Research And Implementation Of Real-time Face Recognition System

In this thesis,we have studied the face recognition system which has been rapidly developed in recent years.Due to the breakthrough in the field of convolution neural network and deep learning,artificial intelligence has becoming increasingly potential that many research achievements has been applied into practice.Face recognition products are not new in the market,but most of them are provided for users' application by way of products and cloud services.These products need to be paid,and their real-time performance may still be compramised.Even worse,they may not be able to meet the use requirements under different backgrounds because of their original product orientation.So it is necessary to develop our own face recognition system.This thesis offers some insights into three important components of face recognition system,face detection,face feature extraction and face classification.Firstly,YOLO(You only look once)is a target detection algorithm that meets the requirements of real-time detection so that we use YOLO to train the face detection model.Since face detection only has a single target,we simplifiedy the YOLO network model.Then,the HNM(hard negative mining)sampling method is introduced to the selection of hard positive samples.According to the error of the samples showed on the detector,we picked out the hard samples to join the training so as to enhance the face detection model.Secondly,since DeepID(Deep hidden identity features)face verification algorithm is among the best at present,and it is a simple and highly-efficient network model for face recognition,we made use of it.To enhance the real-time performance of our system,we based our face feature extraction network on DeepID and made some modifications.Then we added a center loss verification signal,trained a DeepID-liked network model using central loss and softmax loss joint supervisor learning,and used it to extract face features.Thirdly,for face classification,we can use a simple SVM classifier,using the face feature extracted from feature extraction network to train the classifier.At last,we combined the face detection,face feature extraction and classifier into our system.Face detector detects the face from the camera,then input the face to feature extraction network to extract features for the SVM classifier to carry out face recognition.Through the test of the face detection model and the feature extraction network,it shows that the false positive rate of face detection has been significantly decreased,the bounding-box regression fit has proven to be better,and the time performance has also been improved.The feature extraction network was not overfitting,the extracted feature is very distinctinctive and the whole process takes little time.Our face recognition system can reach nearly 25 fps,so that it is enough to meet the real-time requirements.Moreover,its face recognition rate is also beyond the human identification.