一种改进的多尺度分形网络演化分割方法

A Advanced Multi-Scale Fractal Net Evolution Approach

面向对象技术是提高高分辨率图像分类精度的关键技术之一。针对eCognition的分形网络演化分割算法仅仅采用光谱特征和形状特征进行图像多尺度分割的不足,提出了将边缘特征引入其分割过程中,以提高多尺度分割的效果,来获得边缘平滑且分割对象与实际地物更加一致的分割结果,从而改善eCognition的分形网络演化分割方法中存在的过分割、欠分割和边缘粗糙等情况。通过实验,证明了引入边缘特征的分形网络演化分割方法提高了图像分割结果,减小了过分割和欠分割的产生,使得分割结果与实际地物更加一致。

The image analysis software,eCognition has been used more than ten years.This software adopts an object-oriented method to improve classification accuracy of high-resolution data.One of the key technologies is the Fractal Net Evolution Approach(FNEA)to be used for multi-scale segmentation.In this paper,we analyzed some shortages of the FNEA.For example,the FNEA only adopts the spectral and shape features.This paper introduced the multi-scale edge feature into the multi-scale segmentation process to improve the segmentation results,which have smooth edges and are consistent with the targets as soon as possible in the earth's surface.In theory,this improvement can reduce over-segmentation and under-segmentation and improve results of the FNEA.We conduct experiments to segment high-resolution imagery by fusing the edge and spectral features.The result proves that our method can advance the segmentation results and reduce the over-segmentation and under-segmentation.