Eye detection using color information and a new efficient SVM

Eye detection is an important initial step in an automatic face recognition system. We present in this paper a real-time accurate eye detection method using color information and wavelet features together with a new efficient Support Vector Machine (eSVM). In particular, this method consists of two stages: the eye candidate selection and validation. The selection stage rejects 99% of the pixels through an eye color distribution analysis in the YCbCr color space, while the remaining 1% of the pixels are further processed by the validation stage. The validation stage applies 2D Haar wavelets for multi-scale image representation, PCA for dimensionality reduction, and eSVM for classification to detect the center of an eye. The eSVM, based on the idea of minimizing the maximum margin of misclassified samples, is defined on fewer support vectors than the standard SVM, which can achieve faster detection speed and comparable or even higher detection accuracy. Experiments on Face Recognition Grand Challenge (FRGC) database show the feasibility of our proposed method, which can processes 6.25 images with the size of 128*128 per second in average and achieves 94.92% eye detection accuracy.