SMALL TRICKS TO ENHANCE THE ACCURACY OF LICENSE

PLATE CHARACTER RECOGNITION

Balázs Enyedi, Lajos Konyha and Kálmán Fazekas

Budapest University of Technology and Economics, 1111, Goldmann tér 3., Budapest, Hungary

Ján Turán

University of Technology Košice, Letná 9, Košice, Slovakia

Keywords: License plate, character recognition, syntax of license plate, geometry.

Abstract: License plate recognition solutions to date are numerous and quite diverse. It is a complex problem field

that can clearly be separated into two areas: localizing the actual license plate number and recognizing

individual characters. Current professional literature devotes relatively small attention to individual steps of

character recognition, which is exacerbated by the fact that the vast majority of solutions result in severe

data losses due to inconsiderate discarding of information that could significantly enhance the accuracy of

the end result that is, improve recognition reliability. Certain letters and numbers are very easy to mistake

for one another, and some solutions focus too heavily on attempting to differentiate between them,

complicating the recognition algorithm and possibly unnecessarily increasing its computation requirements.

Instead, retaining certain information can result in much faster and more accurate recognition algorithms.

This paper describes tricks to enhance accuracy and presents the points of potential significant data losses

during the recognition process. The solutions described here are applicable along with any recognition

algorithm, enhancing its accuracy and reliability.

1 INTRODUCTION

The structure of license plate numbers is vastly

different in each country. However, they all have

several pieces of information, often ignored by

character recognition solutions, that would make

recognition much easier. The references cited at the

end of this paper belong to the following major

classes (the list is not complete): some referred

license plate recognition solutions actually make use

some of the tricks described hereunder, thus

facilitating recognition (references till

Horng, W-B.

2000); some publications emphasize the fact that

certain letters and numbers are easy to mistake for

one another and attempt to provide a solution to this

problem (references from

Shi, X., 2005.); finally, the

vast majority of current solutions completely ignore

(or at least do not mention) any auxiliary techniques

that would facilitate recognition (these publications

are not listed individually). The optional tricks we

describe in this paper are applicable with minimal or

no modifications to existing solutions.

2 DIFFERENTIATION SIMILAR

CHARACTERS FROM EACH

OTHER

A well-known and often experienced fact is that

some characters in license plates are very easy to

misread and thus be mistaken for one another: 0-O,

(0-D, 0-Q,) 1-I, (1-J), 2-Z, 5-S, 6-G, 7-Z, 8-B, etc.

Such errors typically occur with similar numbers

and letters. Considering the Hungarian scenario

(since this is best known to the authors, although

similar rules apply to other countries), characters in

license plates are arranged as follows (L - letter, N -

number):

LLLNNN, LLLLNN, LLLLLN, LLNNNN,

LNNNNN (the first version is the most ubiquitous,

all the others are much less frequently used, Fig. 1)

Since these rules are well known for each

country, the position of letters and numbers can be

known in advance. Therefore, if such a recognition

error occurs, the solution is to simply replace the

125

Enyedi B., Konyha L., Fazekas K. and Turán J. (2007).

SMALL TRICKS TO ENHANCE THE ACCURACY OF LICENSE PLATE CHARACTER RECOGNITION.

In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 125-128

DOI: 10.5220/0002131801250128

 SciTePress

wrong number with the appropriate letter or vice

versa. Using the most frequent arrangement

(LLLNNN) as an example, if the recognizer sees an

“I” in the 2

position where a letter is expected, this

error is detected in the last stage of the recognition

process and the character is simply replaced with a

“1”. The operation of this decision making and error

correction algorithm is quite reliable, while its

computation power and processing time requirement

is almost zero, which enables a simple and effective

correction method for letters and numbers mistaken

for one another.

3 USING TYPICAL EXTRA

SYMBOL INFORMATION IN

LICENSE PLATE TYPES

A very significant question arises as a consequence

of using the recognition accuracy enhancement

described above: how is it feasible to determine

which of the possible license plate character

arrangements/rules should be applied? For instance,

one of the rules listed above stipulates a number for

position 4, while another rule requires a letter in the

same position, which effectively kills the accuracy

of the above described enhancement process.

Recognition algorithms determine the actual

location of license plate character blocks during the

recognition process with a certain level of accuracy.

In each case, the next step is to identify the location

of each separate character, which makes it possible

to cut/copy and separate each character later. Only

after this step can individual characters be attempted

to recognize (using e.g. a neural network or some

other method). Most copying solutions entail

significant actual data loss, because areas outside

character boundaries are usually discarded, resulting

in the loss of

• various (non-alphanumeric) symbols like flags,

stickers, dashes…; and

• distance and size information (how far

individual characters are from each other and

how large they are).

