The following manuscript was published in
Jun-Wei Hsieh, Shih-Hao Yu, and Yung-Sheng Chen, Morphology-based license plate detection in images of differently illuminated and oriented cars , Journal of Electronic Imaging, Vol. 11, No. 4, 507-516, 2002.
Morphology-based License Plate Detection Jun-Wei Hsieh* , Shih-Hao Yu, and Yung-Sheng Chen Department of Electrical Engineering Yuan Ze University, 135 Yuan-Tung Road, Chung-Li 320, Taiwan, R.O.C.
[email protected]
Abstract This paper presents a morphology-based method for extracting license plates from cluttered images.
The proposed system consists of three major components. At the first, a morphology-
based method is proposed to extract important contrast features as guides to search the desired license plates.
The contrast feature is robust to lighting changes and invariant to different
transformations like image scaling, translation, and skewing. Then, a recovery algorithm is applied for reconstructing a license plate if the plate is fragmented into several parts. The last step of the proposed method is to do license plate verification. The criterion of license plate verification is based on the number of characters appearing in the plate that can be extracted from a clustering algorithm. The morphology-based method can significantly reduce the number of candidates extracted from the cluttered images and thus speeds up the subsequent plate recognition. Since the extracted feature is robust to different camera views and lighting changes, most of the desired license plates can be correctly extracted no matter how the input images are captured. Under the experimental database, 128 examples got from 130 images are successfully detected.
The average accuracy of license plate detection is 98%. Due to the simplicity of the
proposed method, all the license plates can be extracted very fast (less than 0.5 seconds). Experimental results show that the proposed method improves the state-of-the-art work in terms of effectiveness and robustness for license plate detection.
Keywords: Morphological Operations, License Plate Detection, Clustering, Intelligent Transportation System, Rectification.
* To whom all correspondence should be addressed.
1
1. Introduction With the rapid development of public transportation system, automatic identification of vehicles has played an important role in many applications during the past two decades1, 2-3 . For examples, the plate detection method under discussion might be applied in electronic tolling to help identify violating vehicles.
In addition, the system can be utilized for managing park
facilities, monitoring unauthorized vehicles entering private areas, detecting stolen vehicles, controlling traffic volume, ticketing speeding vehicles, and so on. One of the most effective and useful identification methods is the license-plate recognition (LPR) through visual image processing. Due to the widespread application fields, LPR has been an important key function in an intelligent transportation system. A LPR system is mainly composed of three processing modules; that is, license plate detection, character segmentation, and character recognition. Among them, the task “license plate detection” is considered as the most crucial stage in the whole LPR system. Once the license plate has been well located, the result can be fed into the character recognition module for identifying the actual vehicle, which has been explored broadly in optical character recognition applications 4 . In the past, a number of techniques5-20 have been proposed for locating the desired plate through visual image processing.
The major features used for license plate detection
include colors6 , corners7 , vertical edges8 , symmetry9 , projections of vertical and horizontal edges10 , and so on. For examples, K. K. Kim et al.6 used color information and neural networks to extract license plates from images. However, color is not stable when the lighting conditions change.
On the other hand, Dai Yan et al.10 used the projections of edges with different
orientations for determining peaks of the histograms as possible locations of license plates. When the scene is complex, many unrelated edges will disturb the determination of the correct plate locations. Moreover, M. Yu and Y. D. Kim8 proposed a vertical edge-matching algorithm for grouping all possible positions of license plates through edge matching. In this approach, they assumed the vertical boundaries between a license plate and its backgrounds are strong. However, when the colors of the license plates are similar to their backgrounds, the assumption will no longer exist. Other methods using features like corners7 and symmetries9 also made the same assumption. The major problem in these approaches is the used features depend strongly on the intensity differences between the extracted license plate and car colors, which are not
2
stable when the lighting condition, camera orientation, or car color changes. This paper presents a novel approach for detecting license plates from visual images based on morphological operations. In the past, the morphology-based approach has been successfully applied into different applications like face detection21 , noise removing, thinning, etc. However, no paper directly uses morphological operations to extract features to detect the desired license plates.
