Abstract:

Face alignment is a key component in facial processing. It is a challenging task because human facial images in real-world usually contain large variations due to the differences in pose, illumination, etc. Shape init-ialization and feature extraction are crucial in face landmark alignment. This paper proposes a pose-robust face alignment model based on adaptive supervised descent method (SDM). Firstly, in order to reduce the influence of pose differences for face alignment, this paper uses clustering algorithm to cluster the face images into three categories (frontal faces, left faces, right faces) according to pose. Thus, the pose in each cluster is more compact. Secondly, face alignment can be taken as a coarse-to-fine supervised learning process with multi-stage. Therefore, the adaptive block size of feature extraction (from big to small) is used to get discriminative features. Based on the above two strategies, within each cluster, a better initial shape is given and the discriminant regression model is trained for facial landmark localization via adaptive SDM. A series of experiments have been conducted on benchmark datasets LFPW, HELEN and 300W. The experimental results show that this method makes facial landmark localization accurately in large pose images, and demonstrate the superiority of the proposed method by comparing with existing methods.

Key words: face alignment, facial landmark localization, supervised descent method (SDM) model, pose-robust, adaptive feature block size

摘要：

人脸对齐是人脸分析处理中的重要一步。由于现实中的人脸照片通常在姿态、光线等方面存在较大的差异，人脸对齐是一项艰巨的任务。初始关键点的位置以及特征提取对人脸对齐很重要。提出一种自适应监督下降方法（SDM）的姿态鲁棒人脸对齐算法。首先，为了减小姿态差异对人脸对齐的影响，使用聚类算法将图片按照姿态分成三类（正脸，左侧脸，右侧脸），这样每个类别下的姿态更加紧致。其次，考虑到人脸对齐是由粗到细的多阶段监督学习过程，采用自适应特征提取框（由大到小）来提取判别性特征。基于上述两种策略，在每个类别下，提供一个更好的初始关键点位置，通过自适应特征提取的SDM模型来进行回归模型的训练。选用LFPW、HELEN和300W数据集进行评估，实验结果表明，该模型在复杂姿态下能准确定位关键点，并且好于现有的人脸对齐算法。

关键词: 人脸对齐, 人脸关键点定位, 监督下降方法（SDM）模型, 姿态鲁棒, 自适应特征提取框