Adaptive registration for occlusion robust 3D face recognition

Abstract

Occlusions over facial surfaces cause performance degradation for face registration and recognition systems. In this work, we propose an occlusion-resistant three-dimensional face registration method. First, the nose area is detected on a probe face using curvedness-weighted convex shape index map. Then, probable eye and mouth patches are detected and checked for validity. An adaptive model is constructed by selecting valid patches of the average face model. Finally, registration is handled with the Iterative Closest Point algorithm, where the adaptive model is used as the reference. The UMB-DB face database is used to evaluate the registration system: The nose detector has 100% and 93.90% accuracy, for the non-occluded and occluded images, respectively. A simple global depth-based recognition experiment is done to evaluate the registration performance: Our adaptive model-based registration scheme improves rank-1 recognition rate by 16%, when compared with the nose-based alignment approach. © 2012 Springer-Verlag.