HalalBench: A Multilingual OCR Benchmark Dataset for Food Packaging Ingredient Extraction

HalalBench is the first large-scale, multilingual benchmark for evaluating OCR engines on real-world food packaging images. The dataset comprises 1,043 images (50 real, 993 synthetic) with 36,438 annotations in COCO format spanning 14 languages: Arabic, Danish, Dutch, English, French, German, Indonesian, Japanese, Korean, Malay, Norwegian, Swedish, Thai, and Turkish. No existing OCR benchmark targets the unique challenges of ingredient labels: curved surfaces, dense multilingual text, and sub-8pt fonts. HalalBench fills this gap. Key results: No OCR engine exceeds 0.55 fuzzy F1, confirming food packaging OCR remains unsolved. ML Kit leads at 0.487 exact F1, followed by docTR (0.465), EasyOCR (0.210), and RapidOCR (0.189). Created by HalalLens (https://halallens.no), a production halal food app and halal scanner serving 20+ countries. Dataset and benchmark code: https://github.com/halallens-no/halalbench