Numerical experiments with cooperating multiple quadratic snakes for road extraction

Higher-order active contours or snakes show much promise for the extraction of complex objects from noisy imagery. These models provide an elegant mathematical framework for specifying the desired properties of target objects through energy functionals that can be minimized with standard optimization techniques. However, techniques to allow quadratic snakes to change topology during segmentation have not been fully exploited. Additionally, external forces for improving convergence of quadratic snakes have similarly yet to be explored. In this article, we propose a model that allows multiple quadratic snakes to split, merge, and disappear. Although the separate components of our approach have been introduced elsewhere by Cohen (1991), Xu and Prince (1997), and Rochery et al. (2006), this article is the first comprehensive empirical study of their performance on real-world complex network extraction tasks. We analyze the applicability of the model to road extraction from satellite images that vary in complexity from simple networks to large networks with multiple loops. We also analyze the effects of external forces enhanced by oriented filtering, gradient vector flow fields, and Canny edge detection. In a series of experiments, we found that the multiple cooperating quadratic snake model performs well on complex, noisy images. Our experiments also establish a performance improvement when the proposed quadratic model is coupled with the Canny-based gradient vector flow technique.