數值影像中線形地物萃取之研究

Abstract

本論文旨在運用不同類型之特徵及萃取方法，以提昇地物萃取效率及自動化，獲取空間資訊。並針對不同尺度之航攝影像、地形圖、衛星影像等三種資料來源，以線形地物之萃取為例，透過階層式處理達成地物測繪目標。 線形地物萃取之程序分別為：１. 線形地物與非目標物之區分、２. 地物邊緣或骨架特徵偵測、３. 特徵點選取、４. 線形追蹤與連接及５. 點位精進。文中就各階段處理目標及方法比較分析，增進程序中方法之效能。整合二種不同的機率性Hough轉換法之優點，並加入方向約制條件改進轉換點對之組成方法，發展出方向約制機率Hough轉換法。同時應用Hough轉換法於地物萃取中，自動選取道路特徵點，提供道路追蹤檢驗之依據。再以方向追蹤模塊，加快線形追蹤效率。最後，以遺傳演算法，進行點位精進工作。

The primary purpose of this thesis is integrating different features and extract method to promote efficiency and automation for the acquisition of man-made objects information. Aerial images, topography maps, and satellite images, three different kinds of data sources are analyzed. A hierarchical process in extracting the linear features is approached to satisfy the mapping purpose. In this work, there are five major procedures proposed in the linear feature extraction, including: (1) linear object decomposition, (2) edge or skeleton detection, (3) feature points selection, (4) line tracking and linking, and (5) feature points refinement. To progressive the performance of linear feature extraction, the analysis and comparison of various methods in each procedure are made. An orientation constrained probabilistic Hough transform (OCPHT) is proposed and applied to select points automatically on the linear features for the verification of linear feature tracking. The OCPHT algorithm combines advantages of two different probabilistic Hough transform methods and introduces an orientation constrained condition to improve the composition of transformed point pairs. Eight tracking templates are then used to fasten searching efficiency in the tracking. Finally, a Genetic Algorithms (GAs) is approached to refine the location of extracted points for the linear features.