Feng Li

As a significant historical and cultural heritage, the murals of Xu Xianxiu from the Northern Qi Dynasty have long faced complex deterioration including structural ambiguity, partial loss, and chromatic degradation. Traditional single-model restoration approaches (e.g., relying solely on super-resolution or image restoration models) often struggle to simultaneously address multidimensional challenges such as structural ambiguity, local missing data, and color degradation, frequently resulting in imbalanced outcomes in structural integrity, semantic coherence, or chromatic fidelity. Current research predominantly focuses on single-degradation-type restoration, lacking a systematic and progressive framework for tackling the multi-scale composite degradation issues in murals.To address this, this paper proposes a progressive mural restoration framework based on multi-scale separation strategy. Through task- decomposition and modular pipeline design, three deep learning models—Real-ESRGAN, LaMa, and Zero-DCE—are sequentially employed to accomplish large-scale structural reconstruction, local semantic restoration, and global color recovery. Experiments using simulated and authentic mural data were evaluated through quantitative metrics (PSNR, SSIM, FID) and visual perception. Results demonstrate that this progressive workflow significantly enhances restoration quality, achieving notable improvements in structural clarity, content integrity, and color authenticity. The key contributions of this study are as follows: First, we propose a multi-scale separation-based progressive mural restoration framework that decomposes complex restoration tasks into three subtasks: structural reconstruction, local restoration, and color restoration. Second, we develop a modular restoration workflow integrating Real-ESRGAN, LaMa, and Zero-DCE to achieve systematic restoration from structural to semantic and color levels. Third, we design a physical degradation simulation method and construct a mural dataset combining simulated and real-world data. Finally, quantitative and qualitative experiments demonstrate significant improvements in structural clarity, content integrity, and color authenticity achieved by our framework.

This research proposes a virtual restoration system and method for 3D digital cultural relics based on a fuzzy logic algorithm, aiming to solve the problems of the low classification accuracy and poor splicing effect of Terra Cotta Warrior fragments. This method adopts a series of steps to improve the efficiency and accuracy of fragment splicing. Firstly, features such as curvature, torsion, and left and right chord lengths were extracted from the fracture surface contour lines of the cultural relic fragments to form feature vectors. Then, the feature vector was fused and compressed by using the multilayer perceptron. The multilayer perceptron is a neural network model that can process and learn input data via multiple levels of computation, resulting in more expressive feature representations. Next, we used the calculation results of the multilayer perceptron to perform the splicing operation on the fragments. This means that, based on the calculation results of the feature vectors, the system can automatically select appropriate splicing methods to accurately match and splice fragments. Finally, by adjusting the weight of the multilayer perceptron, the error rate of fragment splicing can be reduced, further improving the accuracy of repair. The experimental results show that the method proposed in this article is significantly better than traditional methods in terms of time consumption and can effectively improve the efficiency of fragment matching and stitching. Conclusion: The fragment-stitching algorithm based on multi-feature adaptive fusion improved the speed and effectiveness of stitching in fragment-stitching tasks. In summary, the fragment-stitching algorithm based on multi-feature adaptive fusion can improve the speed and effectiveness of stitching in fragment-stitching tasks. The application of this method is expected to play an important role in the field of cultural relic protection, such as the restoration of Terra Cotta Warrior fragments.

Traditional digital geometry mainly focuses on local geometric features such as curvature and normal of 3D models. Although the curvature can describe the geometric curvature of the model surface, these local geometric properties cannot describe the global functional structure and associated properties of the 3D model. The purpose of this paper was to study the accurate splicing of cultural relic fragments based on the intrinsic structural features of the 3D model, such as geometry, texture, and function, whose method is the key to realizing the virtual reconstruction of damaged cultural relics. In this paper, a method for analyzing the intrinsic structure of 3D point cloud data is proposed, the feature representation method of 3D discrete curves and surfaces is studied, and a method for identifying geometric features of 3D point clouds based on the similarity measure of the principal curvature is proposed, which realizes the effective extraction of geometric features of the 3D point cloud model of cultural relics. By calculating the visual curvature distribution of the model under multiscale constraints, the effective extraction of the structural primitives of cultural relics with rich surface noise is realized. The experimental results in this paper show that the initial matching time is 1.416 seconds, the final matching time is 1.555 seconds, the average number of iterations is 13, the average stitching error is 1.7233 mm, and the standard deviation is 1.0265 mm. The experimental data show that the algorithm proposed in this chapter has good convergence characteristics and effectively avoids the divergence phenomenon that is easy to occur in the existing stitching algorithms.