基于显著图的遥感图像多分辨区域生长分割算法

针对传统区域生长算法对噪声敏感和初始种子过度依赖的问题,本文提出一种基于显著图的遥感图像多分辨区域生长分割方法.该方法利用亮度、颜色、方向三个特征金字塔生成显著图,通过视觉选择注意模型自动选择注意区域作为种子区域.从能分辨种子区域的最大尺度开始区域生长,直到0尺度,从而分割出对遥感图像中感兴趣的区域.实验结果表明该方法能有效地从遥感图像中分割出视觉注意的区域,且有较快的速度.

The Multi-scale Region Growing Method Based on Saliency Map for Image Segmentation

Traditional region growing segmentation algorithm has some problems such as too sensitive to image noise and over-dependence to the initial seed points.In order to solve these problems,a region growing segmentation algorithm is proposed based on the significant map and Gaussian pyramids.By creating the intensity,color and orientation pyramids,significant map which is used to extract seed regions is created.Then map the seed region to the corresponding scale of the pyramids and growing in a regular way.Reserve the edge pixel of region and determine if they are belonged to the same segmented area or not in another fine scale of Gaussian pyramids.The approach can realize the automatic selection of threshold and seed area,and extract meaningful object.The experiment results show that the algorithm is effective.