Research On Image Classification Based On Improved Convolution Neural Network

With the rapid development of technologies such as multimedia,network and storage.The use of image data in various industries is increasing,and meanwhile the sources of digital image people can get from are expanding continuously,so how to choose useful images from large-scale image set has become an urgent problem to be solved today.The traditional image classification method can be divided into two steps:feature extraction and classification.The extraction of image features,which can affects the result of image classification,is the important basis for image classification and recognition.After extracting the image features,the classification algorithm is used to classify the images.At present,new image processing techniques and classification methods are proposed one by one,such as deep learning.Convolution neural network(CNN),which is a new type of neural network of deep learning technology,has been widely applied to various image processing tasks.Although CNN has achieved very good results in some image classification tasks,but there are still some problems such as poor classification ability of the model,low accuracy of training and less convergence.Aiming at these problems,this paper proposes the following two improved convolution neural netw ork algorithm models.The details of our research are as follows:(1)A convolution neural network with improved activation function is proposed in this paper.The main improvement is that we combine the LReLU function and the Softplus function together,and propose a new LReLU-Softplus activation function.LReLU function lack of smoothness,resulting in poor average performance of model.Softplus function has output offset issue which will affect the convergence of the network.The improved activation function inherits advantages of LReLU function when conditional variable less than 0 to solve the problem of numerical offset and neuronal death.It also inherits the smoothness of Softplus function when conditional variable more than 0.Therefore,we propose a LReLU-Softplus activation function,which has the advantages of both the LReLU function and the Softplus function,but none of the shortcomings of the two models,to improve the model’s ability to express.(2)We propose a convolution neural network model with improved Fisher criterion.The main improvement is that we use a new inter-class scatter matrix,which is defined as the absolute value of the difference of two square mean values that come from the projection of two classes of samples,to improve Fisher criteria.Then this improved Fisher criteria is used to improve the cost function of CNN.The main idea of the Fisher criterion is to find a best projection axis,on which the projection interval of two types of samples intersect least,to obtain the best classification effect.However,a drawback of the Fisher criterion,be known as "rank limit" with which the number of discriminate information the Fisher criterion receives is limited by the number of categories,affects the performance of recognition.In this paper,we relieve the"rank limit" problem by improving the Fisher criterion.Then,the cost function is modified to construct a convolution neural network model based on the improved Fisher criterion.Finally we achieve the goal of improving the classification accuracy of CNN as well as fast convergence in the case of a small number of training samples.(3)The two methods we proposed in this paper has been experimentally verified respectively on several data sets such as women’s clothing commodity image library,Mnist handwritten digital library,ImageNet image library and Oxford flowers image library.In addition to the comparison that the CNN algorithm based on improved Fisher criterion(IFCNN)did with CNN,SAE,and FCNN algorithms respectively,we also compare the improved LReLU-Softplus activation function to Sigmoid,Tanh,ReLU,LReLUand Softplus function respectively in the experiments.The experimental results show that the improved methods are feasible and more effective overall.