Face Frontalization For Social Network Users

Nowadays,the amount and the content of images on social networks continue to grow.However,researches on social networks has focused on the analysis of text data streams.But,most of them ignore the rich first-hand pictures uploaded by users.According to statistics,the pictures uploaded by users are rich and there are a large number of face images in it,but most of them are non-frontal faces.This paper takes face information as a starting point and does the following work:Firstly,we got all the pictures in the 120 stars' weibo album by spider.Then we build a dataset to test the accuracy and effect of the generator by semi-manual way.Secondly,to complete the pre-processing work on the picture,ResNet-50 is used to improve the PyramidBox model to make the face detection work better.By this way,we can extract face information from user pictures effectively,and reduce the impact of the accuracy and precision of face detection on face frontalization and face recognition.To reduce manual recognition workload,we use face++ API to detect the pose of faces in the dataset.Thirdly,as the face pose changes,the accuracy of face recognition gradually decreases.To make use of the large number of non-frontal faces in face images,this paper adds the symmetry loss function to improve the generation ability to face frontalization,and improves the recognition accuracy of large-pose face,which makes it possible to use large-pose face as a face tag for social network users.Fourthly,FaceNet is used for face recognition.DBSCAN is used for face clustering.Then we build a frontal face tag library for social network users,improve the efficiency and accuracy of face retrieval and analyze the relationship between users.Through the above work,this paper improves the accuracy and precision of face detection,improves the ability of generator to face frontalization and the accuracy of face recognition.So,we can build a frontal face tag library to users and analyze the relationship between them.