Circle Representation Network for Specific Target Detection in Remote Sensing Images

Compared with natural images, remote sensing images have complex backgrounds as well as a variety of targets. The circular and square-like targets are very common in remote sensing images. For such specific targets, it is easy to bring background information when using the traditional bounding box. To address this issue, we propose a Circle Representation Network (CRNet) to detect the circular or square-like targets. We design a special network head to regression radius and it has smaller regression degrees of freedom. Then the bounding circle is proposed to represent the specific targets. Compared to the bounding box, the bounding circle has natural rotational invariance. The CRNet can accurately locate the object while carrying less background information. In order to reasonably evaluate the detection performance, we further propose the circle-IOU to calculate the mAP. The experiments evaluated on NWPU VHR-10 and RSOD datasets show that the proposed method has excellent performance when detecting circular and square-like objects, in which the detection accuracy of storage tanks is improved from 92.1% to 94.4%. Therefore, the CRNet is a simple and efficient detection method for the circular and square-like targets.