以統計學習的方式從有表情的臉部圖片產生誇大漫畫圖的系統

• Main Authors: Ting-Ting Yang, 楊婷婷
• Other Authors: Shang-Hong Lai
• Format: Others
• Language: en_US
• Online Access: http://ndltd.ncl.edu.tw/handle/59508020547562017396

碩士 === 國立清華大學 === 資訊工程學系 === 96 === In this thesis, we proposed a learning-based system for generating exaggerative caricatures with expression. This system is capable of learning the drawing style of artists from their caricature works as the training data, and automatically creates exaggerative neutral/angry/happy caricatures from neutral/angry/happy images. Most previous works can only deal with frontal-view faces with neutral expression without glasses or hats, and cannot apply more than one drawing prototype learned from the caricatures drawn by a cartoonist at a time. The proposed caricature generation system exaggerates facial images with expression and learns the drawing prototypes from training data as well. The generation process is decomposed into three parts: facial feature exaggeration, texture transformation, and texture mapping. To learn how the cartoonist exaggerates the facial features of distinct facial expressions, the system analyzes the correlation between the photo caricature pairs using LPH (Locality Preserving Hallucination). Then apply Sobel edge detector as well as information of feature points to synthesize the desired texture. After combining exaggerated feature shapes with sketches by RBF (Radial Basis Function) warping, we can obtain caricatures with desired exaggeration. Experimental results show that our system can capture some features selected by the artist and exaggerate them in similar ways.