The proposed system consists of three main stages.
In the first stage, we use
morphological operations to extract the contrast features within a license plate as an important cue to extract license plates.
The contrast feature is invariant to several geometrical
transformations like car color, camera translation, rotations, and scaling. Even through lighting changes, the high intensity difference between the characters and the backgrounds can be still maintained.
Therefore, the proposed method works stable under different image alterations. In
the second stage, the previously selected candidates are then fed into a plate recovery procedure. Usually, due to noises, a license plate cannot avoid being segmented into several fragments. Therefore, a procedure must be developed for reconstructing an intact plate in event it erroneously becomes segmented during the extraction process.
The recovery algorithm is
cluster-based and thus invariant to different geometrical changes. The last stage of the proposed system is to perform plate verification. In this step, each plate candidate is verified according to the number of characters appearing in the candidate. Since the number is also obtained from a clustering algorithm, the criterion works well under different image alterations. Once the set of registration characters has been extracted, a standard optical character recognition system can be applied directly for vehicle identification. The morphology–based segmentation process is able to significantly reduce the number of candidates extracted from a cluttered image. Thus, the subsequent plate recognition can be much sped up since less than five candidates are left for further processing.
More importantly, the proposed detection technique can locate multiple
plates with different orientations in one image.
In the experiments, 130 cluttered images
including different lighting and orientation variations are used to test the effectiveness of the proposed system. 128 plates are successfully located and thus the accuracy rate of detection is approximately 98%. In average, the proposed approach requires less than 0.5 second to finish the detection task on a 640 × 480 test image. Experiments show that the proposed method is a great improvement in terms of effectiveness and robustness of license plate detection. The rest of the paper is organized as follows. In the next section, we will first overview the 3
proposed system.
Then, the details of the morphology-based feature extraction for license plate
detection are described in Section 3.
The whole procedures of the license recovery and
character segmentation algorithms are illustrated in Section 4. Section 5 reports experimental results. Finally, a conclusion will be presented in Section 6.
2. Overview of the Proposed System
Original Image
Feature Extraciton Using Morphological Operations
Selection of License Plate Candidates
License Plate Verification
Final Results
Fig. 1: Flowchart of the proposed system. The paper presents a technique for automatically detecting license plates from color or gray images. Fig. 1 is the flowchart of the whole system. The proposed system is composed of four major parts: image acquiring, feature extraction, license-plate selection, and license plate verification.
Each part is described as follows:
Image Acquisition: Input of the proposed system is an image or an image sequence captured by a general CCD camera; for example, video camera or TV camera, under different uncontrolled environments. Feature Extraction: Characters in a car license plate are specially designed to have distinctive intensities to their backgrounds. The relative high contrast between characters and the background can be used as a key for detecting the wanted plates from images. In this paper the contrast feature is extracted through a morphology-based method to detect the desired plates. In the past, there have been many researchers applying morphology operations into the license plate detection for reducing the undesired noises. None uses these operations for extracting features in license plate detection. License-plate Candidate Selection and Recovering : With the extracted features, several plate candidates can be selected. However, due to noises or lighting condition changes, the whole license plate region may be segmented into several fragments.
Therefore, before
selecting the wanted license plates, a recovery algorithm should be applied for recovering the complete one from the pieces of fragments. License Plate Verification:
Once a set of region candidates has been selected, the
confidence of each candidate for being a license plate is verified according to the geometrical 4
properties and the number of characters appearing in this candidate. The geometrical criteria include the density, height, width, and area of this region. The number of characters in this region is calculated by a clustering method.
3. Feature Extraction Using Morphological Operations In the past, color and vertical edges are two common features used to detect the locations of license plates. However, different lighting conditions will lead to a license plate having various appearances and result in the failure of detecting these features. It is known that a license plate is a pattern composed of several characters that have high distinctive intensities to their background. The high contrast area can be used as a key feature to detect the desired license plates.
