## Publica

Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten. # Field of experts

**Abstract**

Prior models of image or scene structure are useful for dealing with "noise" and ambiguity that occur in many machine vision problems such as stereo, optical flow, denoising, superresolution, and surface reconstruction. This chapter describes a method for learning priors represented as high-order Markov random fields defined over large neighborhood systems using ideas from sparse image patch representations. The resulting Field of Experts (FoE) models the prior probability of an image, or other low-level representation, in terms of a random field with overlapping cliques whose potentials are represented as a Product of Experts [192]. This model applies to a wide range of low-level representations, such as natural images [397], optical flow [395], image segmentations [451], and others. For simplicity of exposition the chapter focuses on applications to modeling natural images and demonstrates the power of the FoE model with two applications: image denoising and image inpainting [33]. (See figure 19.1 for examples.)