Scale Space Segmentation of Color Images Using IFT-Watersheds and Multi-Resolution Region Merging

Color is known to be a strong cue in attracting an observer's attention. Actually, the motivation for using color for segmentation comes from the fact that it provides region information and that when specified appropriately, it can be relatively insensitive to variation in illumination conditions and appearances of objects. This paper presents a multi-resolution segmentation approach for color images using an image foresting transform (IFT)-based watersheds. The scale-space is based on graph morphology and works in color leading to a natural extension to grayscale morphological scale-spaces. The IFT-watershed algorithm is an unified and efficient approach to reduce image processing problems to a minimum-cost path forest problem in a graph and is used in this paper to produce regions in each scale. The relative entropy of color regions are estimated as region dissimilarity features across the scale-space stack, and combined by a merging process. The proposed approach is particularly well suited for the segmentation of complex image scenes or fine-resolution images where very thin and nested regions may have to be retained