Research On Face Detection And Recognition Method Based On Deep Learning

Face detection and recognition technology have been an important research direction in the field of computer vision.In recent years,with the rapid development of related technologies of computer vision,especially the extensive application of deep learning,face detection and recognition have attracted more and more attention.Face detection as a necessary step before face recognition,its detection speed and accuracy affect the performance of the entire process.Under constraint conditions,many existing face recognition algorithms have achieved the accuracy of human eye recognition.However,under non-controllable conditions,there are various complex interferences in the collected face images,the face recognition technology under constraint conditions can no longer meet the needs of the application in reality.For this reason,it is very important to design an identification technique that is robust to complex disturbances.In this paper,a high-performance face detection and recognition algorithm is proposed.The main contents of this paper are as follows:1.The structure and theoretical basis of the convolutional neural network is summarized.This article describes the calculation methods and back propagation algorithms in the network.Afterwards,the key technologies commonly used in deep learning such as Re LU unit,Dropout,and Batch Normalization are introduced as the theoretical basis of the following chapters.2.Research on face detection algorithm based on YOLO detection model.The convolutional neural network is used to regress the confidence score and the position information of the face image,and the convolutional neural network structure in the model is improved.Depthwise separable convolution is used in place of the traditional convolution unit to reduce the size of parameters while deepening the depth of the network.In addition,the face detection speed is accelerated without the loss of detection accuracy.3.Research on the feature extraction method based on convolutional neural network.A network structure for face image feature extraction under complex conditions is proposed.The face verification task is used as a target to optimize the network parameters.Afterwards,it is experimentally proved that when using this feature to measure similarity,it has better robustness than traditional artificial design features and can better express face images.4.Research on the feature matching algorithm based on Gaussian mixture model.The Gaussian mixture model is used to implicitly model the subregions of the face image,and the face image is verified at a finer granularity.The training of the Gaussian mixture model is unsupervised and does not require data annotation,which saves training costs.Afterwards,using the joint Bayesian algorithm to measure the similarity of features on the matching face sub-blocks,which can greatly reduce the impact of the highly nonlinear variations such as postures and expressions and improve the recognition accuracy.