Research On Cross-domain Image Classification Algorithm Based On Domain Adaptation

In the traditional machine learning classification algorithm,the basic assumption is that the training data and the test data satisfy the same distribution,but in actual scenarios,this assumption is often untenable.If there is a large distribution difference between the training data and the test data,the performance of the classifier trained by the traditional classification algorithm will be reduced.Domain adaptive algorithms can solve the problem of different distribution of samples in the source and target domains,and improve the classification performance.Domain adaptive learning can improve the robustness of the classifier against domain differences,thereby improving the accuracy of cross-domain classification.Aiming at the problem domain adaptation in cross-domain image classification,this thesis has conducted in-depth research.In view of the shortcomings of the existing domain adaptation methods,a better improved algorithm is proposed to adapt to the domain adaptation problem in different environments.The main research contents of this thesis are as follows:1.Aiming at the problem that the source domain sample and the target domain sample do not satisfy the basic assumption of the same distribution,which leads to the failure of the classifier,an unsupervised domain adaptive classifier based on kernel extreme learning machine is proposed.First,use the manifold feature learning method to project data into the manifold space to reduce the data deviation between the source and target domains to a certain extent;then introduce joint feature distribution adaptation and manifold regularization terms to the output layer of the extreme learning machine The weights are constrained to improve the robustness of the extreme learning machine against domain deviations.Finally,the RBF kernel function is introduced to construct a kernel extreme learning machine to eliminate the effects of random input weights and improve the stability of the classification model.Experiments conducted on the benchmark dataset prove that the method in this paper is superior to other latest methods in unsupervised cross-domain visual recognition.2.In the study of traditional domain adaptation methods,it is found that both semisupervised and unsupervised domain adaptation require labeled or unlabeled target domain data to participate in training.In real scenes,the test samples from the target domain are unknown or even unavailable during model training.This problem is called the blind domain adaptation problem.In this thesis,the reconstruction classification network(RCN)is used in blind domain adaptation.Only the source domain samples are used to train the source domain RCN model,and the source domain RCN model is used to reconstruct the pipeline to enhance the information of the target sample and narrow the distribution difference between the target domain and the source domain.The enhanced target samples are classified through the classification pipeline of the source domain RCN model.Experiments conducted on benchmark datasets with multiple features prove that the method in this thesis is superior to other latest methods in blind domain recognition.