人脸识别技术借助大数据和深度学习有了长足进步，但存在明显姿态和表情变化的复杂场景下效果仍不理想。三维人脸重建方法被用来矫正人脸姿态和辅助二维人脸识别，一定程度上提高了人脸识别率。然而，现有三维人脸重建方法一般采用参数化模型拟合，复杂度高，对姿态和表情鲁棒性差，且独立于人脸特征表示的学习过程，限制了其对人脸识别的提升作用。本项目针对复杂场景下三维人脸的重建和识别，围绕三维与二维人脸形状/纹理的映射关系、人脸三维重建与身份识别的相关性及联合学习模型等关键问题，采用级联回归和深度学习等技术路线，开展单图三维人脸重建、三面照三维人脸重建、图像集三维人脸重建和人脸三维重建与识别多任务学习等方法研究，构建面向复杂场景的三维人脸重建模型和人脸三维重建与对齐识别的多任务学习模型，提出复杂场景下的人脸三维重建与识别新方法。本项目成果将拓展三维人脸重建的研究、促进无约束人脸识别技术的发展与应用。

3D reconstruction methods have recently been applied to recover 3D face shapes from 2D images, which are then used to assist 2D face recognition. This effectively improves face recognition accuracy. However, existing 3D face reconstruction methods mostly have relatively high complexity and are fragile to pose and expression variations that are commonly seen in complex scenarios. Moreover, they are isolated from the facial feature representation learning process. This project aims to enhance the robustness of 3D face reconstruction to pose and expression variations, and to establish an end-to-end learning framework for joint 3D face reconstruction and face recognition. To this end, it will propose a variety of 3D face reconstruction methods based on single images, mugshot images, and 2D face image sets, and methods for joint 3D face reconstruction and recognition. Cascaded regression technique will be employed to predict 3D face shapes from 2D facial landmarks. Statistical facial texture models will be used to fuse textures in multiple face images. Deep convolutional neural networks will be utilized to implement the multi-task learning framework for joint 3D face reconstruction and recognition. The proposed new methods are believed to be beneficial to extend 3D face reconstruction research and to enhance face recognition performance in real-world complex scenarios.