The Research And Application Of Face Recognition In Low Quality Video

Face recognition is an important biometric technology,which has the advantages of non-intrusive,convenient and non-contact,etc.In recent years,face recognition has achieved high recognition accuracy not only in the laboratory,but also in the wild.There are many applications for face recognition in daily life,such as Alipay face verification payment,train station security,and so on.In these successful experiments or applications,the images are usually involved in high resolution,good illumination conditions,small changes in posture,and inconspicuous occlusion.However,in scenes such as surveillance videos,video face recognition always faces the difficulties of low resolution,poor illumination,large changes in pose,and significant occlusion.Their effects did not meet expectations.Videos with these identification difficulties are low quality,and face recognition in low quality videos is an important challenge we face today.In view of the above problems,this paper proposes a V2V(Video to Video)video face recognition method that combines deep learning and SRC(sparse representation classification),and a S2V(Still to Video)video face recognition method that combines deep learning and CCA(canonical correlation analysis).This paper uses the CenterLoss method to extract depth features,which are more stable than other depth features under changes in illumination and attitude.In V2V face recognition,this paper proposes a compression dictionary and improved voting group sparse representation method to predict video classification.Combining the sparse representation based method enhances the robustness of depth features under occlusion.The compression dictionary and group recognition methods in this paper make sparse representation still applicable in large-scale face recognition.In S2V face recognition,this paper uses canonical correlation analysis to project high-resolution gallery images and low-resolution test videos into a common space,thus solving the heterogeneous matching problem in S2V face recognition.In this paper,V2V face recognition experiments were performed on Honda datasets and COX datasets,and S2V face recognition experiments were performed on COX datasets.The Honda dataset contains a large number of pose changes.The COX dataset has a lot of people,poor lighting conditions and low resolution.Both of them are low quality videos.The experimental results show that the recognition accuracy of our proposed method is competitive with the state-of-the-art video face recognition methods.In addition,this paper implements a real-time video face recognition system,which embodies the effectiveness of the proposed algorithm.