Area-based non-maximum suppression algorithm for multi-object fault detection.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Unmanned aerial vehicle (UAV) photography is crucial for power system inspection, but automated fault detection remains challenging.
• Conventional algorithms struggle with simultaneous processing of diverse objects on power transmission lines.
• Existing deep learning object detection methods face limitations with the non-maximum suppression (NMS) algorithm when objects have multiple labels.

Purpose of the Study:

• To develop an improved object detection method for automated fault detection in power line inspections.
• To address the challenge of redundant annotations caused by multi-labeled objects (e.g., insulators, dampers) using traditional NMS.
• To enhance the accuracy and efficiency of detecting faults in aerial imagery.

Main Methods:

• Proposed an area-based non-maximum suppression (A-NMS) algorithm to resolve multi-label object detection issues.
• Integrated A-NMS within a cropping detection framework for enhanced small object detection.
• Utilized deep learning-based object detection techniques for aerial image analysis.

Main Results:

• The A-NMS algorithm effectively handles objects with multiple labels, such as insulators and dampers.
• The combined A-NMS and cropping detection approach achieved high performance metrics.
• Demonstrated a mean average precision of 88.58% and a recall of 91.23% on aerial image datasets.

Conclusions:

• The proposed A-NMS algorithm significantly improves multi-object fault detection in aerial images.
• This method offers a robust solution for automated power system inspection using UAVs.
• The findings pave the way for more reliable and efficient power line monitoring systems.