Aiming at the shortcomings of facial reconstruction algorithms in terms of detail reconstruction capability, accuracy, and the impact of occlusions, a three-dimensional facial reconstruction algorithm was proposed, incorporating improved multi-level feature loss and global attention. Facial landmarks and facial mask priors were added at the input layer to guide the model to focus on the important facial regions. The global relation-aware pyramid attention module was designed to enhance the model's attention to important features and effectively integrate feature information from different levels. The face mask consistency loss and structural consistency loss were introduced, and the multi-level feature losses were designed to optimize model training, improve robustness to occlusions, make the input image and the reconstructed result approach each other in terms of structure, and enrich the feature representation of the model. Experimental results demonstrate that the reconstructed facial model exhibites more detailed features, significantly enhances facial detail reconstruction under occlusions, and greatly improves the reconstruction accuracy and robustness of existing methods.