Subpixel edge refinement using deformable models

We propose an approach for subpixel edge refinement based on deformable models. In our approach, the result is constrained by two kinds of information: orientation information derived from the gradient and position information. The problem is formulated in a statistical framework: the likelihood function of the observations is computed and used in a classical maximum a posteriori estimator. Two different models are proposed: a parametric model based on B-spline functions and a sampled model. Experiments on both synthetic and natural images are described that show the adequacy and effectiveness of these algorithms.