Original images Adjusted images

#### **2.2 Segmentation character**

This part extracts individual character images from the plate image. The plate window region can contain license surrounding some region that can create resistance in character recognition segment. In order to better extracte characters of plate image. We have to remove this unexpected region so that the image only holds the license number. This can be done by a horizontal segmentation and a vertical segmentation on both sides of the number plate. After segmenting horizontally and vertically, the plate image will be as the Fig. 2(c).

86 Recent Advances in Document Recognition and Understanding

function's discontinuity (Fernandez-Garcia, & Medina-Carnicer 2004). Our method can have potential applications in video retrieval, and in other related areas of video information

This paper is organized as follows. First, Section II introduces methods for image preprocessing. Section III presents the orthogonal Gaussian-Hermite moments and their behaviors in the license plate character image. In Section IV, proposes the GHMs features as the input vector of BP neural network for recognizing characters. Section V shows some experiment results. Finally, conclusions are drawn and with the future work discussed as

The skew correction of license plate is an important step in ALPR. The license plate inclination is determined by the direction of the boundary. In order to find such direction. In our paper (Ma et al., 2009), the license plate image is firstly divided into a set of 55 nonoverlapping blocks. The local orientation of each block is estimated by gradients [Gx,Gy] of pixels in the block. It may reduce much processing time. Next, the direction histogram which can reveal the overall orientation information in the license plate image is counted. The skew angle of license plate is detected by the local maximum of the direction histogram. This approach can solve the direction detection problem in a very straightforward and robust way under various conditions. Fig. 1 gives some images before and after skew

Original images Adjusted images

This part extracts individual character images from the plate image. The plate window region can contain license surrounding some region that can create resistance in character recognition segment. In order to better extracte characters of plate image. We have to remove this unexpected region so that the image only holds the license number. This can be done by a horizontal segmentation and a vertical segmentation on both sides of the number plate. After segmenting horizontally and vertically, the plate image will be as the Fig. 2(c).

processing.

well.

correction.

**2. Image preprocessing** 

Fig. 1. Corrected license plates

**2.2 Segmentation character** 

**2.1 Orientation method for skew correction** 

Fig. 2. Frame removal: (a) Original image; (b) The horizontal cut lines after corrected; (c) Frame removed

Then, calculating for segmentation points by the vertical projection and merging fragments that belong to the same character. The average filter with length s=3 reduce noise. The characters will be extracted from the vertical projection histogram of plate image. The extracted characters are given in Fig. 3.

#### **2.3 Adaptive threshold for image binarization**

All the character images are binarized using an anto-adaptive threshold. We propose many iterations algorithm for obtaining the optimal thresholds for segmenting gray scale images. The image background is black (gray value is 0) and the characters are white (gray value is 255).

Given that F(x,y) is the edge image and T is a predefined threshold Tn is a dynamic threshold. The following equation is used to obtain a local optimal threshold value.

$$\text{IT} = \frac{1}{2} \{ \min \{ F(\mathbf{x}, y) \} + \max \{ F(\mathbf{x}, y) \} \} \tag{1}$$

$$\text{NT}\_n = \frac{1}{2} \left[ \frac{1}{M} \sum\_{F(\mathbf{x}, y)} \sum\_{\leq T} F(\mathbf{x}, y) + \frac{1}{N} \sum\_{F(\mathbf{x}, y)} \sum\_{>T} F(\mathbf{x}, y) \right] \tag{2}$$

where M is number of pixel with its gray value less than T , N is number of pixel with its gray value larger than T.

$$\begin{array}{c} \textbf{Loop}: \\ \text{if } (T \approx T\_n) \\ \text{end loop;} \\ \text{else if } |T \cdot T\_n| \prec 0.6, \\ T = T\_n; \text{ goto loop;} \\ \text{end if} \end{array}$$

If the intensity of every pixel value is greater than T, the pixel is set to white; otherwise, it is set to black. The binary image is given in Fig. 3.

