Small object detection in aerial imagery remains a significant challenge in computer vision due to low resolution, occlusion, and scale variation. To address this, we propose SODM-YOLOv9, a novel model specifically designed for accurate and efficient small-object detection. The proposed architecture refines the YOLOv9 backbone by implementing 16 \(\times \) downsampling to reduce redundant high-level features and removes auxiliary branches in the neck to streamline the network. Additionally, selective fusion of intermediate feature layers (P2, P3, P4) is employed to preserve critical small object details. To further enhance performance, we integrate Ghost convolution modules for lightweight efficiency. We also incorporate a separated and enhanced attention module (SEAM) to improve feature representation of subtle targets and replace the original CIOU loss with WIoU v3 to refine bounding box regression. Extensive experiments on the VisDrone2019 dataset demonstrate the effectiveness of our approach. SODM-YOLOv9 achieves a mean Average Precision (mAP@0.5) of 50.7%, outperforming the baseline YOLOv9c by a significant margin of 7.3 percentage points. The model also shows superior results in precision (60.5%), recall (47.7%), and overall mAP@0.5:0.95, while simultaneously reducing detection latency to 7.9 ms and model size to 83.7 MB. On the TinyPerson dataset, the proposed model achieves an mAP@0.5 of 23.4%, demonstrating its effectiveness in detecting extremely small objects. These results indicate that SODM-YOLOv9 effectively improves detection performance in challenging aerial scenarios, making it suitable for applications such as aerial surveillance, autonomous systems, and remote sensing. This work provides a basis for further research in efficient small-object detection.