In the following, we will describe a morphology-based method to extract the high
contrast area for detecting the desired license plates. 3.1 Preprocessing It is well known that image intensities are very sensitive to different lighting variations. The same object with different illuminations may own considerably different colors or gray intensities. Before feature extraction, a histogram equalization process is applied first to reduce the lighting variations into minimum. The equalization method used here can be found in the textbook of Sonka et al.22 . 3.2 Features for License Plate Detection In this paper, we use several morphological operations to find the high contrast area as an important feature to detect license plates.
Before introducing the proposed method, some
morphological operations should be described first. Let Sm, n denote a structuring element with size m × n , where m and n are odds and larger than zero. Fig.2 shows two kinds of different structuring elements. Let I(x,y) denote a graylevel input image. According to the definition of Sm, n , the smoothing, dilation, erosion, closing, opening, and other operations are defined as follows: Smoothing Operation: ESm× n ( I (x , y )) =
1 n / 2 m/ 2 ∑n/ 2 j =−∑m / 2 I ( x + i, y + j )Sm×n (i, j) , mn i=−
Dilation Operation: I ( x , y ) ⊕ Sm×n =
max
I ( x − i , y − j ) Sm, n ( i , j ) ,
min
I ( x − i , y − j ) Sm, n ( i , j ) ,
Erosion Operation: I ( x , y ) e Sm×n =
|i |≤m /2,| j |≤n / 2
|i| ≤m /2,| j |≤n / 2
5
Closing Operation: I ( x , y ) • S m,n = ( I (x , y ) ⊕ S m,n ) e S m,n , Opening Operation: I ( x , y ) o S m,n = ( I (x , y ) e S m,n ) ⊕ S m,n , Differencing Operation: D( I1 ,I 2 ) = | I1 (x , y ) − I 2 ( x , y) | , 255, if I ( x , y ) > T ; and Thresholding Operation: T (I (x ,y ) = otherwise. 0,
S1,5 =
1
1
1
1
1
S3,7 =
(a)
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
(b) Fig.2: Two kinds of structure elements. As mentioned previously, the license plate is a pattern with high contrast intensities to the background. In this paper, we propose a morphology-based technique to detect these high contrast areas. The whole procedure of morphology-based feature extraction is shown in Fig. 3. First, in order to eliminate noises, a smoothing operation with a structure element S7,7 is applied first. Then, the closing and opening operations with a structure element S7,1 are performed into the smoothed image such that the images I c and I o can be obtained, respectively. In order to detect vertical edges, a differencing operation is further applied into the images I c and I o . All possible vertical edges can be extracted with a thresholding operation. It is known the vertical edges in a license plate are close and adjacent to each other. These adjacent edges can be connected together through a closing operation and then form a connected segment. Therefore, before thresholding, a closing operation is applied to let all adjacent vertical edges form a connected region. segments.
Then a labeling process is executed to extract the license-plate-analogue
After that, a set of potential license plates can be obtained from a cluttered
environment. Fig. 4 shows an example demonstrating the license-plate-analogue segmentation using morphological operations.
6
Source Image
Opening (7*1)
Average (7*7)
Closing (1*7)
Closing (7*1)
Thresholding
Labeling
Differen cing
End
Fig. 3: Details of the proposed method to extract useful features for license plate detection.
(a) Input gray-scale images
(b) Results after applying morphological operations Fig. 4: Results after applying the suggested morphological operations to different images. 4. License Plate Segmentation After labeling, a set of potential license plates can be extracted from the images processed by morphological operations. Due to noise, many incorrect license plates may be extracted from a cluttered environment. For instances, frames of windows, trees, edges among a set of books, etc., are frequently segmented as license-plate-analogue regions. Therefore, some geometrical and textural information are used first to remove impossible regions.
Since the extracted
candidates may be fragments of a license plate, a recovery algorithm is then applied for recovering the whole plates from these fragments. After that, all the extracted candidates are verified according to the number of characters appearing in this candidate region by a clustering 7
algorithm. 4.1 Candidate Extraction Let R denote the extracted plate region with the size w × h where w and h are the width and height of R, respectively. To eliminate impossible plate regions, several criteria are used here. The first criterion is the density of the region R: den =
A , where A is the area of R. hw
second criterion is the ratio r between the width and height of R; that is, r = w / h .
