Research On Algorithm Of Convolutional Network And Its Application In Image Classification

Deep learning has become a frontier of machine learning in recent years.It brings about major breakthroughs in the field of data mining,natural language processing,recommendation and personalization,and so on.Convolutional network is part and parcel of the deep learning framework.It contains convolutional neural network(CNN)in spatial domain,principal component analysis network(PCANet)in principal component transform domain and scattering wavelet network(ScatNet)in wavelet transform domain.Convolutional network is especially designed to solve the recognition problems of two-dimentional images.It can keep the inner invariance properties of translation,rotation,scale transformation and other transformation when processing images.Therefore,it has strong fault-tolerance ability and feature-learning ability.Although convolutional network has already made a lot of achievements in different domains,there exits some problems deserving further research and study.This thesis proposes new advanced convolutional networks in the basis of classical convolutional networks.The new networks merge the algorithms of moments transform,fractional wavelet trasform,and so on,into the classical convolutional network.Main works list as follows:1)A Moments-Net model.This is a kind of improved network based on principal component analysis network.It integrates some invariant properties of moments with PCANet.By using different moments as the filters of convolutional parts of the nework,different moments-nets can be constructed.Moments-Nets based on different kinds of moments have different performance in classification tasks.To verify the effectiveness of the Moments-Net,some experiments have been done on MPEG7-CE1-B dataset,which is a binary-image dataset having rotating features.The experiments include the impact of parameters in the network,the influence of different moments and the comparison between Moments-Net and classical algorithms.Comparing with other algorithms,for example,algorithms of principal component analysis,classical principal component analysis network or the algorithms directly using the values of moments as data features,Moments-Nets generally have better performance when soving specific classification tasks,proving the rationality and effectiveness of Moments-Net.2)A hybrid network model.This is a kind of improved network based on ScatNet+CNN.It is constructed by merging unsupervised learning algorithms with supervised learning algorithms.Firstly,we combine fractional scattering wavelet network with CNN to get a new hybrid network,which named FrScatNet+CNN.The forward FrScatNet can be regarded as the preprocessing process of the structure and the backwardCNN is used as the classifier to get the final classification results.This hybrid network can combine the advantages of FrScatNet and CNN.Experiments done on CIFAR-10 and CIFAR-100 datasets have shown that FrScatNet+CNN has good performance in image classification tasks.On the other side,to verify the effectiveness of hybrid network,we construct other hybrid networks by using ScatNet as unsupervised learning network in the first layer,and other networks such as CNN and FCN as supervised learning network in the second layer.All the hybrid networks are compared with other classical networks,such as full connected network,VGG,NIN(Network In Network),shallow-layer CNN,deep-layer CNN,and so on,by doing experiments in different magnitudes of the subset of CIFAR-10 dataset.The final results have shown the rationality of the hybrid network model.