SLADS: Fast Dynamic Sampling using Machine LearningMS11

In traditional imaging applications, pixels are sampled in raster order. However, it can be argued that such an approach is optimally bad since each new measure pixel gives the least new information. In this talk, we present SLADS (a Supervised Learning Approach to Dynamic Sampling) a method for fast greedy selection of pixels which minimizes the reconstruction distortion for 2D image sampling. SLADS is based on the goal of selecting the pixel at each step that most reduces the absolute reconstruction error. By doing this, we can dramatically reduce the number of samples required to reconstruct an image. A key innovation of SLADS is that it uses machine learning methods to estimate the expected reduction in distortion (ERD) so that the algorithm is very fast. This makes SLADS practical for implementation on real image scanning systems. We present results of SLADS’ application in the problems of synchrotron imaging, scanning electron microscopy, and confocal microscopy, and show how it can be used to reduce scan time and radiation dosage.

This presentation is part of Minisymposium “MS11 - Computational Imaging for Micro- and Nano-structures in Materials Science (2 parts)
organized by: Brendt Wohlberg (Los Alamos National Laboratory) , Jeff Simmons (Air Force Research Laboratory) .

Charles Bouman (Purdue University)