Description »

The purpose of this project is to design and implement algorithms for the reconstruction of a voxel cube from undersampled electron microscope acquisitions, drawing on recent techniques from image processing frequently used for inpainting, denoising, or deconvolution.

In this project, the following issues will be investigated:

• For this specific kind of data, which undersampling design leads to the best reconstruction results? Does a regular pattern suffice? How important is randomization of sampling locations?
• The acquisition proceeds slice by slice in the z direction, and the sampling pattern can be controlled within this sequential process. How can we feed back information already sampled in order to best guide sampling in the next slice? How much can we gain (in the tradeoff between scan time and reconstruction quality) by doing such an actively controlled acquisition?
• Another option is adaptive refinement within one slice. Given we acquired a fast sample, widely spaced, how can we predict interesting regions, so to direct subsequent fine sampling to where it is really needed?

This project will be run in collaboration between the Probabilistic Machine Learning Lab and the Computer Vision Lab. It will provide hands-on experience with powerful sparse reconstruction techniques and novel applications of actively controlled data acquisition.