|
Face verification is applied to our life, which benefits from the high-accuracy of the algorithm based on CNN. However,the performance of face verification is still poor on mobile device since limited computation resources. In this paper, we present a class of extremely efficient algorithm with attention mechanism embedded, the algorithm of 20MB size achieves 96.37% face verification TAR(FAR1e-6) on MegaFace Challenge, which is even comparable to hundrads MB size. We compare our algorithm with similar small size models like MobiFace, MobileFaceNet, Goole-FaceNet, the experimental results show the efficent of our algorithm. |
|
Keywords:Pattern Recognition; Face Verification; Attention Mechanism; Depth wise Separable Convolution |
|