DPCNet: A dual path cross perception network for small object detection in UAV imagery.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Small object detection in UAV imagery faces challenges from tiny scales, dense arrangements, and cluttered backgrounds.
• These factors degrade fine details and destabilize feature representations across multiple scales.

Purpose of the Study:

• To introduce DPCNet, a novel single-stage detector designed for robust small object detection in challenging UAV scenarios.
• To enhance the accuracy and efficiency of detecting small, dense, and occluded targets in aerial imagery.

Main Methods:

• DPCNet employs a dual-path cross perception mechanism separating detail and semantic streams with gated fusion.
• It integrates deep and shallow feature interaction using dynamic sampling and similarity-guided masking for cross-scale consistency.
• A decoupled detection head separates classification and regression with cross-branch guidance, utilizing a geometry-sensitive Shape-IoU loss for bounding-box regression.

Main Results:

• Experiments on VisDrone2019 and HIT-UAV datasets demonstrated significant improvements over the YOLO11n baseline.
• DPCNet achieved mAP@0.5 gains of 2.0% and 5.1%, respectively, showing enhanced precision and recall.
• Performance improvements were particularly notable for small, dense, low-light, and occluded targets.

Conclusions:

• DPCNet offers a compact and robust solution for small object detection in UAV imagery, despite minor computational overhead.
• The proposed method effectively preserves edge details while enriching contextual information through its dual-path design.
• Parameter count reduction by approximately 45% indicates an efficient model architecture.