Research On Face Detection Algorithm In Uncontrolled Environment

Face detection is the process of determining the location and size of all faces in the input image.In recent years,with the rapid development of artificial intelligence,face detection,as one of the hottest research topics,has been widely used in various aspects of life,such as smart communities,campus access control,face payment and so on.Currently,face detection algorithms have achieved excellent results in terms of detection accuracy for face capture image data in controlled environments.The face image in the controlled environment is characterized by high definition,good lighting,unobstructed face and no tilt.However,there are often face capture scenarios in uncontrolled environments in real life.Compared with controlled environments,the images in uncontrolled environments are more complex,and the captured faces in uncontrolled environments are variable in terms of pose,expression,and angle.So face detection in uncontrolled environments is more challenging.The focus of face detection technology has gradually shifted from face detection in controlled environments to face detection in uncontrolled environments.In this paper,we investigate a typical case of face detection in an uncontrolled environment,which is the problem of detecting faces at high viewing angles.We design YOLOv5-Mobile Net V3-CBAM-Bi FPN,which is a high performance face detection algorithm for high top view angle faces.In order to verify the effectiveness of this algorithm,the face detection and recognition algorithm is applied to the captured image under high overhead angle of the elevator to carry out the identity recognition experiment of the person appearing in the image,so as to achieve a high accuracy rate of recognition of the face in high overhead angle.The main research elements of this paper are:(1)Use high top view angle faces for data augmentation to improve the dataset.The specific method is to collect the high top view angle face dataset HOAFD.HOAFD contains a total of 8394 face images of 1399 subjects.In this paper,face images are collected from 6 angles of 0°,15°,30°,45°,60° and 75° for each subject.And 2796 face pictures of 699 subjects with angles of 30°,45°,60°,75° are selected as the test set of this paper to fully simulate the high top view scenario in life.The remaining 5598 face images are added to the network as the training set to expand the original face training set Wider Face to achieve data enhancement and thus improve the generalization ability of the network.The YOLOv5 algorithm are used to carry out face detection experiments before data enhancement and after data enhancement.And the results of the experiments are comparatively analyzed.The results shows that the mean of the average accuracy m AP@.5:.95 of the algorithm is greatly improved by 15.1% by data enhancement.(2)The improvement of face detection algorithm.This thesis takes the YOLOv5 algorithm with outstanding performance in the YOLO series of algorithms as the basic algorithm for research and improves the network structure of the YOLOv5.Firstly,Mobile Net V3 is used to replace the original backbone network of YOLOv5,which greatly reduces the amount of model parameters and calculations and improves the detection speed while maintaining little decrease in accuracy.Secondly,the SENet attention mechanism in Mobile Net V3 is replaced by the ECANet,CBAM,and CA attention mechanisms to retain more effective information and thus further improve the performance of network.Then the performance of the algorithm without replacing the SENet attention mechanism is compared with the algorithm obtained by replacing the SENet attention mechanism by ECANet,CBAM,and CA.The YOLOv5-Mobile Net V3-CBAM algorithm with the best performance among the four is selected as the basic algorithm for the next improvement.Finally,Bi FPN is used to replace the original feature pyramid structure of the Neck part of YOLOv5-Mobile Net V3-CBAM algorithm to improve the efficiency of multi-scale feature fusion and thus improve the accuracy of network detection.The YOLOv5-Mobile Net V3-CBAM-Bi FPN face detection algorithm with the best performance in this thesis is obtained.Its mean of the average accuracy m AP@.5:.95 reaches 93.4% and the detection speed reaches 298 FPS.(3)Face recognition experiment.This thesis makes a face library,which contains 60 face images with known identity information.70 surveillance pictures of people in the elevator containing 90 faces are also taken.The improved YOLOv5-Mobile Net V3-CBAM-Bi FPN face detection algorithm with the best performance obtained and other face detection algorithms in this thesis are combined with the existing Face Net face recognition algorithm to carry out multiple face recognition experiments.The experimental results are compared and analyzed to verify that the YOLOv5-Mobile Net V3-CBAM-Bi FPN face detection algorithm designed in this thesis can detect high top view angle faces with high accuracy.It also can be applied on the ground.It can be combined with the Face Net face recognition algorithm to recognize the high top view angle faces taken in the elevator with high accuracy and thus judge the identity of the high top view angle faces in the elevator correctly.And the technical preparation for the development of the face recognition elevator control system is prepared.