Robust Face Alignment Based on Hierarchical Classifier Network

Abstract

Robust face alignment is crucial for many face processing applications. As face detection only gives a rough estimation of face region, one important problem is how to align facial shapes starting from this rough estimation, especially on face images with expression and pose changes. We propose a novel method of face alignment by building a hierarchical classifier network, connecting face detection and face alignment into a smooth coarse-to-fine procedure. Classifiers are trained to recognize feature textures in different scales from entire face to local patterns. A multi-layer structure is employed to organize the classifiers, which begins with one classifier at the first layer and gradually refines the localization of feature points by more classifiers in the following layers. A Bayesian framework is configured for the inference of the feature points between the layers. The boosted classifiers detects facial features discriminately from its local neighborhood, while the inference between the layers constrains the searching space. Extensive experiments are reported to show its accuracy and robustness.