Coverage segmentation of thin structures by linear unmixing and local centre of gravity attraction

Abstract

We present a coverage segmentation method for extracting thin structures in two-dimensional images. These thin structures can be, for example, retinal vessels, or microtubules in cytoskeleton, which are often 1-2 pixels thick. There exist several methods for coverage segmentation, but when it comes to thin and long structures, the segmentation is often unreliable. We propose a method that does not shrink the structures inappropriately and creates a trustworthy segmentation. In addition, as a by-product a high-resolution crisp reconstruction is provided. The method needs a reliable crisp segmentation as an input and uses information from linear unmixing and the crisp segmentation to create a high-resolution crisp reconstruction of the object. After a procedure where holes and protrusions are removed, the high-resolution crisp image is optionally downsampled back to its original size, creating a coverage segmentation that preserves thin structures.