A robust small-object detection model for UAV aerial imagery under complex background clutter

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Detecting small targets in Unmanned Aerial Vehicle (UAV) imagery presents significant challenges due to scale variations and scene complexity.
• Existing object detection models often struggle with fine-grained feature extraction necessary for small object identification.

Purpose of the Study:

• To enhance object detection accuracy for small targets in UAV imagery.
• To develop an improved model, PF-DETR, based on RT-DETR, that balances performance and efficiency.

Main Methods:

• Introduced a P2 detection head to extend the feature pyramid to finer scales for improved shallow feature utilization.
• Developed a Pyramidal Hierarchical Frequency-Domain Fusion (PHF) module combining wavelet pooling and frequency-domain attention for multi-scale feature fusion.
• Restructured the backbone network using a lightweight BasicBlock_FasterNet_Rep module, integrating FasterNet and RepVGG-style re-parameterization for reduced complexity and enhanced feature extraction.

Main Results:

• PF-DETR achieved a 5.4% increase in mAP@0.5 on the VisDrone2019 dataset compared to the original RT-DETR model.
• Model parameters were reduced by 25.0%, indicating improved model efficiency.
• A 35.0% increase in computational cost was observed, offset by substantial accuracy gains.

Conclusions:

• The enhanced PF-DETR model demonstrates superior performance in detecting small and multi-scale targets within complex UAV imagery.
• The integration of novel modules and network restructuring offers a robust and efficient solution for aerial object detection challenges.
• The study highlights a favorable trade-off between increased computational cost and enhanced detection accuracy, making it suitable for practical applications.