Discriminant Subspace Analysis for Uncertain Situation in Facial Recognition
Pohsiang Tsai, Tich Phuoc, Tom Hintz, J Tony · InTech eBooks · 2008
Recent Advances in Face Recognition 162However, in (Adini et al., 1997), the authors mainly focused their empirical experiments on variations due to changes in illumination.They stated that within-personal variation is larger than between-personal separation.These variations between images of the same individual faces make difficult machine learning.Therefore, in a facial recognition system if the extracted input data contains misleading information (ambiguous regions), classifiers may produce a degraded classification performance (Jan, 2004).Specifically, in this chapter, we will mainly focus our empirical experiments on variations due to changes in facial expression that are less emphasized in (Adini et al., 1997) and deal with the impact of facial expression changes as individuals deform/express their faces either naturally or deliberately in a real-time face recognition system.As Adini et al. (Adini et al., 1997) stated that a facial recognition system should recognize a face insensitive to these within-personal variations.Limited success is reported for face recognition systems that are invariant of facial expression changes (Liu et al., 2002b) (Liu et al., 2003) (Martinez, 2000) (Martinez, 2002) (Seow et al., 2003) (Chen & Lovell, 2004).Our earlier research on a facial expression invariant system demonstrated its challenging nature (Tsai et al., 2005) (Tsai & Jan, 2005).If the number of individuals is increased (along with their varying facial expressions), the facial data will largely overlap.Thus, the variations of individual facial expressions will increase the range of uncertainty.This makes classification difficult.The aim of this chapter was first to address the issue of within-personal variations due to facial expression changes.We then used a kernel-based discriminant analysis technique to reduce the uncertainty (overlapping) in the feature subspace applied before learning so as to improve classification rates.This chapter also examined other linear and nonlinear techniques (PCA, FLD, and KPCA) for comparison.Their transformation effects on a subsequent classification performances were then tested in combination with learning algorithms (multi-layered perceptron neural networks (MLPNNs), radial basis function neural networks (RBFNNs), and support vector machines (SVMs)).The algorithms were then applied to face database with facial expression changes.We found that the transformation of kernel-based discriminant analysis had a beneficial effect to the classification performance.The experimental results indicates that non-linear discrimininant analysis method may robustly deal with the uncertainty problem.It appears that a facial recognition system may be robust to facial expressio changes, and thus be applicable.The structure of this chapter is as follows: First, we provide a concise overview of the facial expression analysis.Second, we discuss the expression variant problem in facial recognition.Third, we introduce a concise overview of the subspace feature extraction methods.Then, in the final part of this chapter, we analyse different subspace transformation methods and their transformation capabilities.We finally present the results of the experiments and discuss them from several aspects, focusing on the advantages and disadvantages of each subspace feature extraction method. Facial expression analysisHuman faces contain abundant information of human facial behaviors (Cohn et al., 1999).According to Johansson's point-light display experiment (Johansson, 1973) (Johansson, 1976), facial expressions can be described by the movements of points that belong to the facial features such as eye brows, eyes, nose, mouth and chin and analyzed by the relationships between those features in movements (Pantic & Rothkrantz, 2000b).Hence, www.intechopen.