Browse Theses

Statistical pattern recognition and computer vision techniques have been successfully applied to many object recognition problems. One typical example is the task of face recognition where a 3D face object may appear dramatically different under different lighting, viewing conditions. In this thesis, we propose combining pattern recognition and computer vision methods for robust face recognition.

The first part of this thesis deals with statistical pattern recognition approaches. Many classifiers such as the Bayesian classifier (theoretically optimal) and nearest-neighbor rule are available. For applications involving high-dimensional patterns such as faces, the demand for a large number of training samples to construct a good Bayesian classifier is difficult to satisfy. In this thesis we propose a statistical framework, subspace discriminant analysis, using which we can construct a practically good classifier (both linear and nonlinear) using only a limited number of training samples. To construct generalizable features for face recognition, a face subspace is constructed. This is motivated by the observation that face recognition is primarily about distinguishing among similar objects-faces. A detailed description of subspace LDA/DCA is presented along with extensive experimental results including the FERET test.

The second part of this thesis deals with taking a computer vision approach to robust object recognition. First, we develop a new shape-from-shading (SFS) theory called symmetric SFS (SSFS) to handle symmetric objects such as faces. One big advantage of SSFS is that we have shown that SSFS not only has a pointwise unique solution for the partial derivatives of the depth map but also a unique solution for albedo. Next for the specific task of face recognition, we propose using SSFS and a generic 3D face model to address the illumination problem and demonstrate significant performance improvement. Finally the problem of pose variation coupled with illumination change in face recognition is addressed. This method is based on a new view synthesis technique. Experiments results using several commonly available databases are reported.