K-means image segmentation on massively parallel GPU architecture

Image segmentation can be computationally demanding, and therefore require powerful hardware in order to meet performance requirements. Recent rapid increase in the performance of graphic processing unit (GPU) hardware, coupled with simplified programming methods, have made GPU an efficient coprocessor for executing variety of highly parallel applications. This paper presents an implementation of k-means image segmentation on the GPU platform with Compute Unified Device Architecture (CUDA). Parallel k-means segmentation is realized in hybrid manner i.e. proposed approach distributes computation load between Central Processing Unit (CPU) and GPU. The emphasis is placed on adaptation of the core algorithm to efficiently process datasets characteristic for image segmentation while exploiting benefits of underlying GPU hardware architecture. Numerical experiments have demonstrated considerably faster segmentation execution with proposed approach comparing to classical CPU-based approach.