<p>This paper proposes a novel visual SLAM 3D surface reconstruction algorithm, which aims to reconstruct high-precision and high-fidelity 3D surface models from RGB-D image sequences. To cope with the sparsity and incompleteness of the depth data, this paper designs a depth-completion network to efficiently fill in the missing depth information so as to recover the complete scene structure. Simultaneously, the camera trajectory estimation is optimized using jointly sampled feature points, which significantly improves the trajectory accuracy, and provides more stable positional information for the subsequent 3D reconstruction. Furthermore, the semantic segmentation information is introduced as the semantic supervisor, which ensures the geometric reconstruction accuracy while further enhancing the semantic consistency, thus improving the accuracy and fidelity of the reconstructed model. Through the comparison test on the Synthetic RGB-D public dataset, the experimental results show that all the methods in this paper outperform the classical algorithms such as iMAP and NICE-SLAM. Even compared the latest Co-SLAM method, the new algorithm achieves improvements in accuracy and completeness of 12.25 and 18.24%, respectively. The high-precision and high-fidelity reconstruction of indoor 3D surface models is successfully realized using real RGB-D measured data.</p>