Table Structure-Aware Learning for Handwritten OCR Correction in Noisy Industrial Documents

초록

This study aims to improve OCR performance on handwritten table-based documents from industrial sites by utilizing structural and contextual features. Inspired by human reading behavior, the proposed framework builds a graph of inter-block relationships using a GNN, capturing layout patterns such as row alignment and repeated inspection phrases. These structure-informed features are integrated into an LLM to perform semantic correction based on contextual flow. The method is suited for inspection reports with irregular handwriting and noisy layouts, where conventional OCR often fails. By combining structural cues and semantic context, the approach enhances text correction accuracy. Performance will be evaluated using Character Error Rate (CER), Word Error Rate (WER), and F1-score for graph-based structure prediction. This framework demonstrates the practical potential of structure- and context-aware correction for digitizing handwritten industrial documents. © 2025 IEEE.

키워드