How to OCR Scanned PDF Drawings

Open the PDF and try to select a word with the text-selection tool. Then search for an obvious sheet title or note. If you can select individual characters and search already finds them, the page has a text layer; OCR may add duplicate or noisy text rather than help. If selection treats the page as one image and search returns nothing, it is a good OCR candidate.

OCR changes text discoverability, not the drawing geometry. It does not calibrate the sheet, validate dimensions, or confirm that a note was recognized correctly.

1. Duplicate the source PDF or use Save As so the original scan remains untouched.
2. Close unrelated large documents if the scan is long or image-heavy, then open the working copy.
3. Choose Document › OCR. OCR starts for the current document; the current dialog does not offer a page-range or language picker.
4. Keep the dialog open to watch progress, or close it if you want the run to continue in the background. Choose Cancel OCR when you need to stop; a cancelled run discards its result.
5. Wait for the completion message before judging search. Large scans and construction sets can take time.
6. Save the recognized document under a distinct filename, close it, and reopen that saved file.
7. Search several terms from different pages and copy short passages into plain text to inspect recognition quality.

Since PolyPDF 1.4.4, OCR also builds a current-session model of likely schedules, title-block fields, drawing and detail labels, dimensions, regions, rows, and cells. This is useful for review and spreadsheet handoff, but it is still inferred from page pixels and selectable text—not an authoritative schedule database.

1. Finish the OCR run and inspect the recognized table names, headings, rows, and dimensions.
2. Choose File › Export › Excel Workbook (Tables + Text)… while the recognized document session remains open.
3. Save the workbook under a name tied to the source drawing and issue.
4. Open the workbook and compare critical quantities, dimensions, tags, and schedule cells with the visible PDF.
5. Treat merged cells, faint rules, handwriting, rotated labels, and dense linework as high-risk review areas.

The structured table model is session data. Saving and reopening preserves the searchable PDF layer, but you should not assume the inferred table model will still be available unless you run OCR again. Export the workbook before closing the session.

Recognition availability comes from the operating system and its installed language support, so the languages offered by one Mac or Windows computer may differ from another. PolyPDF’s current embedded searchable layer is limited to Latin, Greek, and Cyrillic scripts. The operating system may recognize text in additional scripts, but PolyPDF does not promise to embed those characters as a searchable layer in the PDF.

• Treat mixed-script title blocks as a special review case.
• Install and enable the needed OS language support before the project starts, then test with a representative page.
• Do not infer that a displayed language name guarantees equal accuracy across fonts, scan quality, rotations, or handwritten notes.
• When PolyPDF reports that recognized text cannot be embedded, use any offered text export as a review aid, not as proof that the PDF itself is searchable in that script.

A general search test can pass while the identifiers you need still fail. Build a small verification set from the document: a sheet number, a room name, a material abbreviation, a dimension, and a note containing punctuation. Search each value exactly, then try a distinctive fragment. Inspect both true hits and obvious locations the search missed.

• Expect confusion between similar shapes such as O and 0, I and 1, S and 5, or decimal points and scan noise.
• Rotated notes, condensed fonts, faded diazo prints, skewed scans, and text crossing linework are harder inputs.
• Never copy an OCR-derived dimension, quantity, equipment tag, or specification value into downstream work without comparing it to the page image.
• Search results are a navigation aid. The visible drawing remains the source that must be reviewed.

The NACA page shown here is a raster scan with no extractable text before OCR: searching SURFACE returns “No matches.” After the whole-document run finished, we saved the result, quit PolyPDF, reopened the saved PDF, and repeated the same search. PolyPDF returned three results and selected one “Surface of ship” occurrence on the drawing.

Copying text out of the saved PDF returns “REPORT NATIONAL ADVISORY COMMITTEE FOR AERONAUTICS,” “Surface of ship,” and “U.S. S. Akron.” The same text also contains errors in small italic labels and dimension notation, so use OCR hits for navigation and compare every consequential identifier, note, or measurement with the visible scan.

• OCR is best-effort recognition, not a transcription warranty or drawing-validation service.
• A searchable text layer does not make the PDF accessible. Reading order, headings, alternative text, form labels, and other accessibility structure require separate review.
• OCR does not remove confidential pixels. If a scan must be redacted, use an image-aware redaction workflow and verify the output.
• Recognition runs locally through platform capabilities, but OS language availability and results can differ between computers.
• A completed progress bar means the run finished. It does not mean that every word was found or embedded correctly.

For consequential work, define the acceptable use before running OCR: navigation and discovery are reasonable; unreviewed extraction of dimensions, quantities, or compliance language is not.