Non-negative matrix factorization methods for face recognition under extreme lighting variations

Face recognition task is of primary interest in many computer vision applications, including access control for security systems, forensic or surveillance. Most commercial biometric systems based on face recognition are claimed to perform satisfactory when the enrollment and testing process takes place under controlled environmental conditions such as constant illumination, constant pose scale, non-occluded faces or frontal view. More or less deviation from those conditions might lead to poor recognition performances or even recognition system's failure when a test identity has to be recognized under new modified testing conditions. Three non-negative matrix factorization (NMF) methods, namely, the standard one, the local NMF (LNMF) and the discriminant NMF (DNMF) are employed in this paper where their robustness against extreme lighting variations are tested for the face recognition task. Principal Component Analysis (PCA) method was also chosen as baseline. Experiments revealed that the best recognition performance is obtained with NMF, followed by DNMF and LNMF.