Face detection based on template matching and support vector machines

A subspace method for downsizing the training space via template matching filtering is proposed. Two types of templates, eyes-in-whole and face itself from an average face of a set of mugshot photos, are used in template matching for coarse filtration. Only when both eyes-in-whole template matching and face template matching are over corresponding thresholds, a candidate window is regarded as in the subspace. In this template matching constrained subspace, a bootstrap method is used to collect non face samples for SVM training, which greatly reduces the complexity of training SVM. The face detector SVM is trained by John Platt's Sequential Minimal Optimization (SMO) algorithm. During the detection procedure, an image and its scaled images are scanned, each candidate window will first be evaluated by both eyes-in-whole template matching and face template matching, and when both are over corresponding thresholds that candidate will be passed to the SVM classifier for the final decision. The detection results over all scales are then merged into final face detection output by way of fusion that keeps only the maximum one when overlap happens. In this way the training becomes much easier and the speed is improved to be used in practical applications. Experimental results demonstrate its effectiveness.