Dynamic Face Recognition Based On Deep Learning

Compared with static face recognition,dynamic face recognition has more practical value and research value.However,there are still some difficulties and challenges for dynamic face recognition in video.The dynamic face acquired under video and unconstrained conditions is subject to complex interference factors such as pose changes,expressions and side faces,which make it more difficult to recognize.Based on the above problems,this paper further studies the dynamic face recognition method based on deep learning.The main work is as follows:(1)In order to ensure the training effect and data set quality,the face data set is expanded.DBSCAN is used to cluster and clean the original data set.At the same time,face data is enhanced to remove background interference and average pixel filling.(2)Training P network,R network and O network in MTCNN network respectively to realize face detection effect in pictures and videos,and introducing Deep-Sort tracking algorithm to achieve real-time face tracking and detection in videos.At the same time,considering the needs of different scenes,in the face detection task,the living body detection and mask wearing detection tasks are realized by cascading trained MobileNetV2 neural network.(3)To improve the residual network to achieve the task of dynamic face recognition,it is expounded from two aspects of loss function and neural network.Due to the defects of Softmax loss itself and inspired by Center loss and Marginal loss,a joint loss function is proposed to improve the distance between different face classes and reduce the distance between same face classes.In addition,based on resnet34 network,it is optimized from convolution kernel,activation function and other aspects.The deep residual constant mapping layer is introduced to further improve the face recognition rate,and the effectiveness of the algorithm is tested on multiple datasets of MegaFace,FaceScrub,LFW,SLLFW,YTF,IJB-B,IJB-C.(4)a video dynamic face recognition system is set up to realize real-time face detection,recognition and tracking.Compared with other algorithms,the proposed face recognition algorithm based on joint loss and constant mapping has better performance,effectively improves the robustness of dynamic face recognition,and has certain practical value.