Parallel Marker Based Image Segmentation with Watershed Transformation Parallel Marker Based Watershed Transformation

The parallel watershed transformation used in grayscale image segmentation is here augmented to perform with the aid of a priori supplied image cues, called markers. The reason for introducing markers is to calibrate a resilient algorithm to oversegmenta-tion. In a hybrid fashion, pixels are rst clustered based on spatial proximity and graylevel homogeneity with the watershed transformation. Boundary-based region merging is then eeected to condense non-marked regions to marked catchment basins. The agglomeration strategy works with a weighted neighborhood graph representation of the oversegmented image. The throughput of a parallel Bor uvka-like minimum spanning forest (MSF) operator , applied on the considered graph, embodies the desired image partition, reasoning that all regions in a tree fuse into a homogeneous area containing a unique marker. Two gures of merit of the parallel algorithm are worth of mentioning: the local detection of the catchment basins conforming the watershed principle (which strongly depends on the history of the regions' growth); and the parallel computation of the Bor uvka-like MSF which merges, at the same time, partial regions, produced by the local labeling, and non-marked regions to marked basins. Both modules are designed with great concurrency, locality, and reduced software engineering cost, emerging into a scalable algorithm.