A Cultural Relic 3D Reconstruction Method Based on Transformer-Based Inpainting Network and Multi-View Consistency Correction

• 1、School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：3D reconstruction is a key technique for the digital preservation and virtual exhibition of cultural relics. However, when relic surfaces are damaged, the resulting defective images lead to geometric distortions and texture voids in the reconstructed models, while independently inpainting each view introduces texture inconsistencies that further aggravate rendering artifacts. To tackle these challenges, this paper presents a cultural relic 3D reconstruction framework that integrates image inpainting with multi-view consistency correction. In the inpainting stage, a Transformer-based cultural relic image inpainting network is constructed, which leverages an alternating attention fusion Transformer module for interactive feature modeling between channel and spatial dimensions, and employs a sequential attention integration skip connection to mitigate the attenuation of high-frequency details during multi-level downsampling. In the consistency correction stage, a geometry-constrained multi-view texture unification strategy is designed, utilizing anchor view selection and depth-guided cross-view texture mapping to ensure appearance uniformity of inpainted results across different viewpoints. Finally, the corrected multi-view images are fed into COLMAP for pose estimation, followed by high-fidelity 3D modeling via 3D Gaussian Splatting. Experimental results show that the proposed inpainting network achieves an average PSNR gain of 0.81 dB over the runner-up method on the cultural relic dataset, while the consistency correction yields an 8.3\% relative improvement in the cross-view geometric consistency metric TSED and a 0.36 dB increase in reconstruction rendering PSNR, fully demonstrating the effectiveness of the proposed method for high-quality 3D digital restoration of damaged cultural relics.

Keywords： computer application technology cultural heritage digitization image inpainting 3D reconstruction multi-view consistency transformer