Cross-Script Generalization in STR: Solving the Orthographic Diversity Challenge with Global Semantic Segmentation

The proliferation of multilingual digital content necessitates robust scene text recognition (STR) systems that transcend language barriers. While monolingual models achieve superhuman performance in high-resource languages like English, their cross-lingual capabilities remain constrained by orthographic diversity and typological differences. This study challenges three fundamental assumptions in multilingual STR through systematic experimentation: \begin{enumerate} \item The linguistic transferability hypothesis \item The typological similarity principle \item The optimal resource allocation paradigm \end{enumerate} Our investigations reveal that dataset cardinality supersedes language similarity as the dominant factor in cross-lingual adaptation. By developing a meta-transfer learning framework with dynamic resource allocation (Equation \ref{eq:capacity}), we achieve 94\% accuracy in transfer performance predictions while maintaining compatibility with legacy systems. The proposed methodology demonstrates 41\% improvement in cross-lingual robustness indices compared to conventional approaches, establishing new benchmarks for multilingual STR development.