Weather-Robust Foreign Object Detection on Transmission Lines via Physics-Driven Complex Wavelet Unrolling.

Area of Science:

• Computer Vision
• Robotics
• Signal Processing

Background:

• Unmanned aerial vehicle (UAV) grid inspection is crucial for infrastructure monitoring.
• Adverse weather conditions like haze and rain severely degrade visual data, hindering foreign object detection.
• Existing methods fail to distinguish targets from weather noise, causing performance drops.

Purpose of the Study:

• To develop a robust foreign object detection system for UAVs operating in adverse weather.
• To overcome the limitations of current detection methods in visually degraded environments.

Main Methods:

• Proposed the Physics-Prior Complex Wavelet Unrolling Decoupling Module (PCW-UDM) for robust detection.
• Utilized 2D dual-tree complex wavelet transform (2D-DTCWT) to decouple features into sub-bands.
• Introduced Physics-Guided Low-Frequency Dehazing (PGLD) and LISTA-Unrolled High-Frequency Deraining (LUHD) branches.
• Implemented a spatio-temporal cross-domain consistency loss for weather-invariant feature alignment.

Main Results:

• The PCW-UDM effectively decouples degraded features into low-frequency and multi-directional high-frequency sub-bands.
• The PGLD branch physically reverses atmospheric scattering for dehazing.
• The LUHD branch removes rain streaks using deep unrolled sparse optimization without distorting structural phase.
• Achieved state-of-the-art detection precision and robustness on synthetic and real-world datasets (RTTS).

Conclusions:

• The PCW-UDM-equipped network overcomes semantic conflicts in traditional cascaded restoration-detection paradigms.
• The proposed method demonstrates superior performance in extreme weather conditions for UAV foreign object detection.