Network snakes: graph-based object delineation with active contour models

In this paper, a graph-based method of active contour models called network snakes is presented and investigated. Active contour models are a well-known method in computer vision, bridging the gap between low-level feature extraction or segmentation and high-level geometric representation of objects. But the original concept is limited to single closed object boundaries. Network snakes are the method enabling a free optimization of arbitrary graphs representing the geometric position of networks and boundaries between adjacent objects. The main impacts of network snakes are the combination of the image energy representing objects in the real world, the internal energy incorporating shape characteristics, and the topology representing the structure of the scene. The introduction and exploitation of the topology in a comprehensive energy functional turn out to be a powerful technique to cope with complex questions of object delineation from imagery. Network snakes are analyzed and evaluated with both synthetic and real data to point out the role of the required initialization, the benefit of the introduced topology and the transferability. Exemplary investigated real applications are the delineation of field boundaries from remotely sensed imagery, the refinement of road networks from airborne SAR images and bio-medical tasks delineating adjacent biological cells in microscopic images. Concluding remarks are given at the end to discuss potential future research.