To improve anti-noise capability of image segmentation and take full advantage of energy functions in feature and label fields, a regional image segmentation based on energy functions for double random fields is proposed. Firstly, an image domain is partitioned into a set of sub-regions by a geometry tessellation technique. Based on the tessellation, negative logarithm of multi-Gaussian probability distribution is employed to define regionalized feature field energy function to describe the homogeneity of statistical distribution for pixel colors in a homogeneous region. The improved Potts model is the extension of the traditional model for the labels of a pixel and its neighbor pixels, and it is used to define regionalized label field energy function to characterize the relativity of the labels for sub-regions. Combine feature field and label field, and Kullback-Leibler divergence is utilized to define the heterogeneous energy function for describing the heterogeneity of color distributions among different homogeneous regions. The unconditional Gibbs function is adopted to transform the defined energy functions into probability functions for image segmentation. Finally, based on the maximization of probability distribution scheme, Metropolis-Hastings sampler is designed to obtain the optimal segmentation. Synthetic, remote sensing and natural texture images are segmented by several algorithms. Segmentation results show the proposed algorithm realizes image segmentation accurately and efficiently.%为了提高图像分割算法的抗噪性,并充分利用特征场和标号场在能量函数分割模型中的作用,提出基于双随机场能量函数的区域化图像分割方法.首先,利用几何划分将图像域划分为一系列子区域.在此基础上,采用多值高斯分布的负对数定义区域化特征场能量函数,用于描述同质区域内像素颜色的统计分布一致性.扩展传统建模邻域像素标号关系的Potts模型至邻域子区域,定义区域化标号场能量函数,用于表征各子区域标号之间的相关性.联合特征场和标号场,采用KL散度定义异质性能量函数,用于刻画同质区域间颜色统计分布异质性.利用非约束吉布斯表达式将定义的特征场和标号场能量函数转换为描述图像分割的概率分布函数.最后,在最大化上述概率分布函数准则下,设计合适的M-H采样算法,获得最优图像分割.在合成图像、遥感图像和自然纹理图像上进行分割实验,验证文中方法的有效性和准确性.
展开▼
