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

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.