time face detection system

Abstract

In this paper, we mainly investigate a fast algorithm, Extreme Learning Machine (ELM), on its equivalent relationship, approximation capability and real-time face detection application. Firstly, an equivalent relationship is presented for neural networks without orthonormalization (ELM) and orthonormal neural networks. Secondly, based on the equivalent relationship and the universal approximation of orthonormal neural networks, we successfully prove that neural networks with ELM have the property of universal approximation, and adjustable parameters of hidden neurons and orthonormal transformation are not necessary. Finally, based on the fast learning characteristic of ELM, we successfully combine ELM with AdaBoost algorithm of Viola-Jones in face detection applications such that the whole system not only retains a real-time learning speed, but also possesses high face detection accuracy. © 2009 Springer.