高分辨率影像分割的分形网络演化改进方法

分形网络演化是针对高分辨遥感影像的高精度分割方法。它是以像元自下而上进行地物域合并,直至满足区域对象间异质性值大于预设阈值,停止区域合并得到最终分割结果。当对大数据量遥感影像进行分割时,形成初始区域对象的速度较慢,并且数量较多,导致分割时间长,有待在整体分割效率上进一步提高。一种有效的改进措施是采用某种分割方法,快速生成初始区域对象,然后再以初始分割结果区域对象进行区域合并。本文提出一种自动种子点的并行区域生长分割方法,用于快速生成初始区域对象;提出均匀数据划分的并行区域生长策略及消除数据划分线两侧的区域对象方法;采用OpenMP并行技术实现并行区域生长过程。分割效果对比和效率分析结果表明,本文提出的初始分割方法效率较高,并且分割结果可重现,从可信度、通用性角度来看,具有较高的实用价值。

An Improved Method of FNEA for High Resolution Remote Sensing Image Segmentation

Fractal Net Evolution Approach （FNEA） is a high precision algorithm for high resolution remote sens-ing image segmentation. The segmentation algorithm starts with single image objects of one pixel and repeatedly merges them in several loops in pairs to larger units as long as an upper threshold of homogeneity is not exceed- ed locally. However, it＇s high computational cost for large images such as the currently high resolution remote sensing images includes two aspects, one is the time-consuming procedure of creation of initial image objects, and the other is the great amount of initial image objects. In order to boost the speed of the algorithm, generally, an effective improvement measure is to select a faster segmentation method to generate the initial image objects, and then carry out the region merging phase. Thus, the main focus of this paper is to tackle this problem by using a parallel process to segment the original image into subsets as the initial image objects firstly. For the original image data preprocessing, a regular data division way is used to divide the original image data into sub-rectangu-lar data blocks, which will be used as the input data and assigned to different threads for the parallel computing. We introduced an improved method of fractal net evolution approach, the main work as follows： （i） Presenting an automatic seed selection method. The initial seeds are automatically selected from the original image data and the seeded image objects are produced where each image object corresponds to a seed. （ii） Proposing a parallel region growing strategy upon the data paralleled segmentation. Moreover, the strategy solves the problem of merging image objects on both sides of the dividing lines as well. （iii） Using the OpenMP parallel technology. Experimental results, including comparison of final segmentation and assessment of computing efficiency, show that the improved method is more effective and the final segmentation result is reproducible. Thus, the generality and reliability of the method proved the practical value of our work.