`
`
`
`
`Recognition of Vehicle License Plates from a
`Video Sequence
`
`I-Chen Tsai, Jui-Chen Wu, Jun-Wei Hsieh, and Yung-Sheng Chen
`
`
`
`Abstract—This paper proposes a robust system to
`recognize vehicle license plate by multi-frames learning. To
`fast locate the position of a license plate, we adopt a
`morphology-based method to extract important contrast
`features as filters to find all possible license plate
`candidates after calculating motion energy from video
`frames. The contrast feature is robust to lighting changes
`and invariant to different transformations like image
`scaling, translation, and skewing. Due to noise, many
`impossible license regions may be extracted. Hence, a
`Support Vector Machine (SVM) algorithm is adopted for
`verifying license plate regions. After locating license plate,
`the scheme of shape contexts is used to recognize the
`characters in license plate. To improve the correct rate of
`recognition, the verifying technique of multi-frames is
`further involved in our approach. Experimental results
`show that the proposed method
`is robust for the
`recognition of license plate.
`
`Index Terms—license plate recognition, morphology-based
`method, Support Vector Machine, shape contexts.
`
`
`I. INTRODUCTION
`With the rapid development of Intelligent Transportation
`Systems (ITS), license plate recognition (LPR) system has
`being broadly studied up to now. The applications of license
`plate recognition are widely used such as traffic volume, the
`monitoring of unsupervised park, traffic law enforcement, auto
`toll collation on highways, and so on.
`In the LPR system, many researches divided the topic into
`two parts, license plate locating and license plate recognition.
`In the license locating, some researches extracted the license
`plate based on its color [1], shape and gray level; the others
`looked for the region whose features are similar to characters in
`image directly. Duan et al. [2] proposed a method which
`combines Hough transform and contour to detect license plate
`
`
`
`Manuscript received August 20, 2008. This work was supported in part by
`the National Science Council, Taiwan, ROC, under the grant number
`NSC92-2213-E-155-052.
`I-Chen Tsai, Jui-Chen Wu and Jun-Wei Hsieh are with the Department of
`Electrical Engineering, Yuan Ze University, 135 Yuan-Tung Road, Chung-Li,
`Taoyuan 320, Taiwan, ROC.
`Yung-Sheng Chen is with the Department of Electrical Engineering, Yuan
`Ze University, 135 Yuan-Tung Road, Chung-Li, Taoyuan 320, Taiwan, ROC
`(corresponding author, phone: 886-3-4638800 ext 7113; fax: 886-3-4639355;
`e-mail: eeyschen@saturn.yzu.edu.tw).
`
`in static picture. They located the candidates by finding the
`contour in the edge space, and Hough transform is applied to
`filter the fake. However, it is possible that the edge of license
`plate may belong to an imperfect contour under the varied
`environments. In [3], Fujiyoshi et al. found the center of license
`plate in video by means of neural network; the input pattern
`position was random. After many iterations of learning (about
`10000 times), the position of center was decided. The accuracy
`was well but the time consuming was huge.
`In the character recognition, some well-known schemes
`were used such as artificial neural networks [4], fuzzy C-means
`and support vector machine. In [4], a contour tracking
`algorithm was proposed to detect the region of license plate in
`static picture, where an enhanced neural networks was
`designed to recognize the character. The algorithm built a
`hidden layer between input layer and output layer, in which
`similarity was defined by the ratio of the stored pattern and
`input pattern. Chang et al. [1] employed the HSI and color edge
`to find the license plate. The color space possesses a better
`linear independent than RGB. They built H, S, I and E (edge)
`maps and aggregated them by fuzzy operations. They finally
`combined optical character recognition (OCR) with neural
`network method to recognize characters. These methods cost
`much time to recognize character usually.
`As mentioned previously, many researches of license plate
`recognition in video sequence often use single image, which is
`captured by video camera, to achieve the result. Nevertheless,
`the video sequence having a series of variation and containing
`lots of information is worthy of considering since the rich of
`information in video can provide more detailed and powerful
`data being analyzed. In this study, a license plate recognition
`system using video sequence is proposed and simply previewed
`as follows. The morphological operation in [5] is first applied
`to locate candidate positions of a license plate. They are then
`verified by a support vector machine (SVM) method. In order
`to improve the character recognition, the method of shape
`contexts which can resist deformation is adopted since the
`quality of video sequence is usually not clear as the static image.
`Through a recognition process, the system will output a
`recognition result of vehicle license plate by learning technique
`of multi-frames. Fig. 1 shows the flowchart of the whole system.
`The system first uses motion energy calculation to estimate the
`appearance of a vehicle. Then a feature extraction method is
`used to locate license plate region. All possible license plates
`are verified by an SVM. After locating the license plate, the
`scheme of shape contexts is used to recognize the character in
`license plate. We finally refine each character by using a
`verifying technique of multi-frames.
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`AVS EXHIBIT 2005
`Toyota Inc. v. American Vehicular Sciences LLC
`IPR2013-00424
`
`

