An approach to face recognition using Bayesian networks

Abstract

There are many categories of algorithms that tackle face recognition, one of them being based on Bayesian Networks which allow to encode causal relationships between different kind of random variables, thus helping to express correlations between salient facial features(eyebrows, eyes, nose, mouth). Although current algorithms are quite successful on controlled conditions, performance decreases rapidly in case of unconstrained viewing conditions, such as head pose and illumination for instance. In order to diminish the influence of lighting conditions, histogram equalization is used as a preprocessing algorithm. The used algorithm for features extraction from the grayscale image is the two-dimensional Cosine Transform (2D-DCT) and for facial features localization it has been used the Active Shape Models (ASM) which consists in fitting the shape of an object, using a previously learned global shape model, and represented as a set of landmark points on the face. The model of the used Bayesian Network can be explained as follow: the root node on the top represents a face (node F), which is composed of the relationships between eyebrows (node B), eyes (node E), the nose (node N) and the mouth (node M).And finally, these types of facial features generate the corresponding observations. Finally, we compare the proposed system with two popular appearance-based approaches: PCA (Principal Components Analysis or Eigenfaces) and LDA (Linear Discriminant Analysis or Fisherfaces).