Structural-prior-guided Tangut character recognition for digital preservation of historical manuscripts

Tangut manuscripts are valuable historical resources, yet automatic character recognition remains challenging because of complex structures, highly similar glyphs, long-tailed distributions, and limited annotated data. Existing methods mainly rely on holistic appearance and struggle to distinguish characters with similar contours but different component arrangements. To address this issue, we propose SPMTNet, a structural-prior-guided multi-task framework for cropped Tangut character recognition. We construct XiXiaDB, a multi-source dataset containing whole-character images, component annotations, and explicit character-component correspondences. Component composition and spatial relation learning are introduced to enhance structural discrimination, while a Residual Spatial-Channel Attention (RSCA) module strengthens informative local features. Experiments on XiXiaDB and CharSample show that SPMTNet outperforms representative baselines, particularly for highly similar characters and long-tailed categories. Cross-source evaluations further demonstrate its robustness to source variations within Tangut character recognition.