ブログ
LangTo読了まで約6分

How to Translate a Scanned PDF and Keep the Page Layout

A scanned page has no text to translate. See how the page gets rebuilt, what came through on a real transcript, and how to fix anything that did not.

A scanned PDF has nothing to translate. The page is a picture of a document; the words in it are ink, not text. Whatever translates it has to read the page, lift the printed words off the image without disturbing what was printed around them, and set the translation in their place.

Getting the words off is the part that decides whether your document survives. It is worth two minutes to understand, because it also tells you what to expect from the result.

Lifting the words off without taking the page with them

To a computer, the line of a table, the ring of a registry seal and the stroke of a signature are the same thing as the letters: dark pixels on light paper. Erase too eagerly and the table loses its grid. Erase too carefully and the tails of the old words stay behind, sitting beside the new ones.

There is one property that separates them, and it does not depend on recognising anything: a ruled line is a long straight run of ink, and no stroke inside a letter is. Lines can be found by that shape alone and left where they are, while everything else in the area comes away, including the parts of letters that touch a line.

That last clause is what makes it work on real paper. An underline in a scan usually touches the feet of the letters above it, so the whole thing is one joined-up shape; treat it as a shape and sparing the line spares the sentence with it. Treat it as runs and the line is a run while the letters are not. On a test page where the two touch, the letters come away completely and the underline keeps 99% of its pixels.

Ink that carries on past the area is left alone for a different reason: a seal ring or a signature is not part of this line's text. On a page where a registry seal sits across two signature rules, the seal and both rules keep their pixels while the words between them are replaced.

In practice that means the grid of your table, the stamp on your certificate and the signature at the bottom arrive intact, and the words they surround arrive in your language.

What came out on a real document

We ran a Brazilian higher-education transcript through the free preview. It follows the format Portaria MEC nº 1.095 lays down, scanned at 400 DPI in colour, the resolution credential evaluators ask for. Portuguese into English, one page, no glossary and no instructions.

The page is a demanding one: a continued table with no header context of its own, two blocks of label-and-value pairs, a five-line grading legend with numbers inside the sentences, a registry seal overlapping the signature rules, and a long authenticity code in the footer.

Left: the scanned Portuguese source. Right: the English preview, watermarked as free previews are.

  • The seventeen table rows kept their order and their columns. Every course code, workload, credit count, grade, attendance percentage and status stayed on the course it belongs to.
  • The authenticity code came through character for character, as did every course code.
  • The grading legend was translated as sentences, with the numbers inside them unchanged: the 0,0 to 10,0 scale, the 6,0 pass mark, the 75% attendance minimum, the sixty-minute class hour, the fifteen-hour credit.
  • The seal and both signature blocks kept their pixels, and each name stayed paired with its title.
  • The dates were rewritten for an English reader. 18/12/2022 reads 18 Dec 2022, and 09/05/2023 reads 9 May 2023. Same days, and no ambiguity about which number is the month.

The seal crosses both signature rules in the source, and still does in the translation.

And if a line is not the way you need it

Any translation of an official document has a few lines only you can settle: a course title your registrar words differently, a job title with a particular equivalent in the country you are applying to, a name transliterated the way the rest of your paperwork spells it.

So the rebuilt page is not a picture with a translation painted on. It carries real, selectable text, 2,296 characters of it on this page. You can search it, and you can copy an identifier out of it and paste it into a form.

You can also change it. Selecting any line shows what it says in both languages, and the review editor offers every line it set, 76 separate pieces on this transcript, from the table headers to the footer, for you to rewrite.

Every line on the page can be opened this way, in the free preview as well as after payment.

Editing one of them moves nothing else. The page is composed again from the original with only the piece you changed replaced, so everything else keeps its exact pixels. Rewriting the institution's name on this page altered 0.26% of it, all inside that one line; the rest came out identical.

That is the difference between a translation you have to accept and one you can make correct, and correcting a line costs you none of the layout you just checked.

Four things worth a look before you send it

Most of the page checks itself. These four are quick, and they are the ones your own eyes are best at.

Follow one row all the way across. On any table, pick a couple of rows and read them left to right. The grade, the hours and the status should still read together as one record.

Copy an identifier and paste it somewhere. Registration numbers, reference codes and validity strings exist to be typed into other systems. Paste one and see it arrive whole.

Read the small print. Legends and footers carry the rules the rest of the page depends on: the grading scale, the attendance minimum, the validity code. Check the numbers inside those sentences.

Read the two or three lines a decision will turn on. Wording is where your judgment is worth more than any automatic check.

Because a scan begins as an image, every automatic check is made against the reading taken from that image, and anything our review cannot line up with it is listed page by page for you to confirm. On this transcript that list makes a short job of a page that would otherwise take a careful hour.

Start with the page that costs the most to get wrong. LangTo takes PDFs up to 20 pages and 50 MB at $4.95 per page, and translates one page of your own file free, so you can see the rebuild before paying for the rest. Try one page free