A practical FastAPI service for image OCR, PDF table extraction, and table recognition from images or PDFs. The current dependency set uses PaddleOCR 3.7.0, PaddleX OCR 3.7.1, and PaddlePaddle 3.2.0. OCR uses the lightweight PP-OCRv6 detection and recognition models.
The Docker image is based on Python 3.12.
- Image OCR from file upload, URL, Base64 data, or a local path.
- Table recognition from an uploaded image/PDF or a public image/PDF URL.
- Table extraction from uploaded or public PDFs.
- JSON, HTML, or XLSX output for the table-recognition endpoints.
- Interactive OpenAPI documentation at
/docs.
Use Python 3.9 or later; Python 3.12 matches the Docker image.
git clone https://github.com/neozhu/PaddleOCRFastAPI.git
cd PaddleOCRFastAPI
python -m venv .venv
# Linux/macOSsource .venv/bin/activate
# Windows PowerShell# .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 8000Open http://localhost:8000/docs to try the API. PaddleOCR models may be downloaded when an endpoint is first used, so the first request can take longer than subsequent requests.
git clone https://github.com/neozhu/PaddleOCRFastAPI.git
cd PaddleOCRFastAPI
docker compose up --build -d
docker compose logs -fThe Compose configuration exposes port 8000 and persists downloaded PaddleX models in the paddleocr_models Docker volume.
All examples assume the service is running at http://localhost:8000.
POST /ocr/predict-by-file accepts jpg, jpeg, png, bmp, and tiff files through the multipart field file.
curl -X POST "http://localhost:8000/ocr/predict-by-file" \
-F "file=@./receipt.png"The result contains recognized text and its bounding boxes:
{
"resultcode": 200,
"message": "receipt.png",
"data": [
{
"input_path": "...",
"rec_texts": ["Example text"],
"rec_boxes": [[12, 20, 180, 54]]
}
]
}GET /ocr/predict-by-url uses the query parameter imageUrl.
curl -G "http://localhost:8000/ocr/predict-by-url" \
--data-urlencode "imageUrl=https://example.com/receipt.png"Other image OCR endpoints are GET /ocr/predict-by-path (image_path) and POST /ocr/predict-by-base64 ({"base64_str":"..."}). Local paths are resolved by the server, so use that endpoint only for files available to the running service.
POST /table/predict-by-file accepts an image or PDF through file. Select json (default), html, or xlsx using format.
curl -X POST "http://localhost:8000/table/predict-by-file?format=json" \
-F "file=@./report.pdf"For format=json, the response includes both the generated table HTML and simplified rows:
{
"resultcode": 200,
"message": "Success",
"data": {
"html": "<table>...</table>",
"rows": [["Header A", "Header B"], ["Value 1", "Value 2"]]
}
}GET /table/predict-by-url provides the same capability for a public url query parameter. It accepts image and PDF URLs.
POST /pdf/predict-by-file accepts a PDF through file and returns only pages where a table is found.
curl -X POST "http://localhost:8000/pdf/predict-by-file" \
-F "file=@./report.pdf"For a public PDF, use GET /pdf/predict-by-url with the pdf_url query parameter.
| Endpoint | Purpose |
|---|---|
GET /ocr/predict-by-path | OCR an image path visible to the server. |
POST /ocr/predict-by-base64 | OCR Base64 image data. |
POST /ocr/predict-by-file | OCR an uploaded image. |
GET /ocr/predict-by-url | OCR a public image URL. |
POST /table/predict-by-file | Recognize the first table in an uploaded image or PDF. |
GET /table/predict-by-url | Recognize the first table in a public image or PDF URL. |
POST /pdf/predict-by-file | Extract tables from an uploaded PDF. |
GET /pdf/predict-by-url | Extract tables from a public PDF URL. |
- Table recognition requires the version-matched
paddlex[ocr]==3.7.1dependency included inrequirements.txt. - Public URL endpoints require the server to be able to download the source file.
- The service is configured for CPU-compatible defaults. No GPU setup is included in this repository.
PaddleOCRFastAPI is released under the MIT License.