#### **2.4 Normalization**

Characters segmented from different car plates have different sizes. A linear normalization algorithm is applied to the input image to adjust to a uniform size.In our implementation, character blocks are normalized to a fixed size of 32\*16 pixels.Assume the horizontal and vertical projections of the original image F be H and V, respectively. The normalization position (m,n) of (i,j) is obtained by

Application of Gaussian-Hermite Moments in License 89

 

Now we analyze the spatial domain behavior of smoothed GHMs base functions. Because the nth order Hermite polynomial *H t <sup>n</sup>* has n different real roots, the base function of

the nth order GHMs will change its sign n times. In other words, it consists of n oscillations. So GHMs can well characterize different spatial modes as other orthogonal moments. Fig. 4

As to the frequency domain behavior, since Gaussian-Hermite base functions comprise more and more oscillations when the order n is increased, they will thus contain more and more high frequencies. From the spectral analysis viewpoint, the GHMs efficiently separate the signal features in different frequency bands. Fig. 5 shows the Fourier transform

shows the spatial behavior of GHMs base functions of different orders.

Fig. 4. Spatial behavior of 1D GHMs base functions (orders: 0 to 3)

 

 

<sup>0</sup> , , \*

 <sup>1</sup> 2 1 , 21 , 2 , for 0 and 2 *m mm M xS x n M xS x M xS x n nn m n* (10)

2 2 1 exp exp *<sup>n</sup> n n H t <sup>n</sup> t d dt t* (9)

for 0 *m m M xS x g x S x m* (11)

<sup>1</sup> , 2[ ,] \* *m m <sup>M</sup> xS x d <sup>g</sup> x dx S x* (12)

<sup>0</sup> *M xS x* , ,\* *g x Sx* (13)

*mm m S x d S x dx* (15)

<sup>0</sup> *S x Sx* (16)

<sup>1</sup> *M xS x d* , 2[ , \* ] *g x S x dx* (14)

will also have n different real roots. Therefore the base function of

 

The GHMs can be recursively calculated as follows [10]:

with

Where

and in particular,

and \* denotes the convolution operator.

amplitude of some base functions of GHMs.

GHMs *gx H x* (, ) ( / ) *n*

$$m = \sum\_{k=1}^{l} H(k) \times \frac{M}{\sum\_{k=1}^{l} H(k)}\tag{3}$$

$$m = \sum\_{k=1}^{j} V(k) \times \frac{N}{\sum\_{k=1}^{j} V(k)}\tag{4}$$

where m, n is height and width of normalized image.

Fig. 3. Character images before (upper) and after(lower) size normalization

#### **3. Gaussian-Hermite moments (GHMs) and their behaviors in plate character**

Moments, such as geometric moments and orthogonal moments, are widely used in pattern recognition, image processing, computer vision and multiresolution analysis. However, some moments base functions exhibit a great discontinuity [6] at the window boundary. In order to better represent local characteristics of images, particularly for noisy images, one should use orthogonal moments with a smoothing window function. Taking the wellknown Gaussian functions as smoothing kernel, smoothed orthogonal Gaussian-Hermite moments (GHMs) were proposed [7-10]. Moreover, the base functions of GHMs are much more smoothed than other moments, thus less sensitive to noise and avoid the artifacts introduced by window function's discontinuity.

