以真實影像為基礎之全自動水彩風格轉化

Abstract

本研究設計一個由電腦根據使用者輸入目標影像來產生水彩繪畫的濾鏡系統，透過在空間域下的影像處理，以九個步驟模擬下列六個真實水彩畫中的顯著特色： (1)用色差和亮度兩條件來決定是否要將像素設定成白色，製造上色時有留空未塗的狀況 (2)以具有尺度感知特性的濾波器框架將影像重要邊緣保留，造成如同畫家作畫時將部份不需強調之細節予以輕描淡寫而僅強調目標之效果 (3)給像素隨機的位移量並和中值濾波模糊化配合，產生不穩定的繪畫手感和顏料渲染 (4)對邊緣偵測影像進行線段的外觀調整，使之看起來像是鉛筆描線的輪廓圖 (5)將不同尺寸的雜訊影像疊加，以該隨機樣式當顏料推積與擴散的擾流現象 (6)與真實紙張影像的合成，可以讓紙張紋路像是在透明水彩下顯露 結合上述技巧處理後的影像即會擁有水彩韻味且貼近真實畫作之輸出結果。

The thesis presents an image-based watercolor rendering system which is designed to simulate and to fulfill six important characteristics of real watercolor paintings using nine steps of image processing in spatial domain.

These characteristics include: (1) Using the color difference and luminosity conditions to judge whether a pixel is white or not for a situation to preserve blanks. (2) Using a scale-aware filter framework to preserve the most significant edges to reach the effect of salient regions and simplify the regions that need not be emphasized. (3) To give pixels random offset and use median filter for blurring that produces unstable hand drawing and pigments diffusion visualization. (4) In order to assimilate the pencil outline drawings, the appearance of the coherent edge line should be adjusted. (5) To use different sizes of random noise pattern for the turbulent flow. (6) To combine with scanned paper image in order to reveal the paper texture under the watercolor pigments.

After the processing of the above methods, the output image not only mimics an artistic-verisimilitude watercolor painting but also proves that the system is well and fully functional.