Convert PDF Images to Text: Words Locked Inside Pictures

To convert PDF images to text, first find which pages hold pictures with words in them, then run OCR on those pages alone and leave the typed pages untouched. The typed text is already exact. Only the words inside the images need recognizing, and the recognized text needs checking against the picture.

The file in question is not a scan. It was made from a word processor or an email program, you can select its sentences, and a search finds its headings. Yet a search for a name you can plainly see on page 9 returns nothing, because page 9 shows that name inside a pasted screenshot.

Why do some words in a text PDF refuse to be selected?

Those words are part of a picture. A PDF page can hold typed characters and embedded images side by side. Characters can be selected, searched and copied. An image is a block of pixels, and any words in it are shapes with no characters behind them, exactly as on a scanned page.

A PDF made from a word processor stores three broad kinds of content on a page:

When someone pastes a picture into a document and saves it as a PDF, the picture goes into the file as an image. The software that made the PDF does not read it. So a typed motion can contain, in the middle of page 9, a rectangle that behaves like a small scan.

How this differs from a scanned PDF

In a scanned PDF every page is one image from edge to edge, and nothing on any page selects. Here the gaps are local: one figure on a page, or a few pages in a long file. That changes the job. You are not converting the document. You are filling holes in a document that is otherwise sound. Where whole pages of the file are image-only, such as scanned exhibits attached behind a typed motion, the post on unreadable PDFs covers PDFs with some readable pages and some image-only pages.

Why it matters

Three tasks fail quietly on these files. A search misses every name and figure inside a picture, and reports no results as if the term were absent. A copy of the page leaves the picture's words behind. And any program that reads the file's text, such as a review platform or a document comparison, sees a document with pieces missing and gives no warning.

Which legal PDFs carry text inside images?

Look for text inside images wherever a document reproduces another document instead of quoting it: briefs with exhibits pasted into the body, emails with screenshots, contracts with an inserted signature page, reports with charts, and forms that were filled in and flattened. Any PDF assembled from more than one source deserves a check.

Specific places to look:

• Briefs and letters with pasted excerpts. A drafter who wants to show the court the exact look of a contract clause, a text message or a social media post pastes a picture of it into the argument. The words of the excerpt are then in the brief visually and absent from its text.
• Emails saved as PDF. Inline screenshots, photographed receipts, and signature blocks or disclaimers that were set up as images. Check a forwarded chain where earlier messages appear as pictures.
• Agreements with an inserted signature page. The body came from the word processor; the signed page was scanned and swapped in. Names, titles and dates written or typed on that page are in the image.
• Expert and financial reports. Charts carry axis labels, legends and data labels. Check whether a table was pasted from a spreadsheet as a picture, which puts every figure in it out of reach.
• Presentations saved as PDF. Slides lean on images, and a slide's key wording may be part of a graphic.
• Letterhead and footers. A firm name, address or regulatory statement set as a logo image. Worth knowing when a search for a firm's name finds fewer documents than you expect.
• Stamps added as images. A "received" stamp, an exhibit sticker or a page label applied by software as a picture.

None of these is a rule. How a given author builds documents is something you learn from the file in front of you, by testing it.

A medical bill pasted into a filing as a picture hides the provider, the date of service and the charge. Search misses those words until you recognize the image. Convert that page, check the patient name and any insurer line against the picture, and leave the typed pages around it alone.

Finding the image regions in a PDF that search cannot see

Find image regions by selecting everything on a page and looking for what stays unhighlighted, by searching for a word you can see inside a suspect area, and by copying the page into a text editor to see what is missing. Record the page numbers as you go, because the conversion step works from that list.

Test one: select all on the page

Click in the page and press Ctrl+A, or Command+A on a Mac. Text is highlighted. An image with words in it stays its normal colour, an island in the highlighting. This is the quickest way to see the shape and extent of a gap. In a long file, do it on any page that contains a figure, an excerpt or anything that looks pasted.

