How to Copy Text from an Image in Chrome (2026)
You right-click a photo of a document, expecting a "Copy text" option, and get Save image as. Desktop browsers never had that menu item. The letters are on screen, but they are pixels, and nothing between the GPU and your clipboard turns them into a text layer.
Here is the workflow that does: right-click or use the toolbar button, then drag a box over the part of the page you want, get editable text in a small window, copy it. The recognition runs locally. Below, each step and what it actually does.
Why the browser cannot do this
An image element is a bitmap. The DOM knows its width, its height and its alt text; it does not know what the image depicts. A PDF opened in the built-in viewer usually does have a text layer, but a PDF flattened into a single page image, a screenshot in a Slack message, a frame of a video, or text drawn on a <canvas> has no selectable anything. Dragging across it selects nothing because there is nothing to select.
So copying text from an image is two separate jobs: capture a rectangular region of what is on screen, then recognise the characters inside it. The first job is a browser API that extensions can reach; the second needs an OCR engine. That is why this is an extension rather than a setting.
Two ways to start
Any of these arms a region selector over the current tab:
- Right-click menu — the menu entry appears on page, selection, frame, image, link and video contexts. On a
<video>element the entry lets you grab the frame that is currently on screen. - Toolbar popup — the button in the popup starts the same selector, and the popup is also where you set the recognition language and see the history.
Both end up in the same place, so pick whichever is closest to your hands. The popup keeps your language choice between sessions; a Pro language selected there is remembered.
Drawing the box
The page freezes its scroll while the overlay is up — that matters, because a region that scrolls away during a drag is a region that captures the wrong content. Drag to define the rectangle; the box is drawn in the page's own coordinates, and pressing Esc dismisses everything and restores scrolling. A drag smaller than about 8 pixels on either side is treated as a misclick and dropped rather than sent for recognition.
Two practical notes on framing the box:
- Tight beats generous. Recognition reads what is inside the rectangle. Blank margins, a decorative border or the neighbouring column all add noise; a box around just the paragraph you want is faster and more accurate.
- Zoom first if the text is small. The capture is a screenshot, so the on-screen size is the input size. Two clicks of
Ctrl++on a dense table is worth more than any engine setting.
What happens to those pixels
This is the part people worry about, so here is the whole path:
- The visible tab is captured as a PNG. Only what is on screen — the extension never reads the page's files or posts anything anywhere.
- The rectangle is cropped to physical pixels with a per-axis scale measured from the screenshot itself — captured bitmap width ÷ viewport width, and the same for height. That ratio absorbs both HiDPI and page zoom;
devicePixelRatioalone would be wrong at 125% zoom on a 2× screen, and it is used only as a fallback when the viewport is missing. On a 2× HiDPI screen, the crop keeps the extra pixels instead of throwing them away, which is why the same box reads better on a Retina display than on a 1080p one. - The cropped image goes to a popup window sized 580×660, which runs the OCR engine itself.
The engine is WebAssembly, loaded from inside the extension, and the two bundled language packs ship with it. Nothing that is bundled needs a network round trip, which is why the manifest declares the extension usable offline and why the only remote address it can reach at all is the licence endpoint. The crop never leaves the machine.
Reading and copying the result
The text lands in an editable box, not a static paragraph — you can fix one misread character instead of re-running the capture. Two buttons sit under it:
- Copy writes to the clipboard. If the browser refuses the modern clipboard API, the copy falls back to selecting the text and issuing a legacy copy command, so the button keeps working.
- Download as .txt saves the recognition as a UTF-8 file named with the timestamp. This one is part of the licence, and the button nudges rather than downloading when there is no entitlement.
Every successful recognition is also appended to a local history list, newest first, so the quote you pulled off a screenshot an hour ago is still there. Free keeps the last 10 entries; the licence keeps 200.
When OCR is the wrong tool
Recognition is a recovery step, not a habit. Before reaching for it, check whether the text already exists somewhere it can be taken for free:
| Situation | Better first move |
|---|---|
| The page shows real, selectable text | Select and copy. Exact, instant, free. |
| A PDF with a text layer | Copy from the browser's PDF viewer, or save the file. |
| A scanned PDF rendered as images | OCR the region — this is the case it is for. |
| Text baked into a photo or a meme | OCR. |
| A frame of a playing video | Pause, then right-click the video and run the selector. |
| Text drawn on a canvas or WebGL surface | OCR — the DOM holds no text at all. |
Accuracy follows the input. Clean, dark, uniformly lit text at a decent size comes out near-perfect. Long digit strings — invoice numbers, IBANs, tracking codes — are where you should still compare against the image, because similar glyphs are exactly what engines trade on.
Limits worth knowing
- 20 recognitions a day without paying. The count resets at local midnight, and a run is only billed once the engine hands its result back — a capture that fails or a page you aborted with
Escdoes not cost you. - English and Simplified Chinese are bundled and free. The other seven language packs (Traditional Chinese, Japanese, Korean, Russian, French, German, Spanish) are part of the licence and download on first use.
- Ordinary pages only.
chrome://pages, the extensions gallery and the Web Store reject content-script injection, so the menu entry and the toolbar button do nothing there. - The visible tab only. The capture is a screenshot, so text below the fold must be scrolled into view before you box it.
- One region at a time. A two-column page needs two drags; the box is a rectangle, not a smart layout reader.
The licence is a one-time purchase, not a subscription, and it also removes the daily cap, extends history to 200 entries and enables the .txt export.
A short checklist
Summed up in the order you will actually use it:
- Is the text selectable? If yes, select and copy — do not OCR.
- If not, scroll the region fully into view and zoom until the glyphs are comfortable.
- Right-click the page and take the menu entry (or use the toolbar button), then drag tight around the text.
- Pick the language in the popup before you start if the text is not English or Simplified Chinese.
- Fix the box in the result window if one character is wrong, then copy.
- For anything with digits, glance at the image before you paste it somewhere it matters.
FAQ
Can Chrome copy text out of an image by itself?
Not on desktop. Chrome has no context-menu option that runs OCR on an arbitrary image on a page. The OS-level text tools on Android and Windows 11 work on your whole screen, not on a selected web region, and nothing in the browser turns a JPEG into a text layer.
Does the image get uploaded to a server?
No. Recognition runs in the extension window with a WebAssembly OCR engine, and the two bundled language packs ship inside the extension. The only network address the extension is allowed to reach is its own licence endpoint, so the capture stays on your machine.
Why does the extension not open on some pages?
It can only inject into http, https and file pages. Browser pages such as chrome://settings, the extensions gallery and the Chrome Web Store are off limits to every extension, so the menu entry and the toolbar button silently do nothing there.
What is the difference between selecting text and recognising it?
If the page already contains real text, plain selection is faster and always exact. OCR is for the cases where the letters are pixels: a photograph of a sign, a scanned PDF rendered as an image, a frame of a video, text drawn on a canvas. Recognition is a best-effort read of those pixels, so long strings of digits and unusual fonts are the ones to double-check.
More guides in the VKT blog; questions to support@annmax1983.com.
