Prism-OBI: a novel framework for oracle bone inscription recognition via visual perception and feature decoupling

Abstract Oracle bone inscriptions (OBIs) play a vital role in the digital preservation of human civilization, offering profound historical insights. However, its recognition faces significant challenges due to severe degradations, such as blurred strokes and complex geological noise, which generic end-to-end models struggle to disentangle. To address these issues, we propose Prism-OBI, an innovative two-stage, decoupled framework that explicitly separates spatial localization from fine-grained semantic classification. Specifically, we present a degradation-aware visual perception methodology that leverages signal decoupling, dynamic scale adaptation, and explicit spatial encoding to serve as a semantic filter, effectively isolating coherent strokes from background artifacts. Extensive experiments demonstrate that this decoupled paradigm mitigates the localization-classification trade-off, laying a solid foundation for the transferable analysis of diverse historical archives in digital humanities.