Research On Open-set Domain Adaptation Image Classification Based On Adversarial Learning And Subdomain Alignment

In recent years,deep neural networks have achieved great success in many fields.But deep network models need to be trained with a large number of labeled samples.On the one hand,it is costly to manually label datasets directly;on the other hand,some existing labeled datasets with the same category space cannot be directly applied to other tasks because they do not satisfy the IID condition.In order to solve the above problems,researchers propose a domain adaptation algorithm,which transfers the knowledge learned in source domain to target domain.It improves the performance of the model on target domain by reducing the inter-domain deviation.The goal of this paper is open-set domain adaptation image classification,which not only needs to solve the problem of inter-domain differences,but also needs to identify unknown samples in target domain.The main work of this paper is as follows:(1)Open-set domain adaptation image classification based on adversarial learning.First,based on the model of open set domain adaptation by backpropagation,the feature extractor introduces a multi-pooling channel attention module to strengthen the feature mapping of key channels;then,in order to reduce the distribution difference between the known categories of source domain and target domain,the domain discriminator based on the adversarial idea is used to select known samples from target domain to participate in domain adversarial training to realize the extraction of the invariant features of the model domain;finally,aiming at the problem of feature discriminability reduction in the process of adversarial domain adaptation,the discriminability of features is improved by balancing the size of the singular value of the feature matrix,thereby improving the classification accuracy and domain adaptation effect.(2)Open-set domain adaptation image classification based on subdomain alignment.First,for the disadvantage of using a fixed threshold to train unknown categories in the model of open set domain adaptation by backpropagation,the adaptive weight module is used to train unknown categories to reduce the possibility of incorrect alignment caused by the fixed threshold;then,aiming at the problem that the global feature distribution alignment leads to the loss of category difference information,a category-based sub-domain alignment strategy is adopted,which effectively reduces the distribution difference of the same category between source domain and target domain and improves the image classification performance.(3)In order to verify the performance of the model proposed in this paper,various experiments are carried out on commonly used domain adaptation datasets(Office-31,Vis DA-2017)and compared with other open-set domain adaptation models.The experimental results show that the proposed model outperforms other models in open-set domain adaptation image classification.