Face recognition is a difficult problem, especially when pose and illumination vary. The reason is that variations across pose and illumination in a single face can be very large, while variations between different faces are small. In this project, we propose two algorithms that use image synthesis to recognize faces. During the training stage, we synthesize images under different illumination and pose to augment our training set. The first algorithm uses K-nearest neighbor for appearance-based face recognition across pose and illumination. By modeling face space (the set of all face images) using a Gaussian Mixture Model (GMM), the second algorithm estimates the probability density function for each person, then performs a Bayes' classification.

Our idea is an extension of ARENA [T.Sim] , in which a small training set is augmented with additional synthetic images. ARENA does this by generating random 2D geometric (translation, rotation, scaling) perturbations from the original training images. Here, we synthesize images under many different illuminations and poses.