OCR Watermarked PDF Pages: Draft, Copy and Void Marks

OCR can read a watermarked PDF page when the mark is paler than the print, but the mark's own letters come along for the ride: DRAFT, COPY or VOID can land in the converted text, and words it crosses can be damaged. Convert the page, clean the watermark words out of your working copy, and leave the source file alone.

Can OCR read a page with a watermark behind the text?

Yes, when the watermark is clearly lighter than the print that sits on it. Recognition depends on the contrast between dark letters and their background, and a pale gray DRAFT leaves that contrast largely intact. Trouble starts when the mark is dark, colored, or was flattened to solid black by a black and white scan or a photocopy.

Open one watermarked page at 200% and look at a line that the mark crosses. What you see there predicts what the text will look like.

Two things follow from that table. First, the watermark is a different problem from a stamp. A stamp is one patch of ink in one place, and the post on whether OCR can read stamps over text deals with it. A watermark runs across the whole sheet on every page, so whatever it does to the text, it does two hundred times in a two-hundred-page file.

Second, you cannot predict the result from the look of the page alone. Run a test. Schema Law's free plan takes one document a day of up to three pages, so you can upload three watermarked pages and read the parsed result beside the originals. Its homepage says recognition is strongest on clean print, and that nothing marks a passage as uncertain. A word spoiled by a watermark stroke will look as tidy on screen as a word read correctly, so the comparison with the page image is the only check there is.

Why does a DRAFT or CONFIDENTIAL watermark end up inside the converted text?

A watermark is made of letters, and an OCR engine exists to find letters. It has no idea that some words on the page are the document and others are a label laid across it. Large pale capitals that it can make out are recognized and written into the output wherever the layout step decides they belong.

The documentation for the open-source Tesseract engine states the assumption plainly on its page about improving the quality of OCR output: "By default Tesseract expects a page of text when it segments an image." Under that assumption, everything dark enough to be a mark is a candidate for text. Your own tool may behave differently, which is one more reason to test a page.

Where the watermark's letters land

A diagonal word sixty points tall does not sit on any one line of the body text. It cuts through a dozen of them. So the engine has several ways to place it, and you may see more than one in the same file:

• As its own line. The whole word appears once, at the top, the bottom or the middle of the page text, as if it were a heading.
• In pieces. Each body line picks up the one or two capitals that cross it, so a stray "D" ends one line, "R" interrupts the next, and "AFT" turns up three lines later.
• Inside a sentence. The word or a fragment of it is dropped between two words of the clause it crosses.
• Inside a table cell. A fragment lands in a column of figures and looks like a code or a unit.
• Nowhere. The mark is too pale or too steeply angled to be read as letters, and leaves no words behind. It can still damage what it crosses.

Damage to the words under the strokes

The second effect is quieter and matters more. Where a thick gray stroke passes through a small black word, the two merge once the image is reduced to dark and light. A letter gains a bar it never had, or two letters join into one.

Take a fictional example. Calloway Marine Supply, Inc. is selling its assets to Petrakis Holdings LLC, and the last negotiating draft of the agreement carries a diagonal DRAFT on every page. On page 6 the stroke of the "A" runs through the price clause. The line below shows the kind of result to look for. It is an illustration of the failure, not output from any product.

On the page:   The Purchase Price shall be $2,350,000, payable at Closing
In the text:   The Purchase Price shall be $2,35D,000, payable R at Closing

The stray "R" is easy to see and easy to delete. The "D" inside the figure is the one that causes harm, because a reader who cleans up the obvious debris may stop looking.

What stray watermark words do later

Left in place, the watermark's letters cause trouble after conversion as well:

• A search for "draft" or "confidential" hits every page, so the search stops being useful for finding where the parties discussed drafts or confidentiality.
• A comparison of two versions in a word processor reports the watermark fragments as changes.
• A quotation pasted into a brief can carry a stray capital into the middle of a sentence.
• A word count or a table total picks up text that is not part of the document.

Scanned-in watermarks versus watermarks added to the PDF

A scanned-in watermark is part of the page picture: toner on the paper that was scanned, now pixels in the image. A watermark added to the PDF is a separate object that software placed over or under the page content. The two look alike on screen, and three quick checks tell them apart.

1. Try to select the mark. Drag the text cursor across the watermark word. If the word highlights by itself, it is stored in the file as text.
2. Zoom to 800%. An added mark made of text or drawn shapes keeps a crisp edge while the scanned page behind it turns grainy. A scanned-in mark is exactly as grainy as the print around it.
3. Search the file as received for the watermark word. A hit in the same spot on every page means the word is real text in the file.

The answer changes what you do next.

For the middle row, do not assume either way. Whether an added mark reaches the converted text depends on how a given tool turns the page into something it can read. Convert one page and look.

Whichever kind of mark you have, leave it in the file you received. The sender or the records office put it there, possibly because a rule, an order or a licence requires it, and a copy without it is no longer the document you were sent. If you need an unmarked version, ask the source for one.

Paper watermarks and printed letterhead backgrounds

A true paper watermark, the kind you see by holding bond paper to a window, is a variation in the thickness of the sheet. A scanner lights the page from the front, so that mark does not appear in a scan and does not affect recognition. A pale crest or logo printed behind the text of a letterhead is different: it is ink, the scanner records it, and it behaves like any other background mark.

VOID patterns and security paper on certified copies

Security paper carries a printed background that changes when the sheet is copied. A pattern that reads as an even tint on the original can come out as the word VOID, repeated across the page, on a photocopy or a scan. For OCR, that is a second layer of large letters behind the certified text.

The Social Security Administration's operating manual describes the paper in its section on photocopies marked void, GN 00301.280(B)(5): "some public custodians are now issuing documents on a new type of paper stock which when photocopied will show the word 'VOID' on the face of the photocopy." The manual gives the purpose as an attempt "to reduce fraudulent use".

