Figure 1. IKONOS satellite image of a lake.
Bayesian models and algorithms for the perceptual grouping of contours
Our starting point for the development of contour grouping algorithms is the image edge map, attained by application of a multiscale edge detection algorithm. We also make use of the tangent map, in which edgels have been grouped together to form short, linear segments.
See Figures 1, 2 and 3.
We formulate the contour grouping problem as a Bayesian inference task, based upon both photometric and geometric cue likelihood models constructed from the statistics of natural images. This modelling forms the basis of Aaron Clarke’s Master’s project research.
- Contour grouping with prior models (Elder, Krupnik and Johnston, PAMI 2003)
In this paper we investigate how prior knowledge may be incorporated in a probabilistic framework for determining the bounding contour of an object of interest. The maximum a posteriori inference problem is computationally intractable, and thus a constructive algorithm is presented which builds up a set of candidate closed contours that have high probability of representing the true object boundary.
- Efficient computation of closed contours using modified Baum-Welch updating (Johnston and Elder, POCV 2004)
The incorporation of contour closure is a vital and yet difficult aspect of contour grouping for object boundary determination. We extend the previous constructive algorithm approach by way of a dynamic reweighting scheme that influences contour growth toward closure. The re-weighting is determined by an on-line, sequential version of the Baum-Welch algorithm, used for iterative estimation of hidden Markov model parameters.
- Multiscale contour grouping
We are currently investigating multiscale contour representations and the subsequent development of multiscale contour grouping algorithms based on hierarchical hidden Markov models and Belief Propagation.