Eye detection by asymmetric principal component and discriminant analyses

Abstract

This report illustrates in detail the design and implementation of an eye detector by applying the pattern classification technique: Asymmetric Principal Component Analysis (APCA), which is well-known for its outstanding performance in face detection applications. A widely used method of distance measure known as Mahalanobis distance is also applied to facilitate classification. An algorithm for the eye detector is designed and built to handle unbalanced or asymmetric training data for a two-class classification, which consists of an eye class that can be well-defined and a non-eye class which can be anything except the eyes. The asymmetric training data often causes significant differences in the reliability of the two class-conditional covariance matrices. The application of APCA in the development of the algorithm addresses this problem. Furthermore, the Mahalanobis distance is applied to explore the similarity of an unknown sample set to an identified one. The classification of the samples is based on the minimum Mahalanobis distance classifier which is derived from Bayes optimal decision rule. The algorithm is built and tested using Matlab.