Face Detection And Recognition In Natural Scenes Based On Deep Learning

With the development of the information society,information security has received more and more attention.Among them,human faces have their own biological characteristics as the identification of different individuals,so they are used in scenarios such as identity recognition and security monitoring.Due to different lighting,different angles,face occlusion and other factors,the problems of face detection and recognition have encountered great challenges.In this thesis,based on deep learning methods,related research on face detection and face recognition is performed.Compared with traditional target detection algorithms,the target detection algorithms based on convolutional neural networks are classified and analyzed,and the research method is proposed.This thesis first proposes multi-face detection in natural scenes based on YOLOv3 algorithm for face detection,which aims to solve real-time face detection and small face problems in pictures.The model uses Darknet 53 as the backbone network,uses 3 different size feature maps to make predictions,designs the center coordinates,confidence level,and category loss function of the detection frame,and finally directly returns the information of the detected face.The batch normalization processing model is used.At the same time,the loss function is designed for the three angles of face recognition probability,face category,and face detection frame,which effectively solves the image background changes and natural conditions such as lighting and occlusion.The real-time detection problem also solves the problem of multi-face detection in the image.This thesis also proposes an improved model based on the FaceNet algorithm.First,the YOLOv3 model in this thesis is used to extract the detection frame after face detection,and then the face image is segmented for model training.In this thesis,the improved Inception module using the SE module is proposed,and the improved network is used as thebackbone of the FaceNet algorithm to extract face features.The learning rate is segmented with constant attenuation,L2 regularization,and Adagrad optimizer is used to optimize the training to obtain the final algorithm model.Based on the Euclidean distance comparison of the face feature vectors,the model obtains the similarity measure of the face and recognizes it.Finally,the test was performed on the LFW dataset.The experimental results show that the algorithm in this thesis effectively improves the face recognition rate,and the recognition rate is improved by nearly 5% compared with the original FaceNet model.