Face Recognition In Internet Videos Based On Deep Learning

With the rapid development of the Internet and smart phones,the number of Internet videos is growing geometrically.For the needs of Internet video audit,Internet video face retrieval,video intelligent classification and recommendation,there is a demand for face recognition method in Internet videos.Although the current face recognition system has excellent performance in relatively limited circumstances,it is often affected by rich attitude changes,serious image blur and the second creation of Internet users in real Internet videos.Aiming at these problems and difficulties,this paper improves the face recognition algorithm based on the deep learning method.The specific contents and contributions are as follows:(1)In view of the difficulties of face image in Internet videos,such as multi angles,multi poses,blur and so on,a comprehensive model framework IVFRNet(Internet Video Face Recognition Net)is proposed based on the deep learning algorithm.This framework improves the training data,network structure and loss function of deep learning algorithm.Firstly,this paper improves the training data.Considering the lack of large-scale Internet video face data set at present,this paper proposes a variety of methods to manually simulate Internet video face image by using static face image.Static data and simulation data are both used for training to enhance the adaptability and generalization of the model to different kinds of low-quality images.Secondly,this paper improves the network structure of Inception-Res Net,adds a new Inception module,and fine tunes the parameters of model to improve the learning ability of the network.Finally,this paper introduces the center loss function based on the traditional softmax loss function to enhance the representation ability of the features learned by the model.The recognition rate of the original model is 87.5%,and the recognition rate of the improved model is 93.6%.Experimental results show that the improvement proposed in this paper can effectively improve the performance of face recognition algorithm in Internet videos.And the performance of our algorithm is better than the current mainstream algorithms.(2)In view of the characteristics of Internet videos,a face quality assessment net Face Qnet is proposed based on deep learning algorithm.Before face recognition,the quality of face images are evaluated.And then uses the recognition result of highquality image as the recognition result of whole continuous frames,in order to improve the overall recognition rate.The specific work is divided into three parts.Firstly,use deep learning algorithm to evaluate the quality of face images.By using the powerful learning ability and feature representation ability of deep learning algorithm,an overall quality score including many factors is output.By adding branches to the network,the network can learn more abundant and comprehensive features,and enhance its evaluation performance.Secondly,this paper proposes an automatic annotation method to label the training samples based on face recognition.The method is to use the recognition algorithm to calculate the cosine similarity of the image to be calibrated and the reference image.The cosine similarity value is taken as the mass fraction.Finally,this paper adds quality assessment network to face recognition algorithm,and conducts experiments on self-made multi-frame internet video face test set.The recognition accuracy is improved from 75.1% to 90.4%.The results show that the proposed method can effectively improve the accuracy of face recognition algorithm in Internet videos.At the same time,the experiment is designed to compare the method proposed in this paper with the current mainstream method,and the results show that the method proposed in this paper is superior to the current mainstream method in both recognition rate and calculation speed.(3)Based on the face recognition method and quality evaluation method proposed in this paper,a Internet video face retrieval system is constructed.Realize the function of target face retrieval in input video.The validity of the proposed system is verified by testing in the real Internet video test set.