`

`
`
`
`
`
`No
`
`No
`
`Verify by SVM
`
`Load video image
`
`Start
`
`If motion energy
`> threshold
`
`Yes
`
`Feature
`Extractions
`
`Detection of
`License Plate
`Location
`
`Yes
`
`Results
`
`No
`
`If motion energy
`> threshold
`
`Verifying
`Function
`
`Character
`Recognition
`
`Character
`Segmentation
`
`Yes
`
`Find the License
`Region
`
`Fig. 1 Flowchart of the proposed LPR system.
`
`II. PROPOSED APPROACH
`This paper proposes an automatic system to recognize vehicle
`license plate by multi-frames learning. This approach includes
`mainly two parts: license plate detection and license plate
`recognition. The details of the approach are described in the
`following subsections.
`
`A. License Plate Detection
`Video sequence contains background and foreground
`information. In general, background information is more static
`than foreground and the human visual system is usually
`sensitive to the motion of object. In order to simulate the status,
`we extract the motion of object by frame subtraction. Then, we
`may be aimed at processing the part of motion. We use frame
`subtraction in HSI space and the energy of differenced frame is
`defined [6] as
`
`
`
`
`e
`
`=
`
`1N M
`
`−
`
`1
`−
`
`n
`
`=
`
`0
`
`m
`
`=
`
`0
`
`1
`× ∑ ∑
`N M
`
`(
`I n m
`,
`
`)
`
`2
`
`,
`
`
`
`(1)
`
`
`where n and m denote the row and column of differenced frame.
`N and M are the half-height and half-width of differenced
`frame.
`The appearance of vehicle can be estimated by the energy
`e as shown in Fig. 2. The peak of energy which occurs about
`the 400th-frame indicating the appearance of vehicle.
`
`
`Fig. 2 Illustration of motion energy.
`
`
`
`
`
`the
`the car motion, we apply
`After detecting
`morphological operations to extract the position of license plate.
`Because license plates often have high contrast characteristics
`especially in vertical direction, we adopt a morphology-based
`approach to detect license plate regions. The whole steps of the
`morphology-based extraction method are shown in Fig. 3.
`
`
`
`In order to eliminate noise, a smoothing operation with a
`7,7S
`structure element
` is first applied. Then, the closing and
`1,7S are performed
`opening operations with a structure element
`cI and
`oI
`into the smoothed image so that the output images
`can be obtained, respectively. For detecting license plate edges,
`cI and
`oI . To
`a differencing operation is further applied into
`make these edges more compactly and closely, a closing
`operation is used so that all characters embedded in a license
`plate will be connected as a single segment. After that, a
`thresholding operation is used for converting the analyzed
`image into a binary map. Then, a labeling process is executed
`to extract the license plate analogue segments, and thus a set of
`potential license plates can be obtained for further verification.
`
`
`Input
`Image
`
`Average
`(7,7)
`
`Closing
`(1,7)
`
`Opening
`(1,7)
`
`Differencing
`
`Closing
`(5,5)
`
`Feature
`Extraction
`
`Labeling
`
`Thresholding
`
`
`Fig. 3 Flowchart of the proposed method to extract contrast
`feature for license plate detection.
`
`
`
`In order to extract the possible position of a vehicle license
`plate, the connected component and labeling operator are
`applied. Furthermore, based on geometrical characteristics, the
`potential candidates’ positions can be finally detected as shown
`in Fig. 4.
`
`
`(a) (b)
`
`
`
`
`
`(c)
`Fig. 4 The process of morphological operation. (a) the
`differenced
`frame picture,
`(b)
`result after applying
`morphological operators, and (c) marking the region of
`license plate.
`
`
`
`For raising the possibility of the license plate’s region, we
`apply the classifier to distinguish the license plate from
`candidates. The classifier of SVM [7] has been displayed a
`great performance without needing to add a priori knowledge.
`Hence, we adopt the classifier of SVM to distinguish the
`license plate from candidates. The support vectors in our
`system are license plate patterns and non-license-plate patterns.
`The license plate patterns belong to class 1 and the
`non-license-plate patterns belong to class 2. Given a set of
`labeled training examples, let a feature space G including l
`patterns be represented by
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`

