Feature Subspace Using Face Recognition Research

Face recognition is a complex and difficult problem which is important for surveillance and security, telecommunication, digital libraries, video meeting, and human-computer intelligent interaction. Despite the fact that human faces are essentially similar, we are very skilled at recognizing the identities of people from their faces. We can perform this task very easily and it is a basic and important social act although we are still puzzled with the psychological and physiological nature of the process. This thesis studies the theories and methods of FR (face recognition) systematically, focusing on subspace pattern recognition. In the preprocessing stage, we propose to combine pixel averaging, energy-normalization, and the Fourier transform to obtain a vector which has relatively low dimension, low sensitivity to brightness variation and to face shifting in the image plane. In the following feature extraction, we adopt the PCA or K-L transform-based subspace method to avoid the small sample size problem, a general problem in a face recognition system. The extracted eigen features are called eigenspectra or eigenface, depending on whether the Fourier transform is performed or not at the preprocessing stage. Basing on eigenspectra, a local eigenspectra-based scheme for face recognition is addressed in this stage. The classification stage focuses on estimating the parameter of the Gauss kernel for the nonlinear Parzen classifier. Experimental results on the Olivetti Research Laboratory (ORL) face database show that performing energy-normalization and the Fourier transform in preprocessing stage is feasible; better than the traditional eigenspace method, the local eigenspectra-based scheme gets higher recognition rate; comparing with the Euclidean distance rule, the Parzen classifier is more powerful for face recognition.