Unified Probabilistic Model For Face Recognition

Face recognition is different from many other traditional pattern recogntion problems. On the one hand, for a system of face recognition, there are popularly a large number, often over several hundreds, of individuals and only a few images for each person. It is not uncommon to have only a single example image for each person in many applications. On the other hand, the accuracy of a face recognition algorithm is strongly affected by the limitations placed on the problem, such as illumination, head pose, expression, aging, image quality, image scale, background, etc. The primary objective of this dissertation is to investigate large scale (over several hundreds people) frontal facial image recognition from a singe image per person. The maijor distributions of the dissertation are included as follows:(1) A novel unified probabilistic model is proposed for face recognition from a single sample per peron. Based on that faces are similar to each other, the unified probabilistic mode trained on an existing training set with multiple samples per class from a known people group A, which is usually easy to be obtained, can be generalized well to facial images of unknown individuals, and can be used to recognize facial images from an unknown people group B with only one sample per person.(2) A new face recognition technique based Gaussian Mixture Models is proposed. Based on the idea of the unifed problistic model, GMMs, as a more flexible and accurate approach for density estimation than traditional normal density, are used to estimate the distributions of the within-class differences and the between-class differences of facial images. Moreover, a classifier combination is exploit to combine multiple GMMs, and make a notable improvement on performance.(3) A novel face recogntion approach of the unified model in identity subspace (UMIS) is proposed based on the bilateral symmetry of human faces. Bilateral-asymmetric variations of facial images are principally due to the factors irrelevant to identity variation. The face space can be divided into identity subspace and non-identity subspace. The identity subspace...