A durable OCR project preserves the source, records how derivatives were made, measures errors on representative material, and makes correction traceable. The goal is not a perfect-looking export; it is an explainable collection.
Use this collection when OCR must remain explainable across a large or long-lived collection. It separates preservation masters, access copies, OCR derivatives, layout metadata, ground truth, error metrics, sampling plans, and correction records instead of treating one export as the archive.
You should finish with named file roles, durable page identifiers, measurable acceptance rules, representative samples, and a correction trail. Any published accuracy statement should identify its material, ground truth, metric, normalization, sample, and limitations.
Test this cluster with OCR Quality Checker, then begin the detailed review with Preservation Master, Access Copy and OCR Derivative.
When an archive must retain coordinates or reading order, review the layout-preserving OCR formats and methods.
A durable project keeps the source, derivatives, identifiers, processing record, quality evidence, and corrections connected. One polished text file cannot replace that chain.
| Document signal | What it means | Useful next action |
|---|---|---|
| Preservation source | The highest-quality stable capture remains unchanged. | Protect it from routine OCR transformations and replacement. |
| Access and OCR derivatives | Files are optimized for viewing, search, or reuse. | Name their role, settings, source identifier, and creation date. |
| Quality sample | Representative pages are compared with reviewed ground truth. | Document selection, normalization, metrics, and acceptance rules. |
| Correction record | Human changes alter the machine transcript. | Retain what changed, why, by whom or by which process, and against which source. |
Sample by source type and risk, not convenience. A clean modern page should not stand in for faded type, marginalia, tables, mixed languages, or any other class that behaves differently.
Keep preservation masters separate, define measurable acceptance rules before production, retain page identifiers, and record corrections without overwriting the evidence.
Separate preservation masters, access copies, OCR text, and metadata.
Sample type, damage, columns, marginalia, and language changes before choosing settings.
Retain coordinates in ALTO, hOCR, TSV, or another documented representation.
Define ground truth, unit of measurement, sample design, and acceptance threshold first.
Separate masters, access files, layout metadata, accuracy measurements, historical source classes, and representative quality samples.
Keep an unchanged, well-documented preservation master. Create access images and OCR outputs as replaceable derivatives linked back to that master.
Sample each material and layout class separately. Historical OCR quality is usually limited by source variation, not by one universal preprocessing setting.
Use a layout format when future users need to locate text on the page. Document coordinate units, page dimensions, reading order, schema version, and engine metadata.
CER and WER compare OCR with a trusted reference, but the result is meaningful only when sampling, normalization, alignment, and ground-truth rules are documented.
Sample by risk and source type, not convenience. Review every high-consequence field and use randomized checks to detect systematic errors in the remaining collection.
Use a controlled quality sample to record source type, settings, corrections, and acceptance rules before making a collection-wide claim.
Open the OCR Quality Checker