An Improved YOLOv8 Detection Algorithm Based on Screen Printing Defect Images

Micro-defects, such as ink spots, scratches, and sintering formed during the screen printing process of photovoltaic cells, significantly impair module performance. Traditional machine vision methods exhibit limited detection efficiency and high false-positive and missed-detection rates, while existing deep learning algorithms struggle to achieve accurate and adaptive detection of small-target defects and background similar defects in complex industrial environments. This study proposes an enhanced defect detection methodology based on an improved YOLOv8 algorithm. A multi-focus image acquisition platform using primary and auxiliary CCDs was independently developed, integrating a high-frame-rate industrial camera and a high-resolution electron microscope, with an LED ring light employed to suppress reflections, thereby establishing a high-quality dataset covering three defect categories. The algorithm was optimized through multiple dimensions: the RepNCSPELAN4 module was incorporated into the backbone network to improve multi-scale feature fusion, and a novel wavelet transform-based WaveConv module was designed to replace traditional downsampling, thereby better preserving defect edges and texture details. The neck network integrates a lightweight shuffle attention mechanism and a new detail enhancement module to strengthen critical features while controlling model complexity. Additionally, a dedicated auxiliary detection head was added for spotting tiny ink dots. Experimental results demonstrate a marked improvement in performance: on the custom dataset, the improved model achieves a stable mean average precision of approximately 92%. Specifically, ink spot detection reached a precision of 84.9% and recall of 77.7%, effectively reducing missed small-target defects; sintering defect detection attained 98.9% precision and 100% recall, addressing previous misclassifications due to background similarity; and scratch detection precision improved to 92.2%. Visual comparisons confirm that the enhanced model effectively overcomes the limitations of the original approach. By constructing a specialized dataset and implementing targeted, coordinated optimizations to the YOLOv8 architecture, this study significantly enhances the accuracy and robustness of screen-printing defect detection in photovoltaic cells, providing an effective solution for real-time online quality inspection in smart manufacturing lines.