Development Of Face Verification Algorithm On Massive Data Scenario Based On Metrix Learning And Distributed Machine Learning

Face verification has been a spot in the field of computer vision for many years.In recent years,depending on the increasing computer computing performance and training data,in-depth learning has become an important tools of image processing.However,with the rapid growth of face data sets,for large-scale face verification applications,deep learning algorithms encounter problems such as slow model training speed,difficult convergence and difficult deployment of parallel training.To solve these problems,this thesis develops a large-scale face verification algorithm based on metric learning and distributed machine learning,which has good engineering application value.Firstly,a face verification algorithm based on convolution neural network is developed.By comparing and analyzing the computational load,parameters and training time of the mainstream convolution neural network model,a convolution neural network model suitable for large-scale face feature extraction algorithm is designed.The model is used to extract face features,and face comparison is achieved after extracting face features.Then,aiming at the problem of slow training speed and difficult convergence in large-scale face verification application scenarios,this thesis adopts distributed machine learning method,combined with GPU communication technology,develops a multi-machine multi-card parallel model training algorithm based on data parallelism.By increasing the equivalent batch size of training samples and improving the initial learning rate,the number of iterations and training time of convergence of large-scale face comparison algorithm model are reduced,and the difficulty of convergence is solved.In order to solve the problem of large number of parameters in full connection layer of face verification algorithm,this thesis designs an optimization method of aggregation parameters in front of full connection layer,which improves the speed of single iteration and further reduces the training convergence time of large-scale face verification algorithm.Base on above,the model parallelism method is used to distribute the computing parameters of the full connection layer on all parallel computing nodes,which solves the problem that the large-scale face verification algorithm is difficult to deploy training due to the large parameters of the full connection layer.Finally,in order to further improve the result of convolutional neural network in large-scale face verification applications,this thesis develops a large-scale face verification algorithm based on metric learning,and validates it in MegaFace Challenge dataset of millions of scales.The results show that the accuracy rate has been significantly improved.