Abstract: Active Appearance Model（AAM） is the traditional classical method for feature point extraction.However，limited by the linear predicable model，it is very difficult to find the accurate feature points when the initial position is too far away from the destination，thus the single AAM template is hard to meet the requirement of pose-variant face.This paper proposes a face feature point extraction method based on multi-AAM templates：Firstly，template similarity and face feature points are defined and the face pose is classified；then，AAM template for each pose is trained；next，test images are searched with each template and similarity is calculated accordingly；through comparing the similarities，the best template with the highest similarity is chosen to detect the feature points.The experiments on Oriental Face Database prove the essentiality of template selection and the reasonability of similarity defined in this paper.The experimental results show that 92.11% of the whole test set can obtain the correct AAM template thus the feature points can be extracted accurately and 99.92% of test images can get the correct or the neighbor AAM template thus the feature points can be extracted.

摘要： 主动表观模型AAM是经典的特征点检测方法，但是受到其线性预测模型能力的限制，当初始位置偏离目标位置过大时，很难收敛到正确位置，因此传统的单一AAM模板很难满足多姿态人脸特征点检测的要求。针对此问题，提出了一种基于多模板AAM的人脸特征点检测方法：首先定义模板相似度和人脸特征点，将人脸姿态划分为若干类，每一类姿态训练一个特定的AAM模板；然后对测试样本，利用每个AAM模板搜索特征点，并计算搜索结果与模板的相似度，选择相似度最大的AAM模板的搜索结果作为最终的特征点检测结果。通过相似度选择模板的方法可以为AAM搜索提供相对偏离较小的初始位置，因此可以精确地检测特征点。在东方人脸数据库上的实验结果证明了模板选择的必要性和相似度定义的合理性。在整个数据库上的测试结果表明，92.11%的测试图像均可以通过相似度准则选择到正确的AAM模板，从而可以精准地检测特征点；99.92%的图像可选择到正确的AAM模板或其相邻模板，故可以检测到较为准确的特征点。