Figure 1: Valid Hungarian license plate types.

First we will examine the actual structure of license

plates (the Hungarian system will be used in the

example).

Figure 1 shows all valid license plate types currently

in use, including both older and newer, EU-

compliant types. They all contain the country flag

and/or country code (“H”) and dashes as well as

stickers between character blocks above and/or

below the dash, indicating (in Hungary) the presence

of a valid car operating license and green card. The

problem stated above (i.e. which arrangement rule to

apply) can be easily solved by recognizing and using

the position of the dash, regardless of the number of

stickers—actually, the rules described above did not

include dash positions, so the actual rules are as

follows:

LLL-NNN,

LLLL-NN,

LLLLL-N,

LLNN-NN,

L-LNNNN,

L-NNNNN

If the dash position is known and recognized for a

specific license plate, this immediately identifies the

rule to be applied and thus letter-vs-number errors

can easily be corrected. Ostensibly, other symbols

should also be found and recognized for accurate

license plate recognition.

The only question remaining is which country

rules to apply to enable the above described error

correction. The country can be identified by

recognizing either the flag, the country codes (for

EU license plates) or any other country-specific

symbols known in advance.

4 USING TYPICAL GEOMETRY

INFORMATION FOR LICENSE

PLATE TYPES

Referring back to the steps of license plate

recognition, once a character is copied from its

source location, typically all assumed character

frames are scaled to a fixed size, which also results

in data loss unless original dimensions or at least the

proportions of frames are stored prior to scaling. For

example, the country symbol is usually much

smaller (by 50% on average) than license plate

characters, but this information is no longer

available after scaling. Using the Hungarian system

as an example, once the “H” symbol (country code)

is found and recognized, but its relative dimensions

compared to those of other characters is no longer

available, it could easily be mistaken for the first

character of the license plate, resulting in dropping

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

126

the last character due to a false recognition error

alert (more than 6 characters found). A similar issue

may arise if a separate dash is found and scaled to a

fixed size (effectively stretching it), since it can then

easily be mistaken for an “I” character. However, if

its original dimensions are still available after

scaling that is, it can be established that its height

was negligible compared to that of other characters

it can safely be identified as a dash (“-“), enabling

the algorithm to assign an applicable rule to that

specific license plate number.

Figure 2: Various license plate types.

The distance of characters may also be helpful.

In Hungary, characters are spaced evenly, while in

Germany and other countries, for example, character

groupings are applied, resulting in different spacing

(Figure 2). Different inter-character distances

unequivocally identify the position of letters and

number within the license plate number. In this case,

the rule to be applied is selected based on character

spacing rather than the position of a dash.

Based on the license plate types listed in Figure

2, the reliability of the recognizer can also

potentially be enhanced by looking for and detecting

other special symbols in addition to geometry

information, dashes and country symbols.

5 CONCLUSION

The reliability of character recognition can be

significantly enhanced by using the extra

information described above. Such information is

typically ignored and discarded by other solutions.

In addition to letters and numbers, other symbols as

well as character dimensions and spacing can also be

utilized in order to enhance accuracy.

Implementation usually poses no problems, as it

neither requires changes to recognition algorithms

nor adds significant run time penalties. All

references to alternative solutions listed below are

followed by a description indicating which of our

optional enhancements are used in a particular

system. As seen from the list of references, only the

expected position of numbers and letters is used in

the vast majority of cases, with only one significant

exception ([1]) that delves deeper into the individual

enhancement options described above no other

referenced papers focuses on these possibilities.

Figure 3. shows an example for assuming

incorrect syntax of license plate. The algorithm has

found “4ZF-66-VG” and for post processing used

the Hungarian syntax (

LLL-NNN). The bad syntax

resulted “AZF-66-”, because the first character must

be a letter (“A” was the most similar one to “4”), not

a number, and only 3 numbers may follow after the

dash.

Figure 3: Example for incorrect syntax of plate

Figure 4. shows a good example for correct

license plate syntax. The algorithm has found

“&C0R-954I” (“&” represents national flag) and for

post processing used the Hungarian syntax (

LLL-

NNN) which resulted: “COR-954”. So the algorithm

could correct the similarity between “0” and “O”,

and knew that only 3 numbers may follow the dash.

Figure 4: Example for correct plate syntax.

SMALL TRICKS TO ENHANCE THE ACCURACY OF LICENSE PLATE CHARACTER RECOGNITION

127

A general numeric qualification for the

efficiency of the described post processing method

cannot be given, because the syntax applied and

taught to the system influences the efficiency.

