Professional Documents
Culture Documents
Paper 1
Paper 1
255
(b) (c) (b)
256
top and bottom pixels of the selected edge area are recorded candidate plate. Candidates are verified by two stages,
for further edge matching process. In the edge matching evaluating width/height ratio and evaluation based on
step, the ratio of width to height is used to judge whether horizontal crosscuts. In the first stage, only selected
two edge areas are the pair of vertical edges of number candidates are verified against satisfactory predefined
plate. The real ratio may deviate due to the angle of view. constraints on width/height ratio. In the second stage, two
Hence the difference in the ratios of real plate and candidate horizontal crosscuts at 1/3 and 2/3 of candidate’s height are
plate is restricted to be within specified range. Also, two used. The number of objects that cut these crosscuts is
edge areas are matched for equal height. For true number counted. Two approximate constraints are predefined for
plate, the percentage of white pixels is checked for specified this evaluation. Correct accuracy of extraction is 98.76 % by
lower and higher range. This extracted plate is then this algorithm.
normalized and sharpened. Correct plate extraction rate
obtained is 98.45%. III. Methods for Character Segmentation
I. Number plate candidates considered in extraction
Nabeel Khan et al in [6] have developed a new
stage are examined in the phase of character segmentation.
algorithm for plate extraction. After the detection of gray
A wide variety of techniques to segment each character after
scale image, the brightness of image is enhanced by taking
plate localization has been developed. Segmentation
the power of each pixel value as given by (1),
techniques based on projection, Hough transform, region
I(i,j)=I(i,j)p (1) growing, and plate models are considered for addressing.
Where I (i, j) represents gray image, resulting in the image
A. Segmentation by Projection
with brighter edges. For plate segmentation SCAN_MP
In [10], horizontal histogram is used to find out upper
algorithm is implemented which finds out only those objects
and lower bounds of a number plate on the vertical edging
which have horizontal and vertical lines. It checks for
image first. After the area bounded by upper and lower
multiple rows and columns to ensure the correct extraction
bounds is found, the areas above the upper bound and the
of the object. Extracted plates are validated by using specific
areas below the lower bound of a number plate are removed.
features of it like width by height ratio and ratio of black
Area without the upper and lower bounds is considered for
pixels.
character segmentation. Character segmentation phase is
executed in two steps. First the selected area is binarized
C. Morphological based Plate Extraction and secondly vertical projection is performed to find the
In [7], edges in the input images are detected first. Then gaps between characters for a number plate. Based on the
vertical projections of the edge image are found out. Due to results of vertical projection, each number plate is separated
vertical edges in structures such as fences or background, into blocks horizontally starting from the zero points in the
they can be selected as candidate regions. To prevent this, projection histogram. Figure.4 shows the block segmented
authors have proposed the compact factor. A number plate image of number plate. Because of the influence of noise,
has vertical edges in a narrow range but those structures rivet, space mark, rotation of plate, it may happen that one
have vertical edges in a wide range. This feature is used in character region has been separated or two characters are
compact factor. By using it, brightness of pixels in rows is grouped into one single character. Hence a refining
found out and local maximum in that gives the possible algorithm is developed by the authors. This refining is based
candidate area. Plates are extracted by using morphological on width of original number plate characters. The result of
operations. refining is demonstrated in Figure.5. This method has
correct character segmentation rate of 98.82%.
257
C. Region Growing based Segmentation
Six step algorithms is proposed in [13]: image
preprocessing, enhancement of character regions, edge
detection of characters, location of the candidate regions of
the characters, determination of character segmentation
Figure 5. Refining of character ‘A’ connected to the left regions, characters segmentation and binarization. This
boundary of the number plate. algorithm is processed on extracted car number plates.
Character regions are enhanced by contrast stretching
The method developed in [11] has been implemented transformation followed by edge detection using Laplacian
after the processing of captured image and number plate Transformation. Location of the candidate regions are found
extraction. The idea is based on the fact that, number plate out with the help of region growing algorithm. Region
area is usually vertically and horizontally centered. So by Growing is a procedure that groups pixels or sub regions
taking vertical concentration top and bottom positions of the into larger regions based on predefined criteria. The basic
characters is found out. By knowing these positions exact approach is to start with a set of ‘seed points’ and from these
plate area is segmented out. Character segmentation is seed points growing the regions by appending to each seed,
carried out by taking horizontal concentration of the image. to those neighboring pixels that have properties similar to
seed point. In this paper, starting pixels are selected having
B. Segmentation based on Hough Transform the grey value of 255 and region is grown from seeds to
Yungang Zang and Changshui Zhang developed a new with the pixels that are 8-connected to at least one pixel in
algorithm in [12] based on Hough Transform. As a that region. To eliminate the fake regions, standard value of
preprocessing step, extracted plate image is enhanced so that height and width of characters are used for comparisons.
