PaddleOCR online — runs in your browser

How this online PaddleOCR works#

This page loads PaddleOCR.js, the official browser SDK published by the PaddleOCR team as @paddleocr/paddleocr-js. It runs the same two-stage pipeline as the Python library:¹

Both models run through ONNX Runtime Web. The SDK picks WebGPU when your browser exposes it and falls back to WebAssembly otherwise. The model files come from Baidu's official CDN; if that fails, this page fetches the identical ONNX export from the PaddlePaddle organisation on Hugging Face. Either way they are stored in your browser's cache, so the second visit starts in seconds.³

Which model should you pick?#

The server-grade PP-OCRv6 medium model (the default in Python) isn't offered by the browser SDK. In our CPU benchmark small matched medium exactly on three of our five test documents and made one extra character error on each receipt, so it is a fair preview of what the Python library gives you.

Tips for better results#

• Crop to the text. Large photos are downscaled before detection, so small print on a big photo can get lost. Cropping is the single biggest win.
• Text missing? Open Detection settings and lower text_det_box_thresh (for example to 0.4). Faint or low-contrast text often clears the threshold then.
• Letters cut off at the edges? Raise text_det_unclip_ratio to 2.0 so boxes are drawn a little wider.
• Rotated pages work up to a moderate tilt (our receipt photo is turned 7.5° and read perfectly). For pages upside down or sideways, the Python pipeline's use_doc_orientation_classify=True fixes orientation first.
• Use the confidence slider to hide shaky lines. Red boxes are below 0.80, which is where most misreads sit.

Browser vs Python: what's different?#

Put the same OCR on your own website#

Everything this page does is available to you through the official SDK. The minimal version:

import { PaddleOCR } from "@paddleocr/paddleocr-js";

const ocr = await PaddleOCR.create({
  lang: "en",
  ocrVersion: "PP-OCRv6",          // maps to PP-OCRv6_small
  ortOptions: { backend: "auto" }  // WebGPU if available, else WASM
});

const [result] = await ocr.predict(fileInput.files[0]);
for (const line of result.items) {
  console.log(line.text, line.score.toFixed(3), line.poly);
}

Our PaddleOCR.js guide covers bundlers, Web Workers, WebGPU headers and the CDN behaviour we had to work around to make this page reliable.

Questions#