Dirichlet-tree distribution enhanced random forests for facial feature detection

A cascaded approach for facial feature detection in unconstrained environment, e.g., various poses and illuminations, occlusion, low image resolution, different facial expressions and makeup, is proposed in this paper. First the positive facial area is extracted to eliminate the influence of noise under various conditions. Then a Dirichlet-tree distribution enhanced random forests (D-RF) algorithm is proposed to detect facial features using cascaded head pose models in local sub-regions. Meanwhile, multiple probability models are learned and stored in leaves of the D-RF, i.e., a positive/negative patch probability, a head pose probability, the locations of facial features and facial deformation models (FDM). Finally, the composite weighted voting that fuses classification and regression methods is used to decide the locations of facial features. Experiments with the public databases demonstrate the robustness and accuracy of the proposed approach.