Research On Key Problems And Methods Of Visual Domain Adaptation

With the advent of the era of big data,computer vision has been greatly developed.The success of many models depends on large amounts of labeled data.However,in reality,the world contains numerous new scenarios and the model needs to face scenes with different perspectives,different backgrounds and different lighting.In this case,many state-of-the-art models may suffer from performance degradation sharply.For training the new model,it requires a large number of labeled samples.On one hand,it costs a lot of human resources.On the other hand,retraining model will cause waste of time and hardware.Considering that most of the data or tasks are correlative between each other,our thesis leverages domain adaptation to share the existed or learned model parameters to the new model in some way to avoid learning from zero.Domain adaptation can help deal with new situations beyond the limitation of task,and it is a powerful driving force in machine learning.Domain adaptation is a sub-direction of transfer learning,which relaxes the constraint that the distribution of training data and test data must obey independent identical distribution in traditional machine learning.However,due to the discrepancies between samples and domains,negative transfer and under adaptation are always the big challenge throughout the domain adaptation and they are named as domain mismatch in our thesis.First of all,due to the feature differences between the inter domain and the imbalance between the categories of source domain,the phenomenon of bias will occur in the learned feature subspace.Secondly,in different tasks,the discrepancy between two domains is different.When the difference between domains is small,domain adaptation can solve the problem,otherwise,it usually cannot be well solved.Thirdly,unsupervised domain adaptation is a scenario which is most consistent with real world.Under this setting,model bias may occur due to the lack of target labels.In our thesis,the problem of domain mismatch will be reconsidered,and the expected error boundary of target domain samples will be improved based on the theory involved in the feature space bias?distribution discrepancy between inter-domain and model bias respectively according to the transfer learning generalization error upper theory from the aspects of feature matching reconstruction,domain generation and sample level alignment.The main work of this thesis is summarized as follows:(1)In order to improve the issues of negative transfer and under adaptation caused by feature space bias,two domain adaptation methods based on subspace reconstruction are designed from the perspective of feature matching reconstruction.(1)Subspace learning and reconstruction have been extensively studied during recent years in transfer learning.However,existing domain adaptation algorithms ignore class-wise priors,especially in the case of scarcity of some class-wise data,and it will make the model biased.Focus on this problem,we design a kind of class-specific reconstruction transfer learning to align the source and target domain.(2)Considering that most previous subspace projections are linear projections,inspired by deep neural network,nonlinear projections are introduced instead of linear space.Therefore,a reconstruction method based on neural network nonlinear prediction is proposed,and a simple and effective domain adaptation is realized by learning a set of hierarchical nonlinear subspace projection and transfer matrices jointly.(2)Based on the correlation between source domain and target domain,the existed domain adaptation theory requires that the discrepancy between domains should not be too large.However,the difference between domains cannot be controlled in practical application.When large difference exists between domains,reconstruction learning does not always work well.In order to avoid this dilemma,two methods are proposed from the perspective of generative learning.(1)In order to avoid domain mismatches caused by large differences between domains,in our thesis,auxiliary intermediate domain is considered to be generated so that the difference between the generated intermediate domain and the target domain can be controlled theoretically as far as possible.Inspired by the manifold hypothesis,a generative discrepancy metric is proposed.i.e.,manifold criterion.Guided by this criterion,an intermediate domain generative method based on local generative discrepancy metric is proposed,i.e.,Manifold Criterion guided Transfer Learning.Thus the generated domain has the same distribution as the true target domain.(2)In order to alleviate the problem of large differences between domains,besides introducing the generative method to assist the subspace projection,inspired by GANs,our thesis proposes a shallow semi-supervised adversarial domain adaptive network from the perspective of generatve network.The model is named as coupled adversarial transfer domain adaptation and it generates domain aligned features for domain adaptation.(3)Deep domain adaptive method is the mainstream and it usually adopts unsupervised setting.The lack of target labels is one of the main reasons of domain mismatch.In order to alleviate the model bias problem,combining the marginal distribution and conditional distribution,three domain adaptation models are designed from the perspective of class level alignment.(1)Inspired by mutual learning,a strong-weak classifier domain adaptive method is proposed.Through mutual distillation,a deep domain adaptation architecture is proposed to regularize the softmax classifier by the K nearest neighbor(k NN)classifier from the perspective of regularization.(2)The existing domain adaptation methods mainly consider the marginal distribution,and the joint distribution of features and categories is not well aligned.In order to solve this problem,a self-adaptive reweighted adversarial domain adaptation method is proposed,which attempts to enhance domain alignment from the perspective of conditional distribution.(3)Inspired by triplet loss,and considering the importance difference between sample pairs,a Bayesian perspective triplet loss is deduced from the Bayesian perspective.It self-adjusts the weight of intra-domain and inter-domain pair-wise samples and promotes the model training of unsupervised domain adaptation.