Scan degradation metadata for invoice OCR datasets

Real invoices arrive as clean PDFs, office scans, and phone photos. A balanced training mix prevents models from overfitting to pixel-perfect renders.

Difficulty levels

clean: born-digital render, no artifacts
easy / medium / hard: progressive scan and physical effects

Physical artifacts

Punch holes, coffee stains, fold creases, rubber stamps (BEZAHLT, KOPIE), and handwriting overlays. Each sample records artifacts in rendering.degradation.physical_artifacts[].

Filtering with degradation_tags

Export manifest rows include tags such as tilted, stained, stamped, and occluded. Build curriculum learning schedules or hard-example mining without re-generating data.

The Scan Robust tier ships 10,000 fully degraded samples with clean/degraded image pairs.

Next: Get 50 free samples · View pricing · All guides