Faces Recognition Of Hybrid Feature Based On Deep Learning

Deep Learning is the new technology of Artificial Neural Network.Deep Learning in several areas has exceeded the historical record.To improve accuracy and shorten the convergence time is still a goal of neural network.Face recognition as a mature application areas of Deep Learning can be used to check effect of the Deep Learning.Convolution Deep Network is one model of Deep Learning.Convolution Deep Network is also the common model of face recognition,so this paper use Convolution Deep network model as the basis for Face Recognition.In recent years,with the development of the network and the rise of social networking,a large number of people face picture is generated.Facing with a large number of samples,usually requires a large network to be trained,this requires a large network computer cluster and parallel algorithms to support,it will increase the barriers to entry Face Recognition.To solve these problems,this paper study Related models and algorithms of Faces Recognition of Hybrid Feature Based on Deep Learning,main research includes the following aspects:In order to improve the convergence rate of the Convolution Network,this paper presents a Hybrid Adaptive Learning Rate Algorithm.Before network has reached oscillation,Hybrid Adaptive Learning Rate Algorithm use a global learning rate which enables fast convergence rate of network,when the network oscillation,reduce the learning rate,the network slowly approaching extreme point to search optimal solution.Experimental results show that Hybrid Adaptive Learning Rate Algorithm can significantly reduce convergence time in advance to achieve the best recognition rate.Facing the real problem,the Deep Network Construction often faces searching better network architecture process.The network architecture include deep convolution network height and width,network height refers to the number of the network layer,network width refers to the number of feature map of each layer.This process requires a number of different training network architecture,to found relatively better network model.This process not only consume time,but also has no enlightening guidance for solving new problems.After comparing the impact of the width and height on performance of Convolution Deep Network,this paper proposes method of calculating the width of the convolution network architecture,in order to resolve the process of Network Construction,which is often time-consuming process.In order to solve the problem of the high barriers to entry of training a large number face image samples Under limited computing platform,this paper presents a method of training multi-network multi-block.Training Methods of multi-network multi-block divide a large samples to some group.The Training Method use fixed network model to train group.Then network parameter set which to identify each group were obtained.When identifying an unknown sample,First we load each set of network parameters,Then each network is searched to identify the sample,and finally select the best case to result.Experimental results show that this training method to solve the difficult of training a large number face image samples Under limited computing platform is effective,and in the LFW face database,experiment gets good results.