基于改进区域邻接图的遥感图像多尺度快速分割方法

摘要

This paper proposed a multi-scale fast segmentation method. First, we build the improved Region Adjacency Graph (RAG) based on the initial segmentation produced by directly watershed. Then a fast region merging method is applied to the RAG to obtain the multi-scale segmentation result. The merging criterion is the feature vector difference between two adjacent regions. The scale control parameter, which includes the allowed maximum standard deviation and area of a region,is used to produce the multi-scale segmentation results. The proposed method has been successfully tested on QuickBird images. Both the supervised evaluation of segmentation results and the comparison with eCognition have shown that the proposed method performs quite well and it has a full potential in practical applications.%提出一种针对高分辨率遥感图像的多尺度快速分割方法,首先利用直接分水岭分割得到初始分割结果,而后构建改进的区域邻接图,并在其基础上通过较为快速的区域合并策略得到最终分割结果.合并准则为相邻区域间的特征向量差,合并过程中通过设置区域允许最大标准差和区域允许最大面积两个尺度控制参数以生成不同尺度的分割结果.以QuickBird卫星图像进行实验,通过监督评价以及与eCognition软件进行比较,结果表明本文分割方法在精度和效率方面均效果较好,具备实际应用的潜力.