Remote Sensing Image Object Detection Based on Oriented Bounding Box and Yolov5

Target detection on remote sensing images is an important task of data interpretation. Because the targets in remote sensing image are usually arranged in any direction, and the size varies greatly, this makes general target detection algorithms have good detection result in natural scenes, but it is not ideal in remote sensing images. Aiming at the problems of arbitrary arrangement and multi-scale targets, a rotation-based network with oriented bounding box was proposed based on the single-stage detection network YOLOv5. Firstly, an angle prediction regressor is added to the output layer of YOLOv5, which contains an aspect ratio perception weight function to solve the sudden change of angle loss caused by square-like targets. Secondly, BiFPN (weighted bidirectional feature pyramid network) structure is used to achieve simple and fast multi-scale feature fusion. Experimental results show that the proposed algorithm achieves an accuracy of 76.62% on DOTA dataset, which has a higher accuracy compared with other advanced algorithms. And the algorithm significantly improved the detection accuracy of square targets.