Research And Implementation Of Multi-view Face Recognition Based On Deep Learning

Under the controlled environment,face recognition has advantages of non-mandatory,easy to access,easy to accept.And it has been widely applied in the field of public security.However,most of the images collected in the practical are profile faces,which leads to a significant reduction in the accuracy of face recognition and meet the practical application requirements difficultly.Therefore,multi-view face recognition has become a research hotspot.This dissertation,based on deep learning,mainly studies multi-view face detection and recognition,designs and implements the multi-view face detection and recognition software.The specific work is as follows:In terms of face detection,two methods which are based on ResNet and DenseNet respectively are proposed.In the stage of feature extraction,the first method uses the ResNet-50 network as feature extractor.It adopts the skip-connection to ensure the stability at the training stage while increasing the network depth.The second method uses the DenseNet-201 network as feature extractor,which uses the dense connectivity to encourage feature reuse.In the face detection stage,the anchor box is used to predict the face bounding box.And the face detection is performed by using different scales and aspect ratios anchors,so that the network can accurately predict the faces of different sizes.The simulation experiments are based on the CelebA and FDDB face datasets.The experimental results show that the YOLOV2 face detection models based on ResNet-50 and DenseNet-201 are better than YOLOV2 on the CelebA dataset.The recall rate of the models increases by 3% and 5% respectively,and the average accuracy reach 90.8% and 90.9% respectively.On the FDDB dataset,we evaluate the performance of our method on the protocol using ROC curves.The results show that the proposed method has better performance.In the aspect of multi-view face recognition,a multi-view face recognition model based on SE-ResNet-20 is proposed.The model uses SE-ResNet-20 network as the face feature extractor which applies SE block to ResNet-20.It recalibrates channel-wise feature to increase model's sensitivity to informative features and to suppress less useful ones.Using ArcFace as a loss function to train the model and introducing the angular margin on the basis of the cosine space.The ArcFace can enhance the discriminative power of face features,simultaneously enlarging the inter-class differences and reducing the intra-class variations.The simulation results indicate that compared with the ResNet-20,the proposed method improves the average accuracy of multi-view face verification by 1.9%.In terms of the design and implementation of multi-view face detection and recognition software,the software is designed and implemented based on the platform of Pycharm and PyQt5,combining with Caffe,and DarkNet deep learning framework.The software has functions of limiting user permission,training face detection model,detecting face,comparing the face similarity and so on.