Deep Learning for Sparse Scanning Electron Microscopy

High-throughput scanning electron microscopy (SEM) has the goal to acquire large volumes at high resolution [1]. For samples of several millimeters a full scan with a traditional dense scanning pattern would take years. There are various ways to reduce the required time as has been evaluated in [2]. The authors showed that the best strategy for reducing overall acquisition times are sparse scanning methods that record only a small percentage of all possible pixels but with a reasonably high dwell time per pixel. The sparse data pattern must be reconstructed afterwards. As one solution missing data can be reconstructed using inpainting methods [2]. One promising approach is exemplar-based inpainting [3], which can deliver useful results without the introduction of artifacts that have not been observed in known data, as only prior knowledge of already fully scanned images is used for the reconstruction process.