DAGE-YOLO: improved YOLO for oriented object detection in remote sensing imagery
TL;DR
Though this DAGE-YOLO model introduces a moderate increase in the computing overhead, it achieves strong detection accuracy and robustness, offering an effective solution for complex OOD tasks in remote sensing.
Abstract
Remote sensing images are often featured by small object sizes, varying rotation angles, and background interference, which makes oriented object detection (OOD) from these images challenging. Here, an improved YOLO11s model, termed DAGE-YOLO, is proposed to resolve this challenge. Specifically, parallelized patch-aware attention (PPA) and efficient multi-scale attention (EMA) are incorporated into the backbone of the original YOLO11s to preserve features of small objects and improve cross-scale context modeling. Next, the conventional loss is displaced by the Gaussian Wasserstein Distance (GWD) loss. Moreover, an oriented bounding box (OBB) detection head is introduced to allow the model to achieve precise oriented box localization. Experiments on the DOTAv1.5 and HRSC2016 datasets show that DAGEYOLO achieves an mAP50 8.8% and 7.1% higher than that of the baseline model on the two datasets, respectively, and outperforms several mainstream YOLO-based OOD models. Though this DAGE-YOLO model introduces a moderate increase in the computing overhead, it achieves strong detection accuracy and robustness, offering an effective solution for complex OOD tasks in remote sensing.