Research On Face Recognition Algorithm For Surveillance Video And Complex Scenes

With the rapid development of deep learning,the field of computer vision has been well developed in recent years.And as one of the popular research direction,face recognition has made a leap-forward improvement.It has already reached to or even better than the level of human eyes recognition in the test set.And many application products have been realized,such as face recognition gates,unlocking mobile phone by face,face recognition entrance guard,etc.What's more,face recognition is applied to payment scenarios of high security requirements.The face recognition technology for surveillance video and complex scenes still has a large room for improvement,although face recognition has high recognition rate on some test sets.Face recognition technology faces various problems in unconstrained scenes,such as illumination,attitude changes,low resolution,and occlusion.Enriching the diversity of training dataset samples is a solution to improve the robustness of face recognition models.This thesis uses the ResNet framework to train the model with large-scale data set which has been cleaned and integrated and tests in universal test sets,such as BLUFR,YTF,LFW,etc.The main work has been done is as follows:First of all,aiming at the problems of noise and unrichness of sample data in face data set,a method of cleaning and fusion of face databases based on maximum spanning tree is proposed.Calculate similarities between the image and all remaining images in the same class.Exclude noise pictures whose average similaritie are less than the threshold.Calculating deflection angle of human face.Then select standard face of each class to solve the problem of database noise caused by face frame selection strategy based on average similarity and deflection angle.In order to ensure that the weights of the samples in the training set are the same,the same face images are eliminated.Finally,the deduplicated and denoised data sets are merged into a large-scale face data set,and the effectiveness of data set cleaning and fusion is proved by experiments.In order to improve the robustness of face recognition for attitude change,age change and occlusion,this thesis proposes a branch neural network structure.The global features of human face are extracted by backbone network,and the local features of human face are extracted by branch network.The global features and local features of the face are merged to represent the face.So that the network pays attention to both global features while taking into account the local details of the face.For the problem that Angular Softmax Loss training is not easy to converge,it is proposed to use Center Loss optimization method to accelerate the network training process.The effectiveness of the method is verified by experiments.In this thesis,the accuracy of DIR@FAR=1% of BLUFR is improved from 93.93% to 95.41%.The accuracy of LFW,YTF,CFP-FP and AgeDB-30 were improved to 99.60%?95.10%?94.96%?96.02%.Finally,a face recognition system for surveillance video is designed and implemented to test the accuracy of face recognition in actual surveillance scenes.