A Dual Active Contour for Improved Snake Performance

Active contours (snakes) are a sophisticated approach to contour extraction in image interpretation. They challenge the widely held view that low-level vision tasks such as edge detection are 'bottom-up' processes; features are extracted from an image and higher level processes interpolate to find a suitable representation. The principal disadvantage with such an approach is its serial nature; errors generated at a low-level are passed on through the system without the possibility of correction. The principal advantage of snakes is that the image data, the initial estimate, desired contour properties and knowledge-based constraints are integrated into a single extraction process. Snakes incorporate a global view of edge detection by assessing continuity and curvature combined with the local edge strength to determine an edge. The motivation for developing a dual active contour is to enhance the snake model by confronting its principal problems: initialisation -- the final extracted contour is highly dependent on the position and shape of the initial contour as a consequence of many local minima in the energy function. The initial contour must be placed near the required feature otherwise the contour can become obstructed by unwanted features. parameters -- the original technique gives no guidelines for determining the parameters. The values used are critical and must be chosen carefully to obtain meaningful results, due to scale variance and parameter inter-dependence. non-convex shapes -- by its formulation the original technique is poor at extracting non-convex shapes. A dual contour can provide a more balanced approach to contour extraction allowing both convex and non-convex shapes to be extracted. scale invariance -- the snake's internal energy is not scale invariant. This prevents the energy of the snake being used to assess the merit of the solution. A dual contour overcomes the initialisation problems by approaching the desired feature from both sides. This contrasts with other methods which approach a feature from one side and hence have less ability to determine a global minimum. The internal energy of the contour is reformulated to be scale invariant, and allows a relative assessment to reject poor local minima. The new techniques implementation allows any additional local shape information to be integrated within the minimisation process. This shape model is similar to that of Lai and Chin in allowing local control over the contours equilibrium. The problem of determining parameters is simplified by reducing the parameters to a single regularisation parameter which is consistent with the original paradigm