Design And Implementation Of Face Recognition System Based On Deep Learning

At present,due to the rapid development of deep learning,the face recognition problem under constraint conditions has been basically solved,but in the unconstrained scene,face recognition still faces many challenges.Among them,in the unconstrained scene,the multi-pose problem of the face causes serious interference to the face recognition,which becomes the key issue in the field of face recognition in the unconstrained scene.Therefore,it is of great theoretical and practical significance to study the face recognition algorithm with attitude robustness and design and implement the corresponding application system to meet the needs of the industry.After completing the face recognition system based on deep learning in unconstrained scenes,this thesis further studies the feature representation of multi-pose faces and explores the face recognition algorithm with attitude robustness.The main research contents are as follows:A face recognition system based on deep learning in unconstrained scenes is designed and implemented.Firstly,the algorithm functions of each module of the face recognition system are designed,and the tensorflow framework is used as the platform.Python and C++ are the programming languages,and the image acquisition module,the face detection module,the face alignment module,the face recognition module and the alarm module are performed.Programming implementation,a complete face recognition system is implemented.Further,in the unconstrained scene,the phenomenon that "the different people in similar postures are mistakenly judged as the same person,the same person in different postures is wrongly judged as different people" appears,and the formation principle is deeply analyzed.Based on the "front" thinking of the feature level,The multi-pose facial feature representation is mapped to the approximate positive facial feature representation,and the positive facial feature representation is used to realize the recognition to achieve the purpose of robustness to the posture.Focusing on this idea,this thesis adds two processes of pose estimation and pose classification to obtain the pose classification label of the face.The feature extraction network is redesigned,the pose mapping transformation module is constructed by using the gesture tag,and a new attitude loss function is defined to optimize the network.Finally,the test results of three public data sets on LFW,CFP and IJBA-A prove the good recognition effect of this paper.At the same time,the implementation of the face recognition system is tested,which shows that each module of face recognition realizes the design function and the effect is good.The overall system can run smoothly and achieve the expected goal.