改进的Mean-Shift遥感影像分割方法

摘要

Due to low convergence speed of Gauss kernel Mean-Shift, it couldn't satisfy real time application of remote sensing image processing, an improved mean-shift segmentation approach for remote sensing image was proposed. In order to overcome some defects of fixed bandwidths of classic Mean-Shift, referenced spatial bandwidths were given according to the spatial resolution of remote sensing images, and various hue range bandwidths of each band were calculated using plug-in rules. In order to solve the problem of spending much time in Mean-Shift iteration caused by the voluminous imagery data set, some accelerating strategies were proposed. As the number of mode detected by Mean-Shift algorithm were large, global based mode merging method was proposed to stabilize the segmentation result of remote sensing images. Multi-dimension features were used and calculation speed was advanced, color features were used in mode detection and texture features were used in mode merging. Segmentation experiments were processed with Quickbird satellite image. The results show that the presented method is adaptive, and its speed and precision can satisfy application, so it is a robust automatic segmentation algorithm.%基于高斯核Mean-Shift(MS)算法因收敛速度慢难以满足遥感图像处理的实时应用要求,提出一种改进的MS算法应用于遥感图像分割.针对传统MS算法需多次人工试用来确定固定带宽的问题,给出几种类型遥感影像的空间带宽参考值,且不同波段影像用plug-in规则分别计算值域带宽.针对遥感影像数据量大、MS迭代计算时间长的不足,使用一些加速策略来加速收敛；由于采用MS算法检测出的模点数较多,采用基于全局模点融合来稳定遥感影像分割结果；可用于分割遥感影像的特征很多,模点检测时用灰度特征,全局模点融合时用纹理特征,这样充分利用了遥感影像多维特征且不降低计算速度.采用Quickbird影像进行分割试验,研究结果表明:本文算法自适应程度高,速度和精度也能满足应用要求,是一种稳健的自动分割方法.