- Size and proportions. A European plate is not the same size as an American one, and both differ from motorcycle plates, which tend to be smaller and have a different aspect ratio.
- Color and contrast. White, yellow, green, blue or even black backgrounds; light characters on a dark background or vice versa. Color is not decorative: in many countries it indicates the type of vehicle (private, commercial, official, taxi, military).
- Typography and special characters. Specific fonts, embossed characters, stamps, crests or regional emblems that can be mistaken for alphanumeric characters if the engine isn’t trained to tell them apart.
- Number of characters and structure. From short four-digit combinations to long sequences mixing letters, numbers and separators, each with its own country-specific syntax.
- Single-row or two-row plates. Some formats stack information across several rows instead of a single horizontal line, which requires contextual reading rather than simple character segmentation.
- Diplomatic and consular plates. These typically have distinct colors, abbreviations and structures (for example, a green background in some countries, or specific letter combinations such as CD or CC in others) that identify the vehicle’s status.
- Temporary or provisional plates. Some countries issue these on paper or in simplified formats while the permanent plate is being processed, which reduces print and read quality.
- Commercial, official and fleet vehicles. These add prefixes, suffixes or additional markings that also need to be validated within the corresponding national format.
All of this means an OCR engine can’t simply “read characters”: it also needs to know which country a plate belongs to, which format applies, and whether what it read is consistent with that country’s rules.
What is international license plate recognition?
Before getting into how it works, it’s worth clarifying the terminology, since the industry uses several terms almost interchangeably:
- ANPR (Automatic Number Plate Recognition): the term most commonly used in Europe.
- ALPR (Automatic License Plate Recognition): the equivalent term used mainly in the Americas.
- LPR (License Plate Recognition): a shortened form, common in technical documentation and regulations.
- OCR (Optical Character Recognition): the underlying technology that converts the image of the plate into text a computer system can read.
In practice, a complete ANPR/ALPR system combines cameras, lighting, an engine that detects the plate’s position within the image, and an OCR engine that interprets the characters. “International recognition” adds an extra layer: the system has to identify which country the plate belongs to and apply that format’s specific rules before returning a valid result.
The role of deep learning
AClassic license plate recognition worked by segmenting each character in isolation and comparing it against predefined templates (template matching). This worked reasonably well with clean, well-lit plates in a known format, but degraded quickly with difficult angles, dirt, glare or unfamiliar formats.
Deep learning changed the approach. Instead of reading character by character in isolation, convolutional neural networks interpret the plate as a whole image, preserving the spatial relationship between characters and learning contextual patterns: which letter-number combinations are plausible in a given country, how the typography behaves under different lighting conditions, or how to tell a zero apart from an “O” depending on the national format. This approach doesn’t just improve accuracy — it also lets the system generalize and adapt to new formats with far less manual programming effort than classic methods required.
How SIRAM recognizes a foreign license plate
The process SIRAM OCR follows to recognize a plate, regardless of its country of origin, can be summarized in seven stages:
Camera > Pre-processing > AI Detection > OCR > Filter per country > Validation > Result
- 1. Capture (camera). The ALPR camera captures the vehicle image, typically with IR lighting to ensure reliable reading day and night.
- 2. Image preprocessing. Angle, contrast and noise are corrected to maximize the quality of the region where the plate is located.
- 3. AI detection. A deep learning model locates the plate within the scene, regardless of its size, shape or position on the vehicle.
- 4. OCR. The optical recognition engine interprets the detected characters and produces a raw reading.
- 5. Country filter. This is where SIRAM’s own configuration parameter comes in: continuous.pull.filter.countries, which defines which countries the engine should accept in continuous recognition mode. If the list is left empty, the filter is disabled and the system evaluates every available format.
- 6. National format validation. Once a candidate country is identified, SIRAM checks whether the reading matches the corresponding national format using a complementary parameter, continuous.pull.filter.countries.maxerrors, which defines the maximum number of character errors tolerated for a plate to be considered a match for a given national format. This lets the system keep recognizing plates that are slightly dirty, damaged or poorly lit without losing reliability.
- 7. Result. The system returns the recognized plate together with the country (or candidate countries) and the confidence level associated with the reading.
This combination of country filtering and format validation is what allows SIRAM to operate with high accuracy in environments where vehicles from dozens of different nationalities coexist, such as airport parking facilities, ports, border areas or international toll corridors.
