Abstract: Image segmentation methods based on graph cuts have a unified segmentation framework combined with various knowledge, and can get a global optimal solution. Such algorithms have poor efficiency because of massive pixel level processing units and iterative solving model. On the basis of GrabCut algorithm, this paper transforms the image into color-similarity-blocks using the watershed algorithm, and estimates the Gaussian Mixture Model(GMM) parameters with blocks instead of pixels, so sharply decreases the problem scale and significantly improves the algorithm efficiency.

Key words: graph cuts, watershed transform, Gaussian Mixture Model(GMM)

摘要： 基于图割理论的图像分割具有结合多种知识的统一图像分割框架，可获取全局最优解，但海量的像素级处理单元以及为达到一定分割精度而采用的迭代求解模式，导致算法分割效率不高。以GrabCut算法为基础，通过分水岭变换，将图像划分成区域内颜色相似的若干分块，以各个块内像素的RGB均值代表所在分块的全部像素点来估计高斯混合模型参数，使问题规模减小，算法效率得到提高。

关键词: 图割, 分水岭变换, 高斯混合模型

CLC Number: 

• TP391