Block-run-based connected component labelling algorithm for GPGPU using shared memory

摘要

An efficient two-scan connected component labelling (CCL) algorithm is proposed for a general purpose graphics processing unit (GPGPU). Compared to other GPU CCL algorithm, this algorithm has three distinct features. First, block-based and run-based strategies are combined in the first scan to simplify the equivalence label resolving process. Secondly, a novel labelling method for the GPU is introduced by constructing a forest of rooted trees using only 16-bit value for each node. Thirdly, the whole algorithm can be implemented in the GPU shared memory and minimise global memory bandwidth consumption. Experiments show that the algorithm achieves a speedup of between two and five times compared to other state-of-the-art GPU and CPU CCL algorithms.