`

`
`
`
`
`
`x-projection technique. In practice, for characters embedded in
`a license plate, they should satisfy the following requirements:
`
`
`A1: their widths should be similar;
` A2: their heights should be similar;
` A3: the densities of all character are among the determined
`range; and
` A4: the ratios of all character must be greater than one.
`
` region satisfying the above properties is presumed that it is a
`character of license plate. In our study, the letters and numbers
`appearing in a license plate are ordered as shown Fig. 6.
`
`
` A
`
`
`Fig. 5 Example of training patterns.
`
`
`
`(a)
`
`
`
`
`
`(b)
`Fig. 6 (a) The letters and (b) numbers in a license plate.
`
`
`
`
`
`In this paper, we use the shape contexts for recognizing
`the license plate character. In previous method [10], the size of
`. In real case, however the size of
`test images is about 200 300
`×
`) may be too small to sample. The edge
`character (e.g. 7 18×
`points of character are few and they are not representative
`enough. For this reason, we use all points whose value is 1 in a
`character image to replace the edge points. Some characters
`possess a smaller width, it will influence the precise of
`recognition by shape contexts. To avoid this problem, the size
`of character is normalized to its height which is more stable
`usually.
`
`
`
`(a) (b) (c)
`
`
`
`
`
`
`
` (e)
`(d)
`Fig. 7 Recognition using shape contexts. (a) original image,
`(b) shape of original image, (c) shape of test image, (d) the
`feature of original image, and (e) the feature of test image.
`
`
`
`}
`{
`o
`o o
`of
`An object is represented by a set
`,......,
`=
`,
`Ω
`l
`1
`2
`shape points which are sampled from internal and external
`
`,
`
`
`
`(2)
`
`
`
`=
`
`
`
`}
`{
`n∈ix R stands for an input vector and
`iy ∈ + − is
`where
`1, 1
`the desired output. The key of support vector machine is the
`training set selection. The collection of training set will
`influence not only the function of hyper-plane, but also the
`accuracy of distinguishing. Generally, SVM has the following
`form to represent the optimal spearing hyper-plane:
`⎛
`⎞
`∑
`(
`)
`(
`)
`f x
`k x x
`y
`,
`α
`α
`⋅
`⎜
`⎟
`i
`i
`1
`⎝
`⎠
`support vectors
`However, a training set of support vector machine needs a large
`number of patterns. Since, the number of training patterns is m ,
`2m size. Thus, the
`the memory of training an SVM needs
`number of training patterns is generally over 5000. This leads
`to many researches to exploit how to reduce the time and the
`memory size of SVM [8].
`In our method, the morphological operation is applied to
`reduce the number of test patterns. The procedure to build a
`reliable SVM training set is shown as follows, where some
`training patterns are displayed in Fig. 5.
`
`G
`
`=
`
`{
`(
`
`x
`
`i
`
`,
`
`y
`i
`
`} 1
`)
`
`l
`i
`
`
`
`=
`
`sign
`
`−
`
`b
`
`.
`
`
`
`(3)
`
`i
`
`, where z is
`
`Procedure of support vector machine:
`1. Collect the initial patterns that include license plate
`patterns and non-license-plate patterns.
`{ } 1
`z
`
`=L p
`2. Regard all of patterns as a set
`i=
`the number of patterns.
`ip , and obtain a
`3. Calculate the value of hue in each
`, where m and n are
`{
`}
`support vector
`=p
`h h
`h ×
`,..,
`1,
`m n
`i
`0
`the height and the width of pattern, respectively.
`i ∈p L into the above Eq. (3). Then
`4. Transport all of
`find the hyper-plane between these two kinds of
`patterns.
`5. Classify the selected patterns which contain the
`random patterns and training patterns.
`6. Verify the results if most of them are correct? If the
`accuracy is low, iterate step 2-4 until the accuracy
`rate is steady and the training set of SVM is thus
`determined.
`7. According to Eq. (3), check which side of the
`hyper-plane the unknown pattern belonging to. If the
`computing value is greater than 1, the pattern belongs
`to license plate. Otherwise, the pattern should be a
`member of non-license-plate.
`
`
`B. License Plate Recognition
`Before performing recognition, license plate region is
`binarized first into a binary map using a threshold RT . RT is
`found using the “minimum within-group variance” dynamic
`thresholding method [9]. Given a license plate candidate R, the
`verification scheme takes its longest axis as the x-axis and then
`finds its several character geometrical properties using an
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`

`

`
`
`contours of the object. We assume that the shapes of original
`image (b) and test image (c) are given in Fig. 7, where (d) and (e)
`are the statistical graphs of (b) and (c) respectively in the
`log-polar coordinates. Each pixel in Fig. 7(d) and (e) indicates
`the bin of log-polar coordinates. The darker of the pixel
`indicates the more pixels inside this bin.
`io on the original shape, the corresponding
`For each point
`point it on the test shape is our target. Then, there are
`1l −
`vectors from a point on the shape to all the others, and the shape
`is represented by these vectors. The vectors contain much more
`information than the edge points, and the description of shape is
`more detailed. For building a histogram to match, the center
`point is chosen by one of the shape point and the remaining
`1l − points are computed in log-polar space.
`The log-polar mapping is a transformation from the points
`on Cartesian coordinate to the log-polar plane. The functions of
`transformation are shown as
`
`)
`(
`, and
`+
`)
`(
`y x
`.
`
`tan
`θ
`=
`Fig. 8 shows the example for log-polar space.
`
`
`
`r
`
`=
`
`log
`
`2
`
`y
`
`2
`
`x
`
`1
`−
`
`(4)
`
`(5)
`
`
`
`
`Fig. 8 The diagram of log-polar.
`
`
`
`The log-polar space is more sensitive to positions of points
`the shape
`than
`the perpendicular coordinate. The
`on
`information of the log-polar space contains not only the
`position but also the direction to other points.
`To estimate the similarly between two shape contexts
`which come from the same histogram distribution, the
` can be used. Let
`statistical method such as
`test
`2
`χ −
`(
`)
`(
`)
`S and
`S denote the histogram of two
`ah
`bh
`1,2,...,
`1,2,...,
`shape contexts and there are S bins of the histogram. Eq. (6)
`functions the similarity between the two shapes, which is
`ranged between 0 and 1,
`
`2
`2
`χ χ
`≡
`ab
`
`(
`
`h h
`,
`a
`b
`
`)
`
`=
`
`1
`2
`
`S
`
`∑
`
`s
`
`1
`=
`
`2
`
`⎡
`⎣
`
`( )
`( )
`h s
`h s
`−
`⎤
`⎦
`a
`b
`( )
`( )
`h s
`h s
`+
`a
`b
`
`.
`
` (6)
`
`
`According to Eq.(6), the template which contains the minimal
`2χ has the most similar to the test image.
`After performing the above methods, the character can be
`identified and recognized. However, the hypothesis we
`extracted may contain incomplete characters as shown in Fig. 9.
`To improve the correct rate of recognition, the verifying
`technique of multi-frames is further adopted.
`The procedure of verifying process is described as follows.
`First, we build a matrix to store the similarity of character. As
`
`
`
`the result of character recognition in each frame is obtained, we
`compare the similarity in this frame with the matrix and remain
`the smaller one. Here we establish a threshold of similarity for
`checking. All of the characters’ similarity should be less than
`the threshold.
`
`
`Fig. 9 The incomplete segmentation.
`
`
`
`(a)
`
`
`
` (b)
`
`
`(c) (d)
`Fig. 10 The process of verifying technique.
`
`
`
`
`
`
`
`As the illustration in Fig. 10(b), the last character is not
`extracted, and the third character is recognized incorrectly. In
`our method, we keep recognizing the following frames and
`rectifying the result of characters as Fig. 10(c) and (d) display.
`The recognition process is handled until the vehicle disappears
`in scene. The result obtained after verifying technique is much
`reliable because it is solved with unceasing checking.
`
`
`III. EXPERIMENTAL RESULTS
`The proposed system is implemented on MS-Window based
`PC with Pentium VI 2.8G CPU inside and the programming
`environment is under Borland C++ Builder 6.0. The video is
`shot by the digital video camera, and transmitted to the personal
`computer through a captured card.
`In order to analyze the performance of our proposed
`approach, four experiments are demonstrated as follows.
`
`Experiment 1: Fundamental Function
`
`
`
`
`Fig. 11 The motion energy for illustration.
`
`
`The first one shows the fundamental function in our vehicle
`license plate recognition system, where the input of video
`sequence contains one vehicle. According to the motion energy
`as shown in Fig. 11, the system detects the appearance of
`vehicle about the 150th frame. In this experiment, the video
`sequence is shown in Fig. 12 and the license plate number is
`displayed at the left-bottom of the screen.
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`

