Co-parent selection for fast region merging in pyramidal image segmentation

Abstract

The goal of image segmentation is to partition an image into regions that are internally homogeneous and heterogeneous with respect to other neighbouring regions. We build on the pyramid image segmentation work proposed by [3] and [9] by making a more efficient method by which children chose parents within the pyramid structure. Instead of considering only four immediate parents as in [3], in [9] each child node considers the neighbours of its candidate parent, and the candidate parents of its neighbouring nodes in the same level. In this paper, we also introduce the concept of a co-parent node for possible region merging at the end of each iteration. The new parents of the former children are co-parent candidates as if they are similar. The co-parent is chosen to be the one with the largest receptive field among candidate co-parents. Each child then additionally considers one more candidate, the co-parent of the previous parent. Other steps in the algorithm, and its overall layout, were also improved. The new algorithm is tested on a set of images. Our algorithm is fast (produces segmentations within seconds), results in the correct segmentation of elongated and large regions, very simple compared to plethora of existing algorithms, and appears competitive in segmentation quality with the best publicly available implementations. The major improvement over [9] is that it produces visually appealing results at earlier levels of pyramid segmentation, and not only at the top one.