基于BP神经网络的车型识别外文翻译

、外

文

资

料

License Plate Recognition Based On Prior Knowledge

Abstract - In this paper, a new algorithm based on improved

BP (back propagation) neural

network for Chinese vehicle license plate recognition (LPR) is described. The proposed approach

provides a solution for the vehicle license plates (VLP)

which were degraded severely.

What it remarkably differs from the traditional methods is the application of prior knowledge of license

plate to the procedure of location, segmentation and recognition. Color collocation is used to locate the

license plate in the image. Dimensions of each character are constant, which is used to segment the

character of VLPs. The Layout of the Chinese VLP is an important feature, which is used to construct a

classifier for recognizing. The experimental results show that the improved algorithm is effective under

the condition that the license plates were degraded severely.

Index Terms - License plate recognition, prior knowledge, vehicle license plates, neural network.

I.

INTRODUCTION

Vehicle License-Plate (VLP) recognition is a very interesting but difficult problem. It is important in

a number of applications such as weight-and-speed-limit, red traffic infringement, road surveys and

park security [1]. VLP recognition system consists of the plate location, the characters segmentation,

and the characters recognition. These tasks become more sophisticated when dealing with plate

images taken in various inclined angles or under various lighting, weather condition and cleanliness of

the plate. Because this problem is usually used in real-time systems, it requires not only accuracy but

also fast processing. Most

methods [2], [3], [4], [5] reduce the

existing VLP recognition

complexity and increase the recognition rate by using some

constrains on the position, distance from

specific features of local VLPs and establishing some

the camera to vehicles, and the inclined angles. In addition, neural network was used to increase the

recognition rate [6], [7] but the

traditional recognition methods seldom consider the

prior knowledge of the local VLPs. In this paper, we proposed a new improved learning method of BP

algorithm based on

specific features of Chinese VLPs. The proposed algorithm

[8] and remarkable increases the

overcomes the low speed convergence of BP neural network

recognition rate especially under the condition that the license plate images were degrade severely.

II.

SPECIFIC FEATURES OF CHINESE VLPS

A.

Dime nsions

Accord ing to the guideli ne for vehicle in spect ion [9], all lice nse plates must be recta ngular and

have the dimensions and have all 7 characters written in a single line. Under practical environments, the

distanee from the camera to vehicles and the inclined angles are constant, so all characters of the license

plate have a fixed width, and the distanee between the medium axes of two adjoining characters is fixed

and the ratio between width and height is nearly con sta nt. Those features can be used to locate the

plate and segme nt the in dividual character.

B.

Color collocati on of the plate

There are four kinds of color collocati on for the Chin ese

collocatio ns are show n in table I.

TABLE I

Category of lice nse plate

small horse power plate

motor truck plate

military vehicle and police wago n plate

embassy vehicle plate

vehicle lice nse plate .These color

Color collocati on

blue backgro und and white characters

yellow backgro und and black characters

black backgro und and the white characters

white backgro und and black characters

Moreover, military vehicle and police wag on plates contain a red character which bel ongs to a

specific character set. This feature can be used to improve the recog niti on rate.

C.

Layout of the Chi nese VLPS

The criteri on of the vehicle lice nse plate defi nes the characters layout of Chin ese lice nse plate.

All sta ndard lice nse plates contain Chin ese characters, nu mbers and letters which are shown in Fig.1.

The first one is a Chinese character which is an abbreviation of Chinese provinces. The second one is a

letter ranging from A to Z except the letter I. The third and fourth ones are letters or nu mbers. The fifth

to seve nth

ones are nu mbers ranging from 0 to 9

only. However the first or the seve nth ones may be red characters in special plates (as show n in Fig.1).

After segme ntatio n process the in dividual character

is extracted. Taking adva ntage

en ter one of

of the layout and color collocati on prior kno wledge, the in dividual character will

the classes: abbreviati ons of Chin ese provi nces set, letters set, letters or nu mbers set, nu mber set,

special characters set.

(a)Typical layout

(b) Special character

Fig.1 The layout of the Chin ese lice nse plate

III.

