三维到二维:人脸本征形状描述图

3D to 2D:Facial intrinsic shape description maps

基于三维数据的人脸识别克服了二维图像数据受光照和姿态影响较大的问题,但其较高的数据维数约束了它的实际应用。本文针对三维人脸数据的简化描述,提出了将三维人脸映射至二维表示的本征形状描述图方法。该方法首先基于约束离散保形映射将三维人脸数据微分同构映射到一个局部几何特征保持的二维区域。然后基于人脸曲面几何结构特性和表观特性,构建二维本征形状描述图,用于简化对三维人脸数据的描述,并进行识别验证。基于国际公共人脸数据库FRGC2.0和GavabDB的三维人脸识别实验显示,本征形状描述图法在姿态变化大于60°时的识别率达到90.6%,比现有方法高5.9%,单次匹配时间为7.89s。该方法将三维人脸识别问题转换为了二维平面图像的识别问题,有效降低了数据描述的复杂度。得到的结果展示了该方法计算效率高,且对姿态变化有良好的健壮性。

The face recognition based on 3D facial data overcomes the difficulties sensitive to illumination and pose variations in 2D face recognition systems.However,the high computational complexity restricts its practical applications.To simplify the description for 3D face data,a novel strategy to map 3Dface data to 2D ones,called 2D intrinsic shape description map,was proposed in this paper.With the strategy,each 3Dfacial surface was firstly mapped homeomorphically onto a 2D lattice which keeps the local geometrical features based on the constraint discrete conformal.Then,a 2Dintrinsic shape description map was obtained by combining 3D facial geometrical structure and appearance feature for simplifying 3Dface representation and for verifying the recognition.The proposed strategy was compared to state-of-the-art 3D face recognition algorithms in the FRGC 2.0and GavabDB database.The results show that the proposed strategy offers the rankone rate of 90.6% when the pose change is greater than 60°,higher 5.9%than that of the existing method.Moreover,the single matching time is 7.89 s.As the strategy transforms the 3D face recognition into the 2D image recognition,it effectively reduces the complexity of data description and shows higher computation efficiency and robustness to a pose change.