Online Incremental Face Recognition System Using Eigenface Feature and Neural Classifier

Understanding how people process and recognize faces has been a challenging problem in the field of object recognition for a long time. Many approaches have been proposed to simulate the human process, in which various adaptive mechanisms are introduced such as neural networks, genetic algorithms, and support vector machines (Jain et al., 1999). However, an ultimate solution for this is still being pursued. One of the difficulties in the face recognition tasks is to enhance the robustness over the spatial and temporal variations of human faces. That is, even for the same person, captured images of human faces have full of variety due to lighting conditions, emotional expression, wearing glasses, make-up, and so forth. And the face features could be changed slowly and sometimes drastically over time due to some temporal factors such as growth, aging, and health conditions. When building a face recognition system, taking all the above variations into consideration in advance is unrealistic and maybe impossible. A remedy for this is to make a recognition system evolve so as to make up its misclassification on its own. In order to construct such an adaptive face recognition system, so-called incremental learning should be embedded into the system because it enables the system to conduct learning and classification on an ongoing basis. One challenging problem for this type of learning is to resolve so-called “plasticity and stability dilemma” (Carpenter & Grossberg, 1988). Thus, a system is required to improve its performance without deteriorating classification accuracy for previously trained face images. On the other hand, feature extraction plays an essential role in pattern recognition because the extraction of appropriate features results in high generalization performance and fast learning. In this sense, incremental learning should be considered not only for a classifier but also for the feature extraction part. As far as we know, however, many incremental learning algorithms are aiming for classifiers. As for the incremental learning for feature extraction, Incremental Principal Component Analysis (IPCA) (e.g., Oja & Karhunen, 1985; Sanger, 1989; Weng et al., 2003; Zhao et al., 2006) and Incremental Linear Discriminant Analysis (Pang et al., 2005; Weng & Hwang, 2007) have been proposed so far. Hall and Martin (1998) proposed a method to update eigen-features (e.g., eigen-faces) incrementally based on eigenvalue decomposition. Ozawa et al. (2004) extended this IPCA algorithm such that an eigen-axis was augmented based on the accumulation ratio to control the dimensionality of an eigenspace easily. O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg