Read number plates, keep no plate text

anpr-pipeline finds a plate in an image or a video stream, reads it, checks it against the Turkish plate format and stores only a keyed hash of it. One pip install gives you a CLI and an HTTP and WebSocket service.

Install View on GitHub

One read, start to finish

The detector finds the plate, the OCR model reads 34ABC1Z34, and the parser knows that the last block of a Turkish plate is digits, so the Z becomes a 2. The result is checked against the 81 province codes:

TR34 ABC 1234

{
  "raw_text": "34ABC1Z34",
  "normalized_text": "34ABC1234",
  "parsed": {
    "province_code": "34",
    "province_name": "İstanbul",
    "letters": "ABC",
    "digits": "1234",
    "canonical": "34 ABC 1234"
  }
}

On a video stream the same plate is read on many frames. A per-character vote over the best reads settles it, and only then is it stored, as an HMAC-SHA256 hash.

Detection

A YOLOv9-t ONNX model from fast-alpr, trained on plates from more than 65 countries. Runs on a plain CPU.

OCR with confidences

fast-plate-ocr reads the cropped plate and gives a confidence for every character. It can also run on CUDA or OpenVINO.

Turkish plate rules

All 81 province codes and the NN L{1-3} N{2-4} shape. Characters that look alike, such as 0 and O, 1 and I, or 8 and B, are fixed by position.

Tracking and voting

An IoU tracker follows each plate across frames, and a majority vote per character position decides the final reading.

Privacy by default

The database holds a keyed hash, never the plate text, and old rows are deleted after 30 days. See Privacy.

Ready to run

pip install, Docker or uv. REST and WebSocket endpoints, a small web page with webcam support, Prometheus metrics and optional OpenTelemetry.

Start here

InstallFrom PyPI, with Docker, or from source with uv.

Your first plateRead a plate from the command line, then from the web page.

HTTP APISend an image, get the plates back as JSON.

WebSocket streamSend video frames, get confirmed plates as they appear.

Built in the open

anpr-pipeline is free software under the MIT License. It started in 2022 as a YOLOv5, EasyOCR and Flask project and was rewritten in 2026. The old code is still in the repository’s legacy/ folder.

ContributingSet up the project, run the checks and open a pull request.

RoadmapKnown limits, and what comes next.