`

`
`
`we persist for tracking the recognition result by shape contexts
`to rectify the number of license plate. Fig. 13(e) shows the final
`result of the license plate. We record the verifying times for
`every sequence to get the final result of license plate’s number
`in Table 1. Observing Table 1, there are 28 of 41 video
`sequences that pass more than once verifying and achieve at the
`final result. In other words, more than half experiments need
`verifying function to get the final result; this confirms that there
`is the necessary of verifying function to improve the precise of
`system.
`
`
`
`
`
`(a) (b)
`
`
`(c) (d)
`
`(e)
`
`
`
` (f)
`
`
`(g) (h)
`
`
`(i) (j)
`
`
`(k) (l)
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`(m) (n)
`Fig. 12 A video sequence in Experiment 1.
`
`
`
`
`Experiment 2: Verifying Function
`In this experiment, we illustrate the result of verifying license
`plate character.
`The process of verifying function is exhibited in Fig. 13,
`where (a) the initial column is empty; (b)-(d) the numbers of
`license plate are recognized. During the period of car passing
`
`
`
`
`(a) (b)
`
`
`(c) (d)
`
`
`
`
`
`
`(e) (f)
`Fig. 13 A video sequence in Experiment 2.
`
`
`
`Table 1 Verifying times for each sequence.
`Verifying times used for
`Number of video
`obtaining the final result
`sequences
`1
`13
`2
`8
`3
`9
`4
`6
`5
`4
`6
`1
`Average
`2.34
`
`
`Experiment 3: Noise
`Sometimes, the video sequences contain some noises. There are
`many reasons which may cause this situation. For example, the
`shooting environments, the tapes of digital video camera, the
`man-made mistakes and the mechanism of digital video camera
`are probably made it.
`In this experiment, the noise appearing in the video
`sequences is made by the tape of digital video camera. The
`video sequences corrupt some random blemishes which are
`illustrated in Fig. 14. In this case, our proposed method
`demonstrates a good recognition.
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`

`

`
`
`
`
`
`
`Fig. 14 Illustration of a video sequence with noise.
`
`
`
`
`(i) (j)
`
`
`
`In another case, we illustrate a diagram of motion energy
`as shown in Fig. 15 and recognize the license plates in video
`sequence with noise as displayed in
`Fig. 16. Compared to the previous motion energy shown
`in Fig. 2, we notice that the curve of motion energy in Fig. 15 is
`more rigged. Because the frequency of noise is higher than
`signal, the energy of noise is also greater than to the signal.
`
`
`
`Fig. 15 The diagram of motion energy with noise effects.
`
`
`
`
`(a) (b)
`
`
`(c) (d)
`
`
`(e) (f)
`
`
`(g) (h)
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`(k)
`
`
`
` (l)
`
`
`(m) (n)
`
`
`(o) (p)
`
`
`(q) (r)
`
`
`(s) (t)
`
`
`(u) (v)
`
`
`Fig. 16 A video sequence with noise effects.
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`