AMÉRICA: 22 | EUROPA: 49 | ÁFRICA: 15 | ASIA: 18 | OCEANÍA: 3
How AI improves recognition
La evolución del reconocimiento de matrículas puede resumirse en un cambio de paradigma:
The evolution of license plate recognition can be summed up as a paradigm shift:
Before — Classic OCR: character segmentation → comparison against fixed templates → high sensitivity to dirt, angle and lighting → each new country format has to be programmed manually.
Now — Deep learning: contextual reading of the whole plate → robust recognition under adverse conditions → continuous learning from new data → new countries added without redesigning the engine from scratch → higher overall accuracy, in many cases above 99% under favorable conditions.
Comparative studies in the field confirm this difference: engines based on convolutional neural networks achieve significantly higher accuracy rates than classic template-based OCR methods, particularly under variable lighting or with partially damaged plates.
How does the engine learn a new country?
SIRAM OCR’s country coverage is not a closed list. The engine can incorporate new national formats by training on real examples of plates from that country. This means that whenever a new plate design appears or a country requests support, the engine can learn that format and add it to its coverage without needing to redesign the whole system. That’s why the number of supported countries has grown steadily across successive versions of OCR, now exceeding 100 countries across four continents — we cover the full list in the second article of this series.
Frequently asked questions
International license plate recognition is arguably the most demanding technical challenge within ANPR/ALPR: there are no global standards, and every country brings its own combination of size, color, typography, character count and single- or two-row layout. SIRAM OCR solves this complexity by combining deep learning-based detection, a contextual OCR engine, a configurable country filter, and format validation with error tolerance.
Classic OCR vs. deep learning OCR (SIRAM)
| Classic OCR | Deep learning OCR (SIRAM) | |
| Method | Character segmentation + fixed templates | Contextual reading of the whole plate |
| Sensitivity to adverse conditions | High | Low |
| Adding new countries | Manual programming | Training the engine on new data |
| Handling non-Latin alphabets | Limited | Supported (Cyrillic, Greek, Arabic, depending on country) |
| Typical accuracy | Variable, lower under difficult conditions | Up to 99% under favorable conditions |
Nowadays we live in a globalized world where parking operators demand a license plate recognition technologies (ALPR) with several requirements such as a high reliability rate, easy installation and maintenance,but also a wide flexibility in order to support different countries and offer a quick adaptation to any type of integration.
Innova Systems Group´s technology for license plate recognition (SIRAM OCR) allows to respond in a very quick way to all kind of requirements from parking automation systems manufacturers (PARCS or PMS), and the reason relies on a deep learning OCR technology.
Our ALPR solution (SIRAM OCR) allows a fast learning process when adding new unrecognized countries. This means that the license plate recognition engine software responsible of the license plates identification can in a record time learn and support new license plates types from new unrecognized countries. Currently, thanks to deep learning, our Siram OCR ALPR solution allows recognizing more than 100 countries with a high success rate.
There is no doubt that Deep learning technology has demonstrated a great stability and hardiness in the ALPR field in comparison with other older Machine Learning technologies. Specially when talking about adding new types of license plates with no need of software changes, OCR is versatile and a great Deep learning ALPR apprentice. The ability of this license plate recognition software to learn in real time, just like a human being does, allows to recognize patterns and generate knowledge optimizing processes, reducing errors and incorporating new data rules.
For large parking operators and manufacturers of automation systems (PARCS or PMS) with international management, one of the main demands is to find a versatile and effective license plate recognition solution when facing the incorporation of new number plate-based functionalities (#Ticketless , #PaybyPlate, #DynamicExit, #PaybyPlate). This ALPR solution must guarantee a high precision rate when dealing with complicated situations and accesses, as well as a high response speed in the incorporation of new languages, countries, formats and types of vehicles license plates.
Regarding SIRAM OCR the speed of adaptation is assured for being a technology based on Deep Learning, with our ALPR solution is possible to minimize learning periods due to the high automation. The first OCR´s versions offered the chance to incorporate the identification of license plates from 25 countries. This process was achieved within 1.5 months, being a much faster process than using older Machine Learning or Computer Vision technologies, in which the training of a new country involved took between one and two months.
But the most remarkable point has been SIRAM OCR learning speed that allows the incorporation and recognition of a second budge of 50 new countries, since this process has been carried out in less than two weeks.
Nowadays, thanks to Deep learning, our ALPR OCR technological solution allows to recognize vehicles license plates from more than 100 countries with a high accuracy rate, which in many of them is higher than 99%
Looking for an ALPR partner worldwide? Access to full countries list here
Find out more about SIRAM OCR here or contact Innova System Group.
