Multi-core based Parallel N-path labeling HKM clustering algorithm

The detection of useful patterns in large datasets has attracted considerable interest recently. The hierarchical K-means clustering algorithm (HKM) is very efficient in large scale data analysis. It has been extensively used for building visual vocabulary in large scale image/video retrieval. However, the accuracy and speed of HKM still have room for improvement. In this paper, we propose a Parallel N-path labeling HKM clustering algorithm (PNHKM) which improves on the HKM clustering algorithm in the following ways. Firstly, we adopt a Greedy N-best Paths Labeling (GNPL) method to improve the clustering accuracy. Secondly, we focus on developing a parallel clustering algorithm for multicore processors. Our results confirm that the PNHKM is much faster and more effective.