`

`
`
`Experiment 4: The Recognition Rate of Our System
`We perform a series of video sequences in real environments to
`demonstrate the robustness of our system. To evaluate the
`performances of our scheme in vehicle license recognition
`system, the average accuracy of recognition is used in this
`paper. Recognition rate is the ratio of correctly recognized
`Num
`vehicle license plate
` to the total vehicle license
`Correct
`Num in database; that is,
`plates number
`Total
`
`
`Recognition rate =
`
`Num
`Total
`
`.
`
` (7)
`
`Num
`Correct
`
`The video sequences are selected overall from human eye. All
`of the results are obtained by multi-frames. The input of video
`sequences contains 41 license plates, and the recognition of
`license plate character is recorded.
`In 41 video sequences, there are 38 license plates in which
`all characters are recognized correctly, the precision rate is
`92.68%. The result is shown in Table 2. There are 246
`characters which are figured out, and only four characters are
`recognized incorrectly. The accuracy of our system to
`recognize characters is 98.3%. The statistic analysis is shown in
`Fig. 17.
`
`
`Table 2 The accuracy of license plate recognized.
`The number of video sequences
`41
`Characters are recognized correctly
`38
`Accuracy
`92.68 %
`
`
`
`
`
`Fig. 17 The statistic results of character recognized.
`
`Based on our experiments, the main contributions of this
`approach can be summarized as follows.
`1. Some
`traditional methods proposed a vertical
`edge-matching algorithm for grouping all possible
`positions of license plates through edge matching.
`They assumed the vertical boundaries between a
`license plate and its backgrounds are contrasted
`significantly. However, as long as the colors of the
`license plates are similar to their backgrounds, such
`an assumption will no longer exist. Hence a
`morphology-based method was designed in our
`approach for extracting high contrast areas as license
`plate candidates. This feasible feature is invariant to
`different changes like lighting, rotation, translation,
`and complicated backgrounds.
`2. The template-based method was often adopted for
`character recognition. However, the quality of video
`sequence is not usually clear like that of a static
`
`
`
`picture. To improve the recognition of character, the
`method of shape contexts was therefore involved in
`our approach to resist the deformation.
`3. Many researches of license plate recognition in
`video sequence still use the single picture captured
`from a video camera to achieve the result. For the
`sequence contains a lot of information, we proposed
`the verifying technique of multi-frames to enhance
`the correct rate of recognition.
`As a result, the proposed method has good abilities to recognize
`license plates even under cluttered background.
`
`
`IV. CONCLUSIONS
`In this paper, we have presented an approach to recognize the
`license plate according to multi-frames learning. There are two
`schemes used usually to detect license plate in video. The first
`is for the single image trigged when a car passes, and the second
`is to find the license plate from video sequence captured at the
`stop toll collection. In our approach, the input sequences are
`shot without car stopping.
`Besides, our system applies the SVM on the strength of
`high accuracy to locate the position of license plate, and uses
`the shape contexts which can resist the skew and deformed
`situations to recognize characters of a license plate. Our
`experiments show that it is very robust and high accuracy to
`recognize the characters. However, the disadvantages of SVM
`are of huge data and poor speed, we overcome it by using
`morphological operations.
`In the near future, we will consider another classifier
`promising speed and accuracy to make our system more useful.
`Moreover, the current system can only recognize license plates,
`by combining the database and hardware it is possible to build a
`toll collection system for example, for real
`complete
`application.
`
`REFERENCES
`[1] S. L. Chang, L. S. Chen, Y. C. Chung, and S. W. Chen, “Automatic
`IEEE Transactions on
`Intelligent
`License Plate Recognition,”
`Transportation Systems, Vol. 5, No. 1, pp. 42-53, March 2004.
`[2] T. D. Duan, D. A. Duc, and T. L. H. Du, “Combining Hough Transform
`and Contour Algorithm for detecting Vehicles’ License-Plates,”
`Proceedings of International Symposium on Intelligent Multimedia,
`Video and Speech Processing, pp.747-750, October 2004.
`[3] H. Fujiyoshi, T. Umezaki, and T. Imamura, “Area Extraction of License
`Plates Using an Artificial Neural Network,” Systems and Computers in
`Japan, Vol. 29, No. 11, pp.55-64, 1998.
`[4] K. B. Kim, S. W. Jang, and C. K. Kim, “Recognition of Car License Plate
`by Using Dynamical Threshlod Method and Enhanced Neural
`Networks,” Proceedings Lecture Notes in Computer Science, pp.
`309-319, 2003.
`J. W. Hsieh, S. H. Yu, and Y. S. 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.
`[6] Y. J. Wang and B. Z. Yuan, “Robust Face Tracking by Motion Energy
`and Genetic Algorithms,” IEEE ICSP, Vol. 1, pp.695-698, August 2002.
`[7] V. Vapnik, The Nature of Statistical Learning Theory. New York,
`Spring-Verlag, 1995.
`[8] K. M. Lin and C. J. Lin, “A Study on Reduced Support Vector
`Machines,” IEEE Transactions on Neural Networks, Vol. 14, No. 6,
`November 2003.
`[9] M. Sonka, V. Hlavac, and R. Boyle, Image Processing, Analysis and
`Machine Vision, London, U. K., Chapman & Hall, 1993.
`
`[5]
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`

