Font Size: a A A

Research And Application Of In-plane Multi-pose Face Detection Algorithm Based On Deep Learning

Posted on:2020-10-23Degree:MasterType:Thesis
Country:ChinaCandidate:B HuangFull Text:PDF
GTID:2428330590960938Subject:Electronic and communication engineering
Abstract/Summary:PDF Full Text Request
Face detection as an important branch of computer vision is receiving more and more attention.Facial features such as appearance and contour are quite different when the amplitude of head posture varies considerably,which leads to the decrease of accuracy of face detection.Therefore,this paper is devoted to the research of face detection algorithm under multi-pose condition in plane and its application in multi-pose face recognition and authentication system.The main contents of this paper are as follows:(1)In-plane multi-pose face detection algorithm with rotation invariance.Aiming at the problem that the accuracy of face detection decreases under the condition of multi-pose in plane,a convolution neural network named RIN robust to face rotation in plane is designed in this paper.The algorithm is based on three-level convolutional neural network cascade.In order to make the whole network robust to face rotation in plane,orientation estimation tasks are added and the corresponding full connection layer is replaced by the global average pooling layer in the first two-level network.At the same time,in order to solve the problem that shallow network can not make full use of target features,this paper introduces dense connection into feature extraction layer,fully integrate low-level contour information such as eyes,nose and high-level facial spatial structure and orientation semantic information,and improves the detection accuracy of the network under the condition of multi-pose face in plane.(2)Design and implementation of multi-pose face recognition authentication System.In order to realize the face recognition and identification system with better recognition effect in three-dimensional space,this paper applies RIN algorithm to the face detection module of the system.The face rectangular frame and key point information obtained by RIN algorithm are used to align the face to get the vertical face in the plane,which can eliminate the influence of face changes in the plane on face verification.Aiming at the situation of vertical face changing in pitch angle and yaw angle,this paper demonstrates the validity of FaceNet face verification in these two situations by experiments,and obtains the appropriate threshold of multi-pose face recognition and authentication system.This paper designs a corresponding interface for the system,which can visualize the results of face detection and verification.In order to verify the RIN algorithm proposed in this paper,the FDDB dataset is expanded into four kinds of datasets.They are original picture(FDDB-up),clockwise rotation 90 degrees(FDDB-right),clockwise rotation 180 degrees(FDDB-down),clockwise rotation 270 degrees(FDDB-left).Experiments show that the recall rate of FDDB-up dataset is 88.0%,that of FDDB-right dataset is 87.3%,that of FDDB-down dataset is 87.6%,and that of FDDB-left dataset is 87.4% when 100 false alarms occur.The detection speed is 25 FPS on CPU and 41 FPS on GPU.The RIN algorithm proposed in this paper is applied to face recognition and identification system for testing.It is obvious that when the face rotates in a large plane,the system can recognize well.
Keywords/Search Tags:Multi-Pose Face Detection In-Plane, RIN, FaceNet, Face Recognition and Identity Authentication System
PDF Full Text Request
Related items