#### **3.1 Gaussian-Hermite Moments**

GHMs were proposed by J. Shen[11-12]. Given the Gaussian smoothing function *g x*(, ) with

$$g(\mathbf{x}, \sigma) = (2\pi\sigma^2)^{-1/2} \exp(-\mathbf{x}^2 \mid \mathbf{2}\sigma^2) \tag{5}$$

the nth order smoothed GHMs (, ) *Mn xS x* of a signal *S x*( ) is defined as

$$M\_n(\mathbf{x}, S(\mathbf{x})) = \int\_{-\alpha}^{\alpha} B\_n(t) S(\mathbf{x} + t) dt \qquad n = 0, 1, 2, \dots \tag{6}$$

With

$$B\_u\left(t\right) = \mathcal{g}\left(t, \sigma\right) P\_u\left(t\right) \tag{7}$$

Where *P t <sup>n</sup>* is a scaled Hermite polynomial function of order n defined by

$$P\_n\left(t\right) = H\_n\left(t/\sigma\right) \tag{8}$$

With

$$H\_u\left(t/\sigma\right) = \left(-1\right)^u \exp\left(t^2\right) \left(d^u \Big/ 4t^u\right) \exp\left(-t^2\right) \tag{9}$$

The GHMs can be recursively calculated as follows [10]:

$$M\_n\left(\mathbf{x}, S^{(m)}\{\mathbf{x}\}\right) = \mathbf{2}\{n-1\}M\_{n-2}\left(\mathbf{x}, S^{(m)}\{\mathbf{x}\}\right) + \mathbf{2}\sigma M\_{n-1}\left(\mathbf{x}, S^{(m+1)}\{\mathbf{x}\}\right) \quad \text{for } m \ge 0 \quad \text{and } n \ge 2 \tag{10}$$

with

88 Recent Advances in Document Recognition and Understanding

 

 

**3. Gaussian-Hermite moments (GHMs) and their behaviors in plate character**  Moments, such as geometric moments and orthogonal moments, are widely used in pattern recognition, image processing, computer vision and multiresolution analysis. However, some moments base functions exhibit a great discontinuity [6] at the window boundary. In order to better represent local characteristics of images, particularly for noisy images, one should use orthogonal moments with a smoothing window function. Taking the wellknown Gaussian functions as smoothing kernel, smoothed orthogonal Gaussian-Hermite moments (GHMs) were proposed [7-10]. Moreover, the base functions of GHMs are much more smoothed than other moments, thus less sensitive to noise and avoid the artifacts

GHMs were proposed by J. Shen[11-12]. Given the Gaussian smoothing function *g x*(, )

2 22 1 2 *g x*( , ) (2 ) exp( / 2 ) *x* (5)

(, ) 0,1,2,... *M x S x B t S x t dt n n n* (6)

*B t n n g t Pt* (7)

*Pt H t n n* (8)

,

the nth order smoothed GHMs (, ) *Mn xS x* of a signal *S x*( ) is defined as

Where *P t <sup>n</sup>* is a scaled Hermite polynomial function of order n defined by

*<sup>k</sup> <sup>k</sup> <sup>N</sup> n Vk*

<sup>1</sup> <sup>1</sup> ( )

*<sup>k</sup> <sup>k</sup> <sup>M</sup> m Hk*

*i*

*j*

Fig. 3. Character images before (upper) and after(lower) size normalization

where m, n is height and width of normalized image.

introduced by window function's discontinuity.

**3.1 Gaussian-Hermite Moments** 

with

With

With

Before normalization

After normalization <sup>1</sup> <sup>1</sup> ( )

*j*

*j*

( )

( )

*V k*

(3)

(4)

*H k*

$$M\_0\left(\mathbf{x}, \mathbf{S}^{(w)}\left(\mathbf{x}\right)\right) = \mathbf{g}\left(\mathbf{x}, \sigma\right)^\star \mathbf{S}^{(w)}\left(\mathbf{x}\right) \qquad \text{for } m \ge 0 \tag{11}$$

$$M\_1\left(\mathbf{x}, S^{(w)}\left(\mathbf{x}\right)\right) = 2\sigma d[\mathcal{J}\left(\mathbf{x}, \sigma\right)] \big/ d\mathbf{x} \triangleq S^{(w)}\left(\mathbf{x}\right) \tag{12}$$

and in particular,

$$M\_0\left(\mathbf{x}, \mathbf{S}(\mathbf{x})\right) = \mathbf{g}\left(\mathbf{x}, \sigma\right) \* \mathbf{S}(\mathbf{x}) \tag{13}$$

$$M\_1(\mathbf{x}, \mathbf{S}(\mathbf{x})) = 2\sigma d \| \mathbf{g}(\mathbf{x}, \sigma)^\star \mathbf{S}(\mathbf{x}) \| / d\mathbf{x} \tag{14}$$

