Research On High-precision 3D Face Recognition And Its Application On Access Control System

With the development of deep learning and the accumulation of large-scale training data,face recognition technology has achieved great success in recent years.And the tra-ditional 2D face recognition uses facial texture to represent face feature,and compares the similarity of features to identify face.The facial pose,make-up,and the ambient light have a great influence on the 2D texture.They are detrimental to recognition performance and limit application scenarios of face recognition system.3D face recognition uses 3D depth information to identify face.It avoids shortcomings of 2D methods.How to reconstruct the 3D face quickly and accurately and to optimize the 3D face recognition algorithm effec-tively is an urgent problem to be solved in 3D face recognition and its application.In this paper,3D face reconstruction and high-precision anti-spoofing recognition algorithm are systematically studied.The main contributions of this paper are summarized as follows.1.We design and implement the 3d reconstruction algorithm which is based on Struc-tured Light of Stripes and 3D face camera.2D face recognition maybe achieve very poor performance in the application process,because of face pose,the ambient light and so on.And these problems will not affect 3D face recognition which uses the 3D face information only.It is important to capture 3D face data quickly and correct-ly.High-speed and high-precision 3D face camera meets the requirement of 3D face recognition technology.2.In the face recognition application,misrecognition means that the intra-features distance is larger than the inter-features distance.And we design and implement Improve-Center Lost function.It tries the best to minimize the intra-features dis-tance to reduce the probability of misrecognition.In this method,features of all samples belonged to the same identity constitute a subspace.It makes the hard samples contribute more and minimize the size of subspace.This help us optimize models.3.The traditional method of face recognition organizes the training data with sample pair.And only a part of the information from a pair of training images is used to update the model parameters in every iteration.This way to use local information only from a sample pair could make the distribution of intra features unbalanced and the similarity of intra-class different features quite different.We design and implement Cosmos-Loss Lost function.It is based on the feature center decided by all intra-sample features.This method uses the information between sample feature and feature center to optimize model.It means that this method use all samples belonged to some feature center indirectly.This improves the performance of face recognition and makes intra class features more balanced.4.We design and implement a high-precision and anti-spoofing recognition algorithm which is based on Improve-Center and Cosmos-Loss for access control system.And this method improves the anti-spoofing ability for 3D printed mask by bidirectional data augment of training data.In addition,binocular-vision and multi-spectrum face anti-spoofing methods are proposed to enhance the anti-spoofing ability of face recognition system.And these methods are designed and implemented with special improvement for face Mask scenarios like hospital and epidemic situation.