Research On Deep Learning Based Image Classification

The growing image databases are due to the development of computer technology and networks in recent years. So information extraction from a large number of visual image databases has become a hot research point in the field of intelligent visual perception. Therefore, image classification becomes one of the problems to obtain image information.Image classification has a broad application prospects in terms of image retrieval, human computer interaction, intelligent security, UAV platforms and so forth. The thesis focuses on several key techniques in the framework of image classification. The main works are described as follows:1) We analysed the problems of image classification in detail. Those traditional image classification techniques exist several problems; for instance, image features dosen't contain the global structural information, image features can't give a complete index to image information, select and design classifier. The research on these issues provides a valuable reference for future works.2) In order to overcome the defect of feature extraction that image local features can not give a complete index to image structure information, we proposed an image classification algorithm based on nonlinear global coding and spatial pyramid matching. The introduction of global structural information to spatial pyramid matching can avoid the low correct classification rate because of the lack of global structural information. These tests on STL database and other public databases demonstrate the effective of this algotirhm.3) The feature extraction part of traditional image classification methods such as nonlinear global coding always need extract specified local features manually. This way can't represent image information more comprehensive; thereby affect the image classification results. To address this issue, we use sparse autoencoder to do self-learning for image feature, then classify image through convolution neural network which comes from deep learning. Besides, the selection of parameters was analysed and discussed. The self-learning for image feature portion can express image information more fully, while the network structure can express the low abstract feature and high abstract feature at the same time. Experimental results validate this algotithm has a higher classification accuracy while comparison with nonlinear global coding.4) Traditional convolution neural network's input is the original image without taking into accout the multi-scale factor. Therefore, an image classification based on multi-scale convolution neural network is proposed finally. After the pyramid decomposition of input image, each part should be input to convolution neural network for training. Meanwhile, use image pooling method of dimensionality reduction to avoid over-fitting, and finally train a Softmax model for image classification. Detailed experiments on public and self databases demonstrate this multi-scale convolution neural network based image classification method has the best performance. On the one hand, the method takes multi-scale factor of image into consideration. On the other hand, it can express image more complete, including the low and the high abstract features.