Funded by: Afeyan Foundation
PI: Suren Khachatryan
Researcher: Gevorg Nersesyan
The project focuses on the detection and recognition of Armenian epigraphic inscriptions carved in stone, using a combination of classical computer vision and machine learning techniques. A novel adaptable framework for Armenian inscription recognition has already been developed and presented at the 10th Epigraphy.info Workshop, University of Graz (Austria), 24-26 March 2026. It is grounded in a unique combination of field photography, 3D cultural-heritage data, and spatially tuned computer-vision methods. The results show that even in a severely low-resource setting, an elegant adaptive pipeline can robustly segment degraded carved text and support promising early recognition performance. As such, the project lays a serious technical foundation for large-scale Armenian epigraphic analysis and digital preservation. A figure below illustrates the developed preprocessing and detection algorithms:
The research continues in two complimentary directions: a) 3D reconstruction of Armenian inscribed surfaces to separate carved geometry from erosion and improve glyph detection on worn stones; and b) Semantic post-processing for text recovery to infer and reconstruct partially lost inscription content. The achieved results are summarized in two submissions accepted for the presentation at the Digital Humanities: Under-resourced Languages and Armenian International Summer School, Yerevan (Armenia), 6-10 July 2026.
Armenian stone inscriptions often survive in heavily degraded condition, where centuries of erosion, fractures, and surface irregularities obscure carved letterforms and make conventional image-based reading unreliable. We investigate the use of 3D surface analysis for improving glyph detection on worn inscriptional surfaces and present a reconstruction-driven pipeline that models stone geometry at high resolution and analyzes local surface structure to distinguish intentional carved features from damage caused by weathering and material decay. Rather than relying solely on 2D appearance, the method exploits geometric cues such as depth variation, curvature, groove continuity, and local shape regularity to identify likely inscription strokes. Beyond Armenian epigraphy, the work contributes to the broader problem of recovering human-made markings from degraded heritage surfaces and demonstrates the value of geometry-aware analysis for reading inscriptions that are no longer cleanly visible to the eye.
While imaging and geometric methods can improve the visibility of surviving letterforms, the recovery of missing text remains a semantic problem that requires contextual inference beyond the material trace itself. We study computational methods for the semantic reconstruction of incomplete Armenian inscriptions through post-detection language-aware processing and develop an approach that combines epigraphic conventions, lexical and morphological patterns, and contextual constraints to infer plausible restorations of partially preserved text. The proposed framework is designed not to replace expert interpretation, but to narrow the space of plausible restorations and make the reasoning behind suggested completions more explicit and systematic. In the concluding stage of the project the developed methods are applied to two new datasets provided by the partners from EPFL that collect Armenian heritage data from Artsakh and India.