AI‑based multilingual OCR and OCR-enabled collection diversity analytics in libraries

The digitization of multilingual library collections requires text recognition that copes with diverse scripts, writing styles and document structures. Optical Character Recognition (OCR) remains central to this process, but recognition of cursive scripts such as Arabic and Urdu is still difficult. This paper presents a multilingual OCR integration framework for digital library applications that combines classical image preprocessing (Otsu binarization, Hough-based skew correction and morphological filtering; no learned preprocessing), two pre-trained deep learning OCR engines (PaddleOCR and MMOCR) used in inference mode, a fully specified token-level fusion procedure, and a language-sensitive post-processing module. The contribution is at the system-engineering level: the structured integration of existing components into a reproducible multilingual pipeline, with formally defined evaluation protocols and a link to OCR-enabled language coverage analytics. The framework is evaluated on two public benchmarks, the ICDAR 2019 MLT19 scene-text dataset (English, Hindi, Chinese, Arabic) and the synthetic printed Urdu dataset MMU-OCR-21, covering five languages in four script families. Post-processing reduced the average Character Error Rate (CER) from 0.188 to 0.107 (mean per-language relative reduction 42.6%) and the average Word Error Rate (WER) from 0.272 to 0.156, with statistically significant gains for all five languages (paired t-test, p < 0.001). In the fusion ablation, dual-engine fusion alone did not reduce average CER relative to the stronger single engine, so the measured gains are attributed mainly to preprocessing and post-processing. A Language Coverage Index (LCI) is introduced to quantify the share of languages whose OCR output meets a working CER threshold (θacc = 0.15); the LCI rose from 40% to 100% on the benchmark test partition. A supplementary evaluation on 150 digitized library pages from the Internet Archive showed a smaller but consistent CER reduction (37.3%) and an LCI of 60%, with Arabic and Urdu remaining above the threshold, and identified historical typefaces and multi-column layouts as priority challenges. Because the primary evaluation uses scene-text and synthetic word images, the results indicate directional promise rather than deployment readiness; full-scale validation on institutional library collections is the main direction for future work.