Research On Face Recognition System Based On Posture Analysis And Local Features

As one of the important research directions in the field of pattern recognition,face recognition has always been the focus of people's attention.With the in-depth application of deep learning technology in recent years,face recognition technology has also made great progress.However,there are many interference factors in the image samples collected by the camera,such as local occlusion and posture changes,which affect the feature expression ability of the image samples.In addition,the recognition system cannot effectively distinguish between real faces and faces in photos,which is vulnerable to photo spoof,these factors will adversely affect the recognition results.Based on the above problems,this paper proposes a face anti-spoof algorithm based on dual-posture features,and integrates the dualposture features and local features into the face recognition method.A comprehensive face recognition system with a certain anti-spoof function is researched and implemented.main tasks as follows:(1)A face anti-spoof recognition algorithm based on two face samples with different pose is proposed.There is a significant difference between the imaging results of the stereo face and the photo face of the same individual before and after the rigid body transformation.This paper takes two samples of the target face in different poses as network's inputs,comprehensively compares the features of the two samples at different depths of the neural network,and combines the principle of perspective imaging to estimate the rigid body transform features of the target face,also known as dual pose features.The dual pose feature is derived from the imaging difference between the target face before and after the rigid body transformation,and it also contains certain spatial structure information.Experiments have shown that the dual pose feature can be used to identify stereo faces and photo faces for spoof.A certain effect was confirmed.(2)A face recognition algorithm based on fusion of local features and dual pose features is designed.The distribution law and detail characteristics of local features are analyzed.Local features have the ability to assist face recognition and supplement global features,and can improve the robustness of sample occlusion.This has certain practical significance in scenarios such as mask occlusion and sample fouling.This chapter builds a local feature extraction network based on face classification network,inputs local facial samples containing facial features,and outputs the corresponding local feature vectors;a multi-pose feature fusion network with multi-level feature aggregation is designed to obtain the fusion global feature vectors of face features and dual pose features;the two feature vectors are cascaded and input into the face recognition neural network to complete the fusion of local features and dual pose features.The effectiveness of feature extraction and fusion was confirmed by experiments.(3)A comprehensive face recognition system with certain anti-spoof functions is designed and implemented.Its core is two modules of photo spoof recognition and face recognition.The photo spoof recognition module uses the dual pose feature to determine whether the measured object is a real face or a photo spoof face;the face recognition module improves the recognition performance through the fusion of the dual pose feature and the local facial feature,and has a certain local occlusion robustness,which has some practical significance in the scene of people wearing masks daily under the current pneumonia epidemic.After the implementation of the system was completed,the application was tested in a real scenario and the practicability of the integrated identification system was confirmed.