Research On Building Extraction And Region Classification

Automatic extraction of buildings has become a topic of growing interest for computer vision field. Recently, many fruitful applications have been developed in this domain, such as automatic information extraction from images and updating geographic information system databases, three-dimensional building construction, and so on. In this thesis, we focus on automatic building contour extraction and region classification problems from plan and remote sensing image.Main points and innovative proposals are as follows:1. Considering the characteristics of remote sensing image and plan, we analyze the difficulties of image segmentation and contour extraction. In order to overcome some disadvantages of traditional segmentation methods, we combine split &merge and snake model and present a new automatic building contour extraction approach. The experimental results on the images of plan and remote sensing image show that our proposed method outperforms the split & merge and snake model. Moreover, this new approach remarkably significantly reduces the run time.2. After obtaining the contour image, we utilize BP neural network to distinct whether the region is a building or not. The input of BP neural network has two nodes. One is the ratio of the area to the perimeter of a region, and another is the corner of a region that is calculated by Harris descriptor. The output of BP neural network is a binary code that indicates the building and non-building region. The experimental results demonstrate that our proposed method can achieve 95.07 percent on real dataset.