`

`
`
`Jun-Wei Hsieh received the Bachelor’s degree in
`Computer Science from TongHai University, Taiwan,
`in 1990, and received the Ph. D. degree in Computer
`Engineering from the National Central University,
`Taiwan, in 1995. He got the Phai-Tao-Phai award
`when he graduated. From 1996 to 2000, he was a
`Researcher Fellow at the Industrial Technology
`Researcher Institute, Hsinchu, Taiwan, and managed
`a team to develop video-related technologies. He is
`presently an Associate Professor at the Department of
`Electrical Engineering, Yuan Ze University of
`Taiwan. He received the Best Paper Awards from the
`Image Processing and Pattern Recognition society of Taiwan in 2005, 2006,
`and 2007, respectively. His research interests include content-based multimedia
`databases, video indexing and retrieval, computer vision, and pattern
`recognition.
`
`
`
`
`Yung-Sheng Chen was born in Taiwan, R.O.C., on
`June 30, 1961. He received the B.S. degree from
`Chung Yuan Christian University, Chung-Li, Taiwan,
`in 1983 and the M.S. and Ph.D. degrees from
`National Tsing Hua University, Hsinchu, Taiwan, in
`1985, and 1989, respectively, all in electrical
`engineering.
`In 1991, he joined the Electrical Engineering
`Department, Yuan Ze Institute of Technology,
`Chung-Li, where he is now a Professor. His research
`interests include human visual perception, computer vision and graphics, circuit
`design, and website design.
`Dr. Chen received the Best Paper Award from the Chinese Institute of
`Engineers in 1989 and an Outstanding Teaching Award from Yuan Ze
`University in 2005. He has been listed in the Who's Who of the World since
`1998 and awarded with The Millennium Medal from The Who's Who Institute
`in 2001. He is a member of the IEEE, and the IPPR of Taiwan, R.O.C.
`
`
`I-Chen Tsai was born in Taiwan, R.O.C., on August
`17, 1981. She received the B.S. degree and the M.S.
`degree from Yuan Ze University, Chung-Li, Taiwan,
`in 2003 and 2005, respectively, both in electrical
`engineering. She is currently working as a firmware
`engineer in Gueishan industrial park, Taoyuan,
`Taiwan, R.O.C.
`Her research interests include image processing
`and pattern recognition
`
`
`
`Jui-Chen Wu was born in Taiwan, R.O.C., on
`November 22, 1976. She received the B.S. degree
`from Chung Yuan Christian University and the M.S.
`degree from Yuan Ze University, Chung-Li, Taiwan,
`in 2000 and 2002, respectively. She is currently
`working toward the Ph.D. degree in the Department
`of Electrical Engineering, Yuan Ze University,
`Chung-Li, Taiwan, R.O.C.
`Her research interests include image processing
`and pattern recognition.
`
`
`
`
`[10] S. Belongie, J. Malik, J. Puzicha, “Shape Matching and Object
`Recognition Using Shape Contexts,” IEEE Transaction on Pattern
`Analysis and Machine Intelligence, Vol. 24, No. 24, pp. 509-522, April
`2002.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`IAENG International Journal of Computer Science, 36:1, IJCS_36_1_04
`______________________________________________________________________________________
`
`(Advance online publication: 17 February 2009)
`
`