A study on the implementation and the robustness of face verification methods under illumination changes

We focus on the performance of the face authentication system under illumination changes. Four different face authentication methods are tried and compared. These methods include PCA (principal component analysis), GMM (Gaussian mixture models), 1D HMM (1 dimensional hidden Markov models), pseudo 2D HMM (pseudo 2 dimensional hidden Markov models). Experimental results with artificial illumination changes upon ORL (Olivetti Research Laboratory) face database are compared with each others. Face feature vectors are extracted using the 2D DCT (2 dimensional discrete cosine transform). Experiment results shows that the EER (equal error rate) performance degrades in all occasions when the illumination change is introduced. Without illumination change, the EERs are 2.54%, 3.18%, 11.7%, 13.38% for Pseudo 2D HMM, 1D HMM, PCA, GMM respectively. The 1D HMM shows better performance than the PCA without illumination change. But it gets worse according to the illumination changes. Among them, the pseudo 2D HMM shows the best EER performance regardless of the illumination changes.