CMStatistics 2022: Start Registration
View Submission - CMStatistics
Title: Latest developments in Bayesian image analysis in Fourier space (BIFS) models Authors:  John Kornak - University of California, San Francisco (United States) [presenting]
Karl Young - University of California San Francisco (United States)
Eric Friedman - International Computer Science Institute Berkeley (United States)
Ross Boylan - University of California San Francisco (United States)
Abstract: For more than 30 years now, Bayesian image analysis has been a leading approach to image reconstruction and enhancement. The idea of the approach is to balance a priori expectations of image characteristics (the prior) with a model for the image degradation process (the likelihood). The conventional Bayesian modeling approach, as defined in image space, implements priors that describe inter-dependence between spatial locations on the image lattice (commonly through Markov random field, MRF, models) and can therefore be difficult to model and compute. Bayesian image analysis in Fourier space (BIFS) provides for an alternate approach that can generate a wide range of models, including ones with similar properties to conventional models, but with reduced computational burden; the originally complex high-dimensional estimation problem in image space can be similarly modeled as a series of (trivially parallelizable) independent one-dimensional problems in Fourier space. A range of prior models that can be formulated in Fourier space will be illustrated, including MRF-matched models and frequency-selective models, and these will be compared to conventional models. In addition, extensions will be briefly discussed based on a) a data-driven prior approach and b) transforming to the wavelet domain.