Homogeneous Region Merging Approach for Image Segmentation Preserving Semantic Object Contours

Morphological image segmentation is an essential tool for various image analysis tasks. In this paper, we propose two successive region-merging algorithms that mainly aim to represent homogeneous objects simply and clearly. First, we define a marker cluster and extract markers having similar intensity values from each marker cluster. After initial segmentation using the markers, perceptually insignificant small regions are removed to reduce the number of regions based on some heuristics. Then, we classify all regions into three classes according to their homogeneity. Finally, gradient-based region merging is performed for all regions except inhomogeneous regions. Experimental results show that the proposed regionmerging scheme can represent homogeneous objects with few regions while preserving most of the semantic object shapes. Hence, the proposed method can be a useful tool for semantic object segmentation.