Home / Patterns / OCR Fails on Low-Resolution Scans and Fax Images: Causes and Fixes

OCR Fails on Low-Resolution Scans and Fax Images: Causes and Fixes

Goal Accurate Text Extraction Frequency Very Common Category Document Processing Published View source on GitHub ↗

Issue: Low DPI scans, faxes, and compressed mobile photos produce fragmented, unreadable OCR output

Frequency: Very Common

Symptoms

  • Consistent errors on documents from specific sources (fax, mobile upload)
  • Small text (footnotes, fine print) fails completely
  • JPEG compression artifacts cause characters to fragment or merge

Root Cause Low DPI scans, aggressive compression, and small original text create images where characters lack sufficient detail for recognition.

Example

Document: Faxed invoice at 100 DPI
Footer text: "Terms: Net 30 days"
Extracted: "Tenns: Nel 30 drys"

Result: Payment terms not correctly parsed

How to Fix Low-Resolution OCR Failures

Mitigation Strategies

Prevention

  1. Minimum DPI gating with channel-aware thresholds: Measure effective DPI (or an equivalent sharpness proxy for photos/faxes) at intake and reject or flag documents below a channel-specific threshold (e.g., 200 DPI for scans, a lower bar plus extra preprocessing for fax/mobile capture), since fax and mobile-upload channels are the primary source of this failure. Trade-off: overly strict thresholds reject legitimate low-quality-but-readable documents, so thresholds should route to enhanced processing rather than outright rejection.
  2. Super-resolution and denoising preprocessing: Apply a super-resolution model or targeted denoising pass before OCR specifically for documents below the DPI threshold, since small text (footnotes, fine print) fails completely without added detail. Trade-off: super-resolution adds meaningful latency and compute cost, so it should be applied selectively rather than to every document.
  3. Multi-scale ensemble extraction: Run OCR at multiple resolutions/upscale factors on the same region and ensemble or vote across results, since a single fixed scale may not recover enough detail for JPEG-artifact-fragmented characters like “Tenns” for “Terms”. Trade-off: multiplies OCR compute cost per document.

Detection & Response

  1. Per-source/channel accuracy tracking: Break out extraction accuracy by document source/channel (fax, mobile upload, scanner model) and alert when a channel’s accuracy diverges from baseline, since low resolution is a channel-correlated failure rather than a random one.
  2. Image quality metric correlation: Monitor DPI, file size, and noise/artifact levels alongside accuracy for every document, so a quality regression can be causally linked to an accuracy drop rather than investigated blind.
  3. Source quality feedback loop: When a document fails the quality gate, notify the upstream system or sender channel (e.g., prompt a re-scan or re-fax at higher quality) rather than silently degrading output, closing the loop at the source of the problem.

Architecture Patterns

  1. Tiered quality-based routing: Route documents into quality tiers (high-DPI direct-to-OCR, low-DPI enhance-then-OCR, below-minimum reject-or-review) rather than a single binary pass/fail gate.
  2. Targeted high-resolution re-extraction: For critical fields (totals, dates, IDs) that fail confidence thresholds on the first pass, re-extract just that region at higher zoom/resolution rather than reprocessing the whole document.
  3. Confidence-gated human-in-the-loop review queue: Route documents that remain low-confidence after super-resolution and multi-scale ensembling to human review instead of accepting best-effort output on critical fields.

Metrics

  1. accuracy_by_source_channel: Target: within 5% of overall baseline per channel; Alert threshold: > 15% divergence
  2. sub_200_dpi_document_rate: Target: monitored, not necessarily minimized; Alert threshold: sudden increase > 20% week-over-week from a given source
  3. super_resolution_recovery_rate: Target: > 70% of low-DPI documents reach acceptable confidence after enhancement; Alert threshold: < 50%
  4. critical_field_confidence_post_reextraction: Target: > 0.9; Alert threshold: < 0.7

Alerts

  1. Channel Accuracy Drop (P2): Condition - a specific source/channel’s accuracy falls more than 15% below baseline over a rolling window. Action: Sample documents from that channel, check for a quality regression (e.g., fax line degradation), trigger source quality feedback.
  2. Super-Resolution Recovery Failure (P2): Condition - recovery rate after enhancement drops below 50% for a channel. Action: Investigate whether the enhancement model needs retuning for that document population.
  3. Critical Field Low Confidence (P1): Condition - a financial or identifier field remains below 0.7 confidence after targeted high-res re-extraction. Action: Route to human review before the document proceeds downstream.

References

Related Patterns