Reading printed Arabic
Kindi OCR reads scanned Arabic newspapers and books, and returns text you can search and correct.
0.89%Character error rate on MDP-9095% interval: 0.6–1.2
The test
One fixed test that every version and every system is scored on.
- Set
- MDP-90: ninety text blocks from Arabic newspaper pages.
- Scoring
- Each block is scored as a whole, so mistakes in line and column order count too.
- Metric
- Character error rate: substituted, deleted, and inserted characters, divided by the characters in the true text.
- Confidence
- Every system gets a 95% interval next to its number.
The result
The lowest error on this test, out of 5 systems.
| System | Character error rate | 95% interval |
|---|---|---|
| Kindi OCR | 0.89% | 0.6–1.2 |
| Mistral OCR v4.1 | 1.29% | 1–1.6 |
| deep.sa OCR v2.0 | 2.63% | 1.9–3.7 |
| PaddleOCR-VL 1.6 | 5.4% | 4.2–6.5 |
| Tesseract | 6.53% | 4.5–9.3 |
0%2%4%6%8%10% |
The whiskers at each bar show the 95% interval.
Reading the result
Our interval (0.6–1.2) overlaps Mistral OCR's (1–1.6), so we call it the lowest error on this test without calling it a decisive gap. Against PaddleOCR-VL and Tesseract the gap is clear: they sit near 5% and above.
What the number means
On a real page.
In a paragraph of a thousand characters, Kindi OCR gets about 9 wrong. That's what's left for a person to correct.
Have a printed Arabic archive?
Newspapers, books, or scanned documents. We'll read them with you and measure the result on your own sample.