Multi-step Domain Adaption Image Classification Network via Attention Mechanism and Multi-level Feature Alignment

Domain adaption image classification uses source domain task knowledge to enhance the classification effect of target domain tasks. It can reduce the work of data labeling in the target domain and significantly improve the network's adaptive ability. However, the existing methods can not deal well with the performance degradation in large differences in data distribution between the source domain and the target domain. Therefore, based on the idea of reverse gradient layer, a multi-step domain adaptive image classification network with an attention mechanism aligned with multi-level features is proposed in this paper. Specifically, firstly, the attention mechanism is used to combine the source domain and target domain data to generate an intermediate domain. Secondly, several feature alignment strategies are proposed to align the source domain and the target domain from the pixel and global levels.