Study Of Image Recognition Algorithm Based On Theano Deep Learning Framework

With the rapid development of internet technology,network information not only on the number of exponential type rise,form is also becoming more diverse.With the impact of these information to our own,we are also beginning to increasing demand for information extraction.But due to the complex and huge number of network information,just rely on the brain to distinguish these information has become increasingly unrealistic.So how to use the computer replace the human brain to identify,screening and extract the useful information has become one of the most popular the Internet research field.Based on neural network structure,the use of the principle of bionics deep learning arises at the historic moment.With the development of text search engine,and search for the image recognition has become the next development direction,how to improve the speed of image recognition and image recognition rate will be related to the practicality and safety of image recognition,is of great significance.Compared to the shallow network,the depth of learning the multi-layer network structure,can learn a deeper level of feature representation,can be more concise way to express complex functions.In this paper,several multi-layer neural network structures in depth learning are applied to image recognition,which can effectively improve the accuracy of image recognition.The specific contents are as follows:(1)This paper introduces the basic principle of MLP(Multilayer Perceptron),introduces the simple structure of multi-layer perceptron,sums up the function images of Sigmoid function and Tanh function through simple perceptron model,and analyzes the multi-layer perception The regression model used in the output layer of the machine model summarizes the cost function of the logistic regression.The corresponding parameters were optimized using the Batch Random Gradient Descent(MSGD,Minibatch Stochastic Gradient Descent)in the experiment.(2)Introduce the principle of Deep Belief Networks,and analyze the model built by the network of deep confidence.Through the pre-training,we get the optimizationeffect of the partial search of the weight space in the DBN network,and the speed of convergence is accelerated,so that the training time is greatly reduced.(3)and experiment with the Mnist handwritten digital set on the Theano library.Through the comparison between the three algorithms of multi-layer perceptron,Boltzmann machine and convolution neural network training,Image recognition,the traditional convolution neural network image recognition rate of the highest,reaching more than 98.91%.Therefore,in the handwritten numeral,the traditional convolution neural network is used to draw the best conclusion.(4)Introduce the development process of traditional convolution neural network of CNN(Convolutional Neural Network),the structure of traditional convolution neural network and the calculation and training process.On the other hand,it is proposed to improve the convolution neural network,introduce the method of dimension reduction preprocessing of principal component analysis(PCA)data,replace the classifier,and apply the improved convolution neural network to olivettifaces face recognition data set,the effects is good.(5)Introduce the basic principle of the Restricted Boltzmann Machine and analyze the solution of the visible and hidden layers of the restricted Boltzmann machine,and obtain the finite Boltzmann machine(RBM)has two advantages over a multi-layer perceptual(MLP)network model with only one layer of hidden layer.On the other hand,it is proposed to improve the Bohrzmann machine algorithm by introducing an iterative method of constantly recursive new values ??with the old values ??of the variables,to re-train the improved Boltzmann machine,After the limited Boltzmann machine algorithm used in the olivettifaces face recognition data set,the image recognition,the effect is good.