Long CHENG，Ling XIONG，Kai－h(huán)an LI

School of Informatics and Engineering，Wuhan University of Science and Technology，Wuhan 430081，China

Recognition of license plate based on Fourier Descriptors

Long CHENG?，Ling XIONG，Kai－h(huán)an LI

School of Informatics and Engineering，Wuhan University of Science and Technology，Wuhan 430081，China

In order to reduce the complexity of image recognition and im prove the recognition speed for license plate，the gray of the plate image is ad justed before the image recognition.The binary image of license plate is characterized and contrasted with the rotated binary image，and the various uncertainties could be proposed by means of Fourier transform algorithm.By using Fourier descriptors，the standard characteristic Fourier descriptors of license plate could be extracted.The characters of license plate could be recognized and positioned through the feature extraction and matching process of license plate.

Digital image，F(xiàn)ourier Descriptors，License plate recognition

? Long CHENG，E－mail:clonger0216@163.com

1.Introduction

Image processing has become a hot research field in recent years.It has become an important research direction to use computer for image processing and recognition.In practical applications，recognition technology of license plate could be applied to public transportation，including public road toll system，smart parking and so on.During the extraction process of image，the further identification and treatment could be affected due to the external influence factors and inaccurate positioning plate character.

Adjustment algorithms have become a common method for image processing.However，the computational complexity of image distortion exists during the implementation process.The following content presents a character matching positioning method.The core idea is the use of Fourier descriptors［1］which describes the characteristics of the image template matching characters to be treated，and this method has a good advantage in thematching and positioning of characters.This algorithm is not only suitable for the license plate image［2］，but also could be used for the associated binary image after the text，tables，etc.

2.A fter obtaining the license plate image preprocessing

2.1.Binarizing image

For the transportation applications，license plate image is usually a color image.Color image will not only increase the pressure of the system，and also could cause a lot of trouble for image recognition.Therefore，the binary image is required［3］.A simplemethod which based on the color of each pixel of the source image is to set RGB reasonable weight when the sum value of pixel gray is beyond a certain value.The weights are based on different values in different primary response to the same gradation of visual brightness.According to the formula(1)，one could obtain the new gray value generally［3］:

g(r，g，b)=0.299R+0.587G+0.114B(1)

According to the formula(1)，each point could obtain a new color image gray value，and the color image becomes a new gray image in which each pixel has different gray value.By using the image histogram equalization，the histogram could be converged in a wide dynamic range［4］.

In the outdoor natural environment，the image could be fade due to the impact of light and camera exposure，and different aging corrosion，rustand other reasons.These plates could be produced less image contrast defects，detail confusion and partially illegible characters.Especially for a car license plate with high speed，it is easier to bring issues such as fuzzy distortion and result in fuzzy image after binarization process.Therefore，it is demanded for image preprocessing before identifying，including the two processes of restoration and noise reduction.During the pretreatment process，the first is the image of corrosion，and then the inflated，thereby removing burrs and encourages noise，but the image position and size will not be changed［5］.

After the completion of the pre－treatment，the image binary operation will be conducted.To improve the processing efficiency of the computer，the image which has different gray levels becomes into only two colors，i.e.，black and white.In order to change the image into white and black colors，an appropriate threshold value of k need to be selected.When the gray value of pixel is greater than or equal to k，itwill be set to 1，which means it is an object image.Otherwise itbecomes0，whichmeans it is the background image.Therefore，the binarization function could be expressed as follows(2):

Where，g(x，y)stands for the gray image function，h(x，y)stands for the binarization function.During the binarization process，the choice of the threshold value of k is very important.The following will introduce the binarymethod－discriminantanalysis method.Discriminantanalysis of the gray value histogram could be divided into two groups based on the threshold value of k by using the average of the respective groups variance(between group variance)and variance(within group variance)in each group ratio(discrimination ratio)to obtain the maximum value of k than that of discrimination，then the value could be taken as the threshold value［6］.

Assuming the threshold value of k is greater than the gray scale value，it will be divided into two groups，i.e.，the group i(i=1，2)of the number of pixels is ωiand the average gray value is Mi，variance σ2.If the average gray value of all pixels is MT，then the variance could be represented by Formula(3):Between－group variance could be expressed as(4):22

By starting a small initial value of k and then gradually increasing this value，once the cyclemultiple times is about to get themaximum discrimination than their corresponding k value，the threshold value of k is determined.After several experiments，the value of k is determined as0.2 and the effectof binary image has the best and most obvious character traits.

