PorTEXTO: A European Portuguese Benchmark for Visual Text Extraction

2026-06-17

AI Summary

This study addresses the critical absence of European Portuguese data in existing OCR benchmarks tailored to modern application scenarios. To bridge this gap, the authors introduce PorTEXTO, the first benchmark specifically designed for contemporary cultural contexts in European Portuguese, comprising real-world images paired with high-quality text annotations. These annotations are produced through a hybrid pipeline combining automatic generation by large vision-language models with manual verification by native speakers. The work fills a significant void in authentic-scenario OCR evaluation for this language and releases all resources publicly. Experimental results demonstrate that state-of-the-art OCR models suffer substantial performance degradation on real-world samples, and that incorporating dedicated multilingual training data yields greater improvements in recognition accuracy than merely increasing model scale or input resolution.

Abstract

European Portuguese (pt-PT) is largely absent from OCR benchmarks, which skew toward high-resource languages. The few benchmarks that cover pt-PT focus on historical artifacts and literature. This work addresses modern OCR applications, introducing PorTEXTO, the first benchmark for contemporary and culturally relevant pt-PT visual text extraction. To ascertain quality, we employ an annotation pipeline combining transcriptions from a frontier LVLM with exhaustive review by native speakers. We observe a sharp performance drop from synthetic to real world samples in most models, and find that, currently, specialized multilingual data is a better driver for pt-PT performance than model size or resolution budget, motivating the release of open pt-PT OCR resources.

Problem

• European Portuguese
• OCR benchmark
• visual text extraction
• low-resource language
• modern applications

Innovation

• OCR benchmark
• European Portuguese
• visual text extraction
• LVLM annotation
• multilingual data