Photovoltaic panel defect detection algorithm based on infrared imaging and improved YOLOv8.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Industrial photovoltaic panel defect detection faces challenges including high missed detection rates, complex backgrounds, and unclear defect features.
• Existing methods struggle with uneven difficulty levels in target detection tasks, impacting overall efficiency and reliability.

Purpose of the Study:

• To develop an enhanced infrared detection method for photovoltaic panel defects based on the YOLOv8 model.
• To improve feature extraction, reduce model parameters, and address imbalanced difficulty in defect detection tasks.

Main Methods:

• Integration of a multi-channel squeeze-and-excitation network into the YOLOv8 neck for enhanced feature extraction.
• Incorporation of GhostConv and BoTNet into the backbone network to optimize model parameters and performance.
• Application of the Focaler-Complete Intersection over Union (Focaler-CIoU) loss function to manage detection task difficulty.

Main Results:

• The proposed method achieved significant improvements over baseline YOLOv8 on the PV-Multi-Defect-main dataset: +3.6% precision, +10.4% recall, +4.8% mAP50, and +4.5% mAP50-95.
• On the PVEL-AD dataset, substantial gains were observed for dislocation-type defects: +7.8% precision, +17.1% recall, +19.5% mAP50, and +13.2% mAP50-95.
• The enhanced model demonstrated superior detection performance compared to state-of-the-art algorithms.

Conclusions:

• The developed computer vision method offers a robust and effective solution for industrial photovoltaic panel defect detection.
• The integration of advanced network components and a specialized loss function significantly enhances detection accuracy and reliability.
• This approach addresses key challenges in defect detection, paving the way for improved quality control in solar panel manufacturing.