Get text out of scans and image-only PDFs. PP-OCRv5 runs page by page in your browser — Chinese, English, mixed lines and table numbers — and nothing is uploaded.
Drop a PDF here, or click to choose
One file, up to 50 MB. Recognition runs entirely on your device.
Every page of a scan is a single image; the file contains no characters at all. That is not a failed extraction, it is a PDF that never had text in it. Getting words out means looking at pixels: a detection model finds where the text is, then a recognition model reads each box. That is what this page does.
It runs the two PP-OCRv5 mobile ONNX models — 5 MB for detection, 16 MB for recognition — plus a 12 MB inference runtime, downloaded on first use and cached afterwards. Recognition itself is offline, and the file is never uploaded: there is no endpoint that accepts a PDF.
The detector eats a whole page and returns a quarter-resolution probability map: bright means "text lives here". Threshold it, take connected components, grow each box back out by the DB unclip distance, and you have text boxes. The recogniser then takes each box, scales the line to 48 px tall, emits one distribution per time step, and CTC collapses that into a string.
Glyph height matters most. The recogniser was trained on lines 48 px tall; smaller print has to be upscaled, upscaling blurs, and blur becomes mistakes. That is why the page exposes a resolution switch: at 200 dpi a 10 pt line is ~28 px, at 300 dpi it is ~42 px — much closer to what the model expects.
Then skew and noise. A slight scanner tilt is tolerated; past a few degrees characters start dropping. Fourth-generation photocopies and shadows near the binding break the detector into fragments. Those are model limits, not knobs to turn.
Invoices, contracts, medical records, ID cards: the documents that most need OCR are exactly the ones that should not be uploaded. Running locally means the file never leaves the machine, so "does the server keep a copy" stops being a question. No quotas, no queue, and it works offline.
The output is plain text in reading order: tables collapse into a run of text with the columns gone, two-column layouts may interleave, and running headers land in the body. If you need the layout, convert to Word — OCR is not the tool for that.
Blank pages return nothing, and those page numbers are listed for you — same for pages whose print is too faint, too small, or buried under a watermark. Re-run just those pages at 300 dpi.