AUTOMATIC LICENSE PLATE DETECTION IN COMPLEX

CONDITIONS OF ACQUISITION

L. A. D’Amore

1,2

and M. Marengoni

Universidade Metodista de S

ao Paulo, S

ao Bernardo do Campo-SP, Brazil

Universidade Presbiteriana Mackenzie, S

ao Paulo-SP, Brazil

Keywords:

License plate detection, Pattern recognition, Computer vision.

Abstract:

The work presented here shows a robust method for license plate detection. The term robust in this work is

directly related to the efﬁciency of the system as an automated locator of license plates without human inter-

vention and considering speciﬁc characteristics of image acquisition and license plate features. The proposed

method is based on the characters and digits thickness found on the Brazilian license plates. Although the

method was designed for the Brazilian license plate pattern it can be easily adjusted to other patterns. The

results obtained using the proposed method showed a better performance even when compared to commercial

systems.

1 INTRODUCTION

Segmentation techniques and license plate recogni-

tion from digital images are inserted into the area

of computational pattern recognition (Anagnostopou-

los et al., 2008) (Souza, 2000) (Campos et al., 2001)

(Guingo et al., 2004). Pattern recognition is typically

used in commercial applications such as recognition

of: speech, face, iris, hands, ﬁngerprints, characters,

document classiﬁcation, and data mining. This paper

focus on one special case: location and recognition

of license plates from digital images (Guingo et al.,

2004).

The main goal of an automatic license plate recog-

nition system is to detect, to segment and to recognize

the license plate from a digital image (Anagnostopou-

los et al., 2008) (Souza, 2000) (Pedrini and Schwartz,

2008) (de Alencar Lotufo, 2005a). Such a system can

be used either in the public or in the private security

systems, through speciﬁc applications, such as: trafﬁc

violations control, access to restricted areas, access to

public areas and in criminal investigations (Anagnos-

topoulos et al., 2008) (Souza, 2000) (Campos et al.,

2001) (Gao et al., 2007) (Kwasnick and Wawrzyniak,

2002) (Guingo et al., 2002) (Belvesi et al., 1999).

In the trafﬁc violations control these systems can

be used to help recording speed violations, red signal

crossing and trafﬁc in restricted areas, in Sao Paulo

city, for instance, the trafﬁc in the downtown area is

restricted for some vehicles in rush hours, the restric-

tion is based on the last digit of the license plate and

the day of the week. Normally these controls are per-

formed using ﬁxed and mobile cameras capturing ve-

hicles images all day long in different illumination,

speed and weather conditions, (Souza, 2000) (Cam-

pos et al., 2001) (Gao et al., 2007) (Kwasnick and

Wawrzyniak, 2002) (Guingo et al., 2002) (Belvesi

et al., 1999).

Systems for access control in restricted areas have

to verify prior authorization for vehicles to enter or

exit a certain area. This type of system is typically

used in condominiums, businesses and government

ofﬁces (Souza, 2000) (Gao et al., 2007) (Kwasnick

and Wawrzyniak, 2002).

Systems for access control in public areas are used

to check the period a vehicle stay inside, the access

frequency and the trafﬁc ﬂow. This type of system is

typically used in parking areas where there is no re-

quirement for prior vehicle registration (Souza, 2000)

(Campos et al., 2001) (Gao et al., 2007) (Kwasnick

and Wawrzyniak, 2002) (Guingo et al., 2002) (Belvesi

et al., 1999).

The automatic license plate recognition system

has also been used in the public safety area, specif-

ically in criminal investigations. Using images cap-

tured by cameras installed at strategic points, the li-

cense plate is searched and registered in the transit

agencies databases, helping to locate stolen vehicles

and to identify possible escape routes (Kwasnick and

Wawrzyniak, 2002) (Guingo et al., 2002).

445

A. D’Amore L. and Marengoni M. (2010).

AUTOMATIC LICENSE PLATE DETECTION IN COMPLEX CONDITIONS OF ACQUISITION.

In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 445-450

DOI: 10.5220/0002850304450450

 SciTePress

Usually an automated license plate recognition

system has a pre-processing step and three basic steps

(Anagnostopoulos et al., 2008) (de Alencar Lotufo,

2005a) (Gao et al., 2007) (Kwasnick and Wawrzy-

niak, 2002) (Ozbay and Ercelebi, 2005): location and

segmentation of the license plate; location and seg-

mentation of the digits / characters in the license plate,

recognition of the digits / characters in the license

plate.

The pre-processing step is used to enhance the im-

age quality for the following steps. Usually, in this

step, it is applied techniques such as histogram equal-

ization and ﬁltering (Anagnostopoulos et al., 2008)

(de Alencar Lotufo, 2005a) (Gao et al., 2007) (Kwas-

nick and Wawrzyniak, 2002) (Gonzalez and Woods,

2005).

The location and segmentation of the license plate

usually apply techniques based on morphological

characteristics of the plate’s background, the char-

acters color, or a combination of both (de Alen-

car Lotufo, 2005a). The two main techniques used

for the plate’s background are: edge detection with

the Hough transform (Anagnostopoulos et al., 2008)

(Gonzalez and Woods, 2005) and artiﬁcial neural net-

works (Guingo et al., 2004) (Kwasnick and Wawrzy-

niak, 2002). Two of the main techniques based on

the characters/digits color are variation and correla-

tion of image objects (Anagnostopoulos et al., 2008)

(de Alencar Lotufo, 2005a) (Kwasnick and Wawrzy-

niak, 2002) (de Alencar Lotufo, 2005b).

The location and segmentation of digits / char-

acters can be performed using horizontal and verti-

cal projections techniques (Anagnostopoulos et al.,

2008) (Gao et al., 2007) (Belvesi et al., 1999) (Ozbay

and Ercelebi, 2005), searching for an inverted ”L”

(Souza, 2000) and projections about regular polygons

(Guingo et al., 2002).

For the recognition of digits / characters stage the

most widely used techniques are: horizontal and ver-

tical patterns projections comparison (Belvesi et al.,

1999), neural network using pixel values in the in-

put layer (Anagnostopoulos et al., 2008) (Gao et al.,

2007), neural network with the projections obtained

from the regular polygons technique (Guingo et al.,

2004) (Guingo et al., 2002), and subtraction of digits

/ characters patterns (Ozbay and Ercelebi, 2005).

The main goal of this work is to develop a robust

method to locate license plates in images captured

from cameras placed in different roads and under dif-

ferent conditions of: illumination, pose, weather, con-

trast, and outside noises.

The term robust in this work is directly related to

the efﬁciency of the system as an automated locator

of license plates without human intervention and con-

sidering speciﬁc acquisition characteristics and plate

features. Some of the speciﬁc plate features consid-

ered are: license plate tilt, license plate color, license

plate size and position in the vehicle, condition of the

license plate, character/digit readability. Some of the

external interferences considered are: reﬂection, il-

lumination conditions, rain, fog, vehicles registration,

sensor’s noise and license plates standards diversity in

the Brazilian legislation (Guingo et al., 2004) (Kwas-

nick and Wawrzyniak, 2002) (Belvesi et al., 1999).

Some examples of problems addressed in the ro-

bustness are presented in Figure 1 which shows a set

of images from the dataset used in this work . Note

that, only in Brazil, there are eight different patterns

for license plates based on the vehicle usage: rental,

learning, collection, testing, manufacturers, ofﬁcial,

private and diplomatic.

Images with low contrast.

Images with low lighting.

Tilted vehicles.

Vehicles with stickers.

Images with high reﬂections.

Figure 1: Typical images used in the database presenting a

set of problems observed in the image acquisition process.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

446

2 PROPOSED METHOD

One common way to detect license plates in digital

images is based on color change frequency (de Alen-

car Lotufo, 2005a) (Kwasnick and Wawrzyniak,

2002) (de Alencar Lotufo, 2005b), these method, al-

though it can detect license plates positions, it also

creates a large number of hypothesis to be tested.

The method proposed here is based on color change

combined with the thickness of characters and digits

found on the license plates reducing the number of hy-

pothesis to be tested. The method searches the images

looking for regions showing color variations compat-

ible with the thickness of the digits / characters in a

license plate. For each region found in the search pro-

cess a validation technique is applied. The proposed

method was divided into six steps (shown in Figure 2)

which will be discussed below.

Figure 2: The six steps used in the proposed method.

2.1 Pre-processing

The pre-processing step is used to reduce noise in

the acquired images but keeping the edges, mainly in

the characters. This step helps to minimize detection

of false positives in the following steps (Gao et al.,

2007).

In order to deﬁne the techniques to be used in the

pre-processing steps tests were performed using the

mean, Gaussian, median and mean with edge preser-

vation ﬁlters. These ﬁlters were applied using masks

with different sizes (3x3, 5x5 and 7x7). The simple

median ﬁlter with 3x3 mask size showed the best re-

sults, attenuating noise and small inscriptions without

damaging the edges of characters. According to Gon-

zalez (Gonzalez and Woods, 2005) this result was ex-

pected.

The next step shows the result of the pre-

processing step. The original image used as an ex-

ample in this section is shown in Figure 3.

Figure 3: Original image.

2.2 Horizontal Tracks

This step aims to locate horizontal tracks compatible

with the expected characters sizes. This step consid-

ers three parameters: the ﬁrst parameter deﬁnes the

color variation limits between the pixels in the same

line; the second parameter is the minimum expected

width for the characters / digits in a license plate and

the third parameter is the maximum expected width

for the characters / digits in a license plate.

The parameter that limits the gray level variation

is directly related to the image contrast. In the tests

performed the value was empirically determined as

20 for the gray level variation.

The parameters limiting the minimum and maxi-

mum values for the character/digit thickness in a li-

cense plate depends directly on three factors: image

width; vehicle width in the image and the standard

thickness for the character/digit in the Brazilian li-

cense plates.

In this work tests were performed using image res-

olution with 640 pixels in width. The vehicles are

expected to cover from 50% to 100% of this resolu-

tion, depending on the number of lanes covered by

the radar and each lane are, on average, 2 m wide.

The standard thickness of the character/digit is 10 mm

wide, as stated in the National Transit Council reso-

lution 231 (CNT, 2007), as shown as the d measure

in Figure 4. Considering these features an expected

thickness in the acquired image should have some-

thing between 3 and 6 pixels, so these are the limiting

parameters for minimum and maximum thickness val-

ues. The algorithm goes through each horizontal line

AUTOMATIC LICENSE PLATE DETECTION IN COMPLEX CONDITIONS OF ACQUISITION

447

Figure 4: Brazilian license plate.

in the image checking the minimum and maximum

pixel values for adjacent pixels in order to verify if

the difference is within the ﬁrst parameter. When a

valid range is found, the system checks its two neigh-

boring pixels, one at right and one at left, both have

to be either higher or lower tones. Notice that the

same gradient allows considering license plates with

different backgrounds and digit/character color, as far

as they preserve the gradient value. Figure 5 presents

Figure 5: Examples of horizontal tracks located.

three examples of areas that can be found in an image:

region number 1 shows a invalid strip, because it has

gradients of the same sign at the sides of the middle

region. Region number 2 has a central range consid-

ered valid by its width and the gradients of opposite

signs at the side edges. Region number 3 has a cen-

tral strip invalid because its width exceeds the limits

of the expected thickness of characters / digits.

Figure 6: Centers of valid tracks after pre-processing.

When a track meets all the parameters required,

the region’s center is stored in a new image. Figures

6 and 7 show the results of this step with and with-

out the pre-processing step. Notice that the image in

Figure 7 the number of points, possible tracks, which

will be further computed and most likely discarded is

much larger than the image in Figure 6, due to noise

within the image.

Figure 7: Centers of valid tracks without pre-processing.

