Abstract:
|
Electron microscopes provide potentially high resolution videos at an atomic scale, but that resolution comes at a cost. That is, the electrons from the microscope interact with and in turn, alter the experimental sample, necessitating the microscope sparsely radiate the sample to minimize unwanted side effects. Traditional video analysis techniques based on matrix decompositions struggle to deal with sparely sampled data as the missing data typically need to be imputed before a full matrix decomposition is possible. Accurate imputation is difficult and computationally expensive when only, e.g., 1% of the sample is available. We therefore propose an alternative approach to analyzing sparely sampled videos based on a Gaussian mixture regression model that accounts for spatio-temporal trends while simultaneously estimating features of the video that are of interest to the microscopists. We illustrate our method with videos of silver nucleation and the chemical reactions inside a lithium battery.
|