在教堂或其它建築物中常見到玻璃以色塊方式拼湊出圖案，而色塊與色塊之間有一寬度不等之線條區隔。本計畫的主題為發展一個影像藝術化與風格化之自動處理技術，也就是將影像轉為花窗玻璃化(stained glass)的技術，以期能在此領域有所貢獻，研究內容分為三個部分：第一部分為影像物件之分割；第二部分則是將影像中分割好的物件，進行顏色與紋理的比對來轉換顏色或紋理，第三部分提供對分割的邊線加粗呈現出如教堂玻璃的區塊邊框，進而達到影像如中古世紀教堂中的花窗玻璃之藝術風格化。 在影像分割的方法中，我們利用三種方法分割影像區塊，分別是范諾圖(Voronoi diagram) 分割法、以種子點為基礎做自動分割與人工分割。 在紋理與顏色的置換方面，我們從預先建立好的顏色與紋理影像資料庫取得參考影像，再從提出的數種方法之中擇一，以對分割好的輸入影像之每一區塊賦予想要的紋理與顏色。方法包含顏色全部取代或部分取代、紋理全部取代或部分取代、紋理與顏色全部取代或部分取代，而部份取代也有不同加權值可供選擇，紋理粗細亦可根據隨使用者喜好而調整。 對分割的邊線加粗呈現出如教堂玻璃的區塊邊框線，我們以類神經網路來訓練邊框線，以原分割的細線當輸入樣本，變粗後的邊框線當目標輸出樣本做訓練。 本論文所提之方法，提供使用者多種選擇：分割方法之選擇、紋理與顏色的置換法之選擇、選擇保留或不保留邊線、或是將邊線轉成粗邊框線。因而具相當高的彈性，也得到不錯的實驗結果。 We propose a flexible image stylization system for reforming a source image into a "stained glass" style or some other styles the users desire. It consists of three stages including image segmentation, color or texture allocation, and synthesis of calmes. Each stage provides some options for users to choose. There are three different versions of image segmentation available for the users to choose, including automatic, semi-automatic, and manual versions. The automatic version randomly generates seeds in the image region, calculates the distance between each seed and each pixel, and then assigns each pixel to the nearest seed region to accomplish image segmentation. The semi-automatic version segments the image according to the seeds provided by the users, and then repeatedly adjusts the result by deleting or adding some seeds carried out by the user. The last version is simply manually performed by the users. The step next to segmentation is color or texture allocation. Each segmented region is allocated with color or texture of some reference images chosen from a pre-built database of color and texture images. The proposed system provides many options for this allocation, allowing the users to select one from a variety of methods of texture or color allocation. A unique feature of stained glass is that it is held together with lead, zinc, brass, or copper strips called calmes. We use a BPNN to train a calme generator using an image of line drawing as the input together with an image of the corresponding calme of the line drawing as the output. The users can choose to put or not to put the calme on the resulting images. The proposed system is easy to use and very flexible. It provides many options for users to stylize an image into a variety of stylized results. Our experiments show that the proposed system can generate a wide range of satisfactory results.
