An SVM-based Approach to Face Detection

Abstract

This paper presents a face detection system based on support vector machines (SVM). The lambda-coefficients (that correspond to the SVM support vectors) are determined from a set of training images from the Olivetti Research Lab (ORL) database. The support vectors are the archetypes for face and non-face images. Test images (containing both face and non-face images) are presented to the system. The SVM algorithm (now programmed and trained according to the support vector archetypes) maps the test images into higher dimension transform space where a hyperplane decision function is constructed. The hyperplane decision function is based on a degree-one polynomial (i.e. kernel function). The decision function is constructed equidistance between support vector archetypes to give an optimal hyperplane decision function. There are various methods for calculating the lambda-coefficients. In this paper, the coefficients are calculated using the multiplicative update algorithm proposed in (F. Sha, et al., Multiplicative Updates for Nonnegative Quadratic Programming in Support Vector Machines). Following the calculation of these coefficients and implementation of the SVM algorithm, we use the set of test images from the ORL database to demonstrate the SVM's performance as a face detection system. The goal or objective at this juncture is to simply determine the SVM's detection rate for the simplest kernel function (i.e. whether an image presented to the system is a face or non-face image)