GDEIM-SF: A Lightweight UAV Detection Framework Coupling Dehazing and Low-Light Enhancement

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Image degradation (haze, low illumination, occlusion) hinders vehicle and pedestrian detection in complex traffic.
• Reliable detection is crucial for Unmanned Aerial Vehicle (UAV)-based visual perception tasks.

Purpose of the Study:

• To develop an aerial vision framework that integrates multi-level image enhancement with a lightweight detection architecture.
• To improve the reliability and efficiency of object detection in adverse imaging conditions.

Main Methods:

• A cascaded "dehazing + enhancement" module preprocesses images, restoring details and enhancing structural fidelity in low-light regions.
• A lightweight detection architecture, GDEIM-SF, combines GoldYOLO backbone with D-FINE anchor-free decoder.
• Incorporation of CAPR and ASF modules for enhanced edge modeling and multi-scale semantic alignment.

Main Results:

• The proposed method achieved 2.5-2.7 percentage point improvements in mAP@50-90 on the VisDrone dataset compared to similar lightweight models.
• Maintained a low parameter count and computational overhead, demonstrating efficiency.
• Achieved a balanced trade-off between detection accuracy, inference speed, and deployment adaptability.

Conclusions:

• The aerial vision framework offers a practical and efficient solution for UAV-based visual perception in challenging environments.
• The integrated approach effectively addresses image degradation issues for improved object detection.
• The lightweight yet robust design facilitates deployment adaptability for real-world applications.