在本論文中，我們提出以決策樹建構法為基礎的二維多重分閥值法來解決影像分割的問題。此法使用非監督的叢聚樹演算法，將二維灰階統計直方圖當作資料空間，來分析像素值之間的相似程度並予以分群。這個方法包含了二個步驟：叢聚樹的建構以及修剪。在叢聚樹建構的過程中，我們利用資訊增益為準則，採取各個擊破策略法於二維灰階統計直方圖中分別求取各維度的最佳切割點，直到每個節點之資料量或相對密度皆符合自訂的條件為止。為了簡化叢聚樹以得到有意義的群數，因此使用區域與其相鄰區域之相對密度來進行修剪，以獲得最適當的分割結果。為了證明本方法的有效性，我們進行多組影像的實驗，其中包含了人造與真實影像。分割結果與L. Cao所提的快速自動多重分閥值法作比較。由實驗結果顯示，我們的方法確實能夠獲得較好的分割結果。 In this thesis, we propose a two-dimensional multilevel thresholding based on decision tree construction for image segmentation. In this method, we use unsupervised cluster tree method to discriminate the different objects of the two-dimensional gray level histogram. Our method consists of two steps: cluster tree construction and cluster tree pruning. In the cluster tree construction step, we use the gain criterion to select the appropriate cut in the two-dimensional gray-level histogram until every node’s number of data points or relative density conform to user-specify parameters. In order to simplify the tree to find meaningful clusters, we prune tree according to similarity between the region’s relative density and adjacent regions’ relative density. In experiments, we take many samples, including real and synthetic images, to demonstrate the effectiveness of our proposed method. The results compare with the ones of L. Cao et al. that propose the fast automatic multilevel thresholding method. Experimental results reveal that our method is superior to L. Cao's work.
