Research On Image Classification Method Based On Convolutional Network

With the development of science and technology,people receive more and more information from the outside world,and images are one of the most common and important information sources in daily life.Compared with other sources of information,images can contain a huge amount of information,and its complexity and redundancy also make it more difficult for people to process image signals.Initially,people were inspired by the text retrieval and modeling,put forward the bag model.Although the appearance of the bag-of-words model replaces the long-term artificial marking classification,the accuracy of image classification has always been a problem for the bag-of-bag model due to the feature vector extraction and the existence of the classifier itself.With the major breakthroughs in biology in the human visual nervous system,researchers hope to identify images by using artificial neural networks to model the human nervous system.In recent years,the research of deep learning model has made new progress in deep neural networks,becoming a hot topic in the field of artificial intelligence.In this paper,image recognition is carried out based on Let Net5 convolutional neural network,and the accuracy of image classification and recognition is improved by studying its network model structure,parameter meaning and network level.Specifically,firstly,the activation function of deep learning network is improved and the convergence speed of network is improved.Secondly,the convolution ability of convolutional network is improved and the over-fitting of network is optimized.Finally,The network model and support vector machine are combined to construct a hybrid model based on deep convolutional network and support vector machine,and the validity of the above method is verified on the general image data set.The main work and innovation of this paper include:(1)The advantages and disadvantages of the common activation functions Sigmod,tanhx,ReLu and Softplus are studied and analyzed.Based on the merits of ReLu and Softplus functions,a new piecewise function is constructed as an activation function.Based on the depth convolution network in Cifar-10 and Caltech-101.This set of images was validated and the effects of various activation functions on the convergence speed and accuracy of the network were analyzed.(2)By introducing the Dropout layer in the network structure,the over-fittingproblem that seriously affects the generalization ability of the network in deep learning has been solved.The Dropout layer mimics the human nervous system by randomly closing the nodes of the convolutional network,thereby preventing overfitting and reducing the dimensions of the data vector and speeding up the computation.(3)In order to further improve the generalization ability of the network,this paper combines the characteristics of convolutional neural networks and support vector machines,a new hybrid model is designed,that is,the convolution neural network is used to extract the features of the input image set,The last layer using support vector machine classification.The experimental results on image database MNIST verify the effectiveness of the proposed method.