Sparse, variable-representation active contour models

Active contours are a widely used class of models that locate object boundaries in an image by minimizing an energy function which depends on "internal" terms such as the length and curvature of the contour, and "external" terms which are functions of the image values on and near the contour. If we use the inverse rate of change of the image value as the external term, the energy is low when the contour coincides with a strong, short, smooth boundary in the image. It is well known that this basic active contour model has difficulties in detecting object boundaries that are initially far from the contour; in locating boundary shape details; and in avoiding local minima due to image noise. The first two difficulties can be overcome by varying the energy function during the minimization process, and we show in this paper that the third difficulty can be overcome by modifying the contour representation during the process.