Face recognition is one of the most important applications to receive attention in the areas of Computer Vision and Pattern Recognition. However, face recognition has many challenges and difficulties, such as the requirement for high speed search in large datasets and the requirement for high match accuracy under various noise conditions. Currently, as numerous 3D face datasets become available, more and more researchers start to move their concentration to 3D face recognition. Compared with 2D face image, 3D face images contain more explicit information which is very useful for dealing with the head orientation and the facial expression problem. In this thesis, a framework to implement automatic 3D face recognition is proposed and implemented. In the first stage, a key facial feature - the nose has to be extracted for the subsequent face recognition process. In order to exploit the local feature information, we present a face feature extraction methods based on a 3D shape descriptor. Two different 3D shape descriptor Multi Contour Surface Angle Moments Descriptor(MCSAMD) and Multi Shell Surface Angle Moments Descriptor(MSSAMD) are designed and implemented. The nose tip is identified using a binary neural network technique called k-Nearest Neighbour Correlation Matrix Memories(CMM) algorithm. The main face area is localized and cropped based on the nose tip localization with an identification rate of almost 100% on FRGC 3D face database. Secondly, a face aligned approach is implemented by applying a combination of methods including Principal Component Analysis(PCA) face correction, Iterative Closest Point algorithms(ICP) and the alignment using the symmetry of human face. All faces are aligned to a unified coordinate system from the original pose position even under expression variations. The position of the nose tip is also further corrected. After the face alignment, the main face area is divided into several regions with different weights according to the face expression variability. Similarity measurement algorithms based on the pose-invariant 3D shape descriptor MSSAMD are used to match the corresponding regions for different faces. The expression variability weights are applied in the final consideration of face identification and verification. Experiments are performed on the FRGC database which is the largest 3D face database of 4950 faces with different expressions. In the experiments dealing with 4007 faces with different expressions, a 91.96% verification at a false acceptance rate(FAR) of 0.1% and a 97.63% rank-one identification rate are achieved.
展开▼
