Abstract
To address the challenges in drone aerial image target detection-including the loss of critical information on small objects during multiple subsampling operations, the disappearance of minute target features, and insufficient detection accuracy due to dense occlusion interference-we propose PPM-YOLOv11, an improved target detection algorithm based on YOLOv11n. The C3K2_PPA module integrates parallelized patch-aware attention with the C3K2 backbone network to better preserve critical information on small objects. A multi-scale detection head P2 specifically designed for detecting ultra-small objects ranging from 4 × 4 to 8 × 8 pixels is introduced. A high-resolution feature layer is added to the neck network to enhance detection accuracy with respect to ultra-small objects from a drone's perspective. Adding the MultiSEAM module to the neck network enhances detection of occluded small objects by amplifying feature responses in unobstructed regions and compensating for information loss in occluded areas. Experiments on VisDrone2019 and SIMD datasets demonstrate our algorithm achieves a 40.9% mAP50 on VisDrone2019, surpassing the baseline YOLOv11n by 9.3 percentage points. On the SIMD dataset, the mAP50 reached 82.0%, surpassing the baseline network by 3.9 percentage points.