Decomposition and virtualization of eigenface for face recognition under various lighting conditions

It is intended in this paper to construct a recognition method that is more robust against changes of lighting conditions than the conventional method. The method applies orthogonal decomposition and virtualization to the eigenface within the framework of the subspace method. The eigenface is derived by applying principal component analysis (eigenvalue decomposition) to the set of face images for each person under various lighting conditions. In the proposed method, the standard face eigenspace (canonical space) is used, and the face image is orthogonally decomposed into the projection component, which is affected mostly by the lighting variation, and the residual component, which contains other aspects of personality and noise. By analyzing the residual component, the noise component in the input image is eliminated. By the above approach, the eigenspace (eigen-projection) containing the standard face information, and the eigenspace (eigen-residual) containing the individuality, are constructed. To the eigen-projection, the concept of the virtual subspace is applied, and the virtual eigen-projection is constructed. With this procedure, stable recognition is realized using the subspace method, even if only a small number of images (as few as one) are registered.