Watershed Arcs in Hierarchical Image Segmentation
Sampriti Soor, B Daya Sagar · HAL (Le Centre pour la Communication Scientifique Directe) · 2022
Watershed transform, formulated based on robust mathematical morphological operations, has been one of the most reliable image segmentation methods for years. However, it is constrained by the problem of over-segmentation. To encounter this limitation, region merging approaches such as waterfall and P-algorithm exist in the literature, which processes the entire image iteratively. In this paper, we introduce the concept of watershed arcs that acts as the connected partition lines between the segments on an image and explore several properties. We define the operation of removing an arc that results in merging two neighbor segments and apply it to generate hierarchical segmentation of an input image. At each level, a graph is constructed solely from the representatives of the watershed arcs in the previous level, and the set of arcs to be removed at that level is determined by applying a watershed transformation only on the graph. As the cardinality of the graph reduces drastically at each level, a hierarchy of partitions is produced in a comparatively short amount of time than that of the existing methods. The proposed method is applied to a Digital Elevation Model data to extract hierarchical river basins. We also evaluated the performance of this method in image segmentation compared with some state-of-the-art methods.