Learning Blur Invariant Face Descriptors for Face Verification Under Realistic Environment

Face verification technology has been widely used in realist applications such as surveillance, access control and passport authentication. It remains one of the most active research topics in computer vision and pattern recognition. Recently, more research efforts for face verification have focused on uncontrolled environment while current face verification techniques have been proven to be robust and efficient for controlled environment. In this paper, we focus to study on the issue of blur and low resolution (LR), which is common in video surveillance and real application. We propose a descriptor which uses the Fisher Kernel framework to encode the multi-scale absolute phase difference feature of the local image. Then we combine the feature with multiple metric learning approach to achieve a blur robust descriptor that is compact and discriminant. Experiment on blurred ferret dataset and realistic face dataset validates the efficiency of the proposed approach.