Parallel image segmentation with adaptive mesh

Abstract

We propose a general-purpose parallel algorithm for image segmentation, which doesn't require any prior knowledge about the image or region. The algorithm, which is based on our adaptive mesh generation scheme, provides a binary tree structured split-and-merge mechanism to search and localize boundaries along discontinuities, and adapts the partition of the image to those detected discontinuities. This algorithm is independent of the order of processing or starting point, which be an important issue with parallel algorithms. Even though overlapping by only one pixel wide boundary between process blocks, by adopting the smoothness-based local features as homogeneity criteria, consistencies are maintained without the overhead of communication between adjacent process blocks. Our efficient hierarchical stepwise mechanism in merging target evaluation makes the merge process so simple and efficient that only two brother blocks are considered at each merge step in binary fashion.