R-YOLOv5: a lightweight rotation detector for remote sensing

Authors:

Abstract
Recently, detection algorithms for objects with arbitrary orientation have been greatly developed, but traditional object detection algorithms use horizontal boxes to label objects in images, which will lead to the problems such as loss of accuracy and excessive background information. In this paper, a rotation detection algorithm based on YOLOv5 is proposed, which can effectively detects the targets in the remote sensing image. It uses the CSL method to mark the oblique targets, converts the angle of the rotation boxes into a classification problem, and avoids the boundary problem. We also optimize the loss function to speed up the network convergence and adopt the NMS algorithm based on the rotation boxes, which effectively removes the redundant prediction frame. The experimental results demostrate that compared with the original YOLOv5 algorithm, the improved YOLOv5 algorithm can better locate rotation objects in remote sensing images and effectively prevent occlusion problems in dense target scenes. With a guaranteed inference speed of 24.1fps, Our method has a detection accuracy of 72.65% on dataset DOTA 1.0.