RDA-YOLO: A robust dynamic adaptive network for tiny insulator defect detection.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Smart grid operation relies on effective insulator defect detection for safety.
• Drone-based image inspection offers cost-effective and accurate solutions.
• Existing You Only Look Once (YOLO) methods struggle with small defects in complex backgrounds.

Purpose of the Study:

• To develop a high-precision insulator defect detection algorithm (RDA-YOLO) addressing limitations of current YOLO models.
• To improve the detection of small insulator defects against complex backgrounds in smart grids.
• To enhance the robustness of defect detection algorithms for various environmental conditions.

Main Methods:

• Building upon the YOLOv8 algorithm, RDA-YOLO incorporates a reverse large-selection kernel module for enhanced feature extraction.
• A Dynamic Head with a unified attention mechanism is employed to improve classification and localization features.
• A distribution-aware Wise-IoU metric models bounding boxes as Gaussian distributions to boost small target detection.

Main Results:

• RDA-YOLO achieved 91.6% precision and 91.4% mAP0.5 on a proprietary dataset.
• The algorithm demonstrates superior performance compared to other state-of-the-art methods with minimal computational overhead increase.
• Extensive robustness experiments confirmed significantly enhanced performance compared to baseline models.

Conclusions:

• RDA-YOLO offers a high-precision solution for insulator defect detection in smart grids, particularly for small defects and challenging conditions.
• The proposed algorithm enhances feature extraction, classification, and localization capabilities.
• RDA-YOLO demonstrates improved robustness, making it suitable for real-world applications including extreme weather scenarios.