Light Face: A Light Face Detector for Edge Devices

Abstract

Face detection is one of the most important and basic steps in the recognition and verification of human identity. Using models based on convolutional networks such as face detection models is very difficult and challenging due to a large number of parameters, computational complexity, and high power consumption in environments such as edge devices, mobiles with limited memory storage resources, and low computing power. In this paper, a light and fast face detection model is proposed to predict the face boxes with real-time speed and high accuracy. The proposed model is structured based on the YOLO algorithm and CSPDarknet53 tiny backbone. Some tricks such as calculating custom anchor boxes aimed to solve the detection problem of varying face scales and some optimization techniques such as pruning and quantization have also been used to optimize and reduce the number of parameters and improve the speed to make the final model strong and suitable for use in environments with low computational power. One of our best models with a MAP of 67.52% on the WIDER FACE dataset and a volume of 1.7 Mb and a speed of 1.43 FPS on a mobile phone with ordinary hardware has shown significant performance