Research Of Parallel Face Recognition Method Based On Deep Learning Open Source Framework

In the current era of information science,artificial intelligence algorithms are increasingly associated with practice,face recognition has many practical applications scenarios in the direction of information security as a key method for object information recognition,such as video surveillance,access control,identity verification,and so on,which is one of the most popular topics in computer vision.However,face recognition still faces two problems.The first problem is that when dealing with complex and similar face images,the recognition accuracy rate will decrease significantly as data amount increases,resulting in failure to achieve many practical application scenarios.The second problem is that when training neural network models based on massive data,the training takes too much time,the efficiency of testing a variety of different models in a short time is low,a single node and model cannot meet the requirement of identifying a large number of input images at the same time.How to reduce the time of the parallel training of neural network models and the time of using the trained parallel recognition model,and comprehensively improve the recognition accuracy,There is a need to come up with innovative algorithms to achieve this goal,then evaluate and test it.First of all,in order to deal with the above problems,this article starts from a typical engineering application scenario,innovatively proposes an iterative complex residual convolutional neural network model structure.The multi-layer network model structure uses a convolutional neural network principle to construct a new composite model,combing various technical algorithm such as filter connection,residual hop,convolution replacement,batch normalization,pool dimension reduction,gradient descent,etc,which ensure extremely high recognition accuracy for a large number of complex image test sets.Secondly,this article is based on Google's deep learning open-source framework TensorFlow and the parallelism underlying principle,combining the loss function Newton second-order iterative update algorithm theory,in the cluster environment of multiple high-performance GPU graphics processing servers,the parameter server is used to save all parameters that need to be updated for the entire model,and the remaining work nodes analyze the batch data asynchronously and submit the gradientsto realize the distributed parallel training algorithm,the innovation of the algorithm lies the application of the Newton's loss function optimization theory in the GPU cluster asynchronous parameter update process,which reduces the time loss of the data bus transmission and network bandwidth occupation,and greatly accelerates the parallel training.Finally,on the basis of model parallel training,Spark's distributed parallel processing data underlying implementation mechanism is studied,an innovative parallel recognition method using its broadcast variables,elastic data sets,and MapReduce technology methods combined with deep learning abstract feature extraction algorithm is proposed to implement the short-time massive face input image high-throughput analysis and processing in a multi-node cluster environment.Experiments were carried out to test the innovative theories and methods proposed in this paper and verify the performance of the system and algorithms.The experimental results of the above algorithms were compared with the original methods,it was proved that the method proposed in this paper can effectively solve the above problems which improves the accuracy of face recognition and has greatly reduced the time of model training and the identification of massive data sets.