用分水岭方法进行图象分割时，容易造成图象的过度分割，从而使得物体的轮廓线掩埋在杂乱的分水岭线中，为了克服这种过度分割问题，提出了一种保持边界的非线性扩散方法，该方法就是先对图象进行平滑，在去噪声的同时，即减少了梯度图象中区域最小值的数目，而分水岭分割后的图象区域数目与它相同；然后，根据初始分水岭分割的结果，使用区域灰度相似性和边界强度相结合的准则，进行由底向上的层次融合，从而较好地解决了过度分割的问题。实验结果表明，该方法可提供精确且封闭的区域轮廓线。

Image segmentation based on watershed method always results in over segmentation and makes the contours of the objects buried in the irrelevant watershed lines. In this paper, we first smooth the image while preserving the edge by nonlinear diffusion method, which can reduce the noise in the image and at the same time the numbers of the region minimums of the gradient image that is equal to the region numbers of watershed segmentation. Then from the result of the initial watershed segmentation which is organized by the Region Adjacency Graph (RAG), we execute the bottom up hierarchical region merging according to the region average gray value similarity and the edge strength criterions that can settle the over segmentation problem well. The region average gray value is a rough characteristic about the region while the edge strength criterion is local. Similarity criterion of the average gray value of the region is first used in Merging operation. Then upcast the RAG and continue to merge according to edge strength. The results of the experiment using 2D real images show that this method can provide accurately localized and closed region contours.