Research on Domain Adaptive Methods Based on Self-Supervised Learning

In order to solve the problem of training data without labels in a single domain, self-supervised learning tries to train the model by setting auxiliary tasks and providing false labels to the samples. Finally, fine tuning the trained model on the downstream tasks can achieve better results. In this paper, the problem of unlabeled target domain samples in unsupervised domain adaptive is regarded as a self-supervised learning problem, and rotation prediction is proposed as an auxiliary task in the target domain and rotation Angle label is used to provide pseudo-supervised signals for the target domain samples. At the same time, this paper fuses the batch kernel norm maximization algorithm with the target domain sample prediction consistency loss, and proposes the consistency batch kernel norm maximization algorithm. In this paper, visual converter is selected as the basic architecture of the model, and experiments are carried out on three data sets of Office-31, Office-Home and VisDA-2017, and compared with the latest and best algorithms, all of which achieve the best results.