Feature selection by AdaBoost for SVM-based face detection
Duy-Dinh Le · 2004
In this paper, we present a three-stage method to speed up a SVM-based face detection system. In this proposed system, a large number of simple non-face patterns are rejected quickly by two first stage cascaded classifiers using flexible sizes of analyzed windows while the last stage uses a non linear SVM classifier to robustly classify complex 24x24 pixel patterns as either faces or non-faces. For all stage classifiers, an optimal subset of overcomplete Haar wavelet feature set selected by AdaBoost learning is used to achieve both fast and high detection rate. Experimental results show that our system can achieve comparable results to state of the art face detection systems.