Research On Face Detection And Face Recognition Algorithm Based On Mobile Platform

With the improvement of living standard,people pay more and more attention to their personal safety.At present,surveillance cameras have been seen everywhere.In order to reduce the human cost,it is essential to monitor the real-time intelligent analysis of the face in video.In this paper,face detection and face recognition algorithms of mobile platforms are studied in depth.Due to the limited computing resources of mobile platforms,it is a difficult point to improve the speed of the algorithm.In view of the above problems,this paper proposes a face detection and face recognition algorithm based on mobile platform.The specific research contents are as follows:(1)In the research of face detection algorithm based on mobile platform,this paper USES the idea of rough to fine for reference and proposes the face detection algorithm cascaded by traditional method and convolutional neural network.The traditional face detection algorithm NPD(Normalized Pixel Difference)with fast computation was selected as the first classifier to perform fast rough localization of the face in the input image.For the second classifier,we designed a compact and fast convolutional neural network to judge whether the face detected by the first classifier is correct.In the training process,the training images of the convolutional neural network model are processed and generated by the first-level traditional face detection algorithm.The purpose of this is to effectively make up for the shortcomings of the traditional face detection algorithm,so as to improve the detection accuracy and reduce the false alarm rate of the algorithm.This algorithm effectively solves the problems of low detection accuracy and slow speed of face detection algorithm based on mobile platform.(2)In the research of face recognition algorithm based on mobile platform,this paper designs a face recognition network structure with high accuracy and fast speed,which inherits the speed advantage of MobileNetV2 and the precision advantage of MobileFaceNet,and is very suitable for mobile platform with limited hardware resources.In the process of training model,on the basis of traditional Softmax loss function,we designed an improved Softmax loss function.The ultimate goal of training with modified Softmax loss function is to make the difference between the correct category score value and the wrong category score value predicted by the network larger,so that the training model can extract more distinguishing features.The above two algorithms are tested on a variety of open data sets respectively.The experimental results show that these two algorithms not only have high accuracy,but also have fast running speed,and have practical engineering application value.