Incorporate Structure and Content for Devanagari Table Extraction

Table extraction is a crucial component of document image analysis, enabling the transformation of unstructured document content into structured, machine-readable data. Although significant progress has been achieved for English and other Latin scripts, table extraction from Devanagari-script documents remains challenging due to script complexity, limited annotated datasets, and poor-quality scans. This research proposes a hybrid framework that combines structural analysis with content-based validation to enhance table extraction accuracy for Devanagari documents. The framework integrates preprocessing, Devanagari-specific OCR post-correction, semantic consistency checks, and confidence-based feature fusion. Experimental results on over 500 annotated Devanagari documents demonstrate improved performance compared to existing methods, achieving high structural precision, OCR accuracy, and overall TEDS-S score.