AeroLight: A Lightweight Architecture with Dynamic Feature Fusion for High-Fidelity Small-Target Detection in Aerial...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Small-target detection in Unmanned Aerial Vehicle (UAV) aerial images is challenging due to low resolution, dense clusters, and cluttered backgrounds.
• Existing methods struggle with resource-constrained environments common in UAV applications.

Purpose of the Study:

• To introduce AeroLight, an efficient detection architecture for high-fidelity small-target detection in UAV aerial images.
• To address limitations in feature representation and localization precision for minute objects.

Main Methods:

• Optimized feature pyramid with a high-resolution head for minute object sensitivity.
• Dynamic Feature Fusion (DFF) module for adaptive multi-scale feature recalibration.
• Shape-IoU loss function for refined bounding box regression of irregular objects.

Main Results:

• AeroLight improved mAP50 by 7.5% and mAP50-95 by 3.3% on the VisDrone2019 dataset.
• Reduced parameter count by 28.8% compared to baseline models.
• Demonstrated superior performance and generalization on RSOD and Huaxing Farm Drone datasets.

Conclusions:

• AeroLight offers a powerful and efficient solution for real-world UAV applications.
• Sets a new standard for lightweight, high-precision object recognition in aerial imaging.
• Enables improved capabilities in surveillance, inspection, and environmental monitoring.