Near-infrared defect detection method for photovoltaic panels based on the FBCT-YOLOv8 algorithm

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Vision

Background:

• Near-infrared defect detection in photovoltaic panels faces challenges with varying defect scales, morphologies, and small target identification.
• Existing methods struggle to adapt to the complex visual characteristics of defects, impacting overall detection efficiency and accuracy.

Purpose of the Study:

• To propose an advanced defect detection method for photovoltaic panels using near-infrared imaging.
• To enhance the accuracy and efficiency of detecting diverse and small-scale defects in solar panels.

Main Methods:

• The FBCT-YOLOv8 algorithm, built upon YOLOv8, incorporates omni-dimensional dynamic convolution (ODConv) for adaptive feature extraction.
• A lightweight cross-scale feature fusion module (CCFM) is utilized for improved integration of multi-scale features, enhancing adaptability to scale variations.
• Dynamic head (DyHead) and InnerWIoU loss function are employed to refine detection performance and loss calculation.

Main Results:

• The FBCT-YOLOv8 algorithm achieved a GFLOPs reduction to 7.0 and an mAP@50 of 0.91.
• The method demonstrates real-time performance suitable for photovoltaic panel defect detection.
• Ablation and comparative experiments confirmed the proposed method's effectiveness and superiority over existing approaches.

Conclusions:

• The proposed FBCT-YOLOv8 algorithm effectively addresses the challenges in near-infrared defect detection for photovoltaic panels.
• The integration of ODConv, CCFM, DyHead, and InnerWIoU significantly improves detection accuracy, especially for small and varied defects.
• The method achieves a balance between high accuracy and real-time processing, making it suitable for practical applications.