Research On Deep Learning Based Face Reconginion

With the rapid development of artificial intelligence,how to accurately and effectively identify the users' identities and enhance information security has become an important research topic.Face recognition has many advantages over traditional card recognition,fingerprint recognition and iris recognition.Due to the characteristics of non-contact,non-mandatory,concurrency,as well as easy user acceptance,face recognition has been widely used in education,e-commerce and other fields.Deep learning is one emerging branch of machine learning.Different from the traditional shallow network,the deep learning is inspired by the brain's working mechanism,and constructs the deep network structure and the corresponding training methods.Deep Convolution Neural Networks(DCNN)originates from multi-layer forward network and has become a hotspot in the field of image recognition through gradual development.DCNN relies on deep nonlinear network structure and large-scale training data to realize approximation of complex functions,thus obtaining more essential and robust image features,and effectively improving the effect of subsequent classification and recognition.Recent years have seen leapfrog promotion of the accuracy of face recognition with the introduction of DCNN.Nevertheless,models of training set and network structure vary vastly,thus making each model has its own characteristics.In this regard,the thesis studies a face recognition method based on deep multi-model fusion.By fusing the features gained from multiple face recognition models,and constructing the face recognition classifier by using the compound features of Deep Neural Network(DNN)training,the thesis is able to obtain an improved model that integrates the advantages of multiple models.The main works are as follows:1.To analyze and compare the face recognition algorithm based on Convolution Neural Network and open source.The thesis screens two basic models through experiments.The basic features of the basic models are reduced dimensionally,normalized and merged,and the compound features are obtained as the input of the subsequent DNN.2.To construct DNN based on deep multi-model fusion and train the compound features,so as to obtain an improved model to integrate the advantages of different basic models.3.To further analyze the improved model and design a series of experiments,including different training sets,DNN parameters and basic feature weight.The detailed test data of the basic model and the improved model on the LFW dataset are collected,and the reason for improving the model is explored.In the case of using smaller data sets,the accuracy of taking this method in the face recognition authority test set LFW and YTF is 99.1%and 93.32%,which is 0.57%and 0.52%respectively compared with that of the basic model.Moreover,through the further analysis of the LFW test data,this thesis discusses the effectiveness of the improved model in integrating the advantages of different basic models.