面向多源遥感图像的自适应目标检测

近年来，目标检测已经在含有大量标注的数据上展现出了良好的效果，但当真实测试数据与标注数据存在域间差异时，往往会导致训练好的目标检测模型性能降低。由于相比于自然图像，多源遥感图像在成像方式和分辨率等方面存在特有的差异，而传统的方法需要将多源图像数据重新标注，这将消耗大量人力和时间，因此在遥感图像上实现自适应目标检测面临特有的挑战。针对以上问题，本文提出了一种面向多源遥感图像的自适应目标检测算法，在图像级别和语义级别上对网络进行对抗训练。此外，通过结合超分辨网络，进一步缩小了图像级别的差异，实现了自适应目标检测。本文在两个多源遥感数据集上进行实验，结果表明本文方法有效提升了目标域上的检测效果。 

Adaptive object detection for multi-source remote sensing images

In recent years, object detection has shown great performance with large quantities of labeled data, but when a domain discrepancy occurs between the real test data and the labeled data, the performance of a trained object detection model often decreases. Compared with natural images, multi-source remote sensing images have unique discrepancies in imaging methods and resolutions. Traditional methods needed to re-label multi-source images, which spent lots of manpower and time. Therefore, it faced unique challenges to implement adaptive object detection for remote sensing images. In view of the above problems, this paper proposed an adaptive object detection algorithm for multi-source remote sensing images, which conducted adversarial training at the image level and semantic level. In addition, by combining super-resolution networks, we further alleviated the discrepancy at the image level and realized adaptive object detection. We conducted experiments on two multi-source remote sensing image datasets, and the results show that the proposed method effectively improves detection performance on the target domain.