An Improved Face Verification Approach Based on Speedup Robust Features and Pairwise Matching

Human faces are known to present large variability due to factors like pose and facial expression variations, changes in illumination and occlusion, among others, thus making face verification a very challenging problem. In this paper we address the problem of face verification with special interest on how to reduce degradation usually associated with face images acquired under uncontrolled environments. The approach we propose in this paper starts with a preprocessing step to correct in-plane face orientation and to compensate for illumination changes. SURF features are then extracted, which adds to the method a certain degree of invariance to pose, facial expression and other sources of variation. Taking the SURF features as input, an original pair wise face matching procedure is performed. The resulting matching scores are stored in a similarity matrix, which is then evaluated. An experimental study has revealed that the proposed approach produced the best ROC curve when compared to published work regarding the unsupervised setup of the Labeled Faces in the Wild (LFW) [1] face database.