Reading Brazil: A Local Cost-versus-Accuracy Benchmark of Open OCR Engines on Brazilian Documents

• Cristhian Kapelinski UNIPAMPA https://orcid.org/0009-0005-5750-022X
• Aline Lunkes UNIPAMPA https://orcid.org/0000-0002-1846-4853
• Daniel Welfer UFSM https://orcid.org/0000-0003-1560-423X
• Dionatan Schmidt UNIPAMPA https://orcid.org/0009-0006-1453-6860
• Beatriz Machado UNIPAMPA https://orcid.org/0009-0002-2750-0323
• Diego Kreutz UNIPAMPA https://orcid.org/0000-0003-0830-0238

DOI: https://doi.org/10.5753/eniac.2026.31378

Resumo

Privacy rules push Brazilian institutions to read their forms and identity documents on their own hardware. No published benchmark tells them which open OCR engine does that well at the lowest cost: existing ones are English, Chinese, or non-Brazilian Portuguese, and none measures the cost of running locally. We pair a Brazilian-Portuguese accuracy measurement with a cost measurement on one consumer machine, scoring 14 open engines on forms, identity documents, clean prose, and degraded scans. No class of engine wins outright. Only two hold up on all four document types: the OCR engine Surya and the vision-language model Qwen2.5-VL. On identity photos the divide is old engines against recent ones, not OCR engines against vision-language models, and Surya recovers 0.921 of the fields, more than any vision-language model. Engine choice should follow the document type. We also rebuild the forms ground truth from the images and validate it by hand, finding no error in 250 sampled fields.

Referências

Cardeira, J., Glória-Silva, D., da Luz, M. L., Ferreira, R., Tavares, D., Semedo, D., and Magalhães, J. (2026). PorTEXTO: A european portuguese benchmark for visual text extraction. arXiv:2606.19096.

Fernandes, L. et al. (2024). BRIDP: Dataset and validation method for BRazilian identity document parsing. In ISDA.

Jaume, G. et al. (2019). FUNSD: A dataset for form understanding in noisy scanned documents. In ICDAR-OST.

Journet, N., Visani, M., Mansencal, B., Van-Cuong, K., and Billy, A. (2017). DocCreator: A new software for creating synthetic ground-truthed document images. Journal of Imaging, 3(4):62.

Kargaran, A. H. et al. (2026). GlotOCR bench: OCR models still struggle beyond a handful of unicode scripts. arXiv:2604.12978.

Li, Z. et al. (2026). MDPBench: A benchmark for multilingual document parsing in real-world scenarios. arXiv:2603.28130.

Neto, A. F. d. S. et al. (2024). BRESSAY: A Brazilian Portuguese dataset for offline handwritten text recognition. In ICDAR.

Ouyang, L. et al. (2025). OmniDocBench: Benchmarking diverse PDF document parsing with comprehensive annotations. In CVPR.

Poznanski, J. et al. (2025). olmOCR: Unlocking trillions of tokens in PDFs with vision language models. arXiv:2502.18443.

Ribeiro, G. R. et al. (2025). Vision-language models for automated property feature extraction in tax assessment: A case study. In ENIAC.

Rigal, B. et al. (2026). Benchmarking vision-language models for French PDF-to-Markdown conversion. arXiv:2602.11960.

Santos, M. K. et al. (2023a). ESTER-Pt: An evaluation suite for TExt recognition in Portuguese. In ICDAR.

Santos, Y. et al. (2023b). Evaluation of optical character recognition (OCR) systems dealing with misinformation in Portuguese. In SIBGRAPI.

Smith, R. (2007). An overview of the Tesseract OCR engine. In ICDAR. Taghadouini, S. et al. (2026). LightOnOCR: A 1B end-to-end multilingual vision-language model for state-of-the-art OCR. arXiv:2601.14251.

Xu, Y. et al. (2022). XFUND: A benchmark dataset for multilingual visually rich form understanding. In Findings of ACL.

Yang, Z. et al. (2025). CC-OCR: A comprehensive and challenging OCR benchmark for evaluating large multimodal models in literacy. In ICCV.

Zhong, Y. et al. (2026). OCRVerse: Towards holistic OCR in end-to-end vision-language models. arXiv:2601.21639.