Research On Depth Metric Learning Algorithm In Domain Adaptation

In recent years,it is expensive or even infeasible to tag samples for large-scale database.The performance of deep model can be degraded due to insufficient labeled samples during the training.In order to solve this problem,it has been the key method to use the relevant source domain model to improve the performance of target domain without sufficient labeled samples.There exists the domain shift between the source domain and the target domain,the domain adaptation method has been proposed to address this issue.The domain adaptation aims to alleviate the domain shift by joint training of the source and target domain,which can improve the classification performance of the target domain via the supervised leaning.The key of domain adaptation is to domain alignment,as we know,metric learning is used to adjust the distribution of samples in the metric space.The above two methods can be combined together to further improve the domain alignment for domain adaptation,which can enhance the discriminative ability of the model to identify the confused categories.In this thesis,a domain adaptive method based on deep metric learning is proposed.The basic idea is to introduce the triple training of deep metric learning on the basis of domain adversarial generation model.By increasing enough margin for different classes in the source domain in metric space,the margin between similar categories in the target domain can be increased indirectly.Therefore,the better classification performance of the target domain is guaranteed.The procedure of our proposed method involves the following steps.First,we construct an adversarial domain adaptation networks to initially align the data distribution of the source domain and the target domain through domain adversarial training.Furthermore,through the distribution of the label space of the target domain samples,we can identify the easily-confused categories in the target domain.Finally,for these easily-confused categories,the margin in triplet loss in metric learning is utilized to separate the corresponding categories of the source domain in the metric space.Thus,these easily-confused categories in the target domain are separated from each other indirectly,which can obviously improve the classification performance of the target domain.Meanwhile,the experimental results on Office-31,Image CLEF-DA and Office-Home datasets further verify the effectiveness of our proposed method.