SD-YOLO: A Robust and Efficient Object Detector for Aerial Image Detection

Aerial image detection remains challenging due to the varying scales of objects and complex backgrounds. Particularly, when deploying detection algorithms on edge computing platforms like Unmanned Aerial Vehicles (UAVs), it is essential to find out a lightweight network with good trade-off on efficiency and accuracy. Toward this end, we propose a lightweight and robust aerial image object detection network named SD-YOLO in this work, which is built on YOLOv8. Three novel plug-and-play modules: the star operation block (SOB), the dual-path downsampling module with an attention mechanism (DDM), and the low-level detection head (LLDH), are designed in SD-YOLO to enhance its ability to learn discriminative features and recognize tiny targets in aerial images. Experimental results on two public aerial datasets (DOTA v1.0 and VisDrone2019) demonstrate that SD-YOLO achieves superior performance in terms of efficiency and accuracy. Particularly on the VisDrone2019 validation dataset, a lightweight version of SD-YOLO (4.21 MB) outperforms most state-of-the-art methods with larger network structures.