Under unconstrained conditions

the traditional neural network is too sensitive to the change of the individual face expressions but the sensitivity of the face between different individuals is limited

which leads to the phenomenon that the geometric features of the constructed three-dimensional model do not match the individual. In order to solve above problems

a face representation and 3D reconstruction algorithm is proposed which is based on self-supervised deep learning with strong robustness. It can effectively auto map the feature points of 2D faces into3D space to realize 3D face reconstruction. The Efficient-Net is used as the backbone to obtain the facial feature vectors and the 3D deformation model parameters

meanwhile the contrast loss function is introduced on the basis of the twin（ Siamese） neural network to expand the inter-class distance and reduce the intra-class distance. The identity loss function is proposed to preserve the identity information of the same individual in space which enhances the robustness to deformation. Experiment shows that the algorithm has achieved a good performance on the 300 W—LP and AFLW2000—3D database.