GD-DAMNet: Real-Time UAV-Based Overhead Power-Line Presence Recognition Using a Lightweight Knowledge Distillation...

Area of Science:

• Robotics and Automation
• Computer Vision
• Remote Sensing

Background:

• Unmanned aerial vehicles (UAVs) face significant flight safety risks from power line obstacles in urban and rural environments.
• Existing UAV navigation systems require enhanced obstacle detection capabilities for reliable operation.

Purpose of the Study:

• To develop a novel, real-time power-line presence recognition system for UAVs.
• To improve the safety and stability of UAV flights through advanced detection technology.

Main Methods:

• A specialized UAV power-line dataset was created through data cleaning, enhancement, and segmentation.
• A dual-attention mechanism convolutional neural network (CNN) was developed and trained using SGD optimizer and Hard-Swish activation.
• Knowledge distillation transferred CNN knowledge to a Mamba-enhanced GhostNet v2 for efficient, mobile-deployable real-time recognition.

Main Results:

• The proposed method achieved real-time power-line recognition accuracy exceeding 91.4% across diverse regions.
• The dual-attention mechanism proved effective in enhancing CNN performance for this task.
• Knowledge distillation successfully reduced model parameters while maintaining high accuracy.

Conclusions:

• The developed system provides a robust technical foundation for real-time UAV power-line recognition.
• The improved GhostNet v2 model demonstrates suitability for mobile deployment in UAVs.
• This advancement contributes to safer and more widespread UAV applications in remote sensing.