Efficient Image Segmentation In Graphs With Localized Curvilinear Features

In graph-based image segmentation, the arc weights are given by a local edge indicator function based on image attributes and prior object information. In boundary tracking methods, an edge integration process combines local edges into meaningful long edge curves, interconnecting a set of anchor points, such that a closed contour is computed for segmentation. In this work, we show that multiple short-range edge integrations can extract curvilinear features all over the image to improve seeded region-based segmentation. We demonstrate these results using edge integration by Live Wire (LW), combined with Oriented Image Foresting Transform (OIFT), due to their complementary strengths. As result, we have a globally optimal segmentation, that can be tailored to a given target object, according to its localized curvilinear features.