An Attention-Based Bidirectional Feature Fusion Algorithm for Insulator Detection.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Power line insulator defects pose risks to grid reliability, safety, and sustainability.
• Existing You Only Look Once (YOLO) algorithms struggle with complex backgrounds and small defect sizes, leading to detection inaccuracies.
• Accurate insulator defect detection is crucial for effective power line inspection.

Purpose of the Study:

• To develop an improved You Only Look Once version 5 (YOLOv5)-based algorithm for enhanced insulator defect detection.
• To address limitations in existing methods, specifically low detection accuracy in complex scenarios.
• To improve the reliability and safety of power transmission systems through advanced defect identification.

Main Methods:

• Proposed the MC-YOLO algorithm, an enhanced YOLOv5 model incorporating novel modules.
• Introduced a Multi-scale Attention Convolutional Fusion (MACF) module to extract multi-scale features and enhance target representation using attention mechanisms.
• Developed a Cross-Context Feature Fusion Module (CCFM) with a Spatial-Channel Dual Attention Module (SCDAM) for effective fusion of shallow and deep features.

Main Results:

• The MC-YOLO algorithm achieved a mean Average Precision (mAP@0.5) of 67.4% on the test dataset.
• Demonstrated a 4.1% improvement in detection accuracy compared to the original YOLOv5 model.
• While slightly reducing frames per second (FPS), the model maintains practical performance for rapid and accurate defect detection.

Conclusions:

• The MC-YOLO algorithm significantly improves the accuracy of insulator defect detection in power lines.
• The proposed MACF and CCFM modules effectively enhance feature extraction and fusion for better detection performance.
• MC-YOLO offers a practical and accurate solution for maintaining power transmission reliability and safety.