Object Detection of Remote Sensing Image Based on Multi-level Domain Adaption

Due to the limited amount of labeled data, remote sensing object detection is faced with great difficulties. The problem of insufficient labeled data usually can be solved by domain adaption. However, current methods mostly focus on feature alignment, without paying attention to the context information or discussing the level of features, which makes it impossible to effectively apply them to remote sensing images. In this paper, we construct our method based on Faster-RCNN model, and design three domain adaptive components for remote sensing object detection at image-level, instance-level, and pixel-level. Image-level alignment enhances global recognition ability by image weight redistribution. Instance-level alignment makes global awareness possible by combining context information. Pixel-level alignment reduces local differences between domains by focusing on small features and enhancing semantic information. Moreover, we collect domain adaption dataset to verify the proposed method, and the experimental results show that our method is superior to other current methods.