3D Face Recognition using an Adaptive Non-Uniform Face Mesh

3D Face Recognition using an Adaptive Non-Uniform Face Mesh

Wei Jen Chew (The University of Nottingham, Malaysia), Kah Phooi Seng (The University of Nottingham, Malaysia) and Li-Minn Ang (The University of Nottingham, Malaysia)
Copyright: © 2012 |Pages: 12
DOI: 10.4018/978-1-61350-326-3.ch029

Abstract

Face recognition using 3D faces has become widely popular in the last few years due to its ability to overcome recognition problems encountered by 2D images. An important aspect to a 3D face recognition system is how to represent the 3D face image. In this chapter, it is proposed that the 3D face image be represented using adaptive non-uniform meshes which conform to the original range image. Basically, the range image is converted to meshes using the plane fitting method. Instead of using a mesh with uniform sized triangles, an adaptive non-uniform mesh was used instead to reduce the amount of points needed to represent the face. This is because some parts of the face have more contours than others, hence requires a finer mesh. The mesh created is then used for face recognition purposes, using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). Simulation results show that an adaptive non-uniform mesh is able to produce almost similar recognition rates compared to uniform meshes but with significant reduction in number of vertices.
Chapter Preview
Top

Background

A typical mesh consists of many uniform little triangles that cover the whole face, as shown in Figure 1. (Xu et al., 2004) method converts a point cloud face into a mesh, first by using a coarse mesh and then subsequently refining it to a finer dense mesh to represent a face. After that, recognition is done using the face meshes. Although recognition usually concentrates on the face area around the eyes, nose and mouth, the whole face was converted into a finer dense mesh for recognition.

Figure 1.

Uniform face mesh

978-1-61350-326-3.ch029.f01

(Ansari et al., 2007) used a general mesh model of a face and deformed this model according to the range image of the face, estimating the depth of the triangles using plane fitting. To obtain a smoother mesh, they subdivide each triangle into 4 smaller triangles and then deform the mesh again using plane fitting to get a more accurate mesh for the face. After that, recognition is performed using a voting-based classifier. However, the criteria to determine whether the mesh is accurate enough for the face were not discussed.

(Tanaka et al., 1993) also performed subdivision on their triangles. However, they only divided their mesh triangle into 2 triangles instead of 4. The meshes were subdivided according to their surface curvature and will only stop the subdivision at a certain predetermined threshold.

Complete Chapter List

Search this Book:
Reset