ParkVision 05 / 25
Plate OCR for 7 countries (CRNN + CTC)
Reads the alphanumeric plate text, including Kazakhstan plates, with a recurrent recognizer.
Problem
Plates in the region mix Cyrillic and Latin look-alike characters and differ per country.
What it does
- Recognizes plates from Kazakhstan, Russia, China, Uzbekistan, Georgia, Kyrgyzstan and Mongolia
- Correctly distinguishes similar Cyrillic and Latin characters
- Fast plate-only mode via the API
How it works
- Rectified plate → normalization and scaling
- CRNN produces per-timestep logits
- CTC greedy decoding with a country alphabet → plate string
Stack & models
PythonPyTorchCRNNCTCONNX Runtime
Numbers
- ~40–60 ms per frame for the full pipeline on GPU
Figures come from project docs/README; no other benchmarks are claimed.