擴充局部限制模型之臉部特徵定位系統

Abstract

應用在人臉特徵定位上的局部限制模型近年來已經被證明有不錯的效果，但是局部限制模型以及他的其他相關方法只考慮區域性資訊，因而容易陷入最優化過程中的局部極值問題。在本篇論文裡，我們擴充局部限制模型來改善他的缺點並且得到一個較佳的效果。我們加入了先驗知識來得到一個較穩定的結果，並且將組件的概念引進局部限制模型中，由於組件有優於關鍵點的鑑別力，因此受到局部極值問題的影響會比較小。我們利用一個由組件構成的高階層模型來幫助由關鍵點形成的面部形狀跳脫局部極值位置。另外我們利用三維形變模型來生成二維形狀並以此生成資料來訓練出一個形狀模型，由於生成資料含有三維的資訊，對於臉部姿勢變化比較大的資料會有較佳的定位結果。

The constrained local model (CLM) has been showed the good performance recently for the face alignment. However, the CLM and its variety only consider the local information of landmarks and lead a serious local optimum problem. In this paper, we extend the CLM to improve its drawback and gain a better performance. First, a prior knowledge is used for a stable result. Second, the component information is brought in the CLM formulation. Because the discriminative power of components is better than the discriminative power of landmarks, the components have less influence of local optimums. A high level model which consists of facial components helps the face shape jump out the local optimums. Third, we use the 3D morphable model to synthesize the variant 2D shapes which are taken as the training data. Due to the 3D property, the shape model which is trained by synthetic data has better results for the large pose face.