2.2.Positioning p late im age

Image of the license plate contains background information and target information after binarization process.Considering the convenience and subsequent operation，one needs to decrease the original binary image of interest region and locate the position of the plate.Themethod used in this paper is to extend the white area and corrosion firstly，remove the stain noise，and use filling mode to fill the periphery of a large possible area.Therefore，theremight exist several target area on the license plate image.In order to extract the correct license plate location，some judgment needs to be conducted，i.e.，car license plate aspect ratio is about 4.5∶1 generally，and the area and circum ference ratio is fixed.As a feature which sets to match the degree of connectivity domain，the higher the degree match，the more accurate description of the license plate location.So，the license plate in the entire image area could be identified.

3.A lgorithm

Based on the above－mentioned treatments，the target binary image could be expressed as formula(7):Where，Mand N represent the number of rows and columns of the image，respectively.

For the image of license plate，the space between characters of license plate is fixed and in a normal rangewhen the image is not tilted.However，the space will be less than a normal space when the license plate image is tilted.If the image of license plate is not tilted，the space between characters is very close to the normal space，which is the maximum space.Based on thismethod，the tilted license plate could be adjusted by rotating the plate.If the space between characters for not tilted license plate is L1 and the space of unknown license plate is L2，once the value of L1 is greater than that of L2，the unknown license plate needs to be rotated.;for normal and non－normal projection plate(as shown in Figure 1 and Figure 2)，after comparison of Figures 2 and 3，Significant differences could be found，i.e.，when taking normal license plate position，the peak will be distinguished，however there exist adhesions between the crest of the vertical projection of the distribution license plate characters;when the license plate character is tilted，the distribution of the vertical projection of the image between the characters is not clear，and there is no clear transition between the peaks and troughs and there do not exist accurate segmentation and character rotation.

Figure 1.Vertical projection histograms

Therefore，the method of Fourier descriptors is proposed tomatch the characters to achieve the identification［7］.

Typically，computations of the Fourier descriptors need discrete polygon boundary curve，and itwill get discrete pointswith N equally spaced.The calculating of the N points could be done by using the equation(8):

Discrete Fourier transforms coefficients Z(k)are the Fourier descriptors:

4.Plate recognition algorithm design

In the above image processing，the position of license plate could be located.When the license plate is tilted，the processing is required and the image rotation processmightbe considered firstly.However，the calculation of image rotation is relatively complex and itwill cause some distortion for the image.Several post－rotation could result in the distortion of image，there exists a superposition of the cumulative error，and then itwill cause some trouble to identify and deal with the subsequent image.Therefore，the discrete Fourier transform of the image processing is adopted as shown in formula(10)［8］:Given the general character of the license plate on a fairly standard，F(xiàn)ourier descriptorsway could be used to match position on the license plate characters.Fourier descriptors are defined as follows:the boundary of the target image is a closed curve，for a fixed starting point on a closed curve，the change of moving along a boundary curve of the coordinate point is a periodic function.The periodic function could be expanded into a Fourier series after normalization.Since some of the Fourier series coefficients are directly related to the shape of the boundary curves，the shape could be described as an element，

called Fourier descriptors.In the process of position and matching，the character images could be extracted，then pictures of Fourier descriptors calculated by descriptorswillmatch the character images and then identify the characters.

5.Simulation results and analysis

Here we use Matlab to analyze the image with a specific license plate number recognition，to verify the effectiveness of this algorithm.The simulation is divided into two steps:the first is the position process of license plate，and the second is the extraction and character recognition process.

According to the aforementioned algorithm，we will conduct gray－scale image preprocess as follows:

Figure 2.Grayscale image

Once the gray image is obtained，the histogram equalization will be calculated to obtain some useful details and enhance the treatment effect.Meanwhile，in order to improve the processing speed of the computer，the gray image could be binarization by selection the threshold value of 0.3 based on experience.And then，the picture will become a pure black and white image，as shown in Figure 3，to facilitate the further process.

Figure 3.Binary image

After the above steps，the preliminary image processing is completed.Next，the license plate will be located.The first is to find the connected domain boundaries，while preserving this graphic to prepare the back mark on it.Connectivity in all domains，a most likely standard to determine the license plate area is:the actual aspect ratio of themeasured value of the license is about 3.14∶1，and there is the corresponding ratio between area and perimeter:(3.14 ×L×L)/(2×(3.14+1)×L)2≈0.379 2.According to this conclusion，one can take advantage of this obvious feature.

Taking metric=area/0.379 2*perimeter2as a connected domain matching degree plate area，when the value it is approaching to 1，themore likely is the target rectangle with corresponding connected domain.After this treatment，the image is shown as follows:

White areas are corresponding regional connectivity，whose portion ismarked“○”rectanglematching is0.98，which can be determined the area of license plate.

