The Research On Image Classification With Sample Turning Based On Domain Adaptation

Unsupervised domain adaptation aims to generalize a model from the label-rich source domain to the unlabeled target domain.Existing works mainly focus on aligning the global distribution statistics between source and target domains.However,they neglect distractions from the unexpected noisy samples in domain distribution estimation,leading to domain misalignment or even negative transfer.For closed-set domain adaptation,we find that the unexpected noisy samples are mainly semantic ambiguity and background mess samples.In this paper,we present an importance sampling method for domain adaptation,to measure sample contributions according to their “informative” levels.In particular,informative samples,as well as outliers,can be effectively modeled using feature-norm and prediction entropy of the network.The importance of information is further formulated as the importance sampling losses in features and label spaces.In this way,the proposed model mitigates the noisy outliers while enhancing the important samples during domain alignment.In addition,our model is easy to implement yet effective,and it does not introduce any extra parameters.Extensive experiments on several benchmark datasets show that our method outperforms state-of-the-art methods under both the standard and partial domain adaptation settings.For open-set domain adaptation,we find that the unexpected noisy samples are mainly unknown.In this paper,we present an unknown sampling method for open-set domain adaptation,to dynamically learn the decision boundary between known and unknown samples,based on the similarity between each source domain sample and target domain sample.Adversarial learning is more accurate to measure the similarity between two domains.According to the distance to the saddle point,known and unknown samples can be effectively modeled using adversarial learning.In this way,the model proposed in this paper can dynamically predict the decision boundary between known and unknown,and only perform domain alignment for those shared known samples.Extensive experiments on several benchmark datasets show that our method outperforms state-of-the-art methods under open-set domain adaptation settings.