Illumination-robust face recognition

The performance of a face recognition system is closely related to the amount of variation observed in face images. Illumination is a major source of such variations; therefore, compensating for illumination changes is a major requirement for face recognition algorithms. In this thesis, we approach the problem of face recognition by putting illumination compensation to the center of our framework. We describe an illumination-robust face recognition system that can recognize near-frontal human faces with changing facial expressions.; We start developing our approach by first considering the problem of face recognition under varying illuminations with neutral facial expressions and frontal head poses. We develop a Segmented Linear Subspace algorithm that applies a collection of linear subspaces to a segmented image to achieve reliable recognition under varying illumination. We later relax the constraint on the facial expressions by carrying the Segmented Linear Subspace algorithm into the framework of the shape and texture representation. This representation is based on separating the face image into separate shape and texture vectors. We use the Segmented Linear Subspaces to model the texture component of this representation, which provides us robustness to varying facial expressions since the textures are normalized for shape changes. Computation of the shape and texture representation requires locating a number of facial features on the face image automatically, which is in general a difficult task. It is particularly challenging in our case because of the large illumination changes. To address this problem, we propose an Adaptive Active Appearance Model algorithm, which provides a fast and accurate way of locating landmark points on face images. We combine the Segmented Linear Subspace model and the Adaptive Active Appearance Model algorithm into a single system where we compensate for illumination and facial expression changes in a combined framework.