Facial Expression Synthesis from a Single Image

本論文提出了一個基於影像為基礎之臉部表情模擬系統。給定一張正臉、不帶表情的來源人臉影像，此系統能模擬出來源人臉在帶有表情時的影像。在臉部動畫和人機互動的領域，以影像為基礎的臉部表情模擬技術已經是相當普遍的研究主軸。即使已經有很多相關的研究，但大部分的研究仍是在建立在二維空間上，因此忽略了因為表情改變所造成的肌肉變化。不同於之前的研究只著重在二維空間，我們結合二維空間方法和三維空間方法的優點，提出一個整合的系統。一方面，我們利用一個省時的臉部模型重建方法取得三維空間中較準確的臉部幾何資訊。另一方面，我們利用從三維空間取得的資訊在二維空間上模擬表情，產生出一個自然的模擬結果。為了取得帶有表情的影像，我們提出了以位移為基礎的表情模擬方法，此方法只利用臉部的48個特徵點位移，就能估計出所有臉的位移。實驗結果可以證明我們的系統可以產生出令人滿意的模擬結果。This thesis presents an image-based facial expression synthesis system that synthesizes expression on the input neutral facial image with frontal view. Image-based facial expression synthesis has been a popular topic with many applications, such as facial animation and human-computer interaction. Despite many research advances in recent years, most of them are manipulated only in 2D domain and thus neglect the muscle deformation due to expression. Different from traditional 2D methods, the proposed approach is an incorporated system which combines the advantage in 2D and 3D methods. On one hand, the more accurate geometry information is exploited from 3D by applying a time-saving face model reconstruction method. On the other hand, expression on 2D is synthesized using the information from 3D to produce a natural synthesized result. To obtain the expressive image, a displacement-based expression synthesis method is proposed that uses only the displacement of 48 facial feature points to approximate the displacement of all face. Experimental results show that our system can generate satisfactory synthesized results