The
The third
criterion is that the size of a license plate should be larger than a fixed size, for example, 60 × 25 . If the size of a license plate is not larger enough, the characters in the license plate will be too small for recognition. According to the above definitions, if a region R is said to be a candidate of license palate, R should satisfy the following rules: Criterion 1: w > 60 and h > 25 ; Criterion 2: the density den should be larger than 0.25 and less than 0.85; Criterion 3: the ratio r between w and h should be larger than 1 and less than 3.5. These simple rules will significantly reduce the number of potential license-plate-analogue segments into few candidates. These rules are decided according to the features of “Taiwan’s license plate”. However, these thresholds can be easily refined according to different properties of license plates in different countries.
Once the set of possible plate candidates has been
extracted, each extracted regions should be further binarized for verification and recognition. In this paper, a “minimum within-group variance” dynamic thresholding method23 is applied to each extracted candidate region.
After binarization, a conventional labeling process is then
performed to locate each character-like region from the extracted plate candidates. 4.2 License Plate Recovery Due to noises or different lighting conditions, the extracted candidates may be fragments of a license plate. For examples, in Fig. 5, the extracted region in (a) is a fragment of (b). A 8
straightforward method to tackle this problem is to use a vertical intensity projection for calculating the average width and height of each character appearing in the fragment. Then, the whole plate is recovered from the fragment based on this character information. However, this method is not stable under different skewing, scaling, and rotation of the extracted plate. In this paper, a cluster-based method is proposed for calculating the common geometrical properties of characters and then using the information to recover the whole plate. The method is robust to the skewing, scaling, and rotation of the extracted plate.
(a)
(b)
(c)
(d)
Fig. 5: Results after license plate recovery. (b) and (d) are the results of (a) and (c). In this paper, the number of characters appearing in a license plate is considered the primary feature to verify the plate.
For the characters appearing in a license plate, they satisfy the
following geometries: G1. The heights of these characters should be similar. G2. The widths of these characters should be similar. G3. The centers of these characters are arranged along a line. Based on the above geometries, we can develop a character analysis algorithm to calculate the number of characters in a license plate. However, before counting, some wrong characters should be removed first according to the following four constraint rules: Rule A: The area of each character should be larger than 25 pixels;
9
Rule B: Each character’s width should be larger than 3 pixels; Rule C: Each character’s height should be larger than half of the height of the extracted plate. Rule D: Each character’s weight should be smaller than half of the width of the extracted plate. For each character in R, if the character cannot satisfy Rules A-D, it cannot be recognized as a correct character.
Therefore, through checking each character in R, a set Sc of character
candidates, which satisfy Rules A-D, can be selected from R.
Then based on the set Sc , we can
present a clustering method for calculating the common geometrical properties of characters to guide the verification. Initially, assume each character in Sc forms a cluster. Let d i denote the summation of distances of each character in Sc to the ith cluster and si a counter to record the number of characters belonging to the ith cluster. According to the definitions and the geometries G1-G3, the character analysis algorithm to calculate common properties of characters in Sc can be described in Table 1. The common properties include the average character width w and height h , the number Nc of correct characters in Sc , and the set Sˆc of correct
characters in Sc . The main idea of this algorithm is to calculate the average character width w and height h , and then check how many characters in Sc whose width and height are similar to w and h , respectively.
