Subdomain Adaption Network Combining Cosine Distance and Angle Margin

The goal of unsupervised domain adaptation is to transfer knowledge from the source domain to the target domain. Previous adaptive methods have focused on aligning the global distributions of the source and target domains. In this paper, we introduce a subclass domain adaptive network (CASAN) that integrates the Large Margin Cosine Loss and Additive Angular Margin Loss to enhance domain-adaptive classification in scenarios where image quality varies significantly across different domains. In order to optimize the subdomain distribution matching problem, a new metric, cosine distance based subclass domain distribution matching loss, is proposed to better align the features of different domains in the high-dimensional space. Experiments are conducted on three benchmark datasets to compare our model with other mainstream models, and the results achieve higher accuracy.