Segmentation and Long-Tail Recognition for Ancient-Script Classification

Address the unresolved segmentation and long-tail class-imbalance problems in ancient-script symbol recognition systems that classify already-isolated symbol crops, including the need to isolate symbols from degraded source imagery and to recognize rare classes reliably.

Background

The paper observes that several classification-only studies of ancient scripts—including Hebrew paleography, ancient Chinese characters, Oracle Bone Inscriptions, Egyptian hieroglyphs, and Ge'ez—operate on pre-isolated symbol crops. Because these systems do not perform image segmentation and generally lack explicit mechanisms for severe class imbalance, they leave unresolved the broader problem of recognizing symbols directly from noisy or degraded source imagery while maintaining performance on rare classes.

EpigraphNet is presented as addressing these issues for Elamite cuneiform through a segmentation-first pipeline based on zero-shot SAM2-Large masks and inverse-frequency class weighting. The cited passage frames segmentation and long-tail recognition as remaining problems in the related literature, rather than as limitations of EpigraphNet itself.

References

First, several classification-only studies on other ancient scripts (Hebrew paleography, ancient Chinese, Oracle Bone adversarial augmentation, Egyptian hieroglyphs, Ge'ez script) operate on already-isolated symbol crops from published benchmarks and report no explicit imbalance-handling mechanism beyond, at most, data augmentation, leaving open the same segmentation and long-tail problems that motivate our SAM2-plus-inverse-frequency-weighting design.

— Zero-Shot SAM2 Segmentation and Vision Transformer-Based Recognition of Elamite Cuneiform Symbols from Degraded Tablet Images  (2608.18544 - Poudel et al., 19 Aug 2026) in Section 2.4, “Non-Elamite Cuneiform and Other Ancient-Script Studies with Comparable Methodology”