建筑物近景可见光影像中树木遮挡区的非监督提取

Unsupervised Tree Occlusion Removal for Close-Range Building Images Under Condition of Visible Light

针对已有算法在可见光条件下不能完整提取树木或者需要大量人工干预的缺陷,提出了一种基于CIE L*a*b色彩空间的利用冠层与树干空间拓扑关系和关键特征比值的非监督树木分割算法。首先,将RGB转换到CIE L*a*b色彩空间;然后,选择该空间中的关键特征进行Otsu自适应阈值分割;最后,利用数学形态学对分割区域进行去噪,结合冠层与树干的空间拓扑关系和色彩特征比值关系剔除伪树木区域。实验结果表明,本文方法能成功分割出冠层和树干遮挡且自动化程度高,可为近景建筑物三维重建的遮挡区修补提供检测基础。

A new approach based on spatial relationships and key features ratio in CIE L＊ a ＊ b color space is proposed for vegetation extraction to solve current problems in tree occlusion removal algorithms for visible light images, such as incompletely detected tree areas and mass human-computer interaction. First, the color space of an image is transformed from RGB to CIE L ＊ a ＊ b. Second, the classic Otsu method was used to segment the L channel and a channel image. Finally, morphology modification; the spatial topological relationship between crown and trunk and key feature ratio, were applied to acquire the final vegetation areas. Experimental results show that the proposed algorithm can successfully remove vegetation occlusions （i. e. canopies and trunks） with a high level of automation. This study can pave the way for occluded image repair and 3D reconstruction of close-range building images.