Robust and efficient object segmentation using pseudo-elastica

A new object segmentation method based on second-order energy minimisation is proposed. It is called pseudo-elastica   as it relates to the classic Euler’s elastica but resulting contours cannot be expected to converge towards continuous elastica if the resolution is increased. Comparing to prior works, our segmentation technique can be easily applied to both closed contours and open contours with fixed endpoints, and its computational complexity, O(NlogN)O(NlogN), is significantly lower. The efficiency is increased by extending the idea of bidirectional Dijkstra-type search to second-order energies and incorporating heuristics with some sacrifice in exact energy minimisation. Our pseudo-elastica generalises the classic first-order path-based schemes to second-order energies while maintaining the same low complexity. Experiments suggest that it scores similar or better results and usually requires considerably less user input than the state-of-the-art approaches. The algorithm can be made anisotropic in order to allow corners in the contour.

► Pseudo-elastica is an efficient scheme for user-guided image segmentation.
► Our technique imposes curvature regularity on the object boundary contour.
► Efficiency is achieved by dynamic curvature estimation within a bidirectional search scheme.
► Our method compares favourably to related techniques while requiring less user input.
► Due to their efficiency the pseudo-elastica are suitable for time critical applications.