Research And Application Of Image Classification Method Based On Deep Convolutional Neural Network

Image classification is an important research direction of computer vision technology and refers to the image processing method which can realize different categories by acquiring the salient features of images.The traditional image classification method is realized by feature descriptor and classifier.The accuracy of this method depends on the validity of feature extraction.With the development of computer hardware performance,large data and learning algorithms,deep learning methods are widely used in various fields of image processing.Convolutional neural network(CNN)is a deep feed-forward neural network model using the idea of deep learning,which shows good performance in many fields such as speech recognition,face recognition,motion analysis,medical diagnosis and so on.CNN generally consists of convolutional layer,pooling layer,full connected layer and output layer.Compared to traditional artificial neural network,CNN has the characteristics of sparse connection and weights sharing.It simulates the characteristics of the local information of the visual neurons in response to the image.The sparse connection is used to construct the local perceptual field of the image in order to greatly reduce the size of the neural network parameters.By weights sharing,each volume of the convolutional layer is repeated in the whole field,and the local features of images are extracted,which reduces the number of free parameters and thus improves the training efficiency of the model.The image classification method based on CNN avoids the complicated feature extraction.CNN can combine the feature analysis of images into the neural network and realize a true end-to-end image classification by adjusting weights and bias.To design the basic structure of CNN model is a simple process,while the optimization and training of the model is a long and complicated process.Therefore,the research of optimization methods and training efficiency of CNN model is critical.Based on the basic structure of CNN,this paper proposes a set of procedures which make use of optimization methods i.e.dataset expansion,precision enhancement,over-fitting solution and efficiency improvement successively.Through the comparison experiments,it is verified that the diversity enhancement of datasets can improve the accuracy and generalization ability.And in the case of invariable dataset,the scale and depth of network models is improved to promote the accuracy of model classification.Problems of over-fitting for large-scale network models can also be solved by using regularization,sparse optimization,data distribution optimization and other methods.And during training large-scale models and data,optimizing the gradient descent method can effectively improve the convergence speed of the network,while using GPU acceleration will greatly improve the training efficiency.As for the sky cloud image classification at the meteorological industry foundation,a nine-layers CNN model is established and optimized by BN algorithm and dropout method.Trained on the public cloud dataset of the Chinese Academy of Meteorological Sciences,the CNN model,which can achieve an optimal classification accuracy of 97.8%on the test dataset,has better classification accuracy,less parameters and more efficient compared with the classical model Alexnet and GoogLeNet inception v3.With the use of transfer learning,the CNN model is applied to the classification of dataset of self-made all sky cloud instrument.Experiment show that transfer learning can make the network model achieve a rapid convergence in new datasets and a better classification performance in the same training time.After being trained on two types of dataset,the precision can reach to 95.5%on the mixed testsets.Furthermore,the CNN model has a better classification result on the self-made instrument and better adaptability on the size of image and sampling methods of instruments.According to the classification of cigarette cut tobacco composition in the quality inspection industry,the characteristics of each cut tobacco component are studied,and the local characteristic images are established to set up the classified dataset.The cut tobacco identification model is designed based on CNN.The CNN with 11 layers is used in the classification of local characteristic images.In the training process of CNN,the optimal prediction accuracy is up to 93.23%on the local image test dataset,by using multi-scale segmentation method to improve the applicability of the model,using data enhancement method to improve the diversity of the dataset,using the regularization method to enhance generalization ability of the model.In the test process of cut tobacco samples,pretreatment based on image size matching method and the result of CNN output decision method based on statistical analysis are designed.By use of above methods,the recognition model which does not completely depend on the CNN performance has much more tolerance on generalization ability of CNN.The prediction accuracy on cut tobacco training and testing samples can be up to 100%and 98.75% respectively.