10
N : the number of elements in Sc ; ci : a character in Sc with the size hci × wci ; Td : 0.2, the threshold used to determine whether ci and c j are similar; si : a counter to count the number of correct characters matching to ci ; di : a variable to record the sum of distance when calculating si ; Nc : the number of correct characters in Sc ; Procedure Character Analysis Algorithm ( Sc ) { Set all the values of Nc , d i and si to be zero for i=1, … , N; For each pair ( ci , c j ), repeat the following steps: a. calculate the distance d ij between ci and c j as: d ij =
( h − h ) + (w 2
ci
cj
ci
− wc j
)
2
;
b. if d ij ≤ Td , then calculate di and si as follows: si = si + 1 and d i = d i + d ij ; 0.5 Choose the index k such that: k = argmax (si + ); 1≤i ≤ N 1 + di For each element c j in Sc , if d kj ≤ Td , collect c j as an element of the cluster U; Calculate the average width wc and height hc of characters in U, respectively; For each element c j in Sc , repeat the following steps: a. obtain rj , the closest integer to the value wc j / wc ; b. calculate the similarities of hc j and wc j to wc and hc , respectively, by: p j =|
wc j wc
− rj | and q j = 1 − min(
h hc j , ); hc j hc
c. if p j ≤ Td and q j ≤ Td , update Nc as N c = N c + rj and collect c j into the set Sˆc of correct characters; Return w , h , N and Sˆ ; c
c
c
c
}
Table 1 Algorithm for analyzing the common properties of characters from a region R. After analyzing the common properties of characters in R, in what follows, we will present a new method to use the preceding information to recover a whole license plate from its pieces. Let leftR and right R , top R , and bottomR denote the most left, right, top, and bottom coordinates of R in the x and y directions, respectively. 11
Moreover, assume the number of
characters appearing in a standard license plate is a fixed number N p . Then, the license-plate recovery algorithm can be described in details as Table 2. Algorithm for recovering the whole license plate from a region N p : a fixed number; Tb : the threshold got from the “minimum within-group variance” dynamic thresholding method in the previous stage for binarizing R; Tv : 0.2, the threshold used to verify whether R is a correct license plate; Procedure License-Plate Recovery Algorithm (R) { Use the connected component analysis to get the set Sc of characters from R; Use the character analysis algorithm to obtain the average width wc , the average height hc , and the number Nc of correct characters from Sc ; If Nc is less than N p , the new region R new can be obtained from R as follows: a. left Rnew = left R − ( N p − N c ) × wc and right Rnew = right R + ( N p − N c) × wc ; b. if any black pixel in R touches the top of R,
topRnew = topR − hc / 5 ;
c. if any black pixel in R touches the bottom of R, bottomRnew = bottomR + hc / 5 ; Binarize R new using the threshold Tb ; Obtain a new set Scnew of characters from R new using the labeling technique and the filtering process with rules A-D; Use the character analysis algorithm to get the set Sˆcnew and the number Ncnew of correct characters from Scnew ; If | N Cnew / N p − 1| < Tv , then a. R new is recognized as a correct license plate; b. the new boundaries of R new become:
left Rnew = xleft , right Rnew = xright , topnew = ytop , and bottomRnew = ybottom , R where xleft , xright , ytop , and ybottom are, respectively, the x and y coordinates of most left, right, top, and bottom boundary pixels of characters in Sˆnew ; c
return R }
new
Table 2. Algorithm for recovering the whole license plate from a region R. Fig. 5 shows the results when the proposed recovery algorithm is applied for two different fragment pictures. Clearly, the license plates can be completely and correctly recovered. On the other hand, different camera orientation will cause an inclined plate to be captured. For practical applications, an inclined plate will lead to the failure of recognizing the registration
12
numbers in the plate. In order to deal with this problem, a rectification procedure should be applied for transforming the orientation of the extracted license plate into a normal position. The next section will discuss the details of the proposed rectification algorithm. 4.3 Inclined Plate Rectification
WR
DR
+ (xc, yc)
Fig. 6: The geometry of an inclined plate. Due to different camera orientations, it is difficult to guarantee an inclined license plate will not appear in the captured image. When this case happens, the inclined arrangements of registration characters in this plate will lead to the failure of character segmentation and recognition. Therefore, a plate rectification procedure is needed for compensating the inclined effect. Let
RP denote the inclined license plate. Like Fig. 6, let ( xc , yc ) denote as the center of RP and wR its width. In addition, let DR denote as the height difference between the centers of the first character and the last character in RP . Assume RP is the plate of RP after rectification. For each pixel (x, y) in RP , its intensity is compensated as follows: Rp ( x , y ) = R p ( x, y − ( x − x c )
DR ) WR
(1)
According to Eq.(1), the inclined effect of each extracted license plate can be reduced into minimum. Fig. 7 shows an example to demonstrate the effect of rectification, where (a) is the extracted inclined license plate and (b) is the result of (a) after rectification.
