Reflectance-aware structure-preserving multi-image stitching for high-fidelity digitization of Tibetan color paintings
Year: 2026
Type: article
Abstract: Tibetan color paintings are difficult to digitize in museums because mineral pigments and protective glass often introduce specular reflection, while single overview images lack fine spatial details. We propose a reflectance-aware structure-preserving multi-image stitching framework for high-fidelity digitization. The framework constructs an overlap graph for global alignment, uses HHSA-RDNet to predict dense reflectance probability maps, and incorporates reflectance reliability into matching, seam estimation, and fusion. Motion-guided local refinement further reduces micro-misalignments and preserves decorative structures. We construct TCP-MultiStitch with 780 stitching groups and 9,340 local images. Our method achieves 37.06 dB PSNR, 0.910 SSIM, 0.163 LPIPS, 32.74 dB M-PSNR, and 0.853 M-SSIM on TCP-MultiStitch. Direct evaluation on MuralDH without retraining or fine-tuning achieves 36.49 dB PSNR and 0.903 SSIM, showing promising transfer under the stated protocol, although evaluation is currently limited to one external dataset.
Source: npj Heritage Science
Institutions: Qinghai Normal University, Beijing Normal University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225347196