Case study
Real-time license plate recognition with deep learning.
Clait developed a proprietary ALPR (Automatic License Plate Recognition) pipeline that locates the plate, reads its characters and identifies its country in under 200 milliseconds on a standard CPU, with 99.2% accuracy on Italian plates.
Published
- accuracy on Italian plates
- 99.2%
- latency on a standard Intel i7 CPU
- < 200 ms
- country identification accuracy
- 92%
- mean detection precision (mAP)
- 90%
01
The context
As motorway infrastructure goes digital, with Free Flow toll gates and traffic monitoring, automatic and intelligent recognition of passing vehicles is increasingly needed.
02
The challenges
Real-world conditions
Variable lighting, reflections, occlusions, motion blur and perspective distortion.
Accuracy versus latency
Maximising accuracy while staying under 200 ms on edge hardware with commodity CPUs.
Geographic diversity
European plates vary widely in format, font and colour: Italy, France, Germany, Austria and more.
Robustness
Running in real time without failing on the hard cases.
Scalability
Handling thousands of vehicles a day.
03
The solution
A proprietary, high-performance pipeline designed for complex real-world scenarios, handling recognition end to end: from locating the plate on the vehicle to extracting its characters and classifying its country.
Detection and pre-processing
Locates the plate instantly, even under perspective distortion, then crops and rectifies it automatically for maximum legibility.
OCR and validation
Reads character by character, with confidence filters that automatically discard uncertain readings.
Country classification
Identifies the country of origin and returns the three most likely candidates, for more reliable decisions.
04
Data and training
The model is trained on hybrid datasets of synthetic and real frontal images, with advanced data augmentation. The result is strong generalisation: accuracy stays high even on blurred, partly occluded or poorly lit images.
05
The results
A scalable system that processes thousands of vehicles a day with latency under 200 milliseconds: 99.2% accuracy on Italian plates, 92% on country identification and a mean detection precision of 90% (mAP).
06
Technology
- Detection
- RT-DETR v2 Transformer (ResNet-18/50 backbone, Deformable Attention)
- OCR
- CCT-S Compact Transformer, hybrid convolution + Transformer (1.25M parameters)
- Country
- MobileNetV3-Large (5.4M parameters, depthwise-separable convolutions)
- Development
- Python, PyTorch, TensorFlow / Keras
- Inference
- ONNX Runtime with a CPU-optimised backend
FAQ
Frequently asked questions.
Does license plate recognition work without a GPU?
Yes. The pipeline runs on ONNX Runtime with a CPU-optimised backend and stays under 200 ms latency on a standard Intel i7 CPU.
How accurate is the system?
99.2% accuracy on Italian plates, 92% on country identification and a 90% mean detection precision (mAP).
Does it recognise foreign plates?
Yes. The system identifies the plate’s country of origin, such as Italy, France, Germany or Austria, and returns the three most likely candidates.
Keep reading.
- Service · 01Custom Machine Learning modelsClait designs proprietary Machine Learning models, built from scratch on a company’s own data, and runs the whole pipeline: from collecting and preparing the data to production deployment and long-term maintenance.From raw data to model · Complex use cases · Compliance & GDPR
- Service · 02AI consulting: strategy, roadmap and pilot projectsClait guides companies through AI adoption, from the first idea to production: it assesses the organisation’s AI maturity, identifies use cases with a realistic ROI, builds a technical and organisational roadmap and follows the first pilot projects through to go-live.Where AI creates value · Process analysis · Adoption roadmap
Contact
We find a solution for everyone.
We'll find yours too.
Tell us what you'd like to improve, even if for now it's only an idea. We'll start from there and build the rest together.