Few-Shot Learning for Grapheme Recognition in Ancient Scripts

The Indus Valley Civilization (IVC) left behind an undeciphered script containing critical yet inaccessible historical knowledge, posing significant challenges for archaeologists and linguists. Existing automated recognition techniques typically treat Indus graphemes categorically, overlooking their fine-grained spatial configurations and internal structural variations. To overcome this, we introduce FLAIR (Few-shot Learning with Adaptive Inference for Recognition), a novel framework designed to capture subtle visual distinctions in Indus script graphemes through segmentation-guided prototype learning. FLAIR integrates prototypical networks with our segmentation encoder, ProtoSegment , aligning learned spatial prototypes to accurately model intricate grapheme features. We present a new expert-annotated IVC benchmark dataset comprising 39 grapheme classes derived from seal images. Using this benchmark, we show that FLAIR significantly outperforms traditional machine learning methods—especially on rare and visually complex graphemes—while generalizing effectively from minimal data. We further validate FLAIR using controlled experiments on the Omniglot dataset to simulate real-world data scarcity. Our approach achieves state-of-the-art performance in digitizing and recognizing Indus Valley graphemes, laying a foundation for scalable analysis and future decipherment efforts. Code, datasets, and trained models are made publicly available at https://github.com/ParthGaneriwala/few-shot-ivc .