Research Of Unsupervised Domain Adaptation Method Based Discrepancy In Category Distribution

In the era of big data,the performance of deep learning has been greatly improved due to the rapid growth of data volume and continuous improvement of computing power.However,a lot of energy has been spent on manual annotation before model training,which gives rise to transfer learning(TL).Transfer learning aims to help target learning with knowledge from relevant source domain,and unsupervised domain adaptation(UDA)is an important research direction in transfer learning.UDA usually reduces the distribution distance between domains by discrepancy-based or adversarial learning to learn domain-invariant features.However,there may also be differences in the category distribution between domains: 1)each domain usually suffers from class imbalance and different domains may have different or even completely opposite class imbalance ratios.Direct domain alignment may cause misclassification of target samples,leading to negative transfer phenomenon.2)The label set in the source domain may be a subset of that in the target domain.There are target categories that do not exist in the source domain.So it is necessary to correctly classify known categories in the target domain and identify unknown categories.Threshold discrimination is a commonly-used method for unknown class recognition,but it is difficult to describe the confidence of samples and find the appropriate threshold in the absence of target domain labels.As a result,it cannot learn the decision boundary between known and unknown classes effectively.For the first point,this paper proposes a novel UDA method for bi-imbalance scenario called TITo K,by transferring Imbalance-Tolerant Knowledge across domains.First,a class contrastive loss is presented to transfer class contrast knowledge across domains to reduce the influence of imbalanced class distribution to the model.Meanwhile,class correlation knowledge is transferred between domains,and thus the class relationship in the source domain is also transferred to the target domain.Finally,discriminative feature alignment is developed for a more robust classifier boundary.Experiments on several real data sets show that TITo K can effectively improve the learning performance of imbalanced UDA.For the second point,this paper proposes an open-set domain adaption method CRMUD based on contrasting among classes and rectifying by clustering.Both multi-class classifier and open-set classifier are trained to distinguish whether a sample belongs to a known class.At the same time,selfsupervised clustering is adopted to correct the incorrect labels,so as to classify known target sample to the correct classes and identify unknown target samples as well.Experiments on real image datasets demonstrate the effectiveness of the proposed algorithm with reasonable evaluation criteria.