Research On Unsupervised Domain Adaptation Algorithm For Image Classification

In recent years,transfer learning has received much research and attention as an important approach to solve the distribution bias of data sets and few-shot problem.As the main branch of transfer learning research,domain adaptation can mine domain invariant features and structures between different but related domains to achieve knowledge transfer and model reuse.The existing domain adaptation methods still suffer from underfitting of feature distribution,insufficient domain invariant feature representation,and confusion of domainspecific features.Focusing on these issues,this thesis takes image classification as the research context to study the domain adaptation problem under the lack of labeling of the target domain.The idea of feature transformation and feature weighting is mainly utilized,and some improvements are made to the structure and training process of the network.The main work of the thesis includes:Ⅰ.A domain adaptation method based on deep fuzzy maximum mean discrepancy is proposed.Most existing domain adaptive methods simply select all sample features for alignment and do not consider the quality of the sample features.In unsupervised scenarios,this may lead to mis-matching problems.The two ideas of sample weight migration and feature transformation migration are combined in this thesis.The features of the same class of samples on the source and target domains are assigned different weights according to the confidence level of the classifier in the predicted values of the samples.And based on the metric of maximum mean difference,the fuzzy maximum mean difference distance is proposed.By optimizing this distance,the matching degree of samples between domains is enhanced.The classification accuracy of the samples in the target domain was improved by 1.7%on the digits dataset.Ⅱ.A domain adaptation method based on class template matching is proposed.In aligning inter-domain feature distributions,most existing domain adaptation methods perform random matching at the individual level,which is easily disturbed by abnormal samples and the training process is highly volatile and longperiod.Based on the retained sample feature information from each batch,this thesis weights the expected values of the same category of sample features in the feature space to obtain a class template,and establishes a method for non-pointto-point alignment of class templates between domains.In an unsupervised scenario,each target sample can obtain a pseudo label of its category through a source domain classifier,and thus its corresponding class template can be calculated.The class templates reflect the domain invariant features of the class,so aligning the class templates of the source and target domains will reduce the differences in feature distribution between domains.At the same time,the proposed method reduces the computational effort by 60%compared to the maximum mean distance,prompting the model to converge faster.Ⅲ.A domain adaptation method based on domain-specific features is proposed.Although the domain adaptive approach can maximize the alignment of feature distributions across domains,it largely ignores some unique features owned by the domain,and a single classifier forcing the alignment of domainspecific features will confuse domain-invariant features.In this thesis,a new target domain classifier is added to the method,and the classifier corresponding to the sample features from different sources is selected for training.In order to train the target classifier,a new clustering algorithm is proposed to further improve its accuracy based on pseudo-labeling.The classifier’s ability to classify the target domain samples is improved by secondary training of the target domain features.Improved accuracy by 3.7%on the ImageCLEF-DA dataset.The method is easily embedded in existing domain adaptation models and has the advantage of further improving its domain adaptive performance.In summary,this thesis proposes solutions for the different characteristics of the research subjects in the domain adaptation process.And experimental validation and result analysis are conducted on several datasets,which provides ideas for the subsequent research.