A probabilistic grouping principle to go from pixels to visual structures

Abstract

We will describe here how the Helmholtz principle, which is a principle of visual perception, can be translated into a computational tool that can be used for many problems of discrete image analysis. The Helmholtz principle can be formulated as "we immediately perceive whatever has a low likelihood of resulting from accidental arrangement". To translate this principle into a computational tool, we will introduce a variable called NFA (Number of False Alarms) associated to any geometric event in an image. The NFA of an event is defined as the expectation of the number of occurrences of this event in a pure noise image of same size. Meaningful events will then be events with a very low NFA. We will see how this notion can be efficiently used in many detection problems (alignments, smooth curves, edges, etc.). The common framework of these detection problems is that they can all be translated into the question of knowing whether a given group of pixels is meaningful or not. This is a joint work with Lionel Moisan and Jean-Michel Morel. © 2011 Springer-Verlag.