Expression-invariant3D Face Recognition

As an important biological feature recognition technology, face recognition has manyadvantages such as easier data collection, higher acceptance compared to other traditionalBiological feature recognition method, so it received extensive and thorough research.Nowadays2D face recognition method has achieved ideal recognition results under certainrestrained circumstances, but it is vulnerable to the impact of factors such as lighting, pose,facial expression and make-up.With the development of3D scanning devices, it becomes more and more easier to get3D face model. This makes scholars’ research focus shift to3D face recognition, hoping it cansolve the problems2D face recognition method is faced with. Compared to face image,3Dface model has richer shape information, and it is immune to the impact of pose and lighting.So these days it becomes the research focus of face recognition methods. However, shapeinformation is more susceptible to the effects of changes in expression, therefore changes inexpression become current problems faced by3D face recognition technology.This paper proposes an effective method to solve expression problems, and establishes acomplete face recognition system, including face data preprocessing, feature extraction andfeature matching. We mainly make improvements in the latter two parts. According to thefacts that changes of expression can be regarded as isometric transformation of face surface,firstly we use PCA method to correct face pose, and the nose tip is located by minimum leastsquare methods, then we extract several Geodesic stripes based on detected nose tip, andmake sampling in each stripe, solving the defects of slow matching speed between stripes.Then we extract isometric-invariant features on each feature point, and match featuresbetween two stripes. Because facial expression makes different levels of impact on differentparts of face surface, we use SVM to train the results between the stripes, getting optimalweight for each stripe. Finally, similarities are computed by weighed sum of different stripesmatching results.Our experiments use the Gavab Database and the experiment results are better than other3D face recognition algorithms such as MDS method, showing the effectiveness of ourmethod.