Research On Face Detection And Recognition Based On Deep Learning

Last few years,computer vision has developed rapidly as a newly emerging field,and its important potential commercial value,huge application prospect and academic value make it become a hot and difficult point.It has been widely used in areas such as unity of human credentials,security check and finance,etc.More and more scholars,institutions and enterprises have carried out a great deal of research in relevant fields successively.The high-level semantic features of deep learning have led many scholars to apply it in the field of face detection and recognition.Face recognition tasks have a higher level on deep neural networks,which has exceeded the level of human beings on this particular task.However,people still face the problem of illumination,pose variations,face image quality,occlusion and so on,so they can not locate or recognize the accurate face,which gradually becomes the biggest obstacle of face detection and recognition.This thesis has conducted an in-depth analysis and summary of existing research methods and existing problems.The purpose of this thesis is to extract the robust face features of multi-scale and occlusion,and to detect and recognize the face,mainly as follows:(1)Considering the multi-resolution of face,in this thesis,it proposes a multi-view face feature learning method based on deep learning,which mainly uses two modules of deep learning network to train and learn face features.On the one hand,a large viewing angle module is used to learn rough facial features,and on the other hand,a small viewing angle module is used to learn the details of the human face in order to improve the detection accuracy.Firstly,using the large convolution kernel of a network module to perform face feature learning,that is,coarse learning,which can learn the overall structure of face and get the position of face.Then using the small convolution kernel of another network module to perform feature learning,the output of rough learning can be used as auxiliary information for fine learning to obtain more accurate face features.(2)In this thesis,the problems in feature extraction based on classical VGG neural network are analyzed,that is,the full connection layer is directly used as the final feature,the problem that the feature representation of face extraction with different scales is not very good,based on multi-view face feature learning method,this thesis propose a multi-scale deep learning framework based on facial context features.Through feature layer learning at different Deep Learning network stages,the feature layers of different scales at different stages are normalized and concatenated as the final global feature.At each network stage,further learn in detail,learn the face position,context information and face center feature,as a local feature.Finally,connect global features and local features as the final face feature representation,effectively improving the discriminability and expressivity of face,and the performance of face detection and recognition was eventually improved.Finally,the face detection and recognition methods proposed in this thesis are used to detect and identify experiments on three datasets,respectively,and compared with the existing methods,and prove the effectiveness of the proposed method.In this article of the proposed method,from the experimental results of face detection and recognition,we can see that the proposed method improves the detection and recognition accuracy...