Research And Implementation Of Face Recognition System Based On Deep Learning

With the continuous progress of society,Artificial Intelligence(AI)research has developed rapidly,and image recognition has also been fully developed.In image recognition technology,face recognition is noninvasive,easy to obtain,easy to store,high precision and robust compared to other bio-identification technology such as fingerprints,irises,and voice.It has been widely used in various fields such as public security,identification,and entertainment applications,it has become one of the hottest topics in image recognition technology.At present,most face recognition systems are mainly used in fixed scenes with less interference and targets for static images.This paper mainly studies the problem of face recognition in video streams,analyzes the limitations of traditional face recognition methods in video streams,proposes and improves the face recognition method based on deep learning.The identification of faces in the video stream is mainly carried out in two steps: First locate and detect faces through video data,then recognize the detected faces.In the face detection stage,this paper uses a combination of deep U-Net structure and residual block structure to detect faces.This method does not need to set the default detection frame.It is an Anchor-Free face detection method.In the face recognition stage,this paper first extracts the face features,and then recognizes the face by calculating the similarity of the face feature vectors.In this paper,by adding a residual network structure to the deep U-Net network framework,the face feature extraction is realized.U-Net is a feature extraction method based on fully convolutional image segmentation,using the middle layer of the network as the input face image feature vector,which has a comprehensive representation.The face recognition proposed in this paper has achieved good results.The main innovations are as follows:(1)In the face detection stage,an algorithm based on deep residual network is proposed.Compared with other face detection methods,the algorithm is an Anchor-Free method with high accuracy and low overhead.(2)In the face feature extraction stage,based on the fully convolutional networks feature of the deep residual network,this paper implements the encoding and decoding of face images,and extracts the middle layer as the face feature vector.This method does not need to manually set a feature vector extractor,and the feature vector contains almost all face image information,which is comprehensive.