结合最小生成树划分的高分遥感图像分割

High-resolution Remote Sensing Image Segmentation Using Minimum Spanning Tree Tessellation

针对基于几何划分的区域统计图像分割方法计算量大以及对地物复杂边界分割精度低的问题,提出了一种基于静态划分和区域统计的彩色高分辨率遥感图像分割方法。该方法定义了一种静态MST同质区域划分准则,借助MST能较好表达边界的特点,解决复杂地物边界的表达问题,并避免动态划分方法在迭代过程中反复划分造成的计算量大的问题。首先利用MST划分将图像域划分成若干个同质子区域,并假设每个同质子区域内的像素服从独立同一的高斯分布;然后在贝叶斯理论框架下构建基于区域的图像最优分割模型;最后结合M-H/ML算法进行图像分割和模型参数估计。为了验证提出方法,分别对真实及模拟彩色图像进行分割试验,定性定量评价结果表明了提出方法的可行性和有效性。

In this paper,we proposed a new high-resolution color remote sensing image segmentation approach combining static tessellation and the regional statistics algorithm.The approach defined a static regions tessellation rule based on MST model,which solved the expression of the complicated boundary of objects in RS imagery,and avoided huge computation in dynamic tessellation method due to repeatedly tessellation in iterations.Firstly,MST tessellation was employed for partitioning the image domain into some polygons corresponding to the components of homogeneous regions that need to be segmented.The M-H/ML algorithm,which integrated the ML algorithm for parameter estimation with the M-H algorithm for segmentation,was also proposed to address the optimization color image segmentation.Quantitative experience results on a synthetic color image showed the performance of the proposed approach.Experiments were also carried out on real high-resolution color remote sensing images in order to validate the proposed approach.