Research On Image Classification Optimization Algorithm Of Convolutional Neural Network

Image classification is an important research field of the computer vision.Traditional image classification methods require manual design of feature extraction.So,the accuracy of classification is closely related to the method of feature extraction.With the rapid development of network and multimedia technology,the images that need to be classified is becoming more and more complex and large.It is significant that classify the image information effectively to facilitate the management or the application in other aspects.Nowadays,the accuracy of image classification using deep learning method can reach the accuracy of artificial classification,or even surpass that of human beings.Compared with the traditional image classification methods,the advantage of deep learning method is that the image feature extraction does not need artificial intervention.The image features are extracted through the network training to find the optimal parameters adaptively.There are three methods to further improve the classification accuracy of network:deepening the network,increasing datasets,improving the internal soft algorithm of the network.After deepening the network and increasing the dataset to a certain degree,the accuracy will be difficult to improve again,or even decline.Therefore,it is a better choice to further improve the soft algorithm of the network.In this paper,the internal sofi algorithm is improved based on the convolution layer and pooling layer which are basic structures of convolutional neural network.According to the convolution layer,Deeply Supervised SKNet(DS-SKNet)is proposed.The idea of DSN is introduced into SKNet to construct DS-SKNet,and the features of convolution layer are supervised so that the features learned by convolution layer could be more differentiated.Moreover,network training could be accelerated and classification accuracy could be improved.Experiment is conducted on CIFAR10/CIFAR100 which are standard classification datasets.Experimental results show that the classification accuracy can be further improved after the introduction of deeply supervision.Principal Component Analysis Pool(PCAPool)is proposed for pooling layer.PCAPool is easier to retain information than traditional pooling layer methods which take the maximum value or the average value.The specific process includes generating sample matrix,PCA operation and information weighting operation.In addition,the application of PCAPool in convolutional neural network needs to consider the back propagation.This paper gives a mathematical proof and verifies it on datasets CIFAR10/CIFAR100,MNIST,SVHN and Imagenet2012.Finally,PCAPool improves the classification accuracy on multiple models.Finally,the optimization algorithm is applied to the real dataset.In the process of mining,it is dangerous that the shaft wall is cracked due to the action of stress.It is necessary to classify and identify the borehole wall cracks to facilitate the next step of prevention and remediation.The application process includes obtaining real photos,making datasets,training network,optimizing network and other steps.Experimental results show that the proposed optimization methods are also applicable to real dataset.Figure[54]table[10]reference[102]...