Research Of Domain Adaptation In Image Classification Based On Domain Invariant Feature

Image classification integrates technologies in many fields such as image processing,pattern recognition,and machine learning.Traditional machine learning classification models assume that test samples and training samples lie in the same feature space with independent identical distribution.but it is difficult to meet this assumption in real scenarios.Domain adaptation does not require testing and training samples to satisfy the assumption of independent identical distribution.It can mine domain invariant features and structures between different but related domains,so that label information can be transferred and reused between domains.Based on subspace learning,graph embedding,sparse and low-rank constraints to improve the robustness of the model,the thesis applies domain adaptation for the problem of cross-domain image classification.The main contributions of this thesis are summarized as follows:1.In the domain adaptation problem,the source domain and target domain sample distributions are quite different.Traditional subspace learning methods often ignore category information when aligning the overall distribution of the domain.If the original data has insufficient discriminative power,it will be difficult to guarantee the discrimination of samples in the subspace after projection.To alleviate this problem,a semi-supervised domain adaptation via transfer subspace(SSDTS)is proposed.The algorithm makes full use of the prior information of the sample category labels,can obtain the discriminative subspace while fully obtaining the discriminative information contained in the reconstruction matrix,and enhance the discriminative power and robustness of the cross-domain feature expression of the subspace.Experimental results show that the proposed algorithm can obtain more discriminating target subspaces and better classification performance.2.In the domain adaption algorithm,the proximity relationship between the same and different samples should be considered class-by-class,so the process of obtaining the block diagonal structure reconstruction matrix is more complicated.To alleviate this problem,the two stage unsupervised domain adaption Algorithm is proposed.The model includes two stages.In the first stage,the source domain subspace guides the learning of the target domain subspace,so that the two subspaces are close to reduce the domain difference.Then,the label information of the sample is used to impose block diagonal constraints on the reconstruction matrix,which reflects the proximity relationship between the same sample and different samples,which is more conducive to classification.Finally,the label relaxation matrix is introduced to improve the robustness of the model to label noise.In the second stage,the learned target domain subspace is used to update the target domain pseudo-labels,and the updated pseudo-labels are fed back to the first stage.The final classification result is obtained during the two-stage alternate updating process.Experiments on several data sets commonly used in domain adaptation problems have achieved good results.