Fri, Sep 22, 2017 @ 10:00 AM - 11:00 AM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Zeeshan Nadir, Electrical and Computer Engineering, Purdue University
Talk Title: A Model-Based Iterative Reconstruction Approach to Tunable Diode Laser Absorption Tomography
Abstract: Many imaging and sensing problems in the fields of medical imaging, computer vision, machine learning, communications and signal processing etc. can be posed as inverse problems. Broadly, an inverse problem consists of recovering some underlying signal of interest that leads to a directly observable measurement dataset where the measurement dataset may be corrupted by noise. In the presence of sufficient quantity of good quality measurement dataset, the inversion problem can often be solved by direct methods often involving closed form inverse formulas like filtered back projection. However, when the measurement data contains noise or is extremely sparse, then such conventional techniques do not work. Tunable Diode Laser Absorption Tomography (TDLAT) is such an ill-posed nonlinear inverse problem where 2D concentration and temperature images are required to be reconstructed from a handful of projection measurements.
Bayesian methods are a probabilistic approach to reconstruct signals by incorporating prior information about the signals in the form of a prior probability distribution. Typical 2D prior models like Markov Random Field enforce local smoothness on the images by penalizing differences between neighboring pixels. However, the major limitation of such prior models is that they cannot express non-homogeneous and non-Gaussian characteristics of the images and therefore cannot model the long-range correlations between image pixels. In this presentation, I shall present a Gaussian Mixture Model as a prior distribution which can be trained with a few training examples. In order to show the utility of this approach, I shall apply it to Tunable Diode Laser Absorption Tomography problem. I shall formulate the reconstruction problem as a Maximum-aposteriori estimation problem. I shall present an efficient multigrid algorithm to perform the resulting optimization. The results using simulated datasets show that the proposed approach can reduce reconstruction error while also resulting in a computationally efficient algorithm.
Biography: Zeeshan Nadir is a Ph.D. candidate in the school of Electrical and Computer Engineering, Purdue University, West Lafayette, IN. In Summer 2016, he was an intern at MathWorks, Inc., Natick, MA, where he worked on MATLAB coder package. He developed a new functionality in MATLAB Coder which has been incorporated in MATLAB R2017a release. His research interests include statistical signal processing, inverse problems, computational imaging, machine learning and computer vision.
Host: Hosted by Prof. Richard Leahy
Audiences: Everyone Is Invited
Contact: Talyia White