A new preprint on arXiv evaluates OCR systems for multilingual archives, where the choice between cloud and self-hosted options involves trade-offs between accuracy and deployment constraints. The study sets up a comparison across three tasks: bounding box detection, text recognition, and mathematical expression recognition.
The evaluation is based on 48 documents in Arabic, English, and French, though the abstract does not yet report specific results. Instead, it frames the problem: archives with mixed languages need OCR that performs reliably across scripts and layouts, while also fitting operational requirements such as data privacy, cost, or infrastructure.
Because the source is an early preprint and the abstract is truncated, no concrete performance rankings are available. The paper's contribution at this stage is its comparative design, which may help institutions think through the accuracy-versus-deployment trade-off when selecting an OCR system.