YOLO-RMS: A Lightweight and Efficient Detector for Object Detection in Remote Sensing

Abstract

Object detection (OD) plays a critical role in interpreting optical remote sensing images (ORSIs) within the field of Earth observation. Despite significant progress has been made in natural scene OD models, applying these detectors directly to remote sensing images has not achieved the expected results due to the complexity of the scenes and the dramatic variation in shape and scale in remote sensing images. A lightweight and efficient detector based on YOLOv8n is proposed to address these challenges. First, we design Reassembly-PAN (RA-PAN) to guide the reassembly of features in multiscale feature fusion. Second, multiscale dilated attention (MSDA) is introduced after the feature fusion module to make the model focus on effective features in complex backgrounds. Finally, Shape-IoU is employed as the bounding box regression loss to make the model focus on the shape and scale of the bounding box itself and improve its localization ability. Experimental results show that the mean average precision is improved by 4.7% compared with the baseline model on the fine-grained optical remote sensing dataset SIMD, while introducing a few additional parametric quantities. In addition, the proposed method improves the mean average accuracy (mAP) on the optical remote sensing dataset DIOR and RSOD by 1.7% and 2.1%, respectively. Better detection performance is achieved compared with other mainstream models.