Test two: search for a visible word

Pick a distinctive word inside the suspect area, one that does not also appear in the typed text, and search for it. No result on a page where you can read the word confirms the gap. This test also catches the opposite case: an image that already has a hidden text layer over it, in which the word is found.

Test three: copy the page out

Select the whole page, copy, and paste into a plain text editor. Compare the pasted text with the page. Whatever you can read on the page but cannot find in the paste lives in an image. This test is slower, and it is the one to use when the result matters, because it shows exactly what a text-reading program will receive.

When the tests disagree or nothing selects

Zooming in settles a doubtful region, since typed text stays sharp and words in an image turn soft. That check, and the others for a page where nothing selects at all, are set out under how to tell whether a PDF has real text.

Keep a gap list

For a fictional example, take a 38-page summary judgment brief filed for Redfern Logistics against Abel Tran. A paralegal, Mirela, runs the tests and writes down:

Four pages out of 38. That list is the whole job.

Converting the images in a PDF to text while the typed pages stay as they are

Work from a list of the pages that hold images. Extract only those pages into a separate file, run OCR on that file, check the recognized text against the pictures, and keep the result with the original as a companion. The original PDF, and the exact text of its typed pages, is never altered.

1. Leave the original alone. Save it under its received name. If it is a filed or served document, it stays byte for byte as it arrived.
2. Extract the listed pages. With your PDF program, pull them into a new file named for what it is. If the list for a brief reads pages 9, 14, 22 and 38: Redfern MSJ brief - image pages 9, 14, 22, 38 (extract for OCR).pdf.
3. Run OCR on the extract. Upload it to your OCR tool. Schema Law accepts a PDF, PNG or JPEG of up to 20 MB and reads each page into text and tables, which you can view page by page before exporting. The how it works section shows the steps.
4. Export the form you need. Plain text is enough for search and quotation. Choose Word or Excel if a picture contained a table you want to work with.
5. Check each recognized passage against its picture. Names, dates, times, figures. A screenshot's small grey timestamps deserve particular attention.
6. Label the output by original page. The extract's page 1 is the brief's page 9. Write the real page numbers into the text file as headings, such as "From page 9 of the brief (screenshot)".
7. File the companion beside the original. Redfern MSJ brief - text of images (OCR, checked MD 2026-10-07).txt.

What happens to the typed words on those pages

Suppose page 9 of the brief holds typed argument as well as a screenshot. When the extract goes through OCR, expect the tool to read the whole page as a picture, typed paragraphs included. You then have two versions of those paragraphs: the exact text in the original PDF and a recognized reading in the companion file. Use the original for the typed words and the companion only for what was inside the images. Deleting the duplicated paragraphs from the companion keeps that boundary clear.

Converting a single image instead of a page

If only one figure matters, you can take a screenshot of that figure at high zoom, save it as PNG and run OCR on the image alone. The result contains nothing but the picture's words. Enlarge the figure on screen before capturing, since the capture records only the pixels displayed. Check the rest of the document first, though: an appendix or exhibit may hold the same excerpt in a clearer form than the pasted picture, and that is the better source.

If you need one searchable file

If the requirement is a single PDF in which a search finds everything, for your own working set, one way is to run OCR on the extracted pages, export them as a searchable PDF, and put them back in place of the originals in a working copy, clearly named as such. Whether to do this to the whole file instead is a separate question with its own risk.

The same problem in other file types

Word files have the equivalent trap: pages or figures pasted in as pictures inside a DOCX. The steps differ because the container differs, and the post on how to OCR a Word document covers them.

Should OCR run over PDF pages that already have real text?

As a rule, no. Real text records the characters the author typed, with certainty. OCR produces a reading of how those characters look, which can be wrong. Running OCR across typed pages risks replacing exact text with a good guess, in exchange for convenience. Limit recognition to pages that need it.