2.3 Groupings Horizontal Tracks

The third step is responsible for the location of hori-

zontal track groups, and it uses two parameters. The

ﬁrst parameter sets the maximum horizontal spacing

between the centers of horizontal tracks located in the

previous step, this parameter is related to the max-

imum distance between two characters in a license

plate. The second parameter deﬁnes the minimum

amount of horizontal and sequential tracks required

to make a license plate hypothesis.

Again, the National Transit Council resolution

231 (CNT, 2007) helps on deﬁning the maximum hor-

izontal space which is computed based on the average

distance of 50 mm between characters/digits.

Each character might have from one to three

tracks for each character/digit, which gives a mini-

mum of 7 tracks for a license plate in a Brazilian li-

cense plate.

At this step the algorithm stores in a new image

the tracks found and deﬁnes a region which will be

tested for a license plate. Figure 8 shows the result

of this step, the ﬁgure shows three regions with track

clusters.

Figure 8: Groups of horizontal bands.

2.4 Morphology

Notice that some regions at the top and at the bot-

tom of Figure 8 are certainly not license plates. These

regions are formed by grouping one to three single

lines. The fourth step aims to remove these individual

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

448

tracks through the application of morphological oper-

ations of erosion and dilation.

The best results were obtained when it was ﬁrst

applied two erosions, followed by three dilations us-

ing a 3x3 square structural element. Figure 9 shows

the result of this step, it can be observed only two

groups of tracks.

Figure 9: Image after morphological operations.

2.5 Connected Regions

The ﬁfth step ﬁnds connected regions based on a four

neighborhood (Pedrini and Schwartz, 2008) (Gonza-

lez and Woods, 2005) (de Alencar Lotufo, 2005b).

The regions found are placed in new image and these

regions are then called candidates. Figure 10 displays

the results of this step it can be observed that only two

candidate regions were found.

Figure 10: Image of the candidate regions.

2.6 Connected Sub-regions

This last step performs a validation test for each can-

didate regions located in the previous step. The pro-

cess here goes through an exhaustive search using in-

cremental binarizations starting with the value of 10

and going up to the value of 250, thus performing 25

binarizations.

After each binarization the region is searched for six

to eight horizontally connected sub regions, corre-

sponding to the number of characters/digits in the

Brazilian license plates. The regions that meet these

limits are considered valid and, therefore, are consid-

ered license plates. Figure 11 shows three images of

the region considered a valid region. The top image

corresponds to the original image of the targeted re-

gion. The middle image shows the binary image ob-

tained in the search process described above, with the

location of the seven connected sub-regions presented

in the bottom image.

Figure 11: Images of the region considered to be the plate.

3 RESULTS

In order to test the method proposed here a database

with 516 images was used. The images in this

database were acquired from ﬁxed and mobile radar

systems, from four Brazilian highways in the Sao

Paulo state. The images have the following features:

• 640 x 480 pixels.

• Format: jpeg.

• Color and Gray level images.

• Frontal or back license plate images.

• Different types of license plates.

• Tilt posed images.

• Light reﬂections at the license plates, glass or

structure.

• Daylight and night conditions.

• Weather variations such as foggy and rainy condi-

tions.

• Cars with or without stickers with different types

of inscriptions.

• License plates with different conservation condi-

tions.

The database was also used in two other systems,

the SeeCar (HTS, 2009) a commercial system from

Hi-Tech Solutions from Israel and the SIAV, an aca-

demic system available at Universidade Federal do

Rio Grande do Sul (Souza, 2000). Table 1 presents

the results obtained by the three systems in the

database.

AUTOMATIC LICENSE PLATE DETECTION IN COMPLEX CONDITIONS OF ACQUISITION

449

Table 1: Compares the results obtained by the three license

plate detection systems in the Brazilian database.

System Correct license plate detection

Proposed 93,80%

See/Car 75,00%

Siav 54,65%

4 CONCLUSIONS

A new method was proposed to detect license plates

based on the Brazilian license plate legislation in im-

ages acquired in a diverse range of conditions. The

method shows that it is effective on detecting license

plates with a detection rate above 90%, a performance

above other systems, including a commercial system.

The proposed method showed a detection rate about

20% above the commercial system and about 40%

above the academic system used in the comparison.

