The Study Of Image Classification Based On Regularization And Deep Transfer Learning

Image classification,as an important research topic in machine learning,has been widely used in many fields.At the same time,the universal digitization of daily life makes the amount of image data increase rapidly,which makes the research of image classification more and more valuable and practical.With the continuous development of deep learning,the accuracy of image classification has been greatly improved.But its model training relies on a large number of label data,and the training data set and test/application data set should follow the same statistical characteristics to achieve the desired effect.In practical applications,it is difficult or costly to obtain labeled data.But transfer learning can use the knowledge contained in a small amount of labeled data or expired labeled data to solve the problem of lack of labeled data that belonging to similar fields with differences.Convolution or Deep convolutional neural network is usually used in the field of image classification,whose model training process uses the minimum loss function to adjust the model parameters,improper loss function may cause over fitting problem.In this context,this thesis studies the image classification problem based on regularization and deep transfer learning.Following are the main contents.(1)Research on image classification based on improved canonical algorithm.The deep learning network VGGNet was used to study image classification.We add residual module unit to the embedded layer during the process of feature extraction.And the network structure is improved by identity mapping,which deepens the embedded network depth and enables it to acquire more high-level and rich image features.And we add regularizer to loss function to increase the distance between classes and decrease the distance within classes,so to reduce the probability of misclassification.Experimental results on Mini Image Net and Omniglot datasets show that the proposed model performs well in image classification tasks.(2)Research on image classification based on transfer learning and kernel sparse coding.In order to reduce the difference of data distribution between source domain and target domain,we add discriminant constraints to sparse coding to solve the problem of sparse representation of source domain and target domain data in kernel space.Meanwhile the constraints reduce the distribution difference between the source domain and target domain by minimizing the maximum mean difference of image sparse coding between them.And the constraints also maximize the correlation of similar image sparse coding in the two domains.The model learns the common nonlinear dictionary and sparse coding basing on mapping the nonlinear images in source domain and target domain to high-dimensional space.Experimental results on six datasets show that the proposed method is effective in image classification.Finally,we give the summary and prospect,provide ideas for further research on image classification through deep learning and transfer learning.