YOLOv11n-Ghost-P2: An Accuracy–Efficiency Study of Lightweight Small-Vehicle Detection in Aerial Images

Small-vehicle detection in aerial imagery requires fine spatial detail within limited computational budgets. We propose YOLOv11n-Ghost-P2, which combines Ghost-based feature compression with a high-resolution P2 fusion branch. Relative to standard YOLOv11n, the model reduces parameters by 38.5%, floating-point operations (FLOPs) by 39.7%, and single-patch latency by 47.9%. On the source-scene-disjoint validation partition of the Dataset for Object Detection in Aerial Images (DOTA), mean average precision ([email protected]) is 0.6222 for YOLOv11n and 0.5367 for the proposed model, while strict [email protected]:0.95 is 0.3545 and 0.2953, respectively. On an NVIDIA Tesla T4 platform, the model processes 37.7 patches/s; mean full-scene time across three large images decreases from 2.39 to 1.26 s. A five-seed ablation comparing the dual-Ghost architecture with and without P2 shows the largest [email protected] improvement for Tiny vehicles, with a mean gain of 0.01662. On five additional scenes, tiny-vehicle recall increases from 21/34 to 26/34 with P2. The results characterize the accuracy–efficiency trade-off of allocating part of the compression savings to high-resolution detection.