A Sparse Local Feature Descriptor for Robust Face Recognition

A good face recognition algorithm should be robust against variations caused by occlusion, expression or aging changes etc. However, the performance of holistic feature based methods would drop dramatically as holistic features are easily distorted by those variations. SIFT, a classical sparse local feature descriptor, was proposed for object matching between different views and scales and has its potential advantages for face recognition. However, face recognition is different from the matching of general objects. This paper investigates the weakness of SIFT used for face recognition and proposes a novel method based on it. The contributions of our work are two-fold: first, we give a comprehensive analysis of SIFT and study its deficiencies when applied to face recognition. Second, based on the analysis of SIFT, a new sparse local feature descriptor, namely SLFD, is proposed. Experimental results on AR database validates our analysis of SIFT. Comparison experiments on both AR and FERET database show that SLFD outperforms the SIFT, LBP based methods and also some other existing face recognition algorithms in terms of recognition accuracy.