Tree Occlusion Segmentation Of Multispectral Images For3D Building Reconstruction

Tree occlusion removal from an image is a technology to extract an tree from its surrounding scene in an image that is taken under the ground platform, which can provide the basis data and technical support for3D reconstruction and vegetation extraction. Since trees and related surrounding objects are various, the work of tree occlusion removal become sophisticated and difficult. This research is of great utilizable value and practical significance. This paper views the vegetation occlusion removal in building images as a research background and makes research on segmentation method of tree extraction based on Matlab, in term of the unique characteristics of trees in natural images especially the building images. A novel method of vegetation removal is proposed in this paper which is based on CIE L*a*b colour space and iterative self-organizing data analysis technique algorithm (ISODATA) for vegetation auto-extraction. Different approaches, such as spatial topological relationship, key feature ratio (KFR), Otsu, morphology and clustering algorithm etc, are combined to extract vegetation automatically. Three main tasks have been done in this dissertation which are listed as follows:Firstly, the paper researches the application of the image segmentation method in trees occlusion removal. Methodologies of image segmentation can be summed up into four types:threshold segmentation, edge detection, clustering algorithm and segmentation based on region. Classical methods for image segmentation such as Otus, Ncut, Gabor, and CV are utilized to segment trees in building images. This paper also analysis accuracy and performance of these approaches with the criterion of Kappa coefficient.Secondly, research based on feature in CIE L*a*b color space, spatial topological relationship and key feature ratio is made to extract trees in an image. Otsu is utilized to segment a*channel and L*channel imagery to obtain a primary area which contains vegetation. Meanwhile, spatial topology relationship and KFR are presented to eliminate non-vegetation areas which can not be deleted by morphology, and this can extract vegetation from a given building picture primarily. It is worth mentioning that ISODATA is utilized to obtain an adaptive threshold with a promising result. The presented approach which adopts adaptive threshold can be extended in dealing with the removal of yellow vegetation and has an encouraging performance. After the above work CV model can be applied to better according result.Finally, this paper makes full use of CV model to better the result of vegetation extraction. CV model aims to find contours of an object by minimizing an energy function. CV model alone is not applicable for vegetation segmentation since CV can not distinguish vegetation areas and non-vegetation areas only by dividing an image into two parts. In this paper, the harsh contour of tree in an image can be acquired by feature in CIE L*a*b color space, spatial topological relationship and key feature ratio, based on this result, CV is used to extract trees more accurately. Experimental results show that the proposed method is desirable and with low computational complexity and low level of automation.