13
Fig. 7: Result of plate rectification. (a) the inclined license plate; (b) the rectified license plate. 5. Experimental Results
Fig. 8: Results of license plate detection at normal conditions. The first row shows the original test images. The second row shows the final results through detection, binarization, and rectification. The proposed method has been implemented on a personal computer with an INTEL Pentium III 600 CPU by the Microsoft Visual C++ 6.0. In order to analyze the performance of the proposed approach, 130 images are used for testing.
The size of each image is 640 × 480 .
For
increasing the complexity of the test database, the images are acquired at different lighting conditions including the time at a sunny day, a cloudy day, day time, night time, and so on. In addition, in order to examine the robustness and effectiveness of the proposed method, the test database is classified into seven kinds of data sets; that is, (a) normal images, (b) the images when colors between the license plate and background are similar, (c) inclined images, (d) images with smaller or larger license plates, (e) darker or lighter images, (f) images with noises, and (g) images with multiple license plates.
14
Fig. 9: Results of license plate detection when the colors between the license plate and background are similar. The first row shows the original test images. The second row shows the results after detection, binarization, and rectification. Fig. 8 is the results of cars acquired in normal viewing conditions. The original images are shown in the first row and the detection results after binarization are shown in the second row. Fig. 9 shows the detection results of cars when the colors between the license plates and their backgrounds are similar. In such case, edges between license plates and background are not clear.
It will lead to the failure of license plate detection for methods that considers the
boundary of a license plate as an important cue for detection.
However, in this case, our
morphology-based method still works well to detect all the wanted license plates. Fig. 10 shows the results when the license plates are inclined. The first row is the original images and the second row shows the results after morphological operations. The candidate region is drawn with the same intensities as the region in the original image. The third row shows the final results after rectification. In practice, the size of license plates will also affect the detection accuracy. Fig. 11 shows the detection results when the input image owns a larger license. Fig.12 shows the cases when the input image with a smaller license. In this case, many small edge and textures appear in these images. However, the proposed method still works well to detect the desired license plates. 15
Fig. 13 shows the detection results when the license plates are darker or lighter. In the first column, the input image has a darker license plate. Nevertheless, each character in this plate is still well detected and segmented with the proposed method. Fig. 14 shows the results when images have some noises. Fig. 15 shows the results when the input image includes two license plates. No matter what these cases are, all the desired license plates can be well segmented by the proposed methods.
Fig. 10: Results of license plate detection when these plates are inclined. The first row shows the original test images. The second row shows the results of the first row after morphological operations. The third row shows the final results after detection, binarization, and rectification.
16
Fig. 11: Results of license plate detection when these plates are larger. The first row shows the original test images. The second row shows the final results through detection, binarization, and rectification.
Fig.12 Results of license plate detection when smaller plates appear into the input images. The first row shows the original test images. The second row shows the results after morphological operations. The third row is the detection results after binarization and rectification.
17
Fig. 13 The detection results when the lighting on the license plate is too light or dark. The first and second columns are the results of darker license plates. The third column is the case when the license plate is too light.
Fig. 14 The detection results when noisy images are used. The first row shows the original test images. The second row shows the final results through detection, binarization, and rectification.
18
Fig. 15 The detection results when the image includes two license plates. The first row shows the original image and the result after using morphological operations. The second row shows the final results through detection, binarization, and rectification. In order to measure the accuracy of the proposed method, a “correct” detection of license plate is defined as “the located region includes all the registration number and has more than 95% overlapping within its corresponding license plate”. defined as
Then, the average accuracy of detection is
NC , where NC is the number of correct detection and NL is the number of license NL
plates appearing in the input images. Then, under the experimental database, only two examples got from 130 images are failed. In other words, the average accuracy of license plate detection is near 98%. The superiority of the proposed method can be verified through the preceding experimental results. 5. Conclusions In this paper, we have presented a novel approach for detecting the desired license plates from the cluttered images. At the first stage, several morphological operations were performed to extract important contrast features as guides to search the plate-analogue segments. Based on the geometries of registration numbers in a license plate, a recovery algorithm was then applied for grouping different fragments of a license plate into one for locating all the potential license plates.