THE PROPOSED ALGORITHM

This

algorithm

consists

of

four

modules:

VLP

location,

character

segmentation,

character

classification and character recognition. The main steps of the flowchart of LPR system are show n in

Fig. 2.

Firstly the lice nse plate is located in an in put image and characters are segme nted. The n every

in dividual character image en ters the classifier to decide which class it bel ongs to, and fin ally the BP

n etwork decides which character the character image represe nts.

Chin ese

character

Fig.2 The flowchart of LPR system

A. Preprocess ing the lice nse plate

1)

VLP Locati on

This process sufficie ntly utilizes the color feature such as

color collocati on, color cen ters

and distributi on in the plate regi on, which are described in sect ion II. These color features can be

used to eliminate the disturbance of the fake plate

'

s regions. The flowchart of the plate locati on is

show n in Fig. 3.

Fig.3 The flowchart of the plate location algorithm

The regions which structure and texture similar to the vehicle plate are extracted. The

process is described as followed:

(1)

P

x

>th

else

⑵

Here, the Gaussia n varia nce is set to be less tha n W/3

(W is the character stroke width), so

R

gets

its

maximum

value

M

at

the

center

of

the

stroke.

After

convolution,

binarization

is

performed

according to a threshold which equals T * M (T<0.5). Median filter is used to preserve the edge gradie

nt and elimi nate isolated no ise of the bi nary image. An N * N recta ngle median filter is set, and N

represents the odd integer mostly close to W.

Morphology clos ing operati on can be used to extract the can didate regi on. The con fide nce

degree of can didate regi on for being a lice nse plate is verified accord ing to the aspect ratio and

areas. Here, the aspect ratio is set between 1.5 and 4 for the reason of inclination. The prior kno

wledge of color collocati on is used to locate plate regi on exactly. The locat ing process of the lice

nse plate is show n in Fig. 4.

(c)M ed inni filler

i

nf

【八

]

(e > Srinctiire veHtlcation

(f)

Pl

ate ex tt

AC

ting

Fig. 4 The whole process of locat ing lice nse plate

2)

Character segme ntati on

This part presents an algorithm for character segmentation based on prior knowledge, using

character width, fixed nu mber of characters, the ratio of height to width of a character, and so on.

The flowchart of the character segmentation is shown in Fig. 5.

Fig. 5 The flowchart of the character segme ntati on

Firstly, preprocess the lice nse the plate image, such as uneven illu min ati on correct ion, con

correct ion and edge enhan ceme nt operati ons; sec on dly,

trast enhan ceme nt, in cli ne

elim in at ing space mark which appears betwee n the sec ond character and the third character;

thirdly, merging the segmented fragments of the characters. In China, all standard license plates con

tain on ly 7 characters (see Fig. 1). If the nu mber of segme nted characters is larger tha n seve n,

the merging process must be performed. Table II shows the merging process. Fin ally, extracti ng the

in dividual character

'

image based on the nu mber and the width of the character. Fig. 6 shows the

segmentation results. (a) The incline and broken plate image, (b) the incline and distort plate image,

(c)the serious fade plate image, (d) the smut lice nse plate image.

TABLEII

Get Nf

If NF> MaxF

For each character segme nts

Calculate the medium point

M

i

For each two con secutive medium points

Calculate the distanee

D

K

Calculate the mi nimum dista nee

Merge the character segme nt k and the character segme nt k +1

NF = NF - 1

End of algorithm

where Nf is the nu mber of character segme nts, MaxF is the

nu mber of the lice nse plate, and i is

the in dex of each character segme nt.

The medium point of each segme nted character is determ ined by:

+ ^2

(3)

and

S

i2

is the final coord in ate

where is the in itial coord in ates for the character segme nt,

for the character segme nt.

