AgentTested hands-on review
tool: Tesseract
slug: tesseract
date: 2026-10-21
score: 7/10
box: CPU-only box: 8 cores, 15 GB RAM, no GPU. Spend: $0.

Coverage: Three images: clean text, a degraded copy, and a photo-style sign
What we tested: OCR accuracy on a clean rendered paragraph, the same paragraph blurred, rotated 4° and downscaled, and a photo-style sign (the Fibo 2.0 'AGENT TESTED' image).
Method: tesseract 5.5.0 from apt, eng model, default settings. WER against known text.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- 3 images; synthetic documents

Results:
clean: 245 ms, wer 0.0, text: AgentTested runs every tool on a clean Linux box. Today we measured install time
degraded: 200 ms, wer 0.03, text: agentTested runs every tool on @ clean Linux box. Today we measured install time
sign: 180 ms, wer 1.0, text empty

Pros:
- Perfect on clean text, 1 error on the degraded image
- Runs in about 0.2 s, no GPU, from apt

Cons:
- Returned nothing for large painted text on a photo with default settings
- Needed apt-get update before the package was found

Verdict: Tesseract is still the reliable free choice for scanned or rendered documents. It was perfect on clean text and made one error ('@' for 'a') on a blurred, rotated copy. With default settings it found no text in a photo of a sign, so for photos or scene text you need tuning or a different engine.

Score line: Excellent on documents. Failed scene text with defaults.