Accelerated Connected Component Labeling Using CUDA Framework

Connected Component Labeling (CCL) is a well-known algorithm with many applications in image processing and computer vision. Given the growth in terms of inter-pixel relationships and the amount of information stored in a single pixel, the time to run CCL analysis on an image continues to increase rapidly. In this paper we present an accelerated version of CCL using NVIDIA’s Compute Unified Device Architecture (CUDA) framework to address this growing overhead. Our parallelization approach decomposes CCL while respecting all global dependencies across the image. We compare our implementation against serial execution and parallelized implementations developed on OpenMP. We show that our parallelized CCL algorithm targeting NVIDIA’s CUDA can significantly increase performance, while still ensuring labeling quality.