基于谱聚类的高分辨率全色遥感影像分割

High-resolution panchromatic remote sensing image segmentation based on spectral clustering

为了协调高分辨率全色遥感影像区域和边界的最优分割,提出了一种基于像素邻域和光谱特征的谱聚类高分辨率全色遥感影像分割方法。该算法重点着手于构建影像图模型,在其中引入像素邻域作用并充分顾及像素光谱测度差异。假定邻域像素具有连接关系,并在此基础上构建影像连接矩阵,再考虑像素光谱测度差异的影响建模像素间相似性,最终结合像素连接性和相似度构建影像权值矩阵完成图模型建立;而后在图模型的基础上,采用对权值矩阵特征分解并就分解结果进行选择的方式将影像数据变至低维特征空间,进而对获取的新数据执行FCM聚类算法达到影像分割目的。为了验证提出算法的有效性,分别对模拟影像和高分辨率全色遥感影像进行分割实验,定性、定量的评价结果表明了该算法的可行性与优越性。

In order to balance the optimal segmentation of region and edge in high-resolution panchromatic remote sensing image,this paper presents a spectral clustering algorithm based on pixel neighbor and spectral features. The proposed algorithm focuses on the construction of the image graph model in which the influence of neighbor pixels is introduced and the difference in spectral measure of pixels is in view. To build the image graph model,firstly,the paper assumes that neighbor pixels can be connected to build affinity matrix,then considers the effect of differences in spectral measure of pixels to model similarity between pixels. Finally,pixel ＇s connectivity and similarity are combined to build the image weight matrix and at the same time the image graph model is completed. On the basis of the graph model,spectral decomposition is applied to the weight matrix and the proposed algorithm can transform the original image data sets into a lower-dimensional eigenspace by filtering results of spectral decomposition. Ultimately,FCM is performed on this new data sets to obtain the final image segmentation results. Experiments are carried out on simulated images and high-resolution panchromatic remote sensing images. The results are qualitatively and quantitatively evaluated and demonstrate the efficiency and superiority of the proposed algorithm.