Face Recognition Using Kernel Nearest Feature Classifiers

The nearest feature line (NFL), feature plane (NFP) and feature subspace (NFS) classifiers have achieved good results in face recognition. However, in these three methods the facial features need to be extracted before classification can be performed. To overcome this drawback, in this paper we extend these three classifiers to kernel based NFL, NFP and NFS classifiers respectively. In addition, two kinds of KNFS are proposed. One is direct generalization of KNFP, and the other employs kernel principle component analysis to construct nonlinear feature subspace. The advantage of the proposed methods is that original high dimensional face image can be directly classified without the preprocessing step to extract facial features. To overcome the drawbacks of the large computation complexity and possible failure in KNFL and KNFP, these two classifiers are further extended to kernel based nearest neighbor feature line and feature plane. Experimental results demonstrate the feasibility of the proposed methods for directly classifying the high dimensional face images