How to OCR a scanned PDF without uploading it

A scanned PDF is a stack of pictures. You can read it, but you cannot search it, select a sentence or copy a number, and screen readers see nothing at all. Optical character recognition (OCR) fixes that by recognising the text in each image and adding it back to the page as an invisible layer. Most online OCR services do this on their servers, which means your scan is uploaded, stored for some time and processed by someone else's machine. This guide shows how to do it entirely inside your browser with OCR PDF, and what to expect from the result.

How recognition runs on your device

The tool ships its own recognition engine as web code:

• The page images are produced by PDFium, the same PDF engine used in Chrome, compiled to WebAssembly and running in a background worker.
• Text detection and recognition use PP-OCRv5 mobile models, run by ONNX Runtime in the browser. On machines with WebGPU the model runs on the graphics card; otherwise it uses multi-threaded WebAssembly, and if that is unavailable, single-threaded WebAssembly. You do not have to choose; the tool picks the fastest path it can.
• The model files (about 26 MB in total for detection, recognition and the English fallback) are downloaded once from this site, checked against their published checksums and cached in your browser. Later runs start immediately and work offline.
• If the main engine cannot start in your browser, an English Tesseract model takes over. It also runs locally.

At no point are the page images or the recognised text sent anywhere. Your browser only downloads the tool and the model files.

Step by step

1. Open OCR PDF and choose the scanned file. The Recognize text (OCR) dialog opens as soon as the document is ready.
2. Pick a quality. Fast (150 DPI) is enough when you mainly want to search the document. Accurate (300 DPI) renders each page at a higher resolution and takes longer, which helps with small print and poor scans.
3. Choose which pages to process: all, the current page, the selected pages or a range. Pages that already contain text are skipped unless you tick Re-run on pages that already have text.
4. Press Start OCR. On the first run you will see the model download and checksum verification, then a per-page progress indicator. You can cancel at any time and keep the pages that are done.
5. When it finishes, try Search for a word you can see on the scan. Select some text and copy it. Then press Export to download the searchable PDF.

What the output looks like

The tool writes real PDF text objects in invisible rendering mode, positioned on the recognised lines, using an embedded font with a Unicode mapping so copied text comes out as proper characters, including Chinese and Japanese. The scanned image is untouched; the page looks exactly as it did. Any reader that can search PDFs will find the text, and the file works with screen readers.

Recognition is line-based. That is enough for search and copy, but the invisible text follows lines rather than individual words, so double-clicking a single word may select slightly more than you expect.

Getting better results

• Straighten the pages first. If the scan is rotated by 90 degrees, use Rotate PDF before running OCR. Small skew is tolerated; sideways pages are not.
• Use Accurate mode for small print. 300 DPI gives the model more pixels per character.
• Language coverage. The bundled models read Latin-script languages, Chinese and Japanese, with automatic detection. Arabic, Cyrillic and other scripts need models that are not bundled yet.
• Handwriting recognises poorly. The models are trained on printed text.
• Check the text you rely on. OCR makes mistakes on faint characters and unusual fonts. For numbers that matter, compare against the image.

Working offline

After the tool and the model files have been downloaded once, everything needed is cached by the browser. You can open the site without a connection, choose a scan and run recognition. If you want to reclaim the space, the OCR panel shows how much is cached and offers Delete downloaded models.

Privacy, in concrete terms

The tool has no analytics and no telemetry. Its network requests are the tool's own code, the fonts, and the model files, all served from this site. The optional cloud conversion feature elsewhere in the editor is off by default and is not involved in OCR. A working copy of your document is kept in the browser's private storage so an interrupted session can be recovered; you can delete it in Settings.

Related

• Redact sensitive text after making a scan searchable: Redact PDF.
• Combine several scanned files first: Merge PDF.
• Add notes or a signature to the scanned document: Annotate PDF.