Computer Vision-based Handwritten Text-area Detection for Industrial Document Digitization

초록

This study addresses the need for technologies that digitize handwritten documents commonly produced in industrial sites. As a core component of this process, we propose a method to detect text-areas in table-based handwritten documents. Traditional OCR techniques have difficulty recognizing text that crosses table boundaries or is irregularly arranged. To overcome this limitation, we present a computer vision-based approach that accurately detects text-areas within structured tables. Our framework integrates a Faster R-CNN-based object detector with the RLSA algorithm to perform table detection and text-area recognition in sequence. In our experiments using 124 annotated text-areas, the proposed method achieved an F1 Score of 0.93 for online text and 0.67 for offline text. This technique not only improves OCR performance but also lays the groundwork for ontology-based document analysis, contributing to the broader goal of document digitization. © 2025 IEEE.

키워드