In this thesis, we propose a novel system to estimate the intensity of different human facial expressions. Based on the modified optical flow computation, we model the motion of facial pixels instead of relying on complex muscle model and reduce the demand to detect many facial feature points. According to the assumption that faces of the same expression are close in the expression space, this thesis also proposes a novel weighting scheme for the facial optical flow field by using a quadratic programming formulation. Experiments manifest the efficiency of the proposed system on the expression intensity estimation and expression recognition of face images.
在這篇論文中,我們提出一個可在不同人臉分析表情強度的分析系統。由於我們使用整張人臉的光流場向量作表情分析的特徵,所以並不需要依賴過多的人為控制即可做辨識。 首先,我們假設即使是不同個體在表現相同表情時,會有部分的臉部肌群是呈現類似的運動,因此我們設計一個新的特徵去描述該運動的方向性,並且 將整個問題轉換為一個有條件下的二次數學規劃問題( constrained quadratic programming ),透過統計找出對某種表情較具有代表性的區域,在做表情分析及表情分類時,提升該部分特徵的參考價值,進而做出較正確的判斷。