Scanned PDFs: OCR is the beginning, not the end

OCR attempts to recognise characters in an image and create text. A previously image-only page can then become searchable and available to assistive technology. Yet a misrecognised number may still look perfectly correct in the visible scan.

OCR does not automatically create correct headings, lists, table relationships or image descriptions. The text layer and document structure are different parts of the work.

Sources: [1]

Inspect the scan before choosing settings. Are pages rotated, skewed or cropped? Are there shadows near a book spine? Are small numbers blurred? OCR cannot reliably recreate information absent from the image. A better scan or original source can save both work and errors.

Choose recognition languages appropriate to the content and pay particular attention to multilingual documents. Language selection influences the characters and words the tool attempts to recognise. Preserve an unchanged scan for comparison; work on another file and identify which version underwent recognition.

Acrobat’s current workflow places recognition under Scan & OCR, Recognize Text and In this file. Select pages and language before running it. Completion means a text interpretation has been created, not that it is correct or structurally accessible.

Compare recognised text with the page image, especially where one character changes meaning. Check names, dates, amounts, minus signs, decimals and units. In a decision letter, “not” may matter more than an entire decorative heading. Set review depth according to the consequences of error.

Search for a known word and copy difficult lines into plain text. This exposes hidden mistakes, but sampling cannot replace the proofreading warranted by the material’s risk. If 1 became l on one page, the same font or scanning quality may have caused problems elsewhere.

Do not guess when the source is illegible. Flag uncertainty and ask the document owner to establish the content from better evidence. Record corrections affecting numbers or significant words, so an intentional correction can be distinguished from a new recognition error.

Only once the text is sufficiently trustworthy should you build headings, lists and table relationships around it. A correctly tagged incorrect number is still incorrect information. Finish with both content review and a practical reading test of the structured document.

• First, text accuracy: names, amounts, dates, minus signs and word boundaries.
• Then, structure: headings, reading order, tables, links and other content.
• Check for duplicate text layers.
• Try searching, copying and practical reading in the final file.

If the scan is too poor or tables cannot be interpreted reliably, a new checked source file may be better than continued guessing. Preserve the connection to the original and let the document owner resolve ambiguous information.