Eye Detection Using Color, Haar Features, and Efficient Support Vector Machine

1. Introduction

Being an important initial step in an automatic face recognition system, eye detection has become a popular research topic in the last decade. Since face alignment for recognition is usually conducted according to eye positions, incorrect eye location would corrupt the face alignment in both spatial scale and rotation, and thus lead to the extremely poor performance of face recognition. Wang et al (2005) did experiments on FRGC 1.0 database to evaluate the impact of eye detection error on face recognition accuracy. It is shown that only 1% eye location error reduces the face recognition accuracy by over 10% while about 5% error reduces the accuracy by 50%. Phillips et al. (2000) did experiments on FERET database and similar conclusion was also reached. The “partial automatic face recognition algorithm”, in which manual eye locations are given to align the face image, performs much better than the “fully automatic recognition algorithm”.

Generally speaking, current eye detection methods can be classified into three categories (Zhu & Ji, 2005): template based methods, feature based methods and appearance based methods. For the template based methods, a sliding window is moved over the whole image to match with a pre-designed generic eye model in order to detect the eye position (Lam & Yan, 1996; Xie et. al., 1994). Feature based methods focus on the characteristic of eyes, such as the shape, color distribution, and intensity gradient information around eye regions (Kawaguchi & Rizon, 2004; Feng and Yuan, 1998; Zhou & Geng, 2004). Appearance based methods detect eyes based on their photometric appearance. These methods always need to train a classifier at first by collecting a large amount of training data and then detection is achieved via two-fold classification (Nguyen et. al., 2008; Zhang & Zhan, 2006; Wang et. al., 2005; Wang & Ji, 2007). Usually, feature based methods have much faster detection speed since they always only involve some image processing operations while appearance based methods have much higher detection accuracy since statistical learning technology is applied. We will further discuss these methods in details in the following section.