Block-Based connected component labeling algorithm with block prediction

In this paper, we propose a block-based connected component labeling algorithm, which predicts current block's label by exploiting the information obtained from previous block to reduce memory access. By generating a forest of decision trees according to some of previous block's pixels, which are also needed for current block's label decision, we can reduce trees' depth and number of pixels to check. Experimental results show that our method is faster than the most recent labeling algorithms with image datasets which have various size and pixel density.