Then, the potential license plate regions were verified according to the number of 19
characters appearing in the plate based on a clustering algorithm. The morphology-based plateanalogue segmentation process can effectively locate the potential license-plate regions even though an image has multiple license plates. Since only few candidates are extracted from the cluttered environments, the proposed method works much better and faster than other existed systems. In addition, the extracted contrast feature is robust to different camera views and lighting changes. Therefore, no matter how the captured environment is changed, all the desired license plates can be correctly located. The average accuracy of license plate detection is 98%. The experimental results have shown our method is superior in terms of detection efficiency, effectiveness, and stability. References 1.
T. Naito, “Robust recognition methods for inclined license plates under various illumination condition outdoors,” Proc. of IEEE/IEEJ/JSAI International Conference on Intelligent Transportation Systems, pp. 697–702,1999.
2.
C. Busch, R. Domer, C. Freytag, and H. Ziegler, “Feature based recognition of traffic video streams for online route tracing,” Proc. of IEEE Conference on Vehicular Technology, vol. 3, pp.1790-1794, 1998.
3.
R. Zunino and S. Rovetta, “Vector quantization for license-plate location and image coding,” IEEE Transactions on Industrial Electronics, vol. 47, pp.159–167, Feb. 2000.
4.
R. Plamondon and S. N. Srihari, “On-line and off-line handwriting recognition: a comprehensive survey, ” IEEE trans. on Pattern Analysis and Machine Intelligence, vol. 22, no. 1, pp.63-84, Jan., 2000.
5.
M. Shridhar, Miller, G. Houle, and L. Bijnagte, “Recognition of license plate images: issues and perspectives,” Proc. of the Fifth International Conference on Document Analysis and Recognition, p.17–20, 1999.
6.
K. K. Kim, K. I. Kim, J. B. Kim, and H. J. Kim, “Learning-based approach for license plate recognition,” Proc. of IEEE Workshop on Neural Networks for Signal Processing, vol. 2, pp.614–623, 2000.
7.
H.A. Hegt, R.J. Haye, and N.A. Khan, “A high performance license plate recognition system,” Proc. of IEEE International Conference on Systems, Man, and Cybernetics, pp. 4357–4362, 1999.
8.
Y. Mei and D.Y. Yong, “An approach to Korean license plate recognition based on vertical 20
edge matching,” Proc. of IEEE International Conference on Systems, Man, and Cybernetics, pp.2975–2980, 2000. 9.
D. S. Kim and S.I. Chien, “Automatic car license plate extraction using modified generalized symmetry transform and image warping,” Proc. of IEEE International Symposium on Industrial Electronics, vol. 3, pp. 2022–2027, 2001.