That background is doing its job when it shows up on your scan, so the task is to read through it, not to get rid of it.

What to do with a scan that shows VOID

• Compare with the paper if you hold it. If the original shows no VOID and the scan does, the word is a product of copying. Say so in your file note, so nobody later reads the scan as a cancelled document.
• Keep the scan as it came out. Do not adjust the image to make the pattern disappear. The pattern is a security feature of the document.
• Expect VOID in the text. It lands in the same ways a DRAFT mark does: as its own line, in fragments, or between two words of the certificate.
• Proofread the fields that matter. Names, dates, the certificate or file number and the custodian's certification language are short, and a single wrong character changes them.

Suppose a paralegal scans a certified death certificate for the fictional estate of Marguerite Oyelaran. The paper shows a plain green tint. The scan shows VOID eleven times, and the converted text reads "Date of Death VOID March 4" on one line. She deletes the stray word from her working notes, checks the date against the paper, and writes "VOID appears on copies only; original on security paper in the estate file" at the top of the notes.

Whether a copy that shows VOID is acceptable for a filing, a claim or a recording is a question for the court or office that will receive it. Check its rule before you send a copy in place of the certified original.

Other printed backgrounds on certified documents

Security paper is one of several printed backgrounds that sit behind text. Tinted panels behind an amount, a border of fine repeated lettering, and a pale seal printed across the middle of the page all add marks that are not part of the wording. Check stock is the most familiar case, and the post on scanned checks covers security backgrounds and patterned check stock. A raised seal pressed into the paper is a separate matter again, because it has no ink at all. See whether OCR can read an embossed seal.

Unofficial-copy watermarks on records downloaded from court and county portals

Some record portals mark the copies they hand out without charge. The mark tells a reader that the page is a reference copy and not the certified or official one. Whether your portal does this, and what its mark says, is something you can see on your own download and in the portal's terms.

One example you can read for yourself: the Tarrant County, Texas clerk's page on copies of official public records for real property says of its free downloads, "Your document will automatically download with an 'Unofficial Copy' watermark." The same page describes a separate route for ordering a certified copy. Other offices set their own practice, so read the page for the office you are dealing with.

On an unofficial download of a deed, the diagonal mark often cuts the legal calls you need for marketable title. You read lot, block, and book figures from the image, and you note the lien or quiet title language the stroke crosses. Title insurance staff will ask which copy you converted, so the header note should say unofficial.

Converting a watermarked download

A marked reference copy is fine for reading, searching and drafting from. Conversion works the same way as for any other watermarked page, with a few points that are specific to recorded and filed documents.

1. Find what the mark crosses. On a deed, look at whether it runs through the legal description. Bearings, distances and lot numbers are strings of digits and symbols with no surrounding context to catch a mistake. The post on why OCR gets numbers wrong explains why those are the first characters to check.
2. Check the recording data. Instrument number, book and page, and recording date are how everyone else will find the document. Read each against the image.
3. Do not try to remove the mark. The office applied it to distinguish this copy from the official one. A reference copy with the mark taken off is still not the official copy, and it now misstates what it is.
4. Order the official copy when the page matters. If the mark hides a call in a legal description or a figure in a judgment, the fix is the unmarked copy from the office, converted afresh.
5. Use the right copy for the right purpose. What a court, a title company or a recorder will accept is set by their own rules. Check before you attach a download as an exhibit.

Keep a note of which kind of copy your text came from. "Converted from unofficial download, 12 June" tells the next person that a quotation should be checked against the certified copy before it goes into anything filed.

How do you clean watermark words out of converted text without touching the record?

Clean the export, never the source. Keep the PDF exactly as it arrived, convert a duplicate, and take the stray watermark words out of the Word or text file you work in. Then write down that you did it, so anyone reading the working copy knows what it was made from and what was removed.

A cleaning procedure

1. Lock the source. Store the received file under its original name, set it to read-only, and make a duplicate for conversion.
2. Convert and export an editable format. Word or plain text is the easiest to clean. Schema Law's export formats include both, and you can export the same upload again in another format later.
3. Search for the watermark word in capitals, as a whole word, with case matching on. Step through the hits one at a time.
4. Search for fragments. Try the first three letters and the last three, and page through looking for lines that hold only one or two capitals.
5. Read each cleaned line against the page image. A fragment marks the spot where a stroke crossed the line, so the neighboring words are the ones at risk.
6. Add a header note to the working copy. For example: "Converted from Calloway APA draft v7.pdf, received 9 May. Source pages bear a diagonal DRAFT watermark. Watermark text removed from this working copy. Verify quotations against the source."
7. Quote from the source. When a passage goes into a letter or a brief, check it against the page image of the file you received.

Why Replace All is the wrong tool

The watermark word can also be a word the document uses on its own account. A blanket replacement deletes both.

Case matching separates some of these, because the watermark is in capitals and the body text is not. It does not catch a heading in capitals, which is why each hit needs a glance.

What you cannot clean

A searchable PDF export keeps the original page with a hidden text layer behind it. That is the right format for the file, because the page image still shows the watermark exactly as received. It also means the hidden layer may hold the watermark's letters, and a search for "draft" will keep hitting every page. Search for a phrase of two or three words instead, and use the Word or text export when you need clean language.

Removing the stray words does not mean erasing the fact of the watermark. That a page was marked DRAFT or UNOFFICIAL COPY can matter a great deal later, for instance when someone asks which version was signed. The header note keeps that fact attached to the text. If you prefer to mark it page by page, the post on redacted documents sets out a bracket convention for noting in converted text what the page shows and the text cannot. For pages where the mark destroyed words outright, keep a list in the way the triage guide describes under recording what could not be read.