As an important method of face modeling, 3D deformation model (3DMM) has been widely used in 3D modeling, image synthesis and other fields. Due to the influence of training data type, quantity, linear basis and other factors, 3DMM has the phenomenon of over-constraint and cannot provide enough flexibility to represent high-frequency deformation. In order to further improve the problem, the three-dimensional deformation basis function is embedded into the deep neu-ral network, which provides a new idea for improving the representation ability of the model. In order to improve the efficiency of network learning, a two-path neural network is constructed to achieve the balance between global path and local path. By improving nonlinear 3DMM in both learning objectives and network structure, we propose a model that can capture higher levels of detail than linear or previous nonlinear models. The algorithm is compared with the simula-tion experiment, which shows that the algorithm has lower normalized average error in 3D face modeling, and the gen-erated 3D face model has good robustness and accurate reconstruction, and achieves better 3D face reconstruction per-formance.