To address the issues of poor generation performance in single-image-based 3D face reconstruction caused by high model complexity, difficult training, and limited feature extraction accuracy, this study proposes a method for improving the accuracy of 3D face reconstruction that uses MobileNetV2 as the base network architecture. This method integrates the 3D dense face alignment (3DDFA) model and incorporates adaptive spatial feature fusion (ASFF) and the convolutional block attention module (CBAM). Compared with traditional convolutional models, the proposed method reduces model complexity while enhancing the accuracy of local feature extraction. In comparison with the 3DDFA model, the proposed method improves feature extraction accuracy without compromising recognition speed. Comparative experiments are conducted between the proposed method and the 3DDFA model on the AFLW dataset and the AFLW2000-3D dataset. Experimental results demonstrate that the reconstruction accuracy on these two datasets is improved by 11.27% and 11.41%, respectively.