smartface

SmartFace: A Multi-Threaded Face Recognition Framework for IoT EdgeTPU Devices

Abhijeet Boragule and Moongu Jeon

This work is under Under Revision @ NeuroComputing

Most deep learning-based face recognition systems rely on machines and cloud services that result in increased latency, cost, and privacy concerns. Existing face recognition products provide support in constrained environments where face frontalization is necessary, and they are limited to conventional face recognition algorithms. To train face recognition models, a large-scale dataset is required for the training of convolution neural networks (CNN). Although such datasets have been harvested using the Internet, they contain many incorrect class labels that reduce the discriminative ability of the CNN model by projecting different identities into the same latent hypersphere. To address this issue, we propose a face quality assessment and efficient relabeling mechanism during the training phase to alleviate the noisy labels. Additionally, in this paper, from a research model to a real-world application, we present a multi-threaded framework that utilizes Edge Tensor Processing Unit to offer adaptive face recognition, live camera streaming, and face management software in the state-machine environment. Our framework efficiently addresses the most time-sensitive issue associated with face recognition over cloud services and performs on-device face recognition in 100.0 ms. Results on the collected dataset and IJB-B, IJB-C, LFW, AgeDB30, CALFW, CFPFP, and CPLFW public benchmarks demonstrate that our approach achieves state-of-the-art performance.

Wanna know more visit Abhijeet’s Portfolio

Please watch the video for the live demo.