Face Recognition System Based MTCNN

With the advancement of computing power and cloud computing architecture,deep learning algorithms,which heavily rely on massive computing,have developed rapidly.On this basis,many difficult implementation problems in application have obtained feasible solutions.Identification is one of the most eye-catching application and the solution of face recognition has received a lot of attention because of its convenience.At present,deep learning algorithms have achieved applicable results on face recognition,but the massive computing resources cost makes deep learning solutions hard to promote.There are two feasible solutions to this problem.One is to build a local server equipped with powerful GPU,which has huge power consumption,and also is a big waste of equipment resources;the other is to use cloud computing services to perform most of the data processing,which highly depends on the stability of network and holds a serious risk of information leakage.Aiming at these problems,we propose the idea of building the deep learning face recognition system entirely on an embedded platform.Based on the research and analysis on deep learning algorithms and applications,we employ Intel Movidius Neural Compute Stick to undertake the deep learning processing,easing the problem of insufficient computing resources.After experimenting on several representative algorithms,we select SSD as the basic framework of detection algorithm and optimize it.Our detection algorithm can achieve real-time detection speed of 11 fps,and 93.25%accuracy on the standard database FDDB.Sphereface is selected as our face recognition algorithm because of its high resolution.It can achieve 25fps and 93.03%accuracy on LFW database after performing face alignment with our optimized key-point detection model.Finally,the complete processing can achieve 7fps steadily and 99.78%accuracy in application.We also design a multi-device mechanism,which can further improve the operating speed when equipping with more Intel NCS.Compared with existing embedded face recognition systems,our embedded platform face recognition solution has obvious advantages in performance and scalability.We also apply our face recognition solution to the access control application.We add a relay-based gating logic to the face recognition system,a living body recognition to prevent photo fraud,and a whitelist database.At the same time,we build Android andiOS applications to show and record recognition results,and remotely control the whitelist database in system.Finally,our access control system is fully assessed.Our work is highly practical,and it has great reference value for the implementation of high-computation algorithm solutions on embedded platforms with limited resources.