In Maximum Entropy and Bayesian Methods, K.M. Hanson and R.N. Silver, eds. pp. 211-228, (Kluwer Academic, Dordrecht, 1996)

Toward optimal observer performance of detection and discrimination tasks on reconstructions from sparse data

R.F. Wagner, K.J. Myers, D.G. Brown, M.P. Anderson, K.M. Hanson

Abstract

It is well known that image assessment is task dependent. This is demonstrated in the context of images reconstructed from sparse data using MEMSYS3. We demonstrate that the problem of determining the regularization- or hyperparameter alpha has a task-dependent character independent of whether the images are viewed by human observers or by classical or neural-net classifiers. This issue is not addressed by Bayesian image analysts. We suggest, however, that knowledge of the task, or the use to which the images are to be put, is a form of prior knowledge that should be incorporated into a Bayesian analysis. We sketch a frequentist approach that may serve as a guide to a Bayesian solution.

Keywords: observer performance, tomographic reconstruction, MEMSYS, hyperparameter selection

Get full paper (pdf, 139 KB)
Return to publication list Send e-mail to author at kmh@hansonhub.com