Objective This paper aims to address the issue that existing single-image 3D face reconstruction methods based on weakly supervised learning focus on global supervision information and lack constraints on local features which leads to poor reconstruction results in local facial regions. Methods A single-image 3D face reconstruction method based on feature kernel correlation is proposed. On the basis of a global multi-level loss function composed of low-level perceptual loss mid-level facial key-point loss high-level identity loss and shape consistency loss a feature kernel correlation loss function was constructed to better constrain local parts aiming to improve the reconstruction accuracy of 3D faces. Results The proposed loss function was used to constrain the shape of the mouth. Quantitative comparison experiments were conducted on the REALY public benchmark with the mean error of the overall reconstruction being 2. 102 mm and the mouth reconstruction error being 2. 206 mm. Qualitative comparison experiments were carried out on the FFHQ public dataset. Finally ablation experiments and visual demonstrations were performed on the FaceScape dataset and the LYHM dataset to verify the effectiveness of the proposed loss function and the robustness of the model. Conclusion Experimental results demonstrate that the proposed reconstruction method outperforms other methods from different viewpoints in both overall and local reconstruction. Moreover the proposed model also exhibits outstanding performance in various poses expressions and lighting environments.