Research On Convolutional Neural Network In Augment Reality

The development of deep learning is changing speedily.Based on the general neural network,convolution neural network arises at the historic moment,and has become an efficient recognition algorithm in the field of image.Le Cun designed a classic convolution neural network model structure Le Net-5 that includes multiple levels of convolution and sampling process.In this model,effective features selected will be input into the full link layer to be classified and regressed through training and computing.A typical traditional model is used to improve the The structure of network and performance of convolution,so as to effectively improve the efficiency of the network and the recognition effect of the image.Examining in depth the theories related to convolution neural network and reviewing closely domestic and international researches in this field,this paper has accomplished following work: 1)Build their own picture library,and generate a valid data set which is transmitted through the input layer to the optimized convolutional neural network model so that,the image classification results are obtained through training and calculation;2)I compare the experimental results of four commonly used activation functions in the convolution layer,and improve them,with the linear correction function of Tanh.Relu used to speed up the process of convergence,and then test the data set and verify the results;3)employ the network generalization method to maximize the pool of DROPOUT in the hidden layer,with the effectiveness of the method verified by experimental comparison,and the universality and recognition performance of the experimental model improved;4)In the total training,verify the correctness and effectiveness of this method by experiments,adjust the selection of correlation functions considering their influences on corresponding parameters,and compare and analyze the experimental results;5)Combine the augmented reality system with the improved CNN image recognition algorithm,and successfully develop an optimized iOS-based AR system.