What can go wrong

What a tool does with a page that already contains text varies, and you need to know which behaviour yours has. The possibilities:

Look in your tool's documentation for how it treats pages with existing text, then confirm with a test. Take one typed page that has an embedded image, process it, and compare the before and after: search for a word in the typed paragraph and a word in the image, and copy a line containing numbers from each.

Where exactness matters

A typed "$1,250,000.00" in a native PDF is that figure. The same figure after recognition is that figure only if every digit, comma and point was read correctly. For a document you will quote, compare or load into a review platform, that difference is the reason to keep recognition away from good text. The post on e-discovery OCR explains how the two kinds are kept apart, under extracted text versus OCR text.

When a full pass is acceptable

A full pass is reasonable when the output is a working aid and not a record: you want the entire brief, pictures included, as one Word or text file to read and annotate, and you will go back to the original PDF for anything you quote. Name the result so that nobody mistakes it for the document's own text, for example Redfern MSJ brief (full OCR, working copy, not for quotation).docx.

Never overwrite

Whatever you run, run it on a copy. A filed brief, a served production document or a signed agreement must stay exactly as received, and court rules and production agreements can govern how such files are handled. Check the ones that apply to your matter.

What the recognized text of a chart, screenshot or diagram looks like

Recognized text from a figure is a list of the words that were in it, without the picture that gave them meaning. A chart yields labels and numbers with no link between them. A screenshot yields its content plus the interface around it. A diagram yields box labels and loses the arrows.

Charts

Take a bar chart on page 22 of a fictional brief filed for Redfern Logistics. The picture shows six months along the bottom, a count above each bar and a title. The recognized text might read:

Late deliveries by month, 2025
Jan Feb Mar Apr May Jun
14 9 22 31 27 12
0 10 20 30 40
Source: Redfern dispatch log

Every item is there, and you can now find the chart by searching for "Late deliveries". What is missing is the pairing. Nothing in the text says that 22 belongs to March. You know it only because the labels and values happen to be in the same order, and that will not hold for a chart with a legend, stacked bars or values printed out of sequence. Axis numbers appear too, mixed in with data. Vertical labels, such as an axis title printed sideways, may be missed or garbled.

So treat a chart's recognized text as an index entry, not as data. If you need the numbers, look for the table behind the chart in the report or its appendices, or ask for the underlying spreadsheet.

Screenshots

A screenshot of a message thread or web page gives clean, sharp text, and gives you everything in the frame: the clock, the battery level, a contact name, button labels, "Delivered" markers. In a conversation, the left and right alignment that showed who sent each message is gone, so attribution has to be restored by hand from the picture. Add speaker names to the text yourself, in brackets, and say that you added them. The broader technique for these files is in the post on getting text out of a screenshot of a document.

Diagrams and organization charts

An organization chart becomes a list of names and titles. Who reports to whom was carried by lines, and lines are not text. A flow diagram becomes its step labels in whatever order the software read them, which may be left to right when the flow ran top to bottom. A site plan or floor plan yields room names and dimensions scattered without position.

For these, the recognized text serves two purposes: it lets a search find the page, and it saves retyping the labels. The relationships have to be described in words by a person looking at the figure, if they need to be in text at all.

Pasted document excerpts

These are the plainest case, because they are ordinary text that happens to be a picture. Expect the same errors as any scan, and expect them to concentrate where the excerpt was small or was captured at low resolution. A clause pasted at half size into a brief has fewer pixels per letter than the same clause in the agreement itself. If the source document is in the record, take the wording from there and use the pasted picture only to confirm which passage was shown.

What to write down

For each figure you convert, a one-line note in the companion file is enough: the page, what kind of figure it is, and what the text does and does not capture. "Page 22, bar chart. Labels and values recognized; pairing of months to values read from the image by MD." A reader months later then knows which parts came from software and which from a person.

To try this on a short file, Schema Law's free plan takes one document a day of up to three pages and includes every export format, plain text, Word and searchable PDF among them. The pricing page lists the limits that apply to longer extracts.