分水嶺轉換是一種在影像處理與分析領域中，經常被用作區域性影像切割的方法。分水嶺轉換的概念是基於：模擬大水逐漸淹沒一塊崎嶇不平的地形時，建築水壩防止湖泊合併的過程。本篇首先介紹基於上述概念所設計出來的分水嶺轉換演算法的類型，嚴謹地描述這些演算法的定義與流程，以及說明各種分水嶺轉換演算法可能遭遇的各種問題，並提出或整理解決這些問題的方法。 此外，本篇論文提出了兩個新穎的分水嶺轉換的相關方法。首先，在微小且低對比的目標物的偵測問題上，我們提出了一套有效去除雜訊的方法，並搭配適當的分水嶺演算法，能夠迅速並正確地在動態影像中，偵測到微小且低對比的目標物，並完整地萃取其外型輪廓。另外，我們還提出了一個使用分水嶺轉換來作資料分群和分類的方法，稱作「分水嶺分類法」。絕大多數有關分水嶺轉換的應用都是在影像相關的資料上，分水嶺分類法可對任何型態的資料進行分類的動作，並且不需要資料本身相關知識的介入，資料的分類方式透過決策區域來完成，而不須基於決策理論來進行分類，此點有別於傳統的分類演算法。分水嶺分類法分為非監督式和監督式兩種，監督式的分類法可用來強化非監督式的分類結果。 本篇內容介紹了以上所述的兩個分水嶺轉換的相關方法，並以實驗結果證明其可行性及適用性，最後針對這兩個方法作出總結並提出未來可以改進的方向。 Watershed transform is usually adopted for image segmentation in the area of image processing and image analysis. The concept of watershed transform is based on a processing simulating the immersion of a landscape in a lake that is dams have to be built to prevent the merging of different catchment basins. In this dissertation, the algorithms of watershed transform are firstly introduced. The definitions and procedures of watershed transform will also be thoroughly depicted. Problems that might occur in the watershed transform are addressed and solutions are proposed. Two novel methods utilizing watershed transform are proposed in this dissertation. First, we proposed an effective noise removal method to resolve the problem of small object detection with low contrast. By integrating with an appropriate watershed algorithm, our proposed method can efficiently and effectively detect small objects with low contrast, and extract their complete contours. Moreover, we propose a method call “watershed classifier” for data clustering and classification using the watershed transform. Most watershed algorithms are utilized for image data, whereas the proposed watershed classifier is capable of classifying arbitrary data without prior knowledge. Unlike traditional data classifiers, the task of data classification of watershed classifier is carried out through the decision regions directly instead of relying on the decision theory. The watershed classifier can be either unsupervised or supervised. The supervised version of the watershed classifier is also devised to enhance the unsupervised classification performance. Experimental results demonstrate that the feasibility and validity of the proposed watershed classifier in data or image classification.
