A Distribution-based Face/Non-Face Classifier using Gaussian Mixtures
Faculty of Computing, Health and Science
School of Engineering and Mathematics
A key element of many existing approaches to automatic face detection is a classification algorithm that determines if a sub-image of an input image contains a face pattern. In this paper, we present a novel and effective distribution-based face/non-face classifier that detects frontal face patterns with possible in-plane rotation. A 15×15 input image is first processed by a color filter, which verifies the existence of skin colors in the image. Next, the intensity image is extracted from the color image and converted into a vector in a high-dimensional space (R225 ). Principal component analysis is employed to reduce the dimension of this space to 20. In our algorithm, the distributions of face and non-face patterns in the R20 space are modeled using two mixtures of Gaussians. The parameters of the Gaussian mixture models are determined through the Expectation/Maximization algorithm. Finally, classification of a pattern as face or non-face is performed is done through comparison of the estimated probability density functions of face and non face patterns. Experimental results have shown that the proposed algorithm is capable of highly accurate face/non-face classification in terms of both correct detection rate (89.2%) and correct rejection rate (98.9%).