Detect text in any image.

How this works

When you drop an image, the page starts a Web Worker that loads a PP-OCR detection model and recognizer through ONNX Runtime Web. On browsers with WebGPU the models run on your GPU; elsewhere they run as WebAssembly. The worker returns a box, a confidence and the text for each line, and the page draws them over your image.

The detector is a differentiable-binarization segmentation network with a mobile backbone, exported by the PaddleOCR project under the Apache-2.0 licence and packaged for ONNX Runtime by the ppu-paddle-ocr SDK. Details and sizes are on the PP-OCR model page; every project involved is listed on the credits page.

Limits

• Boxes are axis-aligned rectangles, so rotated lines get a loose box.
• The tiny recognizer covers a 6,900-character dictionary (Latin scripts, common CJK and more); rare characters may read as their nearest neighbour.
• Very dense pages (small print, tables) do better with the larger server-grade models in PaddleOCR itself; see the Python guide.
• Images larger than 2,000 px on the long side are downscaled before detection.

Exports

Copy JSON gives you every visible line as { index, text, confidence, box: { x, y, width, height } } in original pixel coordinates, plus the model and timing. Copy text copies the lines in reading order. Save annotated PNG saves your image with the boxes drawn on it.