Can anyone suggest good illumination normalization algos for face authentication. I have tried basic algos like DoG, LBP, ..?

Question

As part of my project, I want to build an illumination invariant face authentication system. The constraint is I have only 1 enrollment image, but I have the option of addin

Answer 1

Have you tried converting to Lab color space? http://en.wikipedia.org/wiki/Lab_color_space

Code: http://robotics.stanford.edu/~ruzon/software/rgblab.html

Answer 2

The illumination can be corrected (there are algorithms such as Gray-World, Retinex...) => try to preprocess an image with them.

By using different color-spaces such as Lab and HSV you will obtain illumination invariant components (H and S in HSV, a,b in Lab) BUT only for WHITE light (it is good for sun light also). If you have blue lamp (for example) those color spaces would not be appropriate.

My suggestion is to use invariant features such as orientation of gradients, LBP features, SIFT and SURF features and then training a classifier with them.

Most common approach for a face recognition is to use PCA (or kernel PCA) with a SVM classifier. You will probably have to use more than one image because person will not be in the same environment with the same head orientation.