Compared with two-dimensional faces, three-dimensional faces contain more feature information and can be applied to more practical application scenarios, such as face recognition, film and television entertainment, medical beauty, etc. Therefore, 3D face reconstruction technology has become a research focus in the field of computer vision. Due to real 3D face data is difficult to obtain, many deep learning-based reconstruction algorithms first use traditional reconstruction methods to construct 3D labels for a large number of 2D face images. These training data may not be accurate which will affect the reconstructive accuracy of these algorithm. To this end, this study proposes a weakly supervised learning model based on a multi-level loss function, which combines traditional 3D morphable model 3DMM and deep learning methods to directly learn 3D face feature from a large number of 2D face images without 3D labels to implement a3D face reconstruction algorithm based on a single 2D face image. In addition, in order to solve the problem that occlusion or large poses in 2D face images often affect the reconstruction of face texture, this paper uses a face parse segmentation algorithm based on the CelebAMask-HQ dataset to preprocess the images to remove the occlusion areas. The experimental results show that the quality and accuracy of 3D face reconstruction based on the proposed method have been improved greatly.