ria-com/nomeroff-net

What it solves

Nomeroff Net is an automatic number plate recognition (ANPR/ALPR) system designed to detect and read license plates from images. It provides a framework for identifying the location of a plate and converting the image of that plate into text.

How it works

The system uses a multi-stage pipeline consisting of:

• Detection: Uses YOLOv8 bounding box and pose networks to locate license plates.
• OCR: A customized OCR module powered by a Recurrent Neural Network (RNN) architecture to read the characters on the plate.
• Classification: A post-processing classifier that determines the country and type of license plate, achieving up to 99% accuracy for Ukrainian plates.

Who it’s for

This project is for developers building applications that require automatic license plate recognition, as well as researchers interested in license plate datasets and OCR training.

Highlights

• Multi-country support: Provides specialized OCR datasets for Ukraine, EU, Russia, Kazakhstan, Georgia, Belarus, Kyrgyzstan, and Armenia.
• YOLOv8 integration: Leverages modern object detection for high precision.
• ** uma own pipeline**: Offers a simple Python API for detection and reading.
• Resource-efficient options: Includes a "short" pipeline for systems with limited GPU memory.
• Open-source datasets: Provides access to the AUTO.RIA Numberplate datasets for training and evaluation.