主動外觀模型(Active Appearance Models，簡稱AAMs)是一種表示影像中在形狀與紋理都具變異性之非剛體物件的方法，並提供此物件一個模型化(Model-based)的表示方式。它利用一個平均值和一組模式(modes)的線性組合，來表示欲分析的非剛體物件，並且藉由改變模型中模的線性組合之係數(或稱模型參數)，可使模型合成出各種變異的非剛體物件。因此，透過AAMs這種表示法，我們便可用模型化的方式，表示人臉的表情。以AAMs進行人臉表情辨識的應用，需要可找到最佳模型參數的搜尋演算法(AAMs search algorithm)，使得模型在這組參數下所合成出的表情，能夠近似於影像中的人臉表情。如此，我們便可透過分析模型參數的方式，進行人臉表情辨識。 本論文提出一個疊代式AAMs參數搜尋演算法，以傳統的AAMs參數搜尋演算法為基礎，將量測模型與測試影像之誤差函數最小化，只採用該方法所搜尋之參數變化量的大小，在參數搜尋的方向上，我們估測每個疊代下誤差函數的梯度，以決定參數的搜尋方向。本論文更進一步提出，將搜尋到之參數做一個微小的擾動，以防止搜尋的結果掉入誤差函數之局部最小值中。 實驗結果顯示，本論文所提出之強健式主動外觀模型搜尋演算法，相較於傳統的尋演算法，平均降低人臉表情形狀的位置搜尋誤差36.41%。在人臉表情紋理的搜尋上，則平均減少30.82%的灰階值誤差。 Active Appearance Models (AAMs) is an image representation method for non-rigid visual object with both shape and texture variations. It is a model-based representation method, and it uses a mean vector and a linear combinations of a set of variation modes to represent a non-rigid object. By adjusting the coefficients of the linear combinations of the variation modes(model parameters), we can synthesize any non-rigid objects. With this, we can express facial expressions using a model-based approach. For the facial expression recognition, an AAMs search algorithm is required to find the optimum model parameters such that the synthesized expression is similar to the facial expression in images. In this paper, we propose a novel iterative AAMs search algorithm. It minimizes the error which measures the difference between a model and a test image. We only adopt the magnitude of the predicted change of the parameters from the traditional search algorithm. However we decide the direction of the change of the parameters by estimating the gradient of the error function at each iteration. Moreover we prevent the local minimum search of the error function at each iteration by disturbing the searched parameters. Our experiments show that the proposed robust AAMs search algorithm reduced 36.41% location error of shape and 30.82% intensity error of texture of facial expressions related to the AAMs search algorithm.
