Face Processing Using Multiscale Geometric Analysis

Biometric Recognition technology is based on the intrinsic properties of the biological characteristics of human being. It is one of the most potential high-tech development directions of this century. Face processing is a very important part of biometric recognition. It includes face detection, face recognition, facial expression recognition etc. and can be applied in human-computer interaction. There is no doubt that the facial feature extraction is crucial to face processing, and wavelet is an important tool in this aspect. However, wavelet is far from perfect. This paper applies the Multiscale Geometric Analysis Tools to the face processing to extract facial feature more effectively. It overcomes the weakness of the wavelet transform which is unable to efficiently extract curve features of face images. Face detection, face recognition and facial expression recognition experiments are carried out.The main work and contribution of this thesis are as follows:1. This paper proposes a new face recognition method based on curvelet transform. Curvelet transform is used for features extraction, PCA/LDA is used for data dimension reduction, and SVM is used for classification in this new method. Coefficients are integrated with each scale's independent data dimension reduction, and the experiment result is very good. Face recognition experiment for Yale database shows that the curvelet method gains 9% recognition rate improvement than wavelet transform. And the new method based independent data dimension reduction can have 3% improvement, the recognition rate reaches 96.6%. The experimental result shows that the new method works very well.2. A new curvelet transform named―half curvelet transform‖is proposed in this thesis. The principle of this new method is to use the symmetry of curvelet domain. When using―half curvelet transform‖to extract features, only half of coefficient is extracted. The experiment result based on―half curvelet transform‖shows that this new method can gain 2% improvement.3. We apply the curvelet transform in the facial expression problem. Some algorithms are similar to those used in face recognition. The facial expression rate for JAFFE database gains 4% improvement than Gabor plus PCA method. So it's suitable to use curvelet for facial feature extraction. It has been shown that curvelet transform is very promising in face processing.