Abstract: Morphological algorithm of watersheds suggested by Vincent and Soille is widely used as a preprocessing stage in image segmentation. To solve the over-segmentation problem, a region merging procedure is generally needed. In this article we present a hybrid region dissimilarity function to perform the RAG based region merging process. It synthesizes many metrics including intensity difference, edge height, texture information, edge complexity and dynamic weight assignment according to merge phases. We further propose a modified watershed transform method, which is able to generate multi-scale information of the gradient image with little additional overheads compared with the original algorithm. We use this information to evaluate the edge height criteria. Experiments show that the algorithm is very robust and adaptive.