Considering only a specific country, if the system is

trained just for domestic number plates, then 100%

of all errors occurring during the recognition of

domestic number plates can be corrected, while it

does not apply to foreign number plates. To improve

efficiency, it may be useful to train the system the

syntax of the frequently occurring foreign numbers

in the given country. We tested approximately 100

erroneously detected domestic number plates, where

we could correct every failure type described in this

paper.

REFERENCES

Oz, C., Ercal, F., Cabestany, J., Prieto, A., Sandoval, D.F.,

2005. A Practical License Plate Recognition System

for Real-Time Environments, (Eds.): IWANN 2005,

LNCS 3512

Vázquez, N., Nakano, M., Pérez-Meana, H., 2003.

Automatic System For Localization And Recognition

Of Vehicle Plate Numbers, Journal of Applied

Research and Technology vol1. No.1.

Brugge, M.H., 2005. Morphological Design of Discrete-

Time Cellular Neural Networks: Chapter 7

Application of the Theory in Car License Plate

Recognition, PhD dissertation

Khan, N.A., 2000. A Shape Analysis Model with

Application to Character and Word Recognition, PhD

dissertation, Technische Universiteit Eindhoven

Duan, T.D., Hong Du, T.L., Phuoc, T.V., Hoang,

V.N.,2005. Building an Automatic Vehicle License-

Plate Recognition System, Intl. Conf. in Computer

Science – RIVF’05

Shapiro, V., Gluhchev, G., 2004. Multinational License

Plate Recognition System: Segmentation and

Classification, Proceedings of the 17th International

Conference on Pattern Recognition (ICPR’04)

Martín, F., Borges, D., 2003. Automatic Car Plate

recognition Using a Partial Segmentation Algorithm,

Signal Processing, Pattern Recognition, and

Applications, SPPRA 2003

C.M.Lee, J., 1999. Automatic Character Recognition for

Moving and Stationary Vehicles and Containers in

Real-life Images, Proc. International Joint Conference

on Neural Networks (IJCNN '99)

Barroso, P., Amaral, J., Mora, A., Fonseca, J.M., Steiger-

Garção, A., 2004. A Quadtree Based Vehicles

Recognition System, 4th International Conference on

Optics, Photonics, Lasers and Imaging (ICOPLI 2004)

Draghici, S., 1997. A neural network based artificial

vision system for licence plate recognition,

International Journal of Neural Systems, 1997, Vol.8

Num.1

Shi, X., Zhao, W., Shen, Y., 2005. Automatic License

Plate Recognition System Based on Color Image

Processing, ICCSA 2005, LNCS 3483

Siah, Y.K., Haur, T.Y., Khalid, M., Ahmed, T., 1999.

Vehicle Licence Plate Recognition by Fuzzy Artmap

Neural Network, World Engineering Congress,

Sheraton Subang, Malaysia

Acosta, B.D., 2004. Experiments In Image Segmentation

For Automatic Us License Plate Recognition, degree

of MSC, 2004, Blacksburg, Virginia

Turner, S.M., Eisele, W.L., Benz, R.J., Holdener, D.J.,

1998. Travel Time Data Collection Handbook:

Chapter 4 - License Plate Matching Techniques, Texas

Transportation Institute, Texas A&M University,

Report FHWA-PL-98-035

Martín, F., García, M., Alba, J.L., 2002. New Methods For

Automatic Reading of VLP's (Vehicle License Plates),

SPPRA-2002 (Signal Processing Pattern Recognition

and Applications)

Chang, S-L., Chen, L-S., Chung, Y-C., Chen, S-W., 2004.

Automatic License Plate Recognition, IEEE

Transactions on Intelligent Transportation Systems,

Vol.5, No.1

Wu, H-C., Tsal, C-S., Lai, C-H., 2004. A License Plate

Recognition System In E-Government, Information &

Security, An International Journal, Vol.15, No.2

Horng, W-B., Lee, C-L., Fan, C-H., 2000. A Study and

Implementation on Automatic Intelligent Vehicle

License Plate Recognition Systems, 2nd Taiwan's

International Conference & Exhibition on Intelligent

Transport Systems

Lee, B.R., Park, K., Kang, H., Kim, H., Kim, C., Klette,

R., Žunić, J., 2004. Adaptive Local Binarization

Method for Recognition of Vehicle License Plates,

IWCIA 2004, LNCS 3322

Dlagnekov, L., Belongie, S., Recognizing Cars, UCSD

CSE Tech Report CS2005-0833., 2005

Ketelaars, N., 2001. Final project: Automated license plate

recognition, Enero 2001.

Xu, Z., Xiang, W., Sun, P., Yu, J., Song, J., 2005. Flowing

Traffic System, Final Report – CSIDC 2005

SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications

128