基于多尺度多特征的高空间分辨率遥感影像建筑物自动化检测

Automatic building detection of high-resolution remote sensing images based on multi-scale and multi-feature

建筑物检测在城市规划、变化检测、地表覆盖等方面均起到重要作用。然而高空间分辨率遥感影像(简称"高分影像")中建筑物朝向不一,形态颜色各异,大小尺寸也有着较大差别,使得建筑物检测成为一道难题。为此,提出一种基于多尺度多特征来自动化检测高分影像中建筑物的方法:首先,对影像降采样构建高斯金字塔模型,固定尺度的滑动窗口在不同层影像中对应着不同的实际地面面积;然后,对影像进行超像素分割并计算滑动窗口中多种描述建筑物特性的特征值,通过多特征融合来衡量建筑物目标在不同尺度影像中的显著性;最后,计算超像素块的显著性均值,结合Otsu算法自动求取阈值,进一步设置长宽比等约束条件,从而准确、自动地提取建筑物目标。分别采用空间分辨率为0. 5 m和0. 2 m的影像进行实验,并和基于颜色和纹理建模的马尔科夫随机场模型算法进行定性和定量的比较。实验结果表明,该方法对高分影像中建筑物的提取有更好的实际效果和检测精度。

Building detection plays an important role in urban planning,change detection,surface coverage and so on.However,in high resolution remote sensing images,buildings vary in shape,color,and size,which makes building detection a difficult problem.Therefore,this paper proposes a method based on multi-scale and multi-feature to automatically extract buildings in high resolution images:Firstly,down sampling images are used to construct Gauss pyramid model,while fixed size windows in different layers of pyramid image represent different ground areas.Then multi features are calculated which describe building characteristics by sliding windows,and multi features are fused to evaluate the saliency of building in different scales.Then the saliency of superpixels is calculated,and Otsu algorithm is used to automatically determine the threshold,and furthermore,some constraints such as the aspect ratio were combined to extract buildings accurately and automatically.Experiments were made by 0.5 m and 0.2 m high resolution remote sensing images in comparison with the markov random field model based on color and texture modeling algorithm for qualitative and quantitative comparison.The results show that the method suggested in this paper can obtain more satisfactory precision and has higher effect on building detection from high-resolution remote sensing images.