折棍变分贝叶斯图像分割算法

为提高图像分割的抗噪鲁棒性并解决分割数目的自适应确定问题,通过在聚类标签先验概率的折棍构造过程中建立Markov随机场,将空间相关性约束引入Dirichlet过程混合模型的概率建模,使聚类的空间平滑性得以增强,并采用变分推断方法获得聚类标签的收敛解析解,提出一种基于折棍变分贝叶斯推断的图像分割算法,实现了对像素聚类标签和分割数目的同步自适应学习,避免了传统方法中因引入空间相关性约束而出现的计算复杂问题.基于Berkeley BSD500图像测试数据集的数值实验结果表明,该算法具有比现有的混合模型聚类图像分割算法更高的PRI值,且在低于0.1的噪声方差条件下表现出了更优的抗噪鲁棒性.

Stick-Breaking Variational Bayesian Based Image Segmentation Algorithm

In order to improve the anti-noise robustness of image segmentation and adaptively determine the number of segmentations,Markov random field is established in the process of constructing the prior probability of clustering label,and the spatial correlation constraint is introduced into the probability modeling of the mixture model of the Dirichlet process,so that the spatial smoothness of clustering can be enhanced,and the convergent analytical solution of clustering label is obtained by the variational inference method.An image segmentation algorithm based on stick-breaking variational Bayesian is proposed,which realizes synchronous and adaptive learning of pixel clustering labels and segmentation numbers,and avoids the computational complexity caused by spatial correlation constraints in traditional methods.The numerical experiment results based on Berkeley BSD500 image test data set showed that the algorithm has better performance than the existing mixture model based image segmentation algorithms.The proposed algorithm has a higher PRI value and a better anti-noise robustness when the noise variance is lower than 0.1.