| Convolution neural network(CNN)is a deep learning model,it is generally applied in difficult artificial intelligence tasks such as image recognition and natural language processing.The architecture of CNN is inspired by the researches on the activation of animal's cortex cells.Therefore,CNN is inherently suitable for processing image data.Compared to other neural models,neurons in CNN are partially connected,and connection weights is sharing among neurons.When inputing high dimension data,the number of parameters in CNN is far less than that in the fully connected network,and therefore it is less likely overfitting than fully connected network when the number of training data is small.Besides,since 2-D images can be sent to CNN directly,the structure and location information of the image is retained.The convolution kernels are feature extractors in CNN,they can extract features by sliding upon the image.Therefore,in this thesis,we study the convolution kernel generation algorithms when the image data is given.The structure of this thesis is as follows: Firstly,we introduce deep learning and it's current research situation,the latest open source deep learning frameworks/toolboxes are given.Then we briefly describe neural network and CNN model.In the part of kernel generation algorithms,we describe the back propagation in CNN.Then,two neural network models that can be trained unsupervisedly are described,these two models' feature extractor can by applied to CNN.We also study the PCANet,a kind of CNN that uses principal component as kernels.Finally,we build our own kernel adaptive CNN--Cluster PCANet,by adding a clustering step in PCANet and generating an over-completed dictionary,thus can extract more features.We do image classification experiments on two general image datasets.The results show that our Cluster PCANet has lower error rates than that of PCANet.Also,our Cluster PCANet has comparable results with other deep neural models. |
PDF Full Text Request