Research On Face Recognition Based On Embedded Platform

Face recognition technology has brought much convenience to people's life and made huge contribution to society safety.In many applications,face recognition technology needs to be run on the embedded device.With the development of embedded platform,face recognition technology with better performance can be realized on embedded platforms.This paper conducts a study on face recognition technology based on embedded platforms.The main contents are as follows:(1)Face detection algorithm.For realizing the real-time face detection on embedded platforms,a face detection algorithm is designed based on cascaded architecture.The core of this detection algorithm is the three-stage-cascaded convolutional neural network.Every stage needs to do the face classification and bounding box regression.And reiterated boxes will be merged by non-maximum suppression.Global average pooling is used for reducing parameters and the net structure is simplified to reduce the calculation.The algorithm can achieve high accuracy with fast detection speed.(2)Face landmarks detection algorithm.One fast face landmarks detection deep network is designed based on multi-task learning,because face landmarks detection task has inner connection with face properties such as expression,age,pose and so on.5 classification tasks are combined including gender,pose,age,opening of mouth,wearing glasses with the landmarks detection task to train the network,and the accuracy of regression task has been increased.After detecting the landmarks,affine transformation and 3d face model are combined to align faces.(3)Face recognition algorithm.A face feature extraction net is designed based on AM-Softmax.This kind of loss can make features become more discrimitive in angle.The model scale and calculation of the network are decreased by depthwise separable convolution and global average pooling.The network can achieve both high efficiency and accuracy,satisfying the need of embedded platforms.After extracting face features,feature matching is conducted based on cosine similarity.(4)Face recognition software based on embedded platform.Based on the face recognition algorithm designed by this paper,face recognition software is developed which is run on the Jetson TX2 platform.The software can recognize faces in real time,with the real time picture and the Identified identity displayed on the interface.