The method also showed that it is robust, changes

in the acquisition condition described in the text such

as weather conditions, illumination, pose, among oth-

ers, apparently do not change the detection rate.

The method is completely automatic and can be

combined with other methods in order to identify the

characters/digits in the license plate.

In a recent survey (Anagnostopoulos et al., 2008)

about license plate detection and identiﬁcation one

can check that the average detection rate is around

95%, so the method presented here has a detection

rate in the same range. Unfortunately the survey does

not mention the type of images used in each algorithm

under consideration presented in the survey.

ACKNOWLEDGEMENTS

The following agencies partially sponsored the work

presented here: Mack Pesquisa and CAPES/CNPQ.

REFERENCES

Anagnostopoulos, C.-N., Anagnostopoulos, I., Psoroulas,

I. D., Loumos, V., and Kayafas, E. (2008). Li-

cense plate recognition from still images and video

sequences: A survey. IEEE Transactions on Intelli-

gent Transportation Systems, 9(3):377–391.

Belvesi, R., Freitas, R., Figueiredo, R., Marcovitch, D.,

Matumoto, G., and Bianchi, R. A. (1999). Um sis-

tema de reconhecimento automatico de placas de au-

tomoveis. Faculdade de Engenharia Industrial FEI;

II ENIA Encontro Nacional de Inteligencia Artiﬁcial,

Rio de Janeiro-RJ.

Campos, T. J., Bampi, S., and Susin, A. A. (2001). Sis-

tema de identiﬁcacao de placas por processamento

automatico de imagens. VII Workshop Iberchip,

Montivideo-Uruguay.

CNT (2007). Resoluo n 231 do conselho nacional de

transito do brasil - sistema de placas de identiﬁcao;

http://www.denatran.gov.br/contran.htm;.

de Alencar Lotufo, A. L. G. R. (2005a). Metodo de localiza-

cao de placa de licenciamento veicular. Universidade

Estadual de Campinas.

de Alencar Lotufo, F. G. D. R. (2005b). Melhorias para sis-

temas de reconhecimento da placa de licenciamento

de veiculos. Dissertacao de Mestrado, Feec, Uni-

camp; Campinas-SP, Brasil.

Gao, Q., Wang, X., and Xie, G. (2007). License plate

recognition based on prior knowledge. School of In-

formation Engeneering, Dalian Maritime University,

China; Proceedings of the IEEE International Con-

ference on Automation and Logistics.

Gonzalez, R. C. and Woods, R. E. (2005). Processamento

de Imagens Digitais. Edgard Blcher.

Guingo, B. C., Rodrigues, R. J., and Thome, A. C. G.

(2002). Faculty division of computer science,

wrockaw university of technology. Federal University

of Rio de Janeiro; 3rd WSEAS - Conference on Neu-

ral Networks and Applications NNA ’02, Fuzzy Sets

and Fuzzy Systems FSFS’02, Evolutionary Computa-

tion EC’02, Interlaken - Switzerland.

Guingo, B. C., Rodrigues, R. J., and Thome, A. C. G.

(2004). Estrategias para o reconhecimento da licencas

de veiculos. Revista Militar de Ciencia e Tecnologia.

HTS (2009). Hi-tech solutions; seecar license plate recog-

nition; http://www.htsol.com.

Kwasnick, H. and Wawrzyniak, B. (2002). License plate lo-

calization and recognition in camera pictures. Faculty

Division of Computer Science, Wrockaw University of

Technology.

Ozbay, S. and Ercelebi, E. (2005). Automatic vehicle iden-

tiﬁcation by plate recognition. Proceedings of World

Academy of Science, Engineering and Technology.

Pedrini, H. and Schwartz, W. R. (2008). Analise de Ima-

gens Digitais : Principios, Algoritmos e Aplicacoes.

Thomson Learning Inc.

Souza, F. P. C. (2000). Localizacao e Leitura Automatica de

Caracteres Alfanumericos, Uma Aplicacao na Identi-

ﬁcacao de Veiculos. PhD thesis, Universidade Federal

do Rio Grande do Sul.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

450