Expression-insensitive3D Face Recognition Based On Facial Inherent Features

Face recognition has become a promising approach to ID authentication with avast number of possible application prospects. It is natural, non-disturbing anduser-friendly. Although2D image-based face recognition have made great leapsover the past twenty years, achieving reliable recognition performance underconstrained conditions, it is still sensitive to variations in illumination, facialposture and expression. The root cause of this is that an2D image is a contractedplane projection of a3D object. Exploiting3D shape information for facerecognition has become a recent research focus. Despite the ability to give anexplicit representation of facial surface, the3D approach is more susceptible tovariable expressions, as facial surface changes induced by variations inexpression.Judging whether two surfaces belong to distinct individuals or aresimply nonrigid deformations of the same subject remains a key challenge in3Dface recognition. This thesis is mainly focused on novel methods of3D facerecognition which is invariable to changeable expressions.The main contents andcontributions are as follows:1. We give a review of face recognition as well as the history and significanceof research, introducing the concept and process of3D face recognition and how3D facial surface data is acquired and represented. It gives a classified summary ofexisting methods based on differences in feature extraction and matchingmethodology, with an emphasis on expression-sensitive3D face recognition and itscharacteristics.2. We propose a novel inherent3D facial feature which is stable againstvariable expressions. Geodesic distances are the inherent features of3D facialsurfaces.We compute the geodesic distances from nose point to every point onfacial surfaces and then extracts iso-geodesic lines to represent geometricproperties of facial surfaces. The experiments demonstrate that iso-geodesicapproache preserves more facial features than the contour line approach underexpression variations.3. We propose a novel3D face recognition method insensitive to varyingexpressions based on inherent facial features. Differential geometry shows that geodesic distances of every pair of point on face surface remain constant againstexpression changes; that is to say, nonrigid deformations induced by variations inexpression are isometric. Face recognition thus translates into an issue of findingan isometric surface eventually. Isometric surfaces of human faces embedded intohigher-dimension Euclidean space have identical presentation.As a result, anexpression-insensitive facial model is built with geodesic distances as inherentfeatures. An efficient improved ICP method is then used for matching the surfacesof above process. The experiments on3D facial depth images demonstrate that thefacial inherent features-based method effectively improve recognition performanceand robustness to variable facial expressions.