C2-YOLO: Rotating Object Detection Network for Remote Sensing Images with Complex Backgrounds

In remote sensing images, the background is complex, the distribution is dense, the target scale varies widely, there are many tiny targets, and the target directions are diverse, which is a challenging detection task. This paper proposes a remote sensing image rotating target detection network (C
²
-YOLO) that integrates the upsampling feature enhancement module and attention mechanism in response to these problems. The network is based on the YOLOV5 target detection algorithm, and the prediction head is added to enhance the ability of small target detection. The content-Aware ReAssembly of Features(CARAFE) module is introduced to design a new feature fusion module. We also integrate the Coordinate Attention (CA) module to focus on object locations in complex scenes. According to the rotation characteristics, we add an angle loss to the loss function to detect the rotation angle of the object. We conduct experiments on two public datasets, DOTA and HRSC2016. Compared with the original YOLOv5 algorithm, the detection accuracy of our algorithm is improved by 2.99% and 3.52%, which can achieve comparable performance to the state-of-the-art detection methods.