3D Face Feature Extraction Based On Depth Image

Face recognition is a kind of biometric technology, which is one of the most active and potential fields in pattern recognition and machine vision. Furthermore, it has broad applications. 3D face research based on 3D data, combining with computer vision and computer graphics, taking full advantage of depth information of 3D face. 3D face recognition methods are in general able to overcome the problems resulting from illumination, expression and pose variations in 2D face recognition.3D face feature extraction is a very important part in 3D face research. Effectively extract face feature is the key to face recognition. Its basic task is to seek the most effective features for classification from all the source features, to realize the dimension reduction of feature space. Therefore, this paper proposed a new face feature extraction method based on 3D face depth image. The main work and innovative achievement are listed as follows:(1) This paper utilize 3D face depth image to extract face feature. We obtain the 3D face point clouds from CASIA 3D face database supported by CBSR, Institute of Automation, Chinese Academy of Science. By preprocessing, make 3D face point clouds project orthogonally, generate 3D face depth image by resample 3D face depth value to regular data. Therefore, 3D face depth image contains adjacency relation to corresponding 3D face point clouds.(2) Because studies showed that face always locate in high dimension and nonlinear structure, this paper extract 3D face depth image feature by adopting Locality Preserving Projections (LPP) algorithm which based on manifold learning, apply manifold learning to 3D face feature extraction. However, face image exist potential lower dimension, and manifold learning make the high dimension data represent to low dimension space, thus, this paper extract face feature at lower dimension by using manifold learning method, it reflect face potential manifold structure feature.(3) Meanwhile, this paper implement a system about 3D face feature extraction and recognition based depth image under MATLAB environment. And using Locality Preserving Projections (LPP) algorithm and classical Principal Component Analysis (PCA) algorithm extract feature in 3D face depth image. Because the result of feature extraction is measured by face recognition rate, in order to fully reflect the performance of face feature, on this paper applying simpler Nearest Neighbor Classification (NNC) to recognize face. The result showed that at the same 3D face depth image sample set, on contrast with PCA, LPP algorithm achieve higher face recognition rate and need less feature dimension. This indicates that LPP algorithm can extract face feature effectively. In this thesis, the preprocessing and feature extraction based 3D face depth image are described in detail. Through experiments, we analyze the advantage and disadvantage of our methods, and propose the direction and objective in the future.