Tiny Face Presence Detector using Hybrid Binary Neural Network

Abstract

Face Detection plays a key role in “always-on” applications such as mobile phone unlock or smart doorbells. Deep learning-based face detection solutions have demonstrated state-of-art performance in terms of accuracy; however generally, the improved accuracy comes with a large computation and memory requirement overhead. This can result in high energy consumption which is a significant cost that can overrun the energy budget especially in battery powered systems. Recent solutions to this problem have advocated the use of a low power always-on sensor running a rudimentary algorithm that can merely indicate the ‘presence’ of a face with low accuracy and in turn ‘wake-up’ a more powerful device executing a high accuracy face detection algorithm. In this paper we present the design of two deeply quantized (binarized) light weight face presence detection deep learning based models that can function as wake up models. The models achieve high accuracy> 98% with a corresponding memory footprint being limited between 3KB and 100KB allowing them to be deployed in highly resource constrained ‘always-on’ embedded platforms.