Improved YOLOv5 for Remote Sensing Rotating Object Detection

Abstract

Remote sensing image object detection is a challenging task owing to high resolution, dense distribution of small objects, and rotating objects at any angle. In this study, aiming to improve the performance of remote sensing image object detection, we propose a lightweight remote sensing rotating object detection model based on YOLOv5. To solve the problem of detecting rotating objects at any angle, we transform the angle regression problem into a classification problem, making the model predict one more angle information, which is used to rotate the horizontal bounding box to enclose the rotating objects at any angle. Aiming at the dense distribution of small objects in remote sensing images, we improve the detection layer structure of the model and add a object detection scale to pay more attention to small objects. Moreover, we also replace the BottleneckCSP module in the YOLOv5 Backbone network with the GhostBottleneck module to reduce the amount of model parameters and lighten the model. We conducted experiments on the public dataset DOTA. Experimental results show that our proposed model has achieved good detection results and the model size is only 13.4M, which can be effectively used for remote sensing object detection and can be easily deployed to the corresponding platform.