基于CPR和CLM的多视角人脸特征点定位方法

A Multi-view Facial Landmark Localization Method Based on CPR and CLM

准确定位人脸特征点在人脸识别、三维人脸模型重建等领域都有重要作用。目前,针对正面人脸的特征点定位已经相当成熟;但是,当姿态偏转角较大时,准确定位人脸特征点依然是一个有待解决的难题。文中针对姿态偏转比较大的特征点定位,提出了一种多视角人脸特征点定位方法。在训练阶段,针对不同姿态角度,分别定义其特征点模板,并通过训练得到用于特征点搜索的CPR(级联姿态回归)模型和用于特征点模板选择的CLM(约束局部模型)模型。在测试阶段,利用每个模型分别对测试样本进行特征点搜索,然后,利用CLM模型计算各特征点模板的拟合度,选择拟合度最高的模型作为最终结果。在FERET公开库上,与当前比较先进的算法进行的对比实验表明,文中方法有效提高了较大偏转姿态下人脸特征点定位的准确性。

Facial landmark localization plays an important role in many face-related applications such as face recognition and 3D face reconstruction. Although existing methods already achieve promising results on frontal and near-frontal face images, their performance on face images with large pose angles is still far from being satisfactory. A multi-view facial landmark localization method is proposed in this paper. It divides head pose angles into a number of non-overlapping ranges. During training,for each range of head pose angles,a facial landmark template is consU＇ucted by using the CPR （ Cascaded Pose Regression） method, and a corresponding texture model is established by using the CLM （Constrained Local Model） method. During testing,given a new face image,all the templates are applied to it,each resulting in a set of facial landmarks. The associated texture models are then used to compute the fitness values of them, from which the one with the maximum fitness is chosen as the final result. Expedmertts on the FERET database with comparison to a state-of-the-art method prove the effectiveness of the proposed method in localizing the facial landmarks on face images with large pose angles.