Research Of Image Recognition Based On Convolutional Neural Network

Convolutional Neural Networks(CNN)is a very important method in deep learning and has solved many complicated pattern recognition problems.It has achieved fruitful results in the image recognition,speech recognition and natural language processing.Compared with the traditional pattern recognition method,CNN has the advantage that it can make convolution computation with pixels directly and automatically extract the image feature,which is similar to the brain's visual system processing mechanism.In addition,compared with the traditional neural network,weight-sharing and sparse-connection of convolutional convolution neural and pooling operation greatly reduce the number of training parameters and narrow the feature map size,which simplify network model and improve the training efficiency.This paper introduces the convolution operation,pooling operation,soft-max classifier and network optimization algorithm in convolution neural network.Then,image recognition experiment on the handwritten numbers data sets and face data sets are researched.Finally,the paper proposes a new method based on the local binary pattern and the convolutional neural network for the research of face recognition.The main work in this paper is as follows:1.According to the network structure of LeNet-5,which is commonly used on handwritten numbers data set,we construct other five different network structures and study the effects on the experimental results of the number of convolution kernel,convolution kernel size,dropout,and the depth of network.Experimen-t results show that the larger number of convolution kernel is,the better face recognition performs.The result using the convolution kernel of 5 x 5 is better than that using the convolution kernel of 3 x 3.The dropout after fully-connected layer is helpful for face recognition.The deeper structure of network is,the higher recognition rate is.2.Because the human face image is more complex,so we use more complex network structure to do experiments.The paper constructs the different network structure according to the network structure in the article centerloss,studies the effect of the number of convolution kernel and dimension of feature on the result of the experiment.The experimental results show that the greater the number of convolution kernels and feature dimension is,the higher human face recognition rate is.3.Because CNN ignores the local structure feature of faces when it extracts face features and local binary pattern of faces is local structure feature,so this paper puts forward a kind of face recognition method based on local binary pat-tern(LBP)and CNN.Firstly,LBP feature of face images is extracted,Then,the combination of LBP images and RGB images were input to convolutional neu-ral network,and optimize network parameters using stochastic gradient descent method.At last,the trained CNN model was used to recognize unseen face samples.The experimental results show that the proposed method has a better recognition performance.