基于3D-2D映射的大姿态人脸识别  

Large-pose face recognition based on the 3D-2D mapping

作　　者：李晓峰[1] 游志胜[1] LI Xiao-Feng;YOU Zhi-Sheng(College of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《四川大学学报（自然科学版）》2022年第4期56-64,共9页Journal of Sichuan University(Natural Science Edition)

基　　金：四川省应用基础研究(2021YJ0079);国家自然科学基金(U20A20161)。

摘　　要：为了解决真实场景下大姿态人脸识别准确率低的问题,本文从数据增广的角度,围绕三维人脸信息数据进行大姿态人脸识别关键技术的研究与探索,提出了一个基于3D-2D映射的大姿态人脸识别算法框架.区别于当前基于3D点云数据的人脸识别算法,本文提出的方法主要利用3D人脸丰富姿态信息,通过3D~2D映射的姿态数据扩充方法,进而训练特定姿态人脸特征提取模型,然后集成到统一大姿态人脸识别框架中.本文提出方法的关键是利用注册3D人脸图像信息来辅助2D人脸多姿态识别,很容易集成现有的2D人脸识别方法到提出的框架中.实验表明,本文提出的方法在无约束真实场景下能够有效提升大姿态人脸识别准确率,同时不显著增加计算负担,为当前基于3D信息解决2D人脸识别问题提供了新思路.To solve the problem of low accuracy of large pose face recognition in real scenes, a large pose face recognition algorithm framework is proposed based on the 3 D-2 D mapping focusing on the exploration of the key technologies of large pose face recognition from the perspective of data augmentation. Different from the current face recognition algorithm based on 3 D point cloud data, the proposed method mainly uses 3 D face data to enrich the posture information via the data expansion method by 3 D-2 D mapping. The specific posture face feature extraction models are trained and integrated into the unified large pose face recognition framework. The key of the proposed method is to use registered 3 D face information to assist 2 D face multi-pose recognition. It is easy to integrate existing 2 D face recognition methods into the proposed framework. Experiments show that the method proposed in this paper can effectively improve the accuracy of large pose face recognition in unconstrained real scenes without increasing the significant computational load and provides a new idea for the current 3 D information to solve the 2 D face recognition problem.

关 键 词：人脸识别 大姿态人脸识别 三维对二维 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]