Virtual University Journals
Search
Search
Preserving Cultural Heritage through AI-based Image Reconstruction Techniques
Muzaffar Ali Khan
Abstract:
The Indus Valley Civilization has left behind thousands of seals bearing the still undeciphered Indus script; most of these survive only in fragmentary or damaged states. Preserving and reconstructing these seal inscriptions is important in cultural heritage and epigraphic research. This paper evaluates the performance of a pre-trained Shift U-Net model for generative inpainting of poor-quality seal inscriptions. Such a Shift U-Net model was trained on a curated dataset comprising thirteen seals with manually marked binary masks to localize erased areas. The generated output presents visually coherent restorations blending with their surrounding textures, which is supported by visualization styles such as blended reconstructions and tint overlays. Quantitative evaluation based on BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator), texture coherence, and DeltaE (CIEDE2000) metrics combined into a composite score yielded indications of modest perceptual gains. The average was 19.6, while only two images reached the threshold value of 30. Meanwhile, a negative correlation between the severity of damage and the quality of restoration confirms the model’s dependence on intact contextual information. Qualitative analysis revealed plausible surface textures with no reconstruction of real graphemes, further underlining the limitations of generic, pretrained models for specialized epigraphic tasks. Future work will consider fine-tuning generative models on domain-specific datasets of Indus seals to achieve visually plausible as well as semantically meaningful restorations.
Keywords:
Indus script, Deciphering, Archaeology, Digital reconstruction, Graphemes
Full Text:
PDF