Efficiency analysis of illumination correction methods for face recognition performance

Face recognition is an important task in the computer vision community leading to multiple applications such as building control access, video surveillance or forensics, to mention only a few. Face images are acquired in the enrollment process to form a database, which is the first stage usually performed off line. Second stage implies real-time test procedure where a new unseen face is captured and the face recognition system authorizes or does not, or recognize the identity of the person, based on similarity matching between the new acquired face image and the ones existing in the database, providing a matching score. When the acquiring conditions provided by the enrollment process highly differ by test environmental conditions, some preprocessing steps may be required. An instance of such step is given by illumination correction. The paper aims at analyzing the efficiency of five recent state-of-the-art normalization approaches in term of illumination correction and their effect on face recognition. Surprisingly, to the best of our knowledge, no systematic comparison exist in the literature to date, fact that had motivated us to carry out such analysis. We should also note that, apart from the face recognition task, other domains, such as medical imaging, could benefit from those preprocessing techniques.