most character pixel is brightened while, background pixels Refining process is also proposed for region growing
are weakened. Character segmentation algorithm is executed algorithm, which then extracts correct segmented characters.
in two stages: horizontal segmentation and vertical After getting the exact position of characters, characters are
segmentation. Horizontal segmentation is processed in three extracted from original image and binarized to get higher
steps: 1) image is divided into many blocks based on contrast. Figure.4 illustrates the execution of this region
vertical projection of the extracted plate image, 2) horizontal growing algorithm. In this method 320 images were tested
segmentation line is found put for each block by analyzing for character segmentation and the success rate obtained is
horizontal projection of the blocks, which is called as 97.2%.
subsection line, 3) using Hough transform on the midpoints
of all subsection lines, all incorrect subsection lines are
eliminated according to prescribed criteria, and then all
correct subsection lines are combined into a whole line.
Advantages committed by this algorithm are; no need of
rotation correction of plate images, influence of background
(a)
is weakened and also the illumination variance. Vertical
segmentation algorithm is based on the projection analysis
constrained by the prior knowledge of the number plate such
as, size of number plate and character, interval between the
characters and inters segment distance. Vertical
segmentation lines are decided on the basis of valleys in the
vertical projection, prior knowledge of the plate, and finding (b)
out the variance of pixel’s gray level in the image. Results
obtained for horizontal and vertical segmentation are
demonstrated in Figure.6.
(c)
258
vertical edge of plate and distance of characters from
horizontal edge of the plate.
V. Conclusions
IV. Methods for Character Recognition This paper has attempted to provide a survey of
Character recognition fully relies on isolated characters. research on LPR algorithms and to offer addressing of some
Inefficiency of character segmentation influences the categories of the methods used in references. This paper
success of character to be recognized. For this work, various discusses the techniques of number plate detection,
techniques have been proposed which are based on character segmentation and character recognition. Though
statistical classifiers, computational intelligence architecture there are several commercial LPR systems, briefing the
and pattern matching. techniques used in these systems is beyond the scope of this
paper, as their operation is strictly confidential.
A. Statistical Classifiers LPR algorithms addressed were country specific. To
To eliminate the inefficiency of OCR and pattern predict the number plate, algorithms presented in the
matching techniques statistical classifiers employ Support literature are restricted to the working conditions such as,
Vector Machine for recognition. Authors in [15] employ distance, background, illumination, vehicle position.
SVM to recognize numbers. The algorithm starts from a LPR can be a measure of vehicle identification, it may
collection of samples of numbers from number plates. Each be further exploited for vehicle model identification and
character is recognized by an SVM, which is trained by speed estimation. The measure of section travel time could
some known samples in advance. In order to recognize a be useful in traffic management, planning and control.
number plate correctly, all numbers are tested one by one Increased mobility and globalization set the challenges
using the trained model. The recognition results are of developing effective LPR system that could handle plate
achieved by finding the maximum value between the from various countries with different character sets and
outputs of SVMs. Similar SVM based algorithm for syntax. Most LPR system focuses on processing of one
recognition is proposed in [16], and [17]. nameplate in the image. But, as, input image may contain
more than one name plates; the future challenging task is to
B. ANN Classifiers detect these plates.
IN [18], authors have formulated joint segmentation and
recognition in a unified framework instead cascaded VI. Acknowledgment
operation or feedback mechanism. A two layer Markov I express my gratitude towards my project guide, Prof.
network is proposed to integrate these two tasks into one S. P. Narote for, giving the valuable guidance, also, keen
statistical framework. Both low level cues and high level and active interest in my efforts, which contributed for the
prior knowledge are incorporated in terms of probabilities. solution of every problem at each stage of my project. I
An efficient non-iterative belif propagation algorithm is would like to thank Prof. S. A. Kulkarni for his valuable
used for state estimation. In [6], recognition algorithm uses support and advice on all the stages of the project.
both techniques of template matching and NN (Perceptron).
Compared with these two techniques, NN have provided REFERENCES
better matching for recognition s compared to template [1] Hao Chen, Jisheng Ren, Huachun Tan, Jianqun Wang,” A novel
matching (91% against 86%). method for license plate localization,” 4th Proc. of ICIG, 2007,
pp. 604-609.
[2] Gisu Heo, Minwoo Kim, Insook Jung, Duk Ryong Lee, Il Seok
C. Pattern /Template Matching Oh, “Extraction of car license plate regions using line grouping
The pattern matching technique is suitable for the and edge density methods,” International Symposium on
recognition of shape invariant, non-rotated characters. P. Information Technology Convergence, 2007, pp.37-42.
