Research On The Man-made Object Extraction From The Aerial Imageries

The ways that we recognize the world have changed fundamentally since the remote sensing technology was invented and applied to acquire information around us. Meanwhile, the depth and width that we know the world is fundamentally enhanced with such technology. However, facing the massive remote sensing images, how to recognize and extract objects from those images is a tough problem in many areas, such as image processing, pattern recognition and computer vision.It is almost more than twenty years for road and building extraction from remote sensing image. However, many models for object extraction are based on hypothesis that building or road had regular shapes or features. As we known, man-made objects are exceptional irregular, and cannot be described by simple models. The active contour model combines both the low and high level vision, and its development in image processing could provide guidance for irregular object extraction.Snake model was put forward almost 20 years ago and widely used in image analysis and computer vision. But it could not deal with curve topological changes automatically. Level set method solved this problem by dimension increasing. However, the Chan-Vese model was based on global information, and could not extract objects accurately and efficiently. In this article, we propose a region-based active contour with narrow band to extract objects of interest. In addition, shadows as the projection of objects, can provide information of man-made objects, but it also influences the extraction of objects. As a result, how to eliminate shadows in the aerial images is also an important research topic in the article.According to the problem of object extraction in remote sensing images, this thesis does some research about building and road extraction, as well as shadow detection. We have the following two contributions: (1) Considering the complicated remote sensing images and the features of shadows, we eliminate shadows based on characteristic approach and HSV color space.(2) We improve the performance of traditional active contours for image segmentation, and propose a region-based model with narrow band to extract objects of interest based on level set method more efficiently and effectively.