Research On Real-time Face Recognition Methods In Videos Based On Deep Learning

With the development of artificial intelligence industry,computer vision has become more and more important in people's lives and various industries,which is widely used in different scenes such as traffic travel,safe cities,drones,financial services,and robots.Face recognition technology in videos is a core research issue in the field of computer vision.At present,face recognition technologies in static images have gradually matured and their accuracy rate is relatively high.However,when they are applied to face recognition in videos,the timeliness of these algorithms is relatively poor,and the demand for real-time recognition of faces in videos cannot be satisfied.Based on the deep learning framework on Caffe,a face recognition framework which is applied on the public face datasets LFW,YTF and the video datasets captured with surveillance cameras is proposed for real-time face recognition in videos in this thesis.Then this thesis continues to improve its accuracy and timeliness.The main research contents are as follows:1.A face recognition framework(named as FR-DL)based on deep learning is proposed,which integrates face detection with MTCNN,face alignment with affine transform,face feature extraction with lightened CNN and face matching with cosine distance.2.This thesis introduces the visual tracking into FR-DL and proposes a real-time face recognition framework in videos based on visual tracking,named as RFRV-VT.This framework identifies the video in groups.Face recognition and face tracking are implemented within the group.Double matching is used to achieve face information connection between two consecutive groups.In order to further improve its timeliness,the hash index is introduced into face matching to convert the face features into hash features,and a two-phase matching method is used for face matching to obtain a new framework(named as RFRV-VT+).The recognition efficiency of FR-DL has been greatly improved,thus satisfying the demand for real-time face recognition.3.A face feature extraction network(named as 32RBSNet)based on ResNet residual structure and a feature fusion method are designed.At the same time,they are combined to generate a facial feature extraction algorithm(named as FFA-32RBSNet).FFA-32RBSNet is applied into RFRV-VT+ in order to get a new framework(named as RFRV-VT++).RFRV-VT++ significantly improves the recognition accuracy of RFRV-VT+,at the cost of slightly reducing the timeliness of RFRV-VT+.Experimental results show that the recognition accuracy of RFRV-VT++reaches 99.48%(LFW),94.2%(YTF),and 99.6%(monitoring video datasets),and the timeliness reaches 27.4 frames per second(YTF)and 30 frames per second.Seconds(monitoring video datasets)in this thesis.It can meet the needs of real-time face recognition in videos.