Extract text from an image without uploading it

OCR PDF takes PDF files only, and its recognizer works on pages: it renders a page, reads it, and writes the words back over that page as invisible text. An image is not a page yet, so you give it one. JPG to PDF wraps each picture in a page, which is all the recognizer needs. Its picker allows JPEG, PNG and WebP, judging each file by its type and file extension, so anything else is turned away with a message like "receipt.heic" is not a supported file type for this tool. JPEG and PNG bytes go in unchanged; a WebP is re-encoded as JPEG through a canvas first.

The PDF in the middle is disposable: feed it to OCR PDF, copy the text, then delete it. Keep the second step's download only if you want the picture with a searchable text layer. If the picture is not a file yet, Scan to PDF wraps camera captures the same way.

In JPG to PDF, Page size offers Match each image, selected by default, and Fixed page size. Keep Match each image when recognition is the point: it makes the page exactly as many points wide and tall as the picture is pixels, so the whole picture is there to render.

Fixed page size centers the picture on the Paper you choose (A4, Letter, Legal, A3 or A5) inside a Margin (pt) that starts at 24, scaling it down to fit, or up if it is smaller than the printable area. That matters because OCR PDF renders each page at 200 dpi and no wider than 2,400 pixels. A 3,000-pixel-wide photo reaches the recognizer at 2,400 pixels across under Match each image, but at roughly 1,500 pixels after being fitted onto A4 portrait with the default margin. A very tall picture, such as a long screenshot, is rendered smaller still, because the page render is also capped at 16 million pixels.

So for anything larger than the printable area, which includes every phone photo, Match each image leaves the recognizer more detail. For a small picture that fits inside the margins untouched, both modes resample the same pixels and the choice barely matters.

Open OCR PDF, click Select file, and add the PDF you just made. Leave Language of the text on English (in your browser), the only entry in the list on the deployed site. Leave Skip pages that already have text ticked: a page made from a photo has no text layer, so nothing gets skipped. Click Make searchable.

The first run downloads about 5 MB of recognition engine and model, which the browser caches. Each page is rendered, then read with automatic page segmentation, which finds headings, columns and captions separately instead of treating the page as one block.

The note in the result says how many pages were recognized and that nothing was uploaded. Below the buttons, a panel labeled Recognized text holds the plain text with a Copy button, which is what you came for. The Download button is named after your file with -searchable.pdf on the end: the same picture, with the invisible text over it.

Recognition follows the picture. A sharp, straight, evenly lit photo of printed text reads well; an angled, dim or blurry one, small text, handwriting and decorative fonts produce errors and gaps. Cropping to just the block of text, or retaking the shot closer and squarer, helps more than any setting in either tool.

A photo your phone stored sideways is a common cause of near-empty output. JPEG bytes are copied into the PDF as they are, and the rotation tag your phone applies when displaying the picture is not applied here, so the recognizer can be looking at sideways text. Turn it upright before step one, or run the PDF through Rotate PDF and recognize it again.

Check numbers and names against the picture. The panel holds everything the engine read, which is why it is the copy to work from; the invisible layer holds less, because words the engine scored below 30 out of 100, and characters its built-in font cannot encode, are left out. A word you can read in the panel is not always findable with Ctrl+F in the download.