Research On Face Image Recognition Technology Based On Spark

In recent years,with the continuous development of artificial intelligence,face recognition technology has become a hot issue in the field of computer vision research.It is widely used in various fields in real life such as national defense security,corporate sign-in,financial business and other fields.However,face recognition technology still faces two problems that need to be solved urgently.Firstly,the accuracy of face recognition is greatly affected by the environment,lighting,facial obstructions and people's aging change.Secondly,when faced with massive face images,face recognition has lower training efficiency and slower recognition speed.In this thesis,face recognition is carried out based on residual convolutional neural network,and the results show that residual neural network has better recognition effect than traditional algorithms in face recognition.Firstly,the thesis makes a detailed study of the residual convolution network.Then it puts forward the improvement scheme according to the problems that the activation function of the residual network been prone to gradient disappearing and the limitation of the learning rate setting.The improved scheme can speed up the convergence of the network by introducing the batch normalization(BN)layer before activating the function layer,improve the accuracy of face recognition by optimizing the Softmax loss function into a A-Softmax loss function.In addition,while ensuring that face recognition rates are improved,this thesis studies the Spark distributed parallel computing framework to a certain extent,adopts the technology method based on elastic distributed dataset,combined with the improved residual network for parallel model training and face recognition,and finally realizes the rapid analysis and processing of the face image data of large data volume.At the end of the thesis,according to the improved face recognition algorithm,more than 99% accuracy can be obtained in LFW face data set,and the results are compared with those of other convolution neural network models,which verifies the superior performance of the improved algorithm.What is more,there builds Spark cluster,then parallel training network model,uses distributed environment to carry on face recognition.By comparing with the time consumed in single-machine environment,it can be concluded that with the increase of cluster scale,both training time and recognition efficiency are greatly improved.