Research On Image Classification Based On Deep Learning With Its Application On Hyperspectral Image Classification

With the development of Internet technology, image recognition and classification technology has been widely used in various fields. The multilayer structure of deep learning network enables it to learn the characteristics of the deeper level, and improve the accuracy of image recognition and classification. In this paper, the theory and application of the two kinds of methods of deep learning are studied which are deep belief network(DBN) and convolutional neural network(CNN).Image feature extraction is the key of image classification. This paper analyzes the existing feature extraction algorithms and the existing problems, then compares them with the methods of deep learning. This paper describes the development process and achievements of deep learning in detail, and makes a detailed study on the construction process and training method of ANN, DBN and CNN are lucubrated. By comparing the experimental results,we find that the method of deep learning is superior to the traditional neural network in image recognition. The DBN is applied to the classification experiments of hyperspectral image recognition. The operation of three kinds of data preprocessing is carried out, including sample expansion, data reduction and reconstruction input. The dimension of data were reduced by utilizing AE and PCA,and reconstructed using spatial correlation of data. Compared with other methods, the classification method of DBN can obtain better classification results than traditional solutions, such as SVM, and simplify the classification process. In this paper, the performance and potential of CNN in the classification of natural scene images are studied. Some improvements are made on the basis of the CNN model, including the expansion of the edge of each layer, the adjustment of the step size, changing the size of the convolution window and the sampling window, using Relu function,,and the addition of the dropout layer, exploring the best effect of the CNN classification model through comparative experiments.The classification results of CNN algorithm and other algorithms on cifar-10 data set are compared and verified, and the advantages and disadvantages of CNN in the classification of natural scene images were analyzed.