Design And Implantation Of Face Detection And Recognition System On Embedded Device Based On Shallow Convolutional Neural Network

With the rapid development of the Computation Power and the scale of annotated datasets,the research of Deep Learning grows rapidly for the second time.Leading the trend,the studies in Computer Vision with Deep Learning has made plenty of benchmarks.In some areas like Object Recognition or Object Detection,the Artificial Intelligence work even surpasses human performance.With the growing accuracy,the algorithms of Face Detection and Face Recognition have been implemented into many commercial works,and enter into people's life.Aiming at the difficulties in applying Face Detection and Face Recognition Algorithms on embedded devices,especially computation complexity,this paper introduces some work from System Design,Algorithm Design and Hardware Research and Design,bringing up a real-time system of Face Detection and Recognition on the embedded platforms.Firstly,the paper gives a review of the development of Face Detection and Face Recognition history.Based on that,on algorithm research,the paper introduces MTCNN as the Face Detection algorithm and makes some improvements.Next,we propose a lightweight CNN model with SE-block,Residual connection and Max Feature Map.The model is trained on CASIA and it gets an accuracy of 98.12%on LFW test.Furthermore,we design and build up a suite of software and two hardware platform,Nvidia Jetson TX2 and Raspberry3B+ncnn+Intel Movidius NCS.Tests show that our system could run real time on our devices.Further work could be down by training models on larger datasets,such as VGGFace2.And modern hardware like TPU could also be involved in our system.