基于局部形状约束网络的人脸对齐

Face Alignment Based on Local Shapes Constraint Network

摘要:

基于级联回归的人脸对齐方法已经取得了很大的成就，但是由于复杂的级联回归器设计、人为设计特征等局限性的影响使得人脸对齐没有找到一个性能更好的解决方案，尤其对于大姿态、大表情等条件下的人脸对齐任务。因此，为解决该问题，本文提出了一种新颖的人脸对齐方法—基于人脸局部形状约束。首先利用卷积神经网络（CNN）初始化人脸整体形状；然后利用人脸局部区域的同质性，将人脸区域进行划分，对每一个区域定义局部形状约束；最后再由整体形状估计做为全局约束，组合各个面部局部形状约束，对整体面部特征点进行回归。实验结果表明，该方法提高了人脸对齐的精确度且速度上达到了实时。

Abstract:

Face alignment method has greatly improved with using cascade regression, however due to the complexity of designing cascade regressors, the limitation of hand-crafted features make it difficult to find a better solution for face alignment task, especially with the big gesture, exaggerated expressions. Therefore, this paper proposes a novel method based on local shape constraint to solve this problem. Firstly it initializes the whole face shape by using deep convolutional neural networks (DCNN), secondly divides face into different regions according to the local regional homogeneity, defining each region constraints on local shapes, Finally takes the whole shape estimation as global constraints, combines each local shape constraints for facial feature point regression. The experiments show that our method based on local shapes constraint results in a strong improvement over the current state-of-the-art.