Lightweight Deep Embeddings Fusion Methods for Face Verification on Mobile Devices

Abstract

With the emergence of deep learning technology, face verification performance on the LFW benchmark outperformed human-level accuracy. In recent, face verification technique is applied to mobile devices to authenticate owner of them. Face verification on mobile devices needs to be considered with both performance and computational cost which is trade-off. In this paper, we propose lightweight network architectures and fusion methods for efficient face verification task on mobile devices. The proposed lightweight networks achieved comparable accuracy on both LFW and YTF datasets with only about 9x fewer size of model and 2x shorter inference time than Inception-resent-v1 which is a deeper network architecture. In addition, proposed score-level fusion method shows improvement of 2.38% on VR@FAR=1e-6 on LFW BLUFR and that of 0.28% on accuracy on YTF than single lightweight network.