Transforming Physical Archives into Searchable Digital Libraries with Optical Character Recognition
Sivankalai Sivankalai, Shanmugam Balachandran · Preservation Digital Technology & Culture · 2025
Abstract This study presents an integrated and extensible OCR framework that transforms physical documents into searchable digital archives, emphasizing system-level engineering over algorithmic novelty. While the core components such as contrast enhancement, PaddleOCR, and BERT-based spelling correction are individually established, their synthesis into a unified pipeline for multilingual and multimodal archival digitization represents a practical contribution to digital preservation. The architecture is implemented using Python and incorporates preprocessing steps such as grayscale conversion, CLAHE-based contrast adjustment, Gaussian blurring, and adaptive thresholding to enhance image quality. Multilingual high-accuracy text recognition is performed using PaddleOCR, while a hybrid spelling correction module combines lexical and BERT-based contextual corrections. A straightforward and easy-to-use graphical user interface (GUI) facilitates interaction and visual diagnostics on the basis of confidence heatmaps as well as pixel intensity histograms. Digitized outputs are retained in a central library repository to facilitate efficient retrieval and organization. The system performs well in high OCR accuracy across different input types 99.12 % in screenshots and 94.81 % in scanned documents. Preprocessing greatly enhances text readability in degraded images. The spelling correction module enhances text readability by more than 95 %. Visualization tools offer useful insights into OCR performance and preprocessing effects. This paper demonstrates a modular pipeline that integrates preprocessing, OCR, error correction, visualization, and repository integration. The application of PaddleOCR with domain-specific preprocessing and contextual correction capabilities brings novelty, particularly for rich archival content.