Image segmentation using linked mean-shift vectors with region attribution analysis

This paper proposes a novel linked mean-shift algorithm that considers region attribution in a cluster merging process. Mean-shift based image segmentation suffers from its extreme computational complexity, despite of its outstanding segmentation accuracy. To resolve this problem, the linked mean-shift algorithm that removes the iterative process in the mean-shift process was introduced. However, the approximation in the linked-mean-shift algorithm gives rise to unwanted merging of the clusters that should not be merged. To prevent the unwanted merging, the proposed algorithm analyzes region attribution, then, in the merging process, applies strict condition to the clusters that have dissimilar attribution than the clusters that have similar attribution. In experiments, the proposed algorithm improved segmentation accuracy than the linked mean-shift algorithm, while retained twenty times faster speed than the mean-shift algorithm. Furthermore, the experimental results for variation of processing time showed the proposed algorithm can provide much settled throughput than the mean-shift algorithm.