Digital Protection and Inheritance of Chinese Cultural Heritage Based on Deep Learning Algorithms
Mingming Zhang, Jiang Cai · Procedia Computer Science · 2026
Chinese cultural heritage carries the historical memory of the nation, but currently faces dual threats of natural erosion and human destruction. The digital methods used in the past, such as manual surveying and single modal storage, often had low efficiency, incomplete information recording, and relatively single inheritance forms. This paper selects the the Mogao Grottoes of Dunhuang murals as the specific research object, mainly analyzes the three key tasks of mural damage repair, color restoration and three-dimensional digitization, and proposes a deep learning technology scheme combining the improvement of U-Net, CycleGAN and PointNet++to meet the challenges mentioned above. The experimental results showed that in the mural restoration task, the improved U-Net model achieved a peak signal-to-noise ratio (PSNR) of 28.5 dB and a structural similarity index (SSIM) of 0.92, which were 11% and 8% higher than traditional U-Net, respectively; The CycleGAN model achieves unsupervised color restoration with an accuracy of 95% and a FID score reduced to 12, outperforming supervised learning methods; The Intersection over Union (IoU) of the 3D reconstruction model based on PointNet++reached 0.89, while the model volume was reduced by 40%.