OCR usually leaves the scan where it is. The page is still a picture, with recognized words placed over it so you can search and select them. That is the right choice when the page has to keep looking exactly as it did. It is the wrong one when you want to work on the document: a table in a scan is still a picture of a table, and adding a row means covering pixels with text boxes.

Digitize does the other thing. It reads a photo of a printed page, rebuilds what it finds as tables and text, and then removes the photo. What is left is a page like one made on a computer: cells you can type in, rows you can add, captions you can move. It is in Papaya PDF on iPhone and Mac.

To see how close it gets, we ran a test with a known answer. We printed a one-page PDF with three tables, photographed the printout with a phone, and digitized the photo three times, once with each model Papaya PDF can run. Then we compared each result with the PDF we had printed.

A phone photo of a printed page with three tables, lying on a wooden table

The same page after Digitize: three tables, their captions and three footnotes as editable content on a clean white page

The first image is the photo. The second is the digitized page from the most capable of the three models. Nothing in it is a picture any more.

What Digitize does

  1. You choose a photo, or a page that is already a scan.
  2. Papaya PDF finds the sheet of paper and marks its four corners. Drag a corner if it missed.
  3. The page is straightened to a standard paper size and its background is turned to clean white.
  4. A model reads the page: the words, where the tables are, and how each table divides into rows, columns and merged cells.
  5. Tables are rebuilt as tables with their ruling lines redrawn straight. Everything else becomes text boxes at the size and weight it was printed.
  6. The photo is deleted from the page.

All of this runs on your device. The photo is not uploaded.

Three ways to read the page

Step 4 is where the results differ. Papaya PDF can use three models on Apple devices, and you choose between them in settings.

  • Apple Vision is the text and document recognition built into iOS and macOS. There is nothing to download, and everyone can use it, signed in or not.
  • Docling is a pair of models: one finds the tables on a page and one splits each table into cells. The words still come from Apple Vision. It is a download of about 370 MB.
  • Surya OCR 2.5 is one model that reads the whole page at once and writes each table out with its structure and formatting. It is a download of about 800 MB.

Docling and Surya are available to Pro accounts, and to free accounts within a weekly allowance of pages.

The same table, four times

The second table on the test page is the hard one. Two of its row headers span several rows, three cells carry small raised footnote marks, and the header row is bold. Here it is in the PDF we printed:

Table 2 in the original PDF, with bold headers, two row headers spanning several rows, and small red footnote marks 1, 2 and 3

Apple Vision

Table 2 digitized with Apple Vision: footnote marks read as ordinary digits, year headers not bold, and an empty extra cell below Remunerated functions

Every word and all fourteen figures are right, and the caption and footnotes are in place. The table's structure is not quite right, and the details are gone:

  • The mark after 2010/11 was read as a digit, so the header says 2010/111.
  • The mark after Information is missing. The one after Agency services became a plain 3.
  • The year headers are not bold.
  • "Remunerated functions" should span four rows. It covers one, with an empty cell under it.

Docling

Table 2 digitized with Docling: the merged row headers are correct, but footnote marks are still read as ordinary digits and the year headers are not bold

Docling fixes the structure: both row headers span the rows they should. The words still come from Apple Vision, so the footnote marks and the bold headers are wrong in the same way.

Surya OCR 2.5

Table 2 digitized with Surya OCR 2.5: bold headers, correct merged cells, and footnote marks 1, 2 and 3 set as superscripts

Surya reads the table as a table. The three footnote marks come back as superscripts attached to the right cells, the header cells are bold, and the merged cells are correct. This is the version you could hand to someone without first going through it against the printout.

Side by side:

  • Words and figures: correct in all three.
  • Captions and footnotes: kept in all three.
  • Merged row headers: Apple Vision got one of the two. Docling and Surya got both.
  • Footnote marks: read as digits or dropped by Apple Vision and Docling. Superscripts with Surya.
  • Bold header cells: Apple Vision and Docling missed two. Surya got all of them.

What none of them reproduce

A digitized page is a reconstruction, and it is honest to say what it does not bring back.

  • The typeface. The original was set in Verdana. Every result uses Helvetica, the closest built-in face Papaya PDF matched from the ink. Widths differ a little, which is why the digitized tables are not pixel-for-pixel the same size as the printed ones.
  • Numeral style. The original figures are old-style numerals from a second font. The results use ordinary lining figures.
  • Colour. The footnote marks were red. They come back black.
  • Exact text sizes. Sizes are measured from the photo, so they are close rather than exact. The small "Table 1" caption came out smaller than it was printed.

Across more pages

One page is an example, not a measurement. We also score every model against 17 pages with hand-checked answers: 14 table PDFs rendered as clean scans and 3 real scans, among them a printed table from 1918 and a phone photo of a receipt.

  • Words read exactly: 92% with Apple Vision, 98% with Surya. Docling uses Apple Vision's words, so it scores the same 92%.
  • Table cells recovered into the right row: 68% with Apple Vision, 77% with Docling, 96% with Surya.

The second figure is the one that matters for Digitize. A wrong word is a quick fix. A table split into the wrong rows has to be rebuilt by hand, which is the work Digitize exists to save.

What Pro adds

Apple Vision is free, needs no download, and reads a page in a few seconds. For a letter or a simple grid it is often enough.

Docling and Surya are Papaya PDF's Advanced OCR. Pro runs them on every page with no limit. A free account gets a weekly allowance of pages, which is enough to try them on a document of your own before deciding. Signed out, Digitize uses Apple Vision.

Surya is the slowest of the three. On the Mac we tested, the page above took about a minute, running on the graphics processor. It is also the only one that returned the page without a list of things to fix, and it does that without the page leaving the device.

Check the result

Digitize removes the photo, so the digitized page is the only copy in the file. Keep the original photo until you have read the result against it. Figures, names and dates deserve a second look whichever model you used: a higher score means fewer mistakes, not none.

For pages that must keep their original appearance, use ordinary OCR instead, which leaves the scan in place. The guide to editing a scanned PDF with OCR covers that path, and editing a table in a PDF covers what you can do with a table once it is a real one.