一种基于概率框架的遥感影像分割方法

A Probabilistic Framework for the Segmentation of Remote Sensing Imagery

基于区域增长的影像分割方法存在着种子选择和分割参数设置的难点。针对这两点,本文提出了一种基于概率合并框架的高分辨率遥感影像分割方法。首先,利用基于标记的分水岭变换和区域合并获得一个初始分割结果,避开种子点的选择;其次,利用初始分割所获得的具有一定大小的区域,计算统计、上下文和形状特征信息;然后,在贝叶斯准则的基础上,计算一个尺度无关的相邻区域间的合并概率,其应用于区域增长合并过程,合并概率具有直观统计意义,可以减少阈值确定的难度;同时由于区域的合并概率具有尺度无关性,可以在一次分割中成功地分割出不同尺度的地物;最后进行了试验,通过目视和定量分析及与e Cognition分割结果的对比证明了本文算法的有效性。

Image segmentation method based on region growing suffers the problem of selecting appropriate growing seeds and segmentation parameters.A segmentation method based on a novel kind of region-merging probability measure is proposed to solve the problem in this paper.First,an initial segmentation is achieved by marker-based watershed transformation and simple region merging.Statistic,context and shape information can be calculated from the initial segmentation regions. Then a scale-independent merging probability of neighboring regions is calculated by the information based on Bayesian criterion,which is integrated into sequential region merging procedure.It is easier to determine a threshold value for the new probability measure because of its intuitive meaning.And multiscale objects can be simultaneously segmented in once segmentation,since the probability is independent of scale. Experiment shows a good performance of the method by visual and quantitative analysis and compared with multi-resolution segmentation results of e Cognition.