Enhanced photovoltaic panel defect detection via adaptive complementary fusion in YOLO-ACF

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Vision

Background:

• Defect detection in photovoltaic (PV) panels using electroluminescence (EL) images is crucial for quality control.
• Challenges include missed detections and false alarms due to similar features between defects and complex backgrounds.

Purpose of the Study:

• To propose an Adaptive Complementary Fusion (ACF) module to enhance PV panel defect detection.
• To improve detection performance, reduce model size, and accelerate detection speed.

Main Methods:

• Integration of an Adaptive Complementary Fusion (ACF) module into the YOLOv5 object detection framework.
• Training and validation using a dataset of 4500 electroluminescence images of photovoltaic panels.

Main Results:

• YOLO-ACF demonstrated improvements of 5.2% in Recall, 0.8% in mAP50, and 2.3% in mAP50-95 compared to YOLOv8.
• YOLO-ACF achieved a 12.9% reduction in parameters, 12.4% in weight, and 4.2% in time, with a 5% increase in FPS compared to the baseline YOLOv5.

Conclusions:

• The proposed YOLO-ACF method effectively balances detection performance, model complexity, and speed for PV panel defect detection.
• The ACF module shows versatility across various defect types and enhances the overall quality control process for solar panels.