Shadows in the Spotlight: Casting Deep Learning Models to Unveil the Geometry of Printed Structures

Knowledge of the geometrical parameters of contact fingers is crucial for monitoring and optimizing solar cells. This paper introduces a modular approach based on deep learning (DL) for a full-area quality inspection of the finger geometry. Our approach is designed for production use, as the underlying optical images can be captured in real time using a top-light and a low-angle laser illumination. The former provides information regarding the width of the metallized region (shading width), and the latter casts a shadow relevant for the computation of structural information as the core width, peak height, and cross-sectional area. Our approach consists of (a) an image processing algorithm for automatic finger detection, (b) a DL model to extract the finger height profile from noisy shadow images, (c) a DL model for generating maps of the metallized regions and high-resolution height images, and (d) a regression model to predict the geometrical parameters. Finally, we convert these parameters into quality maps for visualization and statistical analysis. On comparison with microscopic references, the model achieves a correlation coefficient of 0.93 and a mean absolute error of 20  μm2 for cross-sectional areas ranging from 80 μm2 to 415 μm2 minimizing the need for offline microscopic measurements.