A sampling based approach to facial feature extraction

Facial feature extraction is considered a key step in many biometric applications. We propose a system that performs facial feature extraction in a given grey-scale face image using Hamiltonian sampling. During training stage of the proposed system, the parameters for a hybrid linear Gaussian model are learnt based on the training data. During the testing stage, when a new image is posed to the system, it uses the learnt graphical model together with hybrid (Hamiltonian) sampling to locate and extract the facial features. The system is an appearance based model which uses PPCA that is robust against noise. The modeling of correlations between latent variables by the graphical model helps in making the facial feature extraction both accurate and efficient. The use of the hybrid sampling helps in locating and extracting facial features in fewer iterations.