Illumination Invariant Feature Selection for Face Recognition

We propose a novel hybrid illumination invariant feature selection scheme for face recognition, which is a combination of geometrical feature extraction and linear subspace projection. By local geometry feature enhancement technique, neighborhood histogram equalization (NHE) in our experiment, some illegible edges due to week illumination will be enhanced effectively. Then we applied classic linear subspace projection methods, such as Principle Component Analysis (PCA), subspace Fisher Linear Discriminant (FLD), and Direct Fisher Linear Discriminant (DFLD), on these face images to decrease training samples’ dimension as well as diminish the effect of noise introduced at the first step. Our methods are evaluated on an elaborate selected subset (with large illumination variation) of YaleB database. Experiments on this data set show that the NHE+DFLD yields the best performance. By using only 3-dimensional features (the original face images are 256×256), error rate can be decreased from 0.73 (by DFLD only) to 0.07.