The acquisition of high-quality image data is the foundation of reliable microstructure analysis. Especially in highly automated microscopy workflows, producing sharp images over large sample areas can be challenging because of physical constraints like limited depth of focus, hard-/software constraints that limit the reliability of autofocus methods or even time constraints that limit the use of focus correction overall.
This work focuses on the use of deep learning to build models for estimating the focus deviation in light-optical microscopy and the use of these models to make image acquisistion more reliable by replacing established software-based autofocus routines. The applicability on different materials as well as transferability between different microscopy hardware settings will be discussed. Combined with earlier work on deep learning for image deblurring[1,2], a more robust in-situ image acquisistion strategy is proposed.