Where

$$\mathcal{S}^{\rm un}(\mathbf{x}) = d^{\rm vn} S(\mathbf{x}) \Big/ d\mathbf{x}^{\rm un} \tag{15}$$

$$\mathbf{S}^0(\mathbf{x}) = \mathbf{S}(\mathbf{x})\tag{16}$$

and \* denotes the convolution operator.

Now we analyze the spatial domain behavior of smoothed GHMs base functions. Because the nth order Hermite polynomial *H t <sup>n</sup>* has n different real roots, the base function of GHMs *gx H x* (, ) ( / ) *n* will also have n different real roots. Therefore the base function of the nth order GHMs will change its sign n times. In other words, it consists of n oscillations. So GHMs can well characterize different spatial modes as other orthogonal moments. Fig. 4 shows the spatial behavior of GHMs base functions of different orders.

As to the frequency domain behavior, since Gaussian-Hermite base functions comprise more and more oscillations when the order n is increased, they will thus contain more and more high frequencies. From the spectral analysis viewpoint, the GHMs efficiently separate the signal features in different frequency bands. Fig. 5 shows the Fourier transform amplitude of some base functions of GHMs.

Fig. 4. Spatial behavior of 1D GHMs base functions (orders: 0 to 3)

Application of Gaussian-Hermite Moments in License 91

In order to characterize these features in the license plate character image, we use 231 2D base functions of GHMs of different order(orders: 20) *p q* . The number and the order of

The feature vector with Gaussian-Hermite moments *Mp q*, of license plate character is

For the recognition of segmented characters, the Gaussian-Hermite moments features are extracted from each character are lumped into a vector as input of the BP neural network

A neural network (NN) is an artificial network model, which emulates the cerebral nerve network in the brain. Characters are recognized using a supervised back propagation neural network (BPNN) classifier (see Fig.7). A BPNN is trained by adjusting the weights of the connections between the nodes of the different layers before the BPNN can be used. The input patterns are fed into the input layer and the error between the expected output and the actual output are propagated backwards through the network such that the weights can be adjusted to minimize the errors. This training procedure is repeated until the error is sufficiently small. When learning of neural network complete, we can used that for recognize character of license plate. A major advantage of BPNN is that a trained network is capable of classifying unknown pattern with little computational effort. In this paper, we use a three-layered BPNN architecture. Fig. 7 shows the three-layered BP neural network

The Gaussian-Hermite moments feature obtained from license plate character as the input vector of BPNN. We use GHMs feature extraction as our method for character recognition. Since the base functions of GHMs are much more smoothed and thus less sensitive to noise, GHMs could facilitate the recognition of character in noisy image sequences. This method

The license plate character images is characterized by a vector of Gaussian-Hermite moment Mp+q (in our experiment, = 0.6). The order of used moments are of 0th-20th (N = 20). The character images (including references and noisy images) is then characterized by 231

The number of units in BPNN is shown in Table 1. The three-layered BPNN that contains 231 input units, 120 hidden units. Learning rate was 0.05, and the number of learning cycles was 5000, an error value less than 1.0\*10-7, then learning stopped. It has 34 types of characters, including 24 character and 10 numbers, the maximum sample number for

2D moments of orders (0,0),(0,1),…,(0.20), (1,0), (1,1) ,…, (1,19),…,(20,0).

0,0 0,1 0,2 0,20 1,1 1,19 19,0 19,1 20,0 [ , , , ... , , , ... , , ... , , , ] *MMMM M M M M M M <sup>T</sup>* (19)

**3.2 Representation of feature vector with of GHMs in license plate character** 

GHMs required were empirically determined.