Figure 4.Corrosion expansion image

The second step is the license plate character extraction and recognition.The specific coordination could be obtained based on license plate area.By calculating the specific coordinate position of connected domain，one could get the license plate of a color image on the original image.After grayscale and binary processing，it becomes as follows:

Figure 5.Binary image

Then highlight the license plate image processing，and expand it to a 256 × 256 matrix(as shown below)，carry outa preparation for the following rotation matrix of Fourier transform operation.

Figure 6.Highlight treatment

Read a character from the plate after processing template(with the character“6”，for example，the template image“6”is extracted from the binary image).Computing character template image Fourier descriptors，with pre－defined decision－making function descriptors calculated.The brightness level indicates thematching degree of the corresponding region for the template in the processed image［9］.By checking comparison of C max，an appropriate threshold(240 is more appropriate)could be determined，the display brightness is not less than the value of the threshold point，the highest degree ofmatching template position could be obtained as shown in Figure 7.

By comparing the left and right views，the character“6”could be identified and located.Based on the same method，one could identify and locate the other characters.

In order to identify the Chinese characters by Fourier descriptors，the simulation is conducted for Chinese characters.The recognition rate is slightly lower than that based on the templatematching algorithm.Themain reason contains two key points.Letters and numbers are simply connected region，F(xiàn)ourier descriptors is better than the template matching in simple connected region;however，Chinese characters havemany strokes in the font，there aremultiple connected domains.Fourier descriptors can not be accurate to describe it，so the recognition rate is slightly lower than the templatematching.Thismethod could be used in a combinationmode，a combination between Fourier descriptors and template matching will improve the recognition accuracy and efficiency.

Figure 7.Character recognition positioning

6.Conclusion

Experimental results show that the use of Fourier descriptors to describe the image feature matching is an effective way.In particular，the coarse position is known，the image is not distorted，the search range is not big，and it could greatly reduce the computation time and improve the position accuracy of character recognition.However，there still exist some deficiencies for the character recognition，it needs to be continually studied and improve the performance.In summary，the proposed method in this paper has a good guide for the practical applications.

［1］ Wang Tao.Fourier descriptors of the object shape recognition［J］.Computer Research and Development，2002，12:1714－1719.

［2］ Zhang Yin.Oriented vehicle license plate character recognition preprocessing algorithm［J］.Application Research of Computers，1999(7):85－87.

［3］ Zhang Zheng.Digital image processing and machine vision(first edition)［M］.Beijing:Beijing Posts and Telecom Press，2012.

［4］ Tang Hao－kui，Hu Jing.LPR tilt adjustment algorithm［J］.Jornal of Jinan University(Natural Science)，2007(7):246－248.

［5］ Xiao Jian－liang，Qiu Yu－meng A simple plate recognition algorithm and its implementation［J］.Computers and telecommunications，2011(10):49 －52.

［6］ Zhang Jian－mei，Sun Zhi－tian.Research to improve the segmentation algorithm based on discrete Fourier transform image［J］.Computer Simulation，2012(3):300－303.

［7］ Gonzalez waited.Ruan Qiu－qi Digital Image Processing(Second Edition)［M］.Beijing:Beijing Electronic Industry Press，2007.

［8］ Shan Yao.Discussion image processing technology in Automatic Train Identification Application［J］.Railway Computer Application，2006(9):31－33.

［9］ Shan Yao， Wang Hai－ning， Mu Qun－fei. Automatic Train Identification realized in MATLAB［J］.Chinese Science and Technology Information，2006(10):31－33.

［10］ Shan Yao.Based on CCD technology planar Technical［J］.Railway Computer Application，2006.

基于傅里葉描述子在車牌識(shí)別中的應(yīng)用

程 龍?，熊 凌，李開寒

武漢科技大學(xué)信息科學(xué)與工程學(xué)院，武漢 430081

為降低車牌圖像識(shí)別的復(fù)雜度，提高識(shí)別速度，在圖像識(shí)別前調(diào)整車牌圖像的灰度化。針對(duì)二值化后的車牌圖像進(jìn)行特征分析，對(duì)比旋轉(zhuǎn)的二值化圖像的種種不確定性，提出借助傅里葉算法，利用傅里葉描述子，提取標(biāo)準(zhǔn)車牌字符特征的傅里葉描述子，通過對(duì)車牌字符的特征提取與匹配，實(shí)現(xiàn)字符的識(shí)別與定位。

數(shù)字圖像;傅里葉描述子;車牌識(shí)別

U285.5+33

10.3969/j.issn.1001－3881.2014.24.019

2014－04－12