Check a PDF generated by OCR before trusting its text layer

Where the extra work starts

Test what the text layer actually says so a convenient copy-and-paste result does not hide corrections that will dominate the rest of the job.

ParseMill processes native invoice PDFs and clear printed English PDF scans. Try a typical supplier invoice and a difficult one. Check the extracted fields, correct exceptions and open the workbook before committing to a monthly plan.

Keep the original scan and OCR-prepared copy distinguishable. ParseMill uses English OCR when a PDF page has no native text spans. It does not automatically replace an existing but incorrect OCR layer, so verify both its characters and layout before relying on the invoice output.

A practical way to get it right

Choose checks that matter to the intended output: an identifier, a decimal amount, a signed value, a unit and a heading. Copy each from the OCR layer and compare character by character with the image at readable magnification. Record exact locations and any uncertainty.

Inspect more than one page and include low-quality regions, small print and different fonts. A clean first page does not establish the rest of the file. Keep errors in a review log and decide whether the source quality supports correction, a revised OCR attempt or a request for a better original.

Only after text verification, inspect heading alignment and row relationships. OCR may create readable words whose positions or ordering still confuse a table parser. A bounded trial should retain both character and structural acceptance checks, with recovery effort included in any usefulness assessment.

Adobe: Recognize text in scanned documents explains that OCR creates a text layer and recommends accuracy and completeness review afterward. Use representative field checks before relying on that layer.

What this looks like on a real task

Synthetic example. The people, references and figures below are invented to demonstrate the method.

One correct unit does not offset three material errors. Copy the visible value, OCR text and finding columns into the trial record. Correct only when the image clearly supports the replacement; an unreadable original should remain unresolved rather than being guessed from nearby values or expected totals.

Catch the cases that create rework

Do not repeatedly rerun OCR and choose the output that produces the most plausible totals. Compare against the source image. If several OCR versions differ, retain the selected version and the evidence supporting any corrections.

Editing recognized text can also change layout or obscure the distinction between original and prepared content. Preserve the original and document the preparation rather than overwriting the only source copy.

Adobe: Edit scanned documents recommends checking OCR output and keeping a backup. Preserve the original while evaluating corrections or a new recognition attempt.

Decide whether the change is worth keeping

The readiness decision should state sampled pages, error types, corrections and untested areas. A high error burden may make manual transcription or a better source preferable. If a trial proceeds, compare every required extracted field with the appropriate source evidence. Selectability is a technical property; reliability must be demonstrated through the checks relevant to the task.