Eye detection using discriminatory features and an efficient support vector machine

Accurate and efficient eye detection has broad applications in computer vision, machine learning, and pattern recognition. This dissertation presents a number of accurate and efficient eye detection methods using various discriminatory features and a new efficient Support Vector Machine (eSVM). This dissertation first introduces five popular image representation methods - the gray-scale image representation, the color image representation, the 2D Haar wavelet image representation, the Histograms of Oriented Gradients (HOG) image representation, and the Local Binary Patterns (LBP) image representation - and then applies these methods to derive five types of discriminatory features. Comparative assessments are then presented to evaluate the performance of these discriminatory features on the problem of eye detection. This dissertation further proposes two discriminatory feature extraction (DFE) methods for eye detection. The first DFE method, discriminant component analysis (DCA), improves upon the popular principal component analysis (PCA) method. The PCA method can derive the optimal features for data representation but not for classification. In contrast, the DCA method, which applies a new criterion vector that is defined on two novel measure vectors, derives the optimal discriminatory features in the whitened PCA space for two-class classification problems. The second DFE method, clustering-based discriminant analysis (CDA), improves upon the popular Fisher linear discriminant (FLD) method. A major disadvantage of the FLD is that it may not be able to extract adequate features in order to achieve satisfactory performance, especially for two-class problems. To address this problem, three CDA models (CDA-1, -2, and -3) are proposed by taking advantage of the clustering technique. For every CDA model anew between-cluster scatter matrix is defined. The CDA method thus can derive adequate features to achieve satisfactory performance for eye detection. Furthermore, the clustering nature of the three CDA models and the nonparametric nature of the CDA-2 and -3 models can further improve the detection performance upon the conventional FLD method. This dissertation finally presents a new efficient Support Vector Machine (eSVM) for eye detection that improves the computational efficiency of the conventional Support Vector Machine (SVM). The eSVM first defines a Θ set that consists of the training samples on the wrong side of their margin derived from the conventional soft-margin SVM. The Θ set plays an important role in controlling the generalization performance of the eSVM. The eSVM then introduces only a single slack variable for all the training samples in the Θ set, and as a result, only a very small number of those samples in the Θ set become support vectors. The eSVM hence significantly reduces the number of support vectors and improves the computational efficiency without sacrificing the generalization performance. A modified Sequential Minimal Optimization (SMO) algorithm is then presented to solve the large Quadratic Programming (QP) problem defined in the optimization of the eSVM. Three large-scale face databases, the Face Recognition Grand challenge (FRGC) version 2 database, the BioID database, and the FERET database, are applied to evaluate the proposed eye detection methods. Experimental results show the effectiveness of the proposed methods that improve upon some state-of-the-art eye detection methods