Machine Learning in Imaging and Inverse Problems: Where is it Going?

Tuesday, October 23, 2018 - 4:10pm - 5:00pm
Keller 3-180
Charles Bouman (Purdue University)
We consider the problem of learning models of scattering that decomposes scene returns into a set of scattering centers with limited persistence by utilizing sparsity of scattering coefficients. We consider both mono-static and bi-static radar collection geometries along wide-angle trajectories. The resulting sparse model can be interrogated at arbitrary pose and radar/target geometry for online estimation of t signatures. Based on the prior work on parametric models derived for canonical reflectors, we hypothesize that the scattering coefficient as a function of the viewing angle is embedded in a low-dimensional subspace spanned by a set of functions or dictionary atoms that have a support, which is concentrated in the viewing angle domain. We utilize multivariate Gaussian functions in the bistatic angle and bisector angle domain to approximate the scattering coefficients. Furthermore, we exploit the differentiability of the observation model in the scattering center locations and the parameters associated with the basis functions to jointly estimate the scattering center locations in the continuum to avoid gridding errors and approximate the scattering coefficient as a function of the bistatic angle and the corresponding bisector. Numerical simulations will be presented based on multi-bounce physical optics based electromagnetic simulator for canonical reflectors such as dihedral, trihedral and plates.