A Parallel Framework for Image Segmentation Using Region Based Techniques
Juan C., E.I. David, Francisco F. · 2007
In this work a parallel framework for image segmentation using region based techniques is presented. The algorithm is based on performing several segmentations of the same image using a parallel region-based algorithm. Moreover these segmentations are also obtained in parallel. This way, our proposal presents a two-level parallel layout. Next, an oversegmented image that collects all the information from the previous segmentations is created. A region-merging algorithm, developed previously by the authors, is then applied to this oversegmented image. A relevant aspect is that the information obtained from the partial segmentations will, in fact, guide the merging process, in such a way that the actual characteristics of each region or pixel are not taken into account. The merging algorithm uses the concept of force of repulsion between neighboring pixels that indicates quantitatively their tendency to form part of different regions. The force of repulsion considers several situations in which any two neighboring pixels can be found in all the partial segmentations that are used to create the oversegmented image, including the shadowed zones. The shadowed zones are groups of pixels that differ in their intensity level a certain threshold from the region in which they could be included. Introducing this concept in the region-based algorithms, regions with low levels of homogeneity are avoided, improving the quality of the whole process. Note that, given that the shadowed zones are not treated by the algorithm, the information that can be extracted from these zones is minimum. As stopping criterion of the merging algorithm, we use a function to evaluate the quality of the segmentation. The algorithm has been validated using several artificial and real images demonstrating the benefits of our proposal, and it was tested on a HP Superdome cluster.