Research On Unsupervised Domain Adaptation For Image Classification

Image classification,as one of the core tasks in the field of computer vision,aims to accurately classify objects in images into different categories.However,due to distribution discrepancy between different domains,a model trained on the labeled source domain by traditional methods will face severe performance degradation on unlabeled target domain.To address this problems,we delve into three setting of unsupervised domain adaptation.First,in the case where a single source domain and target domain can be accessed simultaneously,that is,a standard unsupervised domain adaptation scenario,this dissertation proposes a series of methods,including the method based on adaptive mutual learning,and the method based on homeomorphism alignment.Second,the labeled source domain may collect from multiple distribution,so a method for multi-source unsupervised domain adaptation is proposed based on disentanglement then reconstruction.Finally,this dissertation further studies a scenario that requires dealing with data privacy issues,called source-free unsupervised domain adaptation,where source domain data is inaccessible,and the target domain has no labels.In this context,this dissertation proposes a method based on class prototype mining.The main work and contributions of this dissertation are as follows:(1)This dissertation proposes two methods based on mutual learning,which optimizes the unsupervised domain adaptation image classification based on bi-classifier adversarial learning and mean teacher framework respectively.For the method based on bi-classifier adversarial learning,the target domain data is divided into two parts,each of which represents those better predicted by the two classifiers respectively.One classifier maximizes the prediction discrepancy against the other classifier on the subset where it yields better prediction to find the target samples out of the source distribution and minimizes the prediction discrepancy on another subset where it yields worse prediction to yields better prediction and make consistent prediction to reduce the ambiguous target samples.For the method based on the mean teacher framework,the target data also needs to be divided into two subsets according to the prediction of the two networks.Then,we introduce role selection strategy to dynamically set teacher network and student network.Finally,we use the traditional knowledge distillation strategy to update the student network,and we propose the reverse knowledge distillation strategy to update the teacher network,instead of the traditional exponential moving average method,so that both networks can obtain the more discriminative prediction.(2)This dissertation proposes homeomorphism alignment for image classification based on unsupervised domain adaptation.Recent studies have found that directly aligning the distribution between source and target domains tends to destroy the discriminative information of data.To solve this problem,a direct method is to introduce bijection to realize the data transformation between two domains.However,ordinary bijection cannot preserve the topology structure of the map data.Therefore,a more strict bijection,i.e.homeomorphism mapping,is introduced in this work.Our method can be divided into three main steps: firstly,a homeomorphism map is constructed? secondly,train the homeomorphism map by sewing up the transformed feature with original feature? finally,the feature extractor and classifier are trained based on the property of homeomorphism mapping.Extensive experiments show that the proposed method can further improve the performance of the unsupervised domain adaptation method based on distributed alignment strategy.(3)This dissertation proposes a method based on disentanglement then reconstruction,aiming to solve the problem of image classification based on multi-source unsupervised domain adaption.The method first disentangles source features and target features through two disentanglers to obtain domain-invariant features and domain-specific features.At the same time,for these two types of features,we can estimate inherent category prototypes and domain prototypes.Subsequently,we utilize the disentangled domaininvariant features and domain-specific features to train a reconstructor which map them back source features or target features.Through this reconstructor,we are able to build class prototypes for each independent domain by inputting all inherent category prototypes and corresponding domain prototypes.Finally,we require source features and target features to be close to the corresponding class prototypes to make the features more compact and retrain the feature extractor which makes feature extractor more discriminative.Theoretical proof shows that our method can align the distribution between target domain and source domains twice in a iteration,thereby achieving superior classification results.(4)This dissertation proposes a method based on class prototype discovery to solve the problem of image classification based on source data-free unsupervised domain adaptation.The traditional methods usually generate pseudo-source data or prototypes when the source data cannot be accessed,while the generation task itself is a very complicated,and the noise is easily introduced.Instead,our method uses target data and pre-trained source models to build a set of class prototypes.Based on the constructed class prototype,our method can also implement self-supervised training by performing the clustering algorithm to pse-label target data.At the same time,the target features are also aligned to the class prototype to achieve distribution alignment.Therefore,this work transforms the problem of source data-free unsupervised domain adaptation into the problem of discover such class prototypes,which avoids the complicated process of source data or prototype generation and improves the robustness and feasibility.Through our method,we can effectively deal with the challenge of image classification based on source data-free unsupervised domain adaptation,and contribute a novel approach to source data-free unsupervised domain adaptation.To sum up,this dissertation aims to solve the challenges of image classification based on unsupervised domain adaptation.Through in-depth research on unsupervised domain adaptation,multi-source unsupervised domain adaptation and source data-free unsupervised domain adaptation,we propose a series of innovative methods,including adaptive mutual learning,homeomorphism alignment strategy,disentanglement then reconstruction strategy and class prototype discovery strategy.These methods perform well in different scenarios,whether they have access to source data or not.By effectively solving the distribution discrepancy between the source domain and the target domain,the proposed method in this dissertation has improved the generalization performance and classification accuracy of the model and achieved remarkable results in image classification based on unsupervised domain adaptation.These innovative ideas and methods provide a useful reference for future domain adaptation research,and provide a feasible solution to the problem of image classification based on unsupervised domain adaptation in practical applications.