Studies On Extraction Techniques Of Primary Objects From Urban Aerial Images

As an important information resource, primary objects extraction from urban aerial images is an internationally advanced research field in photogrammetry, remote sensing and computer vision. This research has significant academic and practical value, which mainly focused on building and road extraction. The first chapter systematically introduced the state of the art and indicated that, traditional model-matching algorithms cannot effectively extract the high-rise buildings and main roads, which are of great significance for modern cities and have diverse and complex appearances. But human beings can identify and localize these objects just at a glance. So an important way to solve these problems is to explore and imitate the cognitive mechanism and capability of human being's brain.In the second chapter, some fresh achievements in AI about cognitive mechanism of brain were introduced firstly. Furthermore, incorporating with the factual situation, the systematic frame and implemental steps for aerial image's cognition, which is supervised by the brain, were developed. Mainly, we proposed an automatically extracting methodology based on the OAR (Object, Attribute, Relation) model of objects, and expounded implemental steps and methods of the inferentially cognitive process in terms of objects' attributes and relations.The third chapter studied the preprocessing of aerial images based on visual perception model of the eye. The Retinex favorably mimics visually perceptual mechanism of the eye. But there are some deficiencies when Retinex confronted great illuminant variation. In order to solve this problem, a method for shadow detection was presented firstly. Furthermore, a fuzzy Retinex algorithm, which adapts for the illuminant variation, was put forward. Experimental results demonstrated that our algorithms are effective on image enhancement and shadow removing.The fourth chapter elaborated the methods for automatic extraction of buildings. First, the OAR model and the corresponding systematic extraction frame for buildings were developed. Secondly, the vanishing point of vertical lines was inferred, and an adaptive fuzzy Hough transform was proposed to detect building's vertical lines. Then, based on the Radon transform and the time-frequency analysis, an algorithm for analyzing and describing of window textures was brought forward to verify the exact locations of vertical lines and to calculate local horizontal orientations of walls. Lastly, a line Snakes method was advanced to integrate multiple useful information and extractthe whole roof boundary. This procedure accomplished the automatic extraction of high-rise buildings from monocular high-resolution aerial images.In the fifth chapter, automatic extraction of roads was discussed. Incorporating with the OAR model of main roads, a new feature vector of main roads, symmetrical edge orientation histogram (SEOH) was put forward. Then, the systematic frame of main road extraction based on SEOH was developed. After that, a new method of FCM considering the distribution of spatial data was presented to cluster image patches based on their SEOHs. and to extract main road seeds. Finally, the Snakes algorithm was adopted to fuse characteristics of main roads in a global level and to extract the whole main road network. Experimental results demonstrated that the presented algorithm solve satisfactorily the automatic extraction of main roads.