[3] Shen-Zheng Wang, His-Jian Lee, “Detection and recognition of
Comelli et al. used in [19], cross co-relation to find out the license plate characters with different appearances,” Proc. of
matching between normalized characters and templates. 16th International Conference on Pattern Recognition, 2003,
This algorithm is implemented on Italian cars passing vol.3, pp. 979-983.
through tollgate. Templates are designed for two-letter [4] Hsiao-Yun Tseng, Ching-Hao lai, Shyr-Shen Yu, “An effective
license plate detection method for over exposure and complex
province, single prototypes for character string. Algorithm vehicle images,” International Conference on Convergence and
tested more than three thousand real images acquired under Hybrid Information Technology , 2008, pp. 176-181.
different weather and illumination condition obtaining a [5] Mei Yu and Yong Deak Kim, “An approach to Korean license
recognition rate near to 91%. Authors in [20] developed an plate recognition based on vertical edge matching,” IEEE
International Conference on System, Man and Cybernetics,
OCR technique, for recognition of Thai car license plates, 2000, vol.4, pp. 2975-2980.
which are generally shape complicated, using essential- [6] Nabeel Younus Khan, Ali Shariq Imaran, Naveed Ali,
elements-based method. This algorithm recognizes the “Distance and color invariant Automatic license plate
character patterns relying on the essential elements of recognition system license plate recognition based on vertical
edge matching,” International Conference on Emerging
characters. Using this method, majority of characters are Technologies, 2007, pp. 232-237.
reduced to a set of essential strokes. In [21], matching of [7] Farhad Faradji, Amir Hossein Rezaie, Majid Ziaratban, “A
characters is based on edge Hausdorff Distance (HD), also Morphological based License Plate Location,” ICIP, 2007, pp. I
called max-min distance. It is a measure of similarity 57-I 60.
[8] Serkan Ozbay, Ergun Ercelebi, “Auotmatic Vehicle
between two arbitrary point sets. The method possesses all identification by plate recognition,” Proc. of PWASET, 2005,
the mathematical properties of a metric, and its recognition vol. 9, no. 4, pp. 222-225.
rate is very similar that obtained with neural network [9] Tran Dur Duan, Duong Anh Duc, Tran Le, Hong Du,
classifiers but slightly slower. “Combining Hough transform and contour algorithm for
259
detecting vehicle license plate,” Proc. of International
Symposium Intelligent Multimedia, Video and Speech
Processing , 2004, pp. 747-750.
[10] Xiangjian He et al, “Segmentation of characters on car license
plates,” 10th Workshop on Multimedia Signal Processing, Oct.
2008, pp. 399-402.
[11] Choudhary A. Rahman, Wael Badawy, Ahmad Radmanesh , “A
Real time vehicle’s license plate recognition system,” Proc. of
IEEE Conference on AVSS, 2003, pp. 163-166.
[12] Yungang Zhang, Changshui Zhang, “A New algorithm for
character segmentation of license plate,” Proc. of IEEE
Intelligent Vehicles Symposium, 2003, pp 106-109.
[13] Feng Yang, Zheng Ma, Mei Xie, “A Novel approach for license
plate character segmentation,” ICIEA, 2006, pp.1-6.
[14] Chengpu Yu, Mie Xie, Jin Qi, , “A Novel system design pf
license plate recognition,” International Symposium on
Computational Intelligence and Design 2008, pp. 114-117.
[15] Lihong Zheng and Xiangjian He, “Number plate recognition
based on support vector machine,” Proc. of IEEE International
Conf. on AVSS, 2006, pp.13-13.
[16] I Chen Tsai, Jui Chen Wu, Jun Wei Hsieh, Yung Sheng
Chen, “Recognition of vehicle license plate from a video
sequence,” IAENG International Journal of Computer Science,
(Advance online Publication: 17 February 2009).
[17] Xiaojun Chi, Junyu Dong, Aihua Liu, Huiyu Zhou, “A simple
method of Chinese plate recognition based on support vector
machine”, 2006, pp. 2141-2145.
[18] Xin Fan, Guoliang Fan, “Graphical models for joint
segmentation and recognition of license plate characters,” IEEE
Siganl Processing Letters, vol. 16, no. 1, January 2009.
[19] P. Comelli te al., “Optical recognition of motor vehicle license
plate,” IEEE transaction on Vehicular Technology, vol. 44, no.
4, Nov. 1995.
[20] P. Sa-ngamuang, C. Thamnittasana, T. Kondo, “Thai car license
plate recognition using essential elements based method,” Proc.
of Asia-Pacific Conference on Communication, 2007, pp. 41-
44.
[21] TANG shuang-tong, LI wen-ju, “Number and letter character
recognition of vehicle license plate based on edge Hausdorff
distance,” Proc. of 6th International Conf. On PDCAT, 2005,
pp. 1-3.
260
Copyright of AIP Conference Proceedings is the property of American Institute of Physics and its content may
not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written
permission. However, users may print, download, or email articles for individual use.