一种基于姿态先验的鲁棒的人脸对齐方法

摘要

Face alignment is a crucial procedure in facial analysis applications,which has been widely used in various fields. Active Appearance Models are statistical models of shape and appearance,which have shown the capability to achieve exact face alignment. However,it heavily depends on the initialized model and is easily influenced by pose variations,illuminations and occlusions. When the initial model is far from the ground-truth,the performance significantly deteriorates. To avoid such problems,a simple but effective face alignment framework based on pose prior is proposed in this paper. Firstly,we train AAM in-the-wild and use Characteristic Tri-angle to select a preferable initial model for achieving a robust initialization. Subsequently,we employ a robust and accurate simultane-ous Active-Appearance-Model fitting algorithm. Finally,we compare our approach against previous methods and show that it yields the significantly outperformance on challenging LFPW database with large pose variations,expression and illumination.%人脸对齐作为人脸图像分析中的重要步骤,被广泛应用于各个领域.其中,主动表观模型( AAM)因其良好的对齐效果而被大量使用.但AAM对初始模型依赖度极高,且极易受到姿态、光照以及遮挡等因素的影响.当初始形状和标准形状相差较大时,匹配效果并不理想.对此,本文提出一种基于姿态先验的人脸对齐方法.首先,在非限制环境下的人脸库LFPW上进行训练,根据姿态的不同分别建立正脸模型、左偏模型和右偏模型.在搜索阶段,利用特征三角形自动选择合适的模型作为人脸的初始模型,从而避免了姿态变化对初始模型产生影响导致后续匹配效果不理想的问题.其次,利用同时反向合成算法,实现了鲁棒精确的AAM匹配.理论分析与实验证明,所提方法针对遮挡、光照以及姿态变化的有效性.