Effective methods for lightweight image-based and video-based face recognition

Face recognition has achieved significant advances with the rise of deep convolutional neural networks (CNNs) and the development of large annotated datasets. However, how to design deep models in lightweight face recognition is still a challenge when aiming at mobile and embedded devices. In this paper, we focus on recent efficient CNN architectures, speedup skills and reduction methods to design models for lightweight face recognition. We combine octave convolution with MobileNet and ResNet for those models sensitive to computation complexity, replace feature output layer for those models sensitive to memory and explore network scaling for more powerful representation. Further, we extract a subset from the whole training dataset to speed up the performance evaluation of different models. We provide a scaling method on MobileFaceNet to boost the performance with the limit of computational cost, and propose a simple supplementary method for average pooling which throws up those noise frames based on the cluster information in video face recognition. With the upper bound of 1G FLOPs computation complexity and 20MB model size, our best model achieves 99.80% accuracy on LFW, 98.48% on AgeDB, 98% on CFP-FP and 97.67% TAR@FAR 10(-6) on MegaFace.