Study On Image Recognition Technologies Based On Unsupervised Transfer Learning

The recognition accuracy has far surpassed that of humans.However,many machine learning algorithms need to be built on a strict assumption: training data and future test data must belong to the same feature space and have the same distribution(independent identical distribution,i.i.d.).Once the distribution changes,most statistical methods need to be retrained and new data also need to be collected.In many practical application scenarios,recollecting data will consume a lot of manpower and material resources.Once the environment changes again,the newly collected data will become useless again.In tasks where data sets are difficult to obtain,this way is even more unrealistic.Therefore,in this case,research on technologies for knowledge transfer between two domains becomes critical.Transfer learning(TL)has been proposed to solve the above problem.TL aims to use existing data or models to assist the learning of data or models of interest and extract useful "knowledge" in the source domain to the target domain.In the field of image recognition,transfer learning can also be regarded as a domain adaptation,which assumes that the source and target domains have the same tasks,but the distribution of the two domains is different.Domain adaptation can be generally divided into supervised and unsupervised scenarios.The supervised domain adaptation indicates that the target domain has a part of labeled data,while the unsupervised domain adaptation indicates that the target domain does not have any label information.This article focuses on the unsupervised setting,because it is closer to the real applications.Existing unsupervised domain adaptation algorithms are often single-stage methods,which only obtain local optimal solutions,and cannot continue to show the best performances of the models.To this end,this thesis first proposes a two-stage progressive training framework,and simultaneously learns an invariant,discriminative,and domain-agnostic subspace through three guiding terms.Secondly,existing adaptation methods often adopt a pseudolabel strategy,which uses a source classifier to generate the pseudo-label for the target domain.However,they ignore the problem of unreliable distribution alignment caused by false pseudo labels.In view of this,this paper proposes another algorithm to obtain a more reliable distribution alignment while establishing a more sufficient domain difference reduction model.Finally,considering the two fundamental but understudied problems of transfer learning,namely negative transfer and under adaptation,we further propose a domain co-adaptation model to alleviate these two problems.Specifically,this thesis has the following contributions.(1)Guide Subspace Learning(GSL)model.The model learns the invariant,discriminant,and domain-agnostic subspace through three guiding terms,and applies a two-stage progressive training strategy.First,the subspace-guided term reduces the difference between domains by moving the source domain subspace closer to the target domain subspace.Secondly,the data-guided term uses coupled projections to map the data of the two domains to a unified subspace.Finally,in order to improve the discriminability of the subspace,a label-guided term is proposed to predict the labels of the source and target samples.In addition,in order to deal with non-linear domain shift,a non-linear Guide Subspace Learning(NGSL)model is proposed,which utilizes kernel embedding.The contribution of this model mainly consists of three points.1)A model consisting of three guided terms is proposed,and the subspace is learned through a twostage guided learning mechanism.To the best of our knowledge,this is the first work of domain adaptation and transfer learning by formulating guide learning model.2)The proposed method is further extended to the Reproduced Kernel Hilbert Space(RKHS),and a nonlinear guide subspace learning model is proposed.3)By replacing the pseudotarget labels with real labels,our model can easily adapt to supervised and semisupervised settings and degenerate to a "one-stage" method.A large number of experiments show that the proposed model is superior to existing methods.(2)Reliable Domain Adaptation(RDA)model with classifiers competition.Specifically,dual task-classifiers and dual domain-specific projections are introduced to align misclassified and unreliable target samples into reliable samples in an adversarial manner.In addition,domain differences in manifolds and category spaces are eliminated at the same time.The contributions of this model are three-fold.1)A novel reliable domain adaptation model is proposed to solve the problems of insufficient divergence reduction between domains and unreliable distribution alignment.2)A Nonlinear Reliable Domain Adaptation(NRDA)model is proposed.This method can be easily extended to the kernel version to handle nonlinear domain shifts.3)Extensive experiments on many challenging benchmark datasets show that the method can achieve comparable performance to the latest algorithms including shallow and deep learning methods.(3)Domain Co-Adaptation(DCA)model.The dilemma of overfitting(negative transfer)and under-fitting(under-adaptation)is always a basic challenge and understudied problem in transfer learning/domain adaptation.This model will re-examine this tricky issue and propose a more secure transfer learning framework called domain coadaptation model.The framework is essentially a bilateral transfer model with domain correspondence beyond the existing unilateral transfer models.With bilateral coadaptation between domains,the risk of over/under-fitting is therefore largely reduced.Specifically,a symmetrical bilateral transfer(SBT)loss with two terms is proposed under the philosophy of mutual checks and balances.First,each target sample is expressed by the source samples in the common subspace with low-rank constraint,so that the source data with the largest amount of information and transferable data can be retained to avoid negative transfer problems.Second,each source sample is sparsely represented by the target sample symmetrically,so that the most reliable target sample can be used to avoid the problem of under adaptation.Experiments conducted in various cross-domain vision benchmarks show that the performance of the proposed domain collaborative adaptation framework is superior to many recent technologies.