Microscopic Image Analysis of Printed Structures Without a Microscope: A Deep Learning Approach

This paper presents a deep learning approach for the geometrical analysis of printed structures at the microscopic level using inline optical images. The focus is on extracting finger geometry parameters that are essential for monitoring and optimi-zing metallization technologies. Traditionally, obtaining these para-meters requires time-consuming microscopic measurements, limi-ting the evaluation to a few small sample regions. This study intro-duces a method to extract the finger parameters of the full sample directly from inline optical images of lower resolution. Thereby, the height profile is derived from the shadow cast by a lateral laser illumination. The machine learning pipeline contains models (a) for identifying the height profile from noisy shadow images, (b) for generating high-resolution height images and maps of the metal-lized regions and (c) for quantifying geometrical parameters from these images. By combining these models, we can predict the core width, shading width and cross-sectional area of the fingers, which correlate to microscopic reference data. Within seconds, user-friendly quality maps of the full print structure can be generated, allowing not only the identification of print errors, but also pro-viding quantitative data for a detailed process and cell analysis.