Research On Image Classification Based On Fisher Dictionary Transfer Learning

In recent years,dictionary learning methods have shown good performance in the fields of feature learning,image classification and image denoising.However,the image classification algorithms based on dictionary learning are based on the assumption that the training data and the test data have the same feature space and the same feature distribution.But in the actual image classification application,different background and different conditions will make the feature distribution change.The traditional dictionary learning algorithms are no longer applicable,and transfer learning is introduced to solve this problem.Transfer learning learns the essential characteristics of samples from relevant domain data and transfers them to target tasks to solve classification problems with different distributions.However,the existing transfer learning ignores the inter-class differences and intra-class similarities of samples,as well as the linear inseparability of data.In order to solve these problems,this paper proposes Discriminative Fisher Embedding Dictionary Transfer Learning(Discriminative Fisher Embedding Dictionary Transfer Learning,DFEDTL)and the Kernel Fisher Dictionary Transfer Learning(Kernel Fisher Dictionary Transfer Learning,KFDTL)for image classification.Regarding the problem of ignoring the inter-class differences and intra-class similarities of samples in cross-domain,the proposed methods combine the source domain information and part of the target domain information,and construct the discriminative Fisher atom embedding model and the coefficient embedding model to promote the transfer learning.It not only retains the differences between the classes and the similarities within the classes of the training samples,but also improves the discriminative ability of learning the dictionary.In order to minimize the distribution difference between the source domain and the target domain and improve the performance of cross-domain classification,an adaptive maximum mean difference(Adaptive Maximum Mean Difference,AMMD)term is constructed using atoms and profiles(row vectors of encoded sparse matrices).For the linearly inseparable problem,the linearly inseparable data is mapped into a high-dimensional space through nonlinear mapping,so that it is linearly separable in the high-dimensional space.Through the study of dictionary learning and transfer learning,a model with better classification effect is proposed.The model solves the problem of classification with different distributions of source and target domains and the problem of data linear inseparability.Here are the innovations of this paper:(1)Fisher criterion is used to construct discriminative Fisher atom embedding model coefficient embedding model between source domain samples and target domain samples,and samples from the same class are encouraged to have similar coding coefficients.(2)The combination of dictionary learning and transfer learning can not only avoid the problem of low classification accuracy caused by the variation difference between images,but also reduce data annotation tasks.(3)The adaptive maximum mean difference(AMMD)term is constructed by using the relationship between dictionary atoms and profiles to reduce the distribution difference between the source domain and the target domain.(4)By using the kernel method,the dataset that cannot be linearly divided in the low-dimensional space is mapped to the high-dimensional space through the nonlinear kernel,so that the data samples are linearly separable in the high-dimensional space,and the image classification problem under the high-dimensional complex nonlinear relationship is effectively handled.