Scan and fax image enhancement

The problem

Medical scans and inbound faxes arrived in poor states: skewed, noisy, low contrast, illegible. The OCR pipeline downstream choked on inputs that a human eye could read but a model could not. Bad inputs leaked into the clinical record before anyone realized, which created compliance exposure and clinician distrust of the system as a whole.

The approach

We built a server-side enhancement pipeline that deskews, denoises, normalizes contrast, and scores OCR readability before the document hits OCR. Failed inputs route to automatic re-processing with stronger filters; persistent failures escalate for human review. A small ML classifier picks the right enhancement profile per input type (scan, fax, photo of paper). Every transformation step writes an audit record.

Stack and engineering choices

• .NET server-side image pipeline
• Deskew + denoise + contrast normalize
• OCR-readability scoring
• Auto re-processing on failure
• ML classifier for input type
• Escalation queue for persistent failures
• Audit log per transformation

Outcome

The OCR pipeline saw input quality that no longer fluctuated with the scanner or fax line. Documents that previously corrupted the clinical record were caught and either repaired or escalated. The clinical team trusts the pipeline as a black box again, and the audit trail records every enhancement step.