階層式的影像表示法

Abstract

影像表示法在許多電腦視覺應用中是一個重要的步驟. 人類對自然界物體 的視覺能力是具有階層式的結構. 一般而言, 對物體的微觀能獲得物體詳 細的了解. 相反的, 對物體的巨觀能獲得物體大略的認識. 一個真正具有 智慧的電腦應該有對物體做類似人類階層式視覺表示的能力. 對物體大略 的表示只需要較少的特徵, 而詳細的表示則需要較多的特徵. 然而在電腦 視覺中要判斷一個特徵的重要性是非常困難的, 因此多重量度和多重解析 度等方法已被引用來減少特徵的數量.在這篇論文中我們提出一個階層式 的影像表示法. 這個階層式的影像表示是由一群近似體的表示所構成. 每 一個近似體的表示又包含了三個層次:點層, 線層和結構層.這篇論文的主 要貢獻是我們提出一個類似人類視覺能力的階層式表示法,而且定義了一 個方法去衡量所淬取的特徵和實際物體的相似度. 更重要的是, 我們提出 一個有效的結構分割方法. 對電腦而言這是不容易的. Visual concepts involved in real world usually possess graded structures. Generally speaking, macroviews of objects can obtain global perception; conversely, details of objects be obtained by microviews. A really intelligent computer vision system must respond to the visual perception in the similar way as a human does. In graded representation, rough description only needs few important features. The finer the representation would be, the more the features should be included. However, the importance of a feature cannot be easily judged by computer vision systems. Instead of selecting features derived from the primary object, most researcheres employed multiscale or multiresolution approach to reducing the number of selected features. This thesis presents an approach to constructing a graded representation for shapes. A graded shape representation is derived from a set of approximation representations. Each approximation representation is further divided into three levels: pixel level, token level and component level. The contribution of this thesis is that it proposed a new approach to representing object shapes like the perceptive way of human being. The approach also defines a ratio to measure the similarity. Most importantly, an efficient procedure of part decomposition was proposed. As we know, this task is not easy for computer systems.