A Study on Connected Components Labeling algorithms using GPUs

Connected Components Labeling (CCL) is a wellknown problem with many applications in Image Processing. We propose in this article an optimized version of CCL for GPUs using GPGPU (General-Purpose Computing on Graphics Processing Units) techniques and the usual UnionFind algorithm to solve the CCL problem. We compare its performance with an efficient serial algorithm and with Label Equivalence, another method proposed in the literature which uses GPUs as well. Our algorithm presented a 5-10x speedup in relation to the serial algorithm and performed similar to Label Equivalence, but proved to be more predictable in the sense that its execution time is more image-independent. Keywords-Image processing; Connected Components Labeling; Connected Components Analysis; GPU; CUDA