EpiSAM: Character Segmentation in Challenging Stone Inscriptions

Stone inscriptions are invaluable sources of historical and linguistic knowledge, yet their automated analysis remains a major challenge due to surface irregularities, erosion, and low visual contrast. Conventional document and handwriting analysis techniques fail to perform well in these scenarios. In this work, we propose character detection as a core strategy for robust inscription analysis. We introduce EpiSAM, a prompt-guided transformer framework for character segmentation in stone inscriptions. Rather than treating characters in isolation, EpiSAM employs a novel neighbor-aware strategy, explicitly predicting adjacent characters alongside the target. These contextual cues resolves boundary ambiguities, improving mask generation and enabling more accurate character segmentation. Furthermore, we expand an existing stone inscription dataset by adding dense polygonal annotations for characters, thereby enabling comprehensive research on Southeast Asian epigraphy. Experimental results shows that EpiSAM achieves consistent improve- ments over existing baselines, while also exhibiting strong zero-shot generalization in challenging epigraphic scenarios.

Rather than predicting each character in isolation, we exploit the structured nature of inscriptions, where characters appear in locally coherent spatial sequences. We reformulate character segmentation as a localized, neighbor-aware prediction task: estimating a target character jointly with its immediate left and right neighbors.

This formulation introduces strong contextual constraints that reduce boundary ambiguities and improve separation between adjacent characters, especially under heavy noise and texture ambiguity. By leveraging local spatial continuity, the model achieves more reliable instance-level character segmentation without requiring any explicit line-level modeling.

Based on this formulation, we propose EpiSAM, a prompt-guided transformer framework that integrates neighbor-aware contextual awareness for robust character segmentation. Our method leverages the strong visual representation and prompt-based localization capabilities of the Segment Anything Model (SAM) to produce fine-grained masks guided by sparse prompts.