Face Recognition via AAM and Multi-features Fusion on Riemannian Manifolds

We develop a novel face recognition algorithm which is robust to random position perturbations of key points and does not require face alignment, e.g. resizing, rotating, cropping, etc. In our proposed method, a well trained Active Appearance Model (AAM) is first divided into several regions by special landmarks, and each region is given a label by a template. This model is then fed to new cooling facial images to segment the images into irregular regions. In these regions, multi-features fusion matrices are calculated and embedded to related Riemannian manifolds to train classifiers which are combined to construct a final classifier. Our experiment results show its accuracy, efficiency, and robustness on FERET and A-R human face database.