多尺度对象高空间分辨率遥感影像谱聚类分割

Spectral clustering segmentation of high spatial resolution remote sensing imagery based on multi-scale object

针对基于像素模型的单尺度或多尺度谱聚类影像分割方法在相似矩阵存储、特征分解效率及分割精度方面存在的不足。该文首先通过给定多组空间及光谱带宽参数,利用mean-shift初分割生成不同尺度的超像素对象层;然后联合像素与超像素对高空间分辨率影像中的不同类别地物进行的多尺度建模表达其空间拓扑关系,即在图割理论框架下建立“像素-超像素”联合的多尺度无向权图模型G(V,E,W),同时根据遥感影像纹理特征丰富的特点,在顶点相似性计算过程中融合纹理特征;最后使用基于normalized cut准则的谱聚类算法,对图模型划分得到最终分割结果。该方法较好地降低了基于像素的谱聚类分割方法的计算复杂度,同时提高分割结果准确率。标准测试数据集和“高分2号”遥感影像分割结果表明了该方法的有效性。

According to the shortcomings in storage of similarity matrix,inefficiency of eigenvector decomposition and the accuracy of segmentation using pixel based single-scale or multi-scale spectral clustering?First,different space and range domain bandwidth parameters were given to generate different scale super-pixel layer by mean-shift algorithm.Then,pixels and multi-layer superpixels were used to model the high spatial resolution remote sensing imagery to represent its topology structure,that is construct a“pixel-superpixel”based undirectional weighted graph model under the graph cut theory,and we also optimized the method to calculate the vertex whose connected with each other to finish similarity matrix.Finally,a spectral clustering algorithm based on normalized cut criterion was used to partition the graph model and get the final segmentation result.This method reduces the computational complexity while improve segmentation accuracy of pixel-based spectral clustering.The segmentation results of standard images database and GF2 remote sensing images show that the method is effective.