Multiscale estimation of regular image contours via a graph of local edgel hypotheses

The problem of estimating the regular portions of the contours in an image is formulated in a probabilistic and multiscale framework. The objective is to compute a small set of polygonal lines which, with high probability, contains an approximation to every contour in the scene. These polygonal lines are represented by paths in a graph whose arcs represent local contour hypotheses. The main difficulty of this problem is that, in order to achieve high probability of accurate reconstruction according to a global metric, it is necessary to deal with a combinatorially large number of contour hypotheses. To control the complexity of the search, the notion of a compressible graph is introduced and an efficient contour estimation algorithm based on graph compression is proposed. *Research supported by US Army grant DAAH04-95-1-0494, Center for Imaging Science, and MURI grant DAAH04-96-1-0445, Foundations of Performance Metrics for Object Recognition.