The Research Of Image Classification Algorithm Based On Convolutional Neural Network

Recently,with the rapid development of computer and Internet technology,image data has come into sight and played an important role in our daily life.Because of the explosive increase of the image data,the classification of these data becomes notably necessary while the data content turned to be increasingly complex.Traditional method for image classification no longer meeting the requirement of practical applications.Thusly,the solution to improve the accuracy of image classification under big data is of great significance.Convolutional Neural Network(CNN)is a novel type of artificial neural network based on two-dimensional image classification.Not only the advantages of traditional neural network,good fault tolerance,adaptability,and self-learning ability can be obtained,but also the ability of automatic feature extraction and the characteristic of weight sharing are fulfilled.During the process of CNN,the network structure is an important factor to affect the accuracy and efficiency of the data classification.And it is essential to pay more attention on the optimization of the CNN network structure.In this paper,we analyzed the basic concepts and improved the algorithms of CNN when compared with classical CNN,two aspects of the work are mainly carried out as follows:(1)Based on the classical PCA network(PCANET)structure,the maxout neural network is introduced before the nonlinear activation function,and the SVM classifier is replaced by the softmax classifier.Moreover,we propose a PCA unsupervised pre-training maxout CNN.In the process of solving the network parameters,adjust the parameters can be easily handled without specific adjusting techniques,the training time is decreased,the convolutional kernel solution needs no iteration,and different image classification tasks can be adapted.The whole process of the network can be divided into five stages:The first stage: PCA unsupervised pre-training is executed onto filter,followed by the convolutional of already--learned filter and image in order to extract the feature of images;The second stages: the extracted features are input to the non-linear activation function Relu via the maxout neural network;The third stages: the output of non-linear activation function is binarized,so that a new feature map can be obtained;The fourth stages: new feature map is applied to count block histogram and input the columns quantized block histogram to the fully connected layer;The fifth stages: softmax classifier to classification is functioned.As for handwritten MNIST,as well as the deformable database and CIFAR-10 database,the experimental results show that the classification accuracy of PCA unsupervised pre-training maxout CNN is improved to some extent.(2)Based on the classical Network in Network(NIN)network structure,the input image pixels are reconstructed,and a multi-path CNN based on bilateral filtering is constructed.The network reduces the loss of texture and shape information of foreground objects in the process of complex image feature extraction.Two paths are input into the CNN,one path is input with the original image,another path is input with the image that after the pixel reconstruction of the original image.Two paths extract features independently.Then,the feature vectors extracted from the two paths are merged together after the last pooling layer and finally transferred to the softmax classifier for the classification process.On the natural CIFAR-100 image database,we analyze the complexity of the image and the learning curve under different complexity of image.It can be concluded that data-missing of foreground object sampling texture and shape information leads to the classification mistake during the feature-vector extraction process of convolutional layer and pooling layer.On the basis of CIFAR-10 and CIFAR-100 database,the experimental results show that the accuracy of image classification is better than that of traditional single path CNN.