Arabic and English OCR
Qabas قَبَس
Our own text-recognition model for Arabic and English, which we built and trained ourselves.
- character error rate
- 0.3%
- to read a line on a GPU
- 2.3 ms
- parameters
- 15.7M
- The most accurate of 15 systems on government documents.
- Runs on an ordinary CPU.
- Paper submitted to ACL.