水平集分层分割遥感图像中的建筑物

针对高分辨率遥感图像，结合建筑物特征，提出水平集分层模型分割图像中的建筑物。首先，学习植被样本得到其在HSV空间中色调与饱和度的联合分布函数，利用阴影灰度方差通常小于非阴影区域的特点，将植被和阴影剔除以简化背景利于后续分割。然后，根据灰度级高低将一幅图像看作多层图像层，把建筑物的屋顶灰度特征和边缘特征融合到传统Chan-Vese（C-V）水平集算法中，分割出每层中灰度级相似的建筑物候选区域，从而将不同灰度级建筑物候选区域分层分割出来再整合。最后利用建筑物面积、建筑物与阴影位置关系等先验知识排除误分割，得到最终结果。实验表明:该方法能更好地分割出形状各异、各个灰度级的建筑物，甚至是灰度不均匀的建筑物，分割漏检率较传统C-V法降低了25%，虚检率降低了22%。有效减少了漏分割和过分割。

Automatic building segmentation from remote sensing images using multi-layer level set framework

Towards high resolution remote sensing images, combining with features of buildings, a novel method to extract buildings based on multi-layer level set framework was proposed. Firstly, as far as the impact of shadow and vegetation was concerned, it should be removed on the basis of the separation of gray value thresh and the joint distribution of hue and saturation. Then, an improved C-V level set segmentation algorithm combining with building features of roof's gray and obvious boundaries was applied to extract building regions of similar gray-scales on each gray layer, and thus all building regions of different gray-scales could be extracted layer by layer, followed by layers of segmented regions integration. Finally, the non-building regions were excluded by using normal areas of buildings and related position between buildings and shadows. The experiment results demonstrate that, compared with the traditional level set methods, this one can detect each single building of gray heterogeneity and buildings of multiple shapes and different gray-scales. Meanwhile, compared to the traditional C-V method, it largely reduces the leakage segmentation ratio by 25% and over-segmentation by 22%.