Face recognition under varying illumination

For tackling the problem of face recognition when illumination varied in direction, we proposed two face recognition algorithms. The first is illumination subspace method. We constructed different subspaces that correspond respectively to different illumination directions. We projected the test face image to the subspace having the same illumination direction and perform feature extraction. We then completed face recognition through feature matching between test image and the corresponding subspace. When applicable, illumination subspace method is quite effective. The second method is more general than the first. In the second method, we produced face images under virtual illumination, which is made possible through training RBFN (radial basis function network) with images whose illumination directions are known. Thus we can implement feature matching between test images under any illumination direction and produce virtual image having the same illumination direction. Experimental results show that the illumination subspace method achieves a recognition ratio higher than that achieved by the standard eigenface method. Experimental results also show that the produced face images under virtual illumination can be used effectively as training images without adversely affecting the recognition ratio.