中文摘要

影像分割技術（Image segmentation）在許多影像處理的工作上扮演著基礎且關鍵的角色，例如物件的識別與辨識（Object recognizing）、影像內容搜尋（Content-based image retrieval）等等之應用，都必須將影像分割為具有某種意義的基本單位，再藉由這些基本單位做更進一步的處理。在近期影像分割技術的發展中，分水嶺分割法（Watershed transform）是一個相當先進的技術，可得到良好的影像分割效果，所以本計畫著重於此一技術的改進及加強。
相較於其他的影像分割方法，分水嶺分割法提供了快速且有效的影像分割，在完成區域分割之後，可保證區域之封閉性；但是，在這個方法中，過度切割(Over-segmentation)的問題常伴隨著發生，造成區域切割的結果過於瑣碎，使得影像分割的成效不彰。因此，本計畫提出一個新方法，利用人類視覺評量系統(Human vision system)針對邊界進行評估，對於間隔差異不顯著的區域進行合併，此法更可容許使用者藉由調整控制參數，動態地修改切割的結果。上述方式可以切割出影像中符合人類視覺的區域，並可使其他影像處理應用能有更進一步的發展。

關鍵詞：分水嶺演算法，視覺化，區域整併

Abstract

Image segmentation techniques are important but difficult in many image processing topics, such as object recognition and content-based image retrieval. In order to solve those problems, a successful image segmentation method is essential for splitting an image into meaningful regions. Recently, the watershed transformation develops quickly and applied successfully in image processing applications. Several approaches have been proposed for image segmentation using the watershed techniques and performed very well.
The watershed transformation is one of the most efficient unsupervised segmentation algorithms. Moreover, the watershed transform always produces closure segmented regions in the image. However, the conventional watershed transform may split a meaningful region into many unnecessary small regions. That denotes the over-segmentation problem. Hence, in this project, we evaluated the dams between regions by using the properties in the human vision system (HVS). Regions will be endeavored to merge when the separated dams are unremarkable. This work deconstructs some weak dams to diminish over-segmentation in watershed transform. Furthermore, the proposed method gave the distinct segmentation results by dynamically modifying some user-controlled parameters. This interactive scheme would perform more meaningful regions by conforming to the HVS. This work may become a fundamental and useful tool for the further develops in image processing applications .

Keywords: immersion watershed algorithm, watershed accuracy, human perceptual