10. Y. Dai, H. Q. Ma, J. Liu, and L. Li, “A high performance license plate recognition system based on the web technique,” Proc. of IEEE Proceedings Intelligent Transportation Systems, pp.325–329, 2001. 11. R. Parisi, E. D. Claudio, G. Lucarelli, and G. Orlandi, “Car plate recognition by neural networks and image processing,” Proc. of IEEE International Symposium on Circuits and Systems, vol.3, pp. 195-198, 1998. 12. J.A.G. Nijhuis, et al., “Car license plate recognition with neural networks and fuzzy logic,” Proc. of IEEE International Conference on Neural Networks, vol. 5, pp. 2232-2236, 1995. 13. G.D. Lee, K. S. Kim, and D. S. Jeong, “Rough edge detection of low contrast images using consequential local variance maxim,” Proc. of IEEE Region 10 Conference on TENCON, vol. 1, pp. 734–737, 1999. 14. E. R. Lee; P. K. Kim; and H. J. Kim, “Automatic recognition of a car license plate using color image processing,” Proc. of IEEE International Conference on Image Processing, vol. 2, pp.301-305, 1994. 15. K. Yamaguchi, et al., “A method for identifying specific vehicles using template matching,” Proc. of IEEE/IEEJ/JSAI International Conference on Intelligent Transportation Systems, pp.8-143, 1999. 16. T. Naito, et al., “Robust license-plate recognition method for passing vehicles under outside environment,” IEEE Transactions on Vehicular Technology, vol. 49, pp.2309-2319, Nov. 2000. 17. L. Salgado, et al., “Automatic car plate detection and recognition through intelligent vision engineering,” Proc. of IEEE 33rd Annual 1999 International Carnahan Conference on Security Technology, pp.71–76, 1999. 18. S.H. Park, K. I. Kim, K. Jung, and H. J. Kim, “Locating car license plates using neural network,” Electronics Letters, vol. 35, pp.1475–1477, Aug. 1999. 19. T. Sirithinaphong and K. Chamnongthai, “Extraction of car license plate using motor vehicle 21
regulation and character pattern recognition,” Proc. of IEEE Asia-Pacific Conference on Circuits and Systems, pp.559 –562, 1998. 20. Da-Shan Gao and Jie Zhou, “Car license plates detection from complex scene,” Proc. of 5th International Conference on Signal Processing, vol. 2, pp.1409-1414, 2000. 21. Chin-Chuan Han, Hong-Yuan Mark Liao, Gwo-Jong Yu, and Liang-Hua Chen, “Fast face detection via morphology-based pre-processing,” Pattern Recognition, vol.33, no.10, pp.1701-1712, 2000. 22. M. Sonka, V. Hlavac, and R. Boyle, Image Processing, Analysis and Machine vision, Brooks/Cole Publishing Company, 1999. 23. P. K. Sahoo, et al., “A survey of thresholding techniques,” Computer Vision, Graphics, and Image Processing, vol.41, pp. 233-260, 1988. 24. N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Transactions on Systems, Man and Cybernetics, vol.9, pp.62-66, 1979.
22
FIGURE CAPTIONS Fig. 1:
Flowchart of the proposed system.
Fig. 2:
Two kinds of structure elements.
Fig. 3:
Details of the proposed method to extract useful features for license plate detection.
Fig. 4:
Results after applying the suggested morphological operations into different images.
Fig. 5:
Results after license plate recovery. (b) and (d) are the results of (a) and (c).
Fig. 6:
The geometry of an inclined plate.
Fig. 7:
Result of plate rectification. (a) the inclined license plate; (b) the rectified license plate.
Fig. 8:
Results of license plate detection at normal conditions. original test images.
The first row shows the
The second row shows the final results through detection,
binarization, and rectification. Fig. 9:
Results of license plate detection when the colors between the license plate and background are similar. The first row shows the original test images. The second row shows the results after detection, binarization, and rectification.
Fig. 10: Results of license plate detection when these plates are inclined. The first row shows the original test images. The second row shows the results of the first row after morphological operations. The third row shows the final results through detection, binarization, and rectification. Fig. 11:
Results of license plate detection when these plates are larger. The first row shows the original test images.
The second row shows the final results through detection,
binarization, and rectification. Fig. 12: Results of license plate detection when smaller plates appear into the input images. The first row shows the original test images. The second row shows the results after morphological operations. The third row is the detection results after binarization and rectification. Fig. 13: The detection results when the lighting on the license plate is too light or dark. The first and second columns are the results of darker license plates. The third column is the case when the license plate is too light.
23
Fig. 14: The detection results when noisy images are used. The first row shows the original test images. The second row shows the final results through detection, binarization, and rectification. Fig. 15: The detection results when the image includes two license plates. The first row shows the original image and the result after using morphological operations. The second row shows the final results through detection, binarization, and rectification.
TABLE CAPTION Table 1: Algorithm for analyzing the common properties of characters from a region R. Table 2: Algorithm for recovering the whole license plate from a region R.
24
