基于超像素的Grab cut前景提取算法

Super-pixel based foreground extraction algorithm using Grab cut

针对海量的像素以及迭代更新高斯混合模型参数导致的Grab cut很难兼顾精确分割和实时交互的现象,提出基于超像素的交互式快速分割框架,精简了问题规模,保证了分割的准确度和实时性.首先,采用融合边缘自信度的均值漂移算法将图像分割为一系列保持颜色、空间信息以及边界特性的同质区域;其次,以每个同质区域的RGB均值为结点,构建加权网络图,建立前景和背景两个高斯混合模型;然后,采用EM算法结合一种新的max flow/min cut算法来逼近高斯混合模型参数;最后,本文引入区域项的自适应调整参数来提高分割的准确度.实验结果表明,文中算法对用户交互可以及时做出分割响应,并且分割的结果更加合理.

To solve the difficulty to ensure both accuracy and real-time interaction in Grab cut resulting from the massive pixels and iterative updated Gaussian mixture model parameters,a fast interactive segmentation method based on super-pixel is proposed in this paper.The method reduces the scale of the problem while keep accurate and real-time segmentation simultaneously.It is achieved through the following four steps.Firstly,it uses mean shift algorithm with embedded confidence to segment the image into some homogeneous regions which maintain the color information,spatial information as well as the feature of boundaries.Secondly,it constructs the foreground and background GMM with the mean RBG of the homogeneous regions as the network nodes.Then,it combines EM algorithm with a new max flow/min cut to estimate the real GMM parameters.Finally,it automatically adjusts the weights of the region term in order to improve the segmentation accuracy.The experiment results show the algorithm presented can response to user interaction timely and the segmentation results are more reasonable.