Research On Single Stage Multi-level Face Detection Algorithm Based On Convolutional Neural Network

Face detection has always been a hot topic in computer vision.Especially in the past decade,with the rapid development of deep learning,face detection algorithms based on deep learning emerge in endlessly,and related products are widely used in real life.Based on the current face detection algorithm,this thesis proposes a single-stage multi-level face detection model,which can effectively solve the problem of dense,multi-scale,low-quality face image detection.In particular,using this face detection algorithm in face recognition system can effectively improve the accuracy of authentication.In order to extract face features efficiently,a face feature extraction module based on convolutional neural network is constructed.The module is mainly composed of three parts:the main feature extraction module based on the split attention network(Res Ne St),the feature fusion module based on the feature pyramid network(FPN)and the feature enhancement module based on the context module in SSH.Among them,the face feature extraction module based on Res Ne St benefits from the attention splitting mechanism,which can effectively fuse channel domain features and show good feature extraction ability.FPN structure can effectively extract face features of different scales and perform multi-level feature fusion.This structure is conducive to the detection of faces of different sizes.Applying context module to feature pyramid can improve the receptive field of convolution kernel and enhance the context details of feature graph.Secondly,this thesis constructs a multi task branch of face detection based on face features.In order to solve the problem of low-quality face image detection,this thesis constructs a high-resolution face prediction branch based on face enhancement,which uses selfsupervised learning method to mine the supervision information of the image itself.Another branch of self-supervised learning predicts 3D face information,which uses graph convolution(GCN)method to decode the 2D feature vector of face and complete the 3D reconstruction process.At the same time,3D face structure information is integrated into the high-resolution face prediction process to improve the face super-resolution quality.In addition,in order to improve the performance of face detection algorithm,this thesis combines the face key point regression branch based on supervised learning,which effectively improves the face localization ability of the model.Finally,in order to verify the effectiveness of the face feature extraction algorithm based on Res Ne St and the multi task face information prediction branch,as well as the influence of face detection algorithm on face authentication,this thesis makes an experiment.Experimental results show that the proposed single-stage multi-level face detection algorithm has good detection performance.The experimental results show that the factors that affect the performance of face detector mainly come from two aspects;The first is the feature extraction ability of the model;The second is how to use effective features to complete the task of face detection.