Purpose
Convert the enhanced image into text that is useful enough to clean and speak.
Role in the complete pipeline
Recognition compares a custom CRNN-CTC path with Tesseract fallback output, then uses confidence and readability checks to select or reject a candidate.
INPUT
→Enhanced single-frame and stacked image candidates.
METHOD
- Extract image features
- Decode a character sequence with CRNN-CTC
- Generate fallback candidates with Tesseract
- Check confidence, readability, spacing, and agreement
Selected text, its provenance, and validation signals; or a rejection reason.
Common failure cases
- Similar glyphs are confused
- Line ordering is wrong
- Low confidence still looks superficially plausible
- Stylized fonts fall outside training data
How VisoraAI responds
- Compare alternative OCR paths
- Reject implausible text
- Ask for a new capture when no candidate is dependable
Current limitations
- Confidence is not guaranteed correctness
- Verified evaluation data is still needed before accuracy claims
Planned improvements
- Evaluate on a documented dataset
- Expand controlled blur training
- Add language support after English stabilizes