Enhanced YOLOv11n for small object detection in UAV imagery: higher accuracy with fewer parameters

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Object detection in Unmanned Aerial Vehicle (UAV) imagery is hindered by high-altitude capture challenges.
• These include densely packed targets, a high proportion of small objects, and limited onboard processing power, resulting in poor accuracy and increased false/missed detections.

Purpose of the Study:

• To propose an improved YOLOv11 model addressing the limitations of object detection in UAV imagery.
• Enhance detection accuracy, particularly for small and densely distributed objects, while optimizing computational efficiency.

Main Methods:

• Designed a Multiscale Edge-Feature Adaptive Selection (MSEAF) module in the backbone to handle small objects and weak edge information.
• Reconstructed the neck with ScalCat and Scal3DC modules, incorporating a P2 small-object detection head for better high-resolution information utilization.
• Developed a shared, reparameterized lightweight detection head (SRepD) to reduce computational redundancy and improve feature fusion.

Main Results:

• On VisDrone2019, the model improved mAP50 by 4.6% and Precision by 4.6% over YOLOv11n, with an 8.5% parameter reduction.
• On TinyPerson, mAP50 and Precision increased by 5.5% and 5.6% respectively, with a 7.7% parameter reduction.
• Outperformed YOLOv11s on mAP50 (3.8% gain) and Precision (3.2% gain) using only 25% of its parameters, and showed superior performance on the HazyDet dataset.

Conclusions:

• The proposed improved YOLOv11 model effectively addresses key challenges in UAV object detection.
• The integration of MSEAF, ScalCat, Scal3DC, P2 head, and SRepD modules leads to significant improvements in accuracy and efficiency.
• The model demonstrates superior performance across various datasets, including those with small objects and hazy conditions, highlighting its practical applicability.