Reconstruction of intricate surfaces from scanning electron microscopy

This PhD thesis is concerned with the reconstruction of intricate shapes from scanning electron microscope (SEM) imagery. Since SEM images bear a certain resemblance to optical images, approaches developed in the wider field of computer vision can to a certain degree be applied to SEM images as well. I focus on two such approaches, namely Multiview Stereo (MVS) and Shape from Shading (SfS) and extend them to the SEM domain.
The reconstruction of intricate shapes featuring thin protrusions and sparsely textured curved areas poses a significant challenge for current MVS techniques. The MVS methods I propose are designed to deal with such surfaces in particular, while also being robust to the specific problems inherent in the SEM modality: the absence of a static illumination and the unusually high noise level. I describe two different novel MVS methods aimed at narrow-baseline and medium-baseline imaging setups respectively. Both of them build on the assumption of pixelwise photoconsistency.
In the SfS context, I propose a novel empirical reflectance model for SEM images that allows for an efficient inference of surface orientation from multiple observations. My reflectance model is able to model both secondary and backscattered electron emission under an arbitrary detector setup. I describe two additional methods of inferring shape using combinations of MVS and SfS approaches: the first builds on my medium-baseline MVS method, which assumes photoconsistency, and improves on it by estimating the surface orientation using my reflectance model. The second goes beyond photoconsistency and estimates the depths themselves using the reflectance model.