Patch-based image segmentation of satellite imagery using minimum spanning tree construction
Patch-based image segmentation of satellite imagery using minimum spanning tree construction
复制标题
DOI：
--
复制
发表时间：
2010
期刊：
影响因子：
--
通讯作者：
A. Skurikhin
中科院分区：
文献类型：
--
作者：
A. Skurikhin
提出了一种分层图像分割和特征提取方法。该方法结合使用Canny边缘检测和图像拉普拉斯检测来检测图像的光谱不连续，然后构建连续降低细节级别的分割图像的层次结构。这些图像被表示为具有光谱和结构特征的多边形像素块（多边形）集。这个层次结构构成了面向对象图像分析的基础。为了建立原始图像的精细细节表示，种子分区（多边形）建立在由不规则大小的三角形组成的三角形网格上，其空间排列与图像内容相适应。这是通过在检测到的光谱不连续性的顶部构建三角形网格来实现的，这些不连续性形成了Delaunay三角剖分的约束网络。多边形化图像以图形的形式表示为空间网络，图形的顶点对应多边形分区，图形的边缘反映成对分区关系。图像图划分基于Boruvka最小生成树算法的迭代图收缩。该方法的一个重要特点是分割的聚集受到检测到的光谱不连续的约束；因此，聚集分区的形状更有可能与现实世界对象的轮廓相对应。«少
We present a method for hierarchical image segmentation and feature extraction. This method builds upon the combination of the detection of image spectral discontinuities using Canny edge detection and the image Laplacian, followed by the construction of a hierarchy of segmented images of successively reduced levels of details. These images are represented as sets of polygonized pixel patches (polygons) attributed with spectral and structural characteristics. This hierarchy forms the basis for object-oriented image analysis. To build fine level-of-detail representation of the original image, seed partitions (polygons) are built upon a triangular mesh composed of irregular sized triangles, whose spatial arrangement is adapted to the image content. This is achieved by building the triangular mesh on the top of the detected spectral discontinuities that form a network of constraints for the Delaunay triangulation. A polygonized image is represented as a spatial network in the form of a graph with vertices which correspond to the polygonal partitions and graph edges reflecting pairwise partitions relations. Image graph partitioning is based on the iterative graph oontraction using Boruvka's Minimum Spanning Tree algorithm. An important characteristic of the approach is that the agglomeration of partitions is constrained by the detected spectral discontinuities; thus the shapesmore » of agglomerated partitions are more likely to correspond to the outlines of real-world objects.« less