Can OCR Read Stamps Over Text? What to Expect
By Schema LawPublished Updated
OCR can read text with a stamp over it when the stamp is lighter than the print, or a different colour, and the scan kept that difference. It usually fails where dark ink covers the letters, or where the page was scanned in pure black and white. The stamp's own wording is a separate reading job.
What does a stamp do to the text underneath?
A stamp adds a second layer of marks on top of the printed letters. OCR sees one picture, so wherever stamp ink and print overlap, the shapes merge. A letter with a line through it or a blot beside it no longer looks like any letter, and the software guesses or skips.
Colour and greyscale scans keep the two layers apart. A red RECEIVED stamp over black type stays red, or a lighter grey. The stamp is paler than the print, so the letters still stand out.
Black-and-white scans fuse them. Many office copiers default to a one-bit mode in which every dot is either black or white. A red or blue stamp dark enough to register comes out exactly as black as the type. From that point no tool can tell ink from print, and the overlap is a solid shape. A black-and-white photocopy made before the scan does the same damage, and a later colour scan cannot undo it.
The usual errors fall into three kinds:
- Dropped words. The software finds nothing it recognises under the stamp and leaves a gap, often without any sign that something is missing.
- Wrong characters. A stamp border through a line turns "l" into "t", "c" into "e", or a "3" into an "8".
- Stamp words mixed into the sentence. "The tenant shall FILED pay" is a typical result when a stamp sits mid-paragraph.
The first kind is the dangerous one, because a silent gap reads as if the page said less than it does. If the scanner settings are still yours to choose, the guide to digitizing law firm paper files covers resolution and colour mode.
On a stamped exhibit the dangerous gap often sits under the clerk's mark or an exhibit sticker. You still need the date in the stamp if you will use it to corroborate a timeline, and you read any evidence label the ink crosses. Compare those words with the image before they go into a chronology.
Which stamps cause the most trouble?
The worst stamps are dark, dense and placed over the words you care about. A pale date stamp in a margin costs nothing. A heavy black notary seal across a signature block, or a COPY stamp struck diagonally through a paragraph, can remove whole lines.
| Stamp | Where it usually lands | Risk to the text under it |
|---|---|---|
| Received or filed date stamp in coloured ink | Top corner or margin | Low on a colour scan, moderate in black and white |
| Clerk's file stamp on the first page | Over the caption or title | Moderate, and the caption is text you need |
| Diagonal COPY, DRAFT or CONFIDENTIAL in large outline letters | Across the whole body | High, because it crosses every line |
| Round notary or corporate seal in ink | Over signatures and the notarial wording | High where the ring and lettering are dense |
| Embossed seal with no ink | Over signatures | Low for the text, but the seal itself may not show on the scan at all |
| Exhibit sticker | Bottom corner, sometimes over the last line or a footer | Total for whatever it covers, since a sticker is opaque |
One case is different. Digital stamps are not ink. A watermark or stamp added in PDF software to a file that already has real text often leaves that text intact. Try selecting the words under it before you assume the worst. The post on why you can't copy text from a PDF gives the quick test.
Numbering stamps raise their own questions about sequence and searching. Those are covered in the guide to OCR on Bates numbered documents.
Reading the stamp itself: dates, received marks, exhibit stickers
A stamp is often evidence in its own right. A received date may record when a notice arrived, and a file stamp when a pleading was lodged. OCR can read clean stamp text, but stamps are small, tilted, curved or unevenly inked, so treat every stamp reading as unverified until you have compared it with the page.
What to expect from each kind:
- Mechanical date stamps. Adjustable band stamps print the day, month and year from separate wheels, and one wheel is often lighter or out of line. Expect a wrong digit more often than a missing date. Read the year first: a worn "3" and "8" look alike.
- Stamps with handwriting in them. A RECEIVED box with a pen-written date or initials combines print and handwriting. The printed word may come through and the handwritten part may not. The post on what OCR can do with handwritten notes explains why.
- Round seals. Text that runs around a circle is not on a straight line, and recognition is built around lines. You may get fragments, the words in the wrong order, or nothing.
- Exhibit stickers. A printed sticker with a typed number usually reads well. One filled in by hand often does not. Either way the label will appear somewhere in the page text, not as a separate field.
Where the stamp text ends up matters as much as whether it was read. It can appear at the top of the page text, at the bottom, or in the middle of the paragraph it overlaps. If you plan to search a set for a received date, search a few pages you have already checked by eye first, so you know how that stamp tends to come out.
When the stamp date is the fact you are relying on, record it by hand from the page image. Do not copy it from the converted text.
Entry and exit stamps on passport pages are a common case in immigration files. The guide to immigration document OCR covers identity pages and seals.
What to do when stamped text is unreadable
When a stamp has made the text under it unreadable, work down this list, starting with the cheapest fix. The aim is a better picture of the page if one exists, and an honest record of the gap if one does not.
- Look for an unstamped copy. The same contract may sit elsewhere in the file without the stamp: the version attached to an email, the client's own copy, an earlier production.
- Rescan the original in colour. If you hold the paper, scan the stamped pages again in colour or greyscale. You do not need to redo the whole document, only the pages that failed.
- Ask for a better copy. If the page came from the other side or from a records office, a request for a clean or colour copy of two named pages is small and specific.
- Type the passage yourself. If a person can read through the stamp, a person can transcribe it. Mark the passage as transcribed by hand so the next reader knows.
- Record what is illegible. If nobody can read it, say so in your notes with the page number, and use a bracketed "[illegible]" in any quotation. Do not fill the gap with what the clause probably said.
Whatever happens to the extracted text, keep the image. A searchable PDF holds the original page picture with the text behind it, so the stamp, its date and its position stay visible to anyone who opens the file.
In Schema Law you can try this on your own pages before you commit a whole set. Upload a PDF, PNG or JPEG of up to 20 MB, open the parsed pages and read the stamped areas against your original. The free plan takes one document a day of up to three pages, which is enough for the worst two or three stamped pages in a file. The list of export formats shows what you can download afterwards.
Then check the rest of the page as you would any conversion. Names, dates and amounts near a stamp deserve a second look, and the post on why OCR output is full of errors sets out a method for comparing output with the page.