License plate recognition with hot-swappable models

What it does

FastALPR is a Python framework that finds license plates in images and reads the characters. By default, it pairs a YOLOv9 plate detector from the author’s open-image-models repo with an OCR model from fast-plate-ocr, both running under ONNX Runtime. You hand it an image, and it returns bounding boxes plus the plate text.

The interesting bit

The whole pipeline is modular. You can subclass BaseOCR to drop in Tesseract, or point the detector at your own ONNX model. It is essentially opinionated glue: it handles the boring pre- and post-processing, but the actual inference brains are replaceable.

Key highlights

• Supports multiple ONNX Runtime backends—CPU, CUDA, OpenVINO, DirectML, and Qualcomm QNN.
• Ships with ready-to-use detection and OCR models, but neither is hard-coded.
• Includes a draw_predictions helper to annotate frames with OpenCV.
• Live demo hosted on Hugging Face Spaces.

Caveats

• No ONNX runtime is included by default; you must install a backend (CPU, GPU, etc.) separately before it will run.

Verdict

Worth a look if you need an ALPR pipeline you can tune or deploy on edge hardware without fighting a monolithic model. Skip it if you want a single-command, batteries-included executable with no Python dependencies.