AURORA OCR Character recognition for difficult real-world documents
A proprietary character-recognition and document-understanding engine for Latin and Cyrillic documents: scans, photos and PDFs. Our newest model also reads Greek.
Recognition performance
How to read these numbers. They are character-level and text-line results, not full-page or end-to-end document OCR accuracy. Upper- and lower-case letters that share one shape (c/C, o/O, s/S and similar) are scored as correct by the character model; their case is decided afterwards by a separate rule from the geometry of the text line.
The model
- Architecture
- ConvNeXt-Small, ImageNet-pretrained, fine-tuned in PyTorch
- Parameters
- ≈ 49.5 M
- Training data
- ≈ 1.1 M glyph samples, each validated by a person (the model measured above)
- Test set
- internal test set of real PDF documents, never used in training
- Newest model
- 1,643 characters: Latin, Cyrillic and Greek letters with accents, digits and signs. Its accuracy will be published after the same evaluation.
How we evaluate
We test on independent public benchmarks, not only on our own data: olmOCR-bench (real PDFs) and UNLV (scanned business documents). Our methods are compared with 7–8 established PDF readers and OCR tools, category by category:
- Small print
- Scan quality
- Tables
- Formulas
- Multi-column pages
- Headers and footers
Benchmark results are being prepared for publication.
What we are working on
Reaching at least 99.8 % under realistic degradation: displaced characters, background artefacts, blur, low contrast, varying stroke weight and phone photos of paper. The training pipeline uses hard-example mining, targeted retraining and simulated scanner and camera damage.
The next stage is line-level OCR: preliminary text-line accuracy is about 93–94 %.
Need OCR for a document workflow?
Tell us about your documents, languages and processing requirements. We will discuss whether AURORA OCR fits your use case.