Face Recognition by Coarse-to-Fine Landmark Regression with Application to ATM Surveillance

While ATM provides us convenient banking services, it has great security risks. The authentication of only password requiring is not safe enough. With the rapid development of face recognition technology based on deep convolutional neural network (CNN), undoubtedly, applying it into ATM authentication will improve security further. In this paper, we explore a new authentication mode combine face recognition and basic password for ATM. We think that it would prevent the economic crime on ATM fundamentally. However, computational and storage costs of CNN based methods are still high. To this end, we propose a new face recognition method by landmark regression. Our pipeline integrates a landmark localization network with a light face recognition network. For landmark localization, we employ a fully convolutional neural network to produce facial landmark response maps directly from raw images in a coarse-to-fine manner. For face recognition, we train a light CNN to obtain a compact representation, where the rectified linear unit (ReLU) is replaced by max-feature-map (MFM). Our approach shows good performance on several datasets. And it is practicable due to its high speed, good accuracy, and low storage space requirement.