License Plate Recognition Algorithms and Technology

License Plate Recognition Algorithms and Technology

Automatic license plate recognition has two essential technological issues:
· the quality of the license plate recognition software with its applied recognition algorithms, and
· the quality of the image acquisition technology, the camera and the illumination.

The very key factor is the license plate recognition software. The sophistication of the recognition software, the intelligence and quality of the applied license plate recognition algorithms, the mathematical knowledge and the years of experience behind determines the capabilities of the recognition software. The better the algorithms are, the highest the quality of the recognition software is:
· the highest recognition accuracy it has,
· the fastest processing speed it has,
· the most type of plates it can handle,
· the widest range of picture quality it can handle,
· the most tolerant against distortions of input data it is.

In early years of LPR available software were bound to specific countries. One software could read - for example - Spanish plates only, other could read plates from Hong Kong only, etc. This was not accidental: the geometrical structure of the plate as well as its syntax were essential parts of the plate reader software. Without the presumption of a fixed plate geometry (character ratios, character distribution, font type, plate colour, etc.) and a well defined syntax the algorithm may not even found the plate on the picture.


Plate geometry and basic syntax

The most advanced algorithms today read plates without such presumptions. For example, a good algorithm should read all plates from Europe with the same level of quality. There are indeed a wide variety of plate types in Europe:
· black (dark) characters on white (light) plate,
· white characters on black plates,
· one-row plates,
· two-row plates,
· plates with different character-size,
· latin and cyrillic fonts,
· plates with our without region's shield or special mark, etc.


Reading plates of different type is a measure of technology level

If a license plate recognition algorithm can not utilise such additional information like the prior knowledge of the plate structure or plate syntax, it looses a very helpful part of its input data. This loss results in reduction of maybe the most important quality measure, the plate recognition accuracy.

Without using additional information about the plate the remaining part of the recognition algorithm should be significantly better than it was before, when the additional information could be used. Otherwise it would not be possible to gain back the same recognition accuracy.

We believe that there are two key technological parts of a license plate recognition algorithm that basically determines its quality level:
· a robust, very high accuracy and intelligent optical character recognition technology, and
· a technology that allows intelligent structural analysis of complex higher structures.

The robust, very high accuracy Optical Character Recognition (OCR) technology is a very essential requirement.


The OCR task

To get better perception of the nature of recognition accuracy, consider the following example:

Assume the plates have an average of 7 (seven) characters as license plate number. If the overall plate recognition accuracy is required to be above 96%, than the recognition accuracy of the individual characters should be at least 99.5%. Out of 1000 characters not more than 5 could be misread/misrecognised:

(99.5%)7 = 0.9957 = 0.995 · 0.995 · 0.995 · 0.995 · 0.995 · 0.995 · 0.995 = 0.9655 = 96.5%

If someone speaks about 99% overall recognition accuracy, than the recognition rate of the individual characters has to be at least 99.85%.

But the above calculation is only a very simple estimation of the maximum acceptable OCR error rate: it is not the real error rate the OCR can have! The real OCR error rate has to be much lower than the one given by the above estimation, as there are several other parts of the entire algorithm than can make mistake. And the overall recognition accuracy is the multiplication of the accuracy of the individual (and independent) sub-algorithms.

For example, suppose that there are three additional sub-algorithms before the OCR:
· a plate localisation sub-algorithm, responsible for finding the plate on the picture, having 98.7% accuracy,
· a contrast/brightness normalisation sub-algorithm, responsible to equalise the plate picture, having 99.2% accuracy, and
· a character segmentation sub-algorithm, responsible for finding and cutting out the individual characters on the plates and pass them to the OCR, having 99.6% accuracy.
The OCR has a 99.5% recognition accuracy on the individual characters. The overall license plate recognition accuracy is then only 94.2%, as:

0.987 · 0.992 · 0.996 · 0.9957 = 94.2%

The image acquisition technology determines the average image quality the license plate recognition algorithm has to work on. Needless to say that the better the quality of the input images are, the better conditions the license plate recognition algorithm has, and thus the higher license plate recognition accuracy can be expected to be achieved.

What does good image quality mean?

In order to expect reasonable results from a plate recognition algorithm, the processed images should contain a plate
· with reasonable good spatial resolution,
· with reasonable good sharpness,
· with reasonable high contrast,
· under reasonable good lighting conditions,
· in a reasonable good position and angle of view.

Indeed, 'reasonable' is not an exact definition, still it has a well understandable meaning. Here are some problematic images:

Low spatial resolution (too small characters on the plate)

Blurred image

Low contrast

Overexposure

Bad lighting conditions (shadow and strong light)

High distortion

An image acquisition system is considered to be good if it provides a stable, balanced, reasonable good image quality under all of its working conditions. If an LPR system has to work outdoor 24h/day, 7days/week in Middle-Europe, than it has to handle quite a wide range of lighting and weather conditions. Under the chapter Technical and Quality Issues of License Plate Recognition we provide further technical details regarding image quality.

There are some interesting sites where you can find samples of plates all around the World:
· Automobile License Plate Collectors Association
· Eugen Winklharrer collector's page
· Alessandro Libanore collector's page
· PL8S the license plate collectors website

Keywords: license plate recognition algorithm, license plate recognition technology, recognition accuracy, license plate recognition accuracy, LPR technology, car number ocr, car plate ocr, license plate structure, license plate syntax, license plate localisation, license plate localization, character segmentation, contrast brightness normalisation, contrast brightness normalization, license plate recognition