Image Classification Algorithm Based On Convolutional Neural Network

With the emergence of large annotational data sets and efficient computer processors,the fild of Deep Learning has made rapid progress.Computer Vision(CV)is one of the active research directions in Deep Learning applications,and image classification as a typical classification task is the basic and important research topic in the field of CV.In the process of image classification,feature extraction plays an important role in improving the accuracy of classification.Therefore,for the task of image classification,Deep Learning has been widely studied and applied because of its powerful ability of feature extraction.Among the algorithms,Convolutional Neural Network(CNN)develops the most rapidly.The performance of CNN is improved at the cost of deeper network,more training parameters and huge consumption of computer resources.However,with the increase of the number of network layers,the parameters of the shallow network can not be updated effectively,resulting in gradient dispersion;more complex network structure and huge resource consumption make it unable to be successfully deployed in mobile devices.In addition,the activation function in CNN can directly affect its performance,and the selection of activation function usually depends on cross validation or experience.Therefore,how to optimize the network structure,reduce the complexity of the model and select high-performance activation function has great research significance and application value.Based on this,we propose a series of improvement strategies,the specific work is as follows:1.Maxout hidden layer is introduced as the activation function of CNN,and three sparse Maxout structures are proposed to solve the problem that Maxout is not sparse.Among them,the new segment correction unit sw Re LU increases the sparsity of Maxout and improves the vanishing gradient phenomenon in the negative numberfield of Re LU rectified unit.In three open datasets,four kinds of Maxout structures are compared in terms of feature sparsity and accuracy.The results show that the three kinds of sparse Maxout structures have strong feature sparsity and high classification accuracy,and the accuracy of Maxout is improved by 1.3%,0.7% and 1.6%respectively by sw Re LU-Maxout.2.In order to reduce the amount of parameters,we optimize the flatten layer and propose the spatial pyramid weighted average pooling(RSPP).On the open dataset,RSPP is compared with spp and flatten in terms of accuracy,space complexity and training time.The experimental results show that RSPP has higher accuracy than spp and lower complexity than flatten.3.We design two kinds of Integrated CNN,namely Bagging_CNN and SE_CNN.The Maxout_CNN based on SVM classifier is constructed to increase the number of base classifiers.Two kinds of integrated convolutional neural networks are compared in two multi classification tasks.The results show that Bagging_CNN performs better when the accuracy of the base classifier is high,while the SE_CNN performs better when the accuracy of the base classifier is low.4.We apply the proposed algorithm to fine-grained image classification task(Flower Recognition).The 32 Flowers database was constructed by using the reptile technology.According to the difficulty of the task,Batch-normalization(BRN)was introduced and the fusion loss function was proposed.A comparative experiment was carried out for Flower Recognition task.The results show that in most cases,the fusion loss function is better than the cross entropy loss function and hinge loss function.BRN can effectively reduce the over fitting phenomenon of the model.Finally,comparing the two integration algorithms,SE_CNN has the highest accuracy of 89.1%.