一种无需初始化的遥感图像无监督分割方法

An Unsupervised Segmentation of Remote Sensing Images Without Initialization

遥感图像自动分割通常为有监督分割或带初始化的无监督分割,算法的性能受先验统计知识和初始点选择的影响较大.本文提出一种既无需统计先验也无需初始化的全自动分割方法.该方法基于图割理论和遥感图像自身数据特点建立模型,在迭代应用快速能量最小化方法的过程中融入一种自动初始化方法,实现全自动的分割.实验结果表明了该分割方法的有效性,有利于遥感图像进一步的目标检测识别.

Automatic segmentation methods of remote sensing images are either supervised apriori information or unsupervised with initial seeds/contours. The performance of these methods depends highly on the reliability of statistic apriori information or initial choice. Therefore a new segmentation method without apriori information or initialization is proposed. According to the remote sensing image data, an energy function model is firstly built up based on graph cuts. Second, an automatic initialization approach is then combined with swap moves algorithm iteratively, which is a fast approximation energy minimization algorithm, to achieve the final segmentation of the remote sensing image. Experiment results show its feasibility and efficiency. They also demonstrate the great significance in the following detection and recognition of interested objects.