Interactive image segmentation methods have been proposed based on region constraints (user-drawn markers) and boundary constraints (anchor points). However, they have complementary strengths and weaknesses, which can be addressed to further reduce user involvement. We achieve this goal by combining two popular methods in the Image Foresting Transform (IFF) framework, the differential IFT with optimum seed competition (DIFT-SC) and live-wire-on-the-fly (LWOF), resulting in a new method called Live Markers (LM). DIFT-SC can cope with complex object silhouettes, but presents a leaking problem on weaker parts of the boundary. LWOF provides smoother segmentations and blocks the DIFT-SC leaking, but requires more user interaction. LM combines their strengths and eliminates their weaknesses at the same time, by transforming optimum boundary segments from LWOF into internal and external markers for DIFT-SC. This hybrid approach allows linear-time execution in the first interaction and sublinear-time corrections in the subsequent ones. We demonstrate its ability to reduce user involvement with respect to LWOF and DIFT-SC using several natural and medical images.
展开▼
