Rotation invariant face detection using a model-based clustering algorithm

We present a model-based clustering algorithm for locating frontal views of human faces with in-plane rotation in complex scenes, which can describe the arbitrary shape of the distributions efficiently in a feature space. An optimization technique is employed for selecting representative face and nonface models from the sample images. Image invariance properties on human faces and Hausdorff distance are used for finding the orientation of a face candidate, and the Euclidean distance and normalized correlation coefficient are used for the similarity measures between features. Three different types of feature spaces are used for the matching; binary image, graylevel image, and frequency information. Binary similarity is used for the reduction of the processing time in detecting candidate faces and their orientations in a scene, while the correlation measures of graylevel images and frequency domain features obtained by DCT (Discrete Cosine Transform) are used for the verification. Experimental results show that proposed face detection algorithm gives very high detection ratio compared to the conventional ones