The dista nee betwee n two con secutive medium points is calculated

by:

(4)

(u)

Fig.6 The

segme ntati on

results

B. Using specific prior knowledge for recognition

The layout of the Chin ese VLP is an importa nt feature (as described in the sect ion II),

which can be used to con struct

a classifier for recog nizing. The recog nizing procedure adopted

conjugate gradie nt desce nt fast lear ning method,

which is an improved lear ning method of BP

neural network[10]. Conjugate gradient descent, which employs a series of line searches in weight or

parameter space. One picks the first

desce nt directi on and moves along that direct ion

un til the minimum in error is reached. The sec ond desce nt direct ion is the n computed: this direct

ion the

“

conjugate direct ion

”

is the one along which the gradie nt does not cha nge its direct ion

will not

“

spoil

”

the con tributi on from the previous desce nt iterati ons. This algorithm adopted

topology 625-35-N as show n in Fig. 7. The size of in put value is 625 (25*25 ) and in itial weights are

with ran dom values, desired output values have the same feature with the in put values.

Input X XI

X2

…

Xi

…

K625

Fig. 7 The n etwork topology

As Fig. 7 shows, there is a three-layer n etwork which

contains worki ng sig nal feed forward

operati on and reverse propagati on of error processes. The target parameter is t and the len gth of n

etwork output vectors is n. Sigmoid is the

itialized

with ran dom values, and cha nged in a directi on that will reduce the errors.

The algorithm was trained with 1000 images of differe nt backgro und and illu min ati on most of

which were degrade severely. After preprocess ing process, the in dividual

characters are

non li near tran sfer fun cti on, weights are in

stored. All characters used for training and test ing have the same size (25*25 ).The in tegrated

process for lice nse plate recog niti on con sists of the follow ing steps:

1) Feature extract ing

The feature vectors from separated character images have direct effects on the recog niti on

rate. Many methods can be used to extract feature of the image samples, e.g. statistics of data at

vertical directi on, edge and shape, framework and all pixels values. Based on exte nsive experime

nts, all pixels

values method is used to con struct feature vectors. Each character was

reshaped into a column of 625 rows

'

feature vector. These feature vectors are divided into two

categories which can be used for training process and testing process.

2) Training model

The layout of the Chin ese VLP is an importa nt feature, which can be used to con struct a

classifier for training, so five categories are divided. The training process of nu mbers is show n in

Fig. 8.

i

n pul

fknluix

vector

Fig. 8 The architecture of a n eural n etwork for character recog niti

on

As Fig. 8 shows, firstly the classifier decides the class of the in put feature vector, and the n

the feature vector enters the

neural network correspondingly.

After the training process the

training and testi ng process is

optimum parameters of the net are stored for recog niti on. The

summarized in Fig. 9.

(a) Training process

(b)Test ing process

Fig.9 The recog niti on process

3)

Recog nizing model

After training process there are five n ets which were completely trained and the optimum

parameters were stored. The untrained feature vectors are used to test the n et, the performa nee of

the recognition system is shown in Table III. The license plate recognition system is characterized by

the recog niti on rate which is defi ned by equati on (5).

Recognition rate =(number of correctly read characters)

/

(number of found characters) (5)

TABLE III

Class

Number

Letter

Chin ese character

Number and letter

Recog niti on

99.5%

97.4%

96%

97.3%

98.2%

Special character

IV. COMPARISON OF THE RECOGNITION RATE WITH OTHER

METHODS

In order to evaluate the proposed algorithm, two groups of experime nts were con ducted.

One group is to compare the proposed method with the BP based recog niti on method [11]. The result

is show n in table IV. The other group is to compare the proposed method with the method based on

SVM [12].The result is show n in table V. The same training and test data set are used. The comparis

on results show that the proposed method performs better tha n the BP n eural n etwork and SVM

coun terpart.

TABLE IV

Method

Our method

Chin ese character

96%

94

.

5%

TABLE V

Number

99.5%

97.6%

Letter

97.4%

89.8%

BP

Method

Our method

SVM

Chin ese character

96%

Number

99.5%

Letter

97.4%

95.7%

93

.

7%

99.5%

V. CONCLUSION

In this paper, we adopt a new improved learning method of BP algorithm based on specific