[13,14]. The Bp neural network is a three-layer structure.

represented by

**4.1 BPNN model** 

architecture.

**4.2 Input of BPNN with GHMs features** 

training and testing is 200. The number of units in BPNN

distinguishes characters by their unique features.

**4. Character recognition** 

Fig. 5. Frequency behavior of 1D GHMs base functions (orders: 0 to 3)

Moreover, from the recursive calculation of GHMs, we see that these moments are in fact linear combinations of the derivatives of the signal filtered by a Gaussian filter. As is well known, the derivatives have been extensively used for image representation in pattern recognition.

2D orthogonal Gaussian-Hermite moments of order (p,q) of an input image *I x*(,) *y* can be defined similarly

$$M\_{p,q} = \left[ \int\_{-\infty}^{\infty} \mathbf{G}\left(t, \upsilon, \sigma\right) \mathbf{H}\_{p,q}\left(t \;/\ \sigma, \upsilon \;/\ \sigma\right) \mathbf{S}\left(\mathbf{x} + t, y + t\right) dt d\upsilon \tag{17}$$

Where *Gtv* (, , ) is the 2D Gaussian function, and *Ht v p q*. /,/ , the scaled 2D Hermite polynomial of order (p,q) , with

$$H\_{p,q}\left(t\mid\sigma,\upsilon\mid\sigma\right) = H\_p\left(t\middle|\sigma\right)H\_q\left(v\middle|\sigma\right) \tag{18}$$

Obviously, 2D Gaussian-Hermite moments are separable, so the recursive algorithm in 1D cases can be applied for their calculation. Fig. 6 shows the Fourier transform amplitude of a bidimensional GHMs kernels of different orders. We use GHMs to efficiently recognition the character plate image.

Fig. 6. Frequency behavior of 2D base functions of GHMs (orders: (0,1), (1, 0), (0, 3) and (3, 0))

#### **3.2 Representation of feature vector with of GHMs in license plate character**

In order to characterize these features in the license plate character image, we use 231 2D base functions of GHMs of different order(orders: 20) *p q* . The number and the order of GHMs required were empirically determined.

The feature vector with Gaussian-Hermite moments *Mp q*, of license plate character is represented by

$$M = \begin{bmatrix} M\_{0,0}, M\_{0,1}, M\_{0,2}, \dots, M\_{0, \mathfrak{D} \prime} & M\_{1,1 \prime} \dots, M\_{1, 1 \mathfrak{D} \prime} & \dots & M\_{19,0 \prime} M\_{19,1 \prime} & M\_{\mathfrak{D} \mathfrak{D},0} \end{bmatrix}^{\intercal} \tag{19}$$

#### **4. Character recognition**

For the recognition of segmented characters, the Gaussian-Hermite moments features are extracted from each character are lumped into a vector as input of the BP neural network [13,14]. The Bp neural network is a three-layer structure.

#### **4.1 BPNN model**

90 Recent Advances in Document Recognition and Understanding

Moreover, from the recursive calculation of GHMs, we see that these moments are in fact linear combinations of the derivatives of the signal filtered by a Gaussian filter. As is well known, the derivatives have been extensively used for image representation in pattern

2D orthogonal Gaussian-Hermite moments of order (p,q) of an input image *I x*(,) *y* can be

, . ,, / , / , *Mp q Gtv H t v Sx t p q <sup>y</sup> t dtdv* (17)

*H t v Ht Hv p q*. /,/ *<sup>p</sup> <sup>q</sup>* (18)

 *Ht v p q*. /,/ , the scaled 2D Hermite

 

Fig. 5. Frequency behavior of 1D GHMs base functions (orders: 0 to 3)

is the 2D Gaussian function, and

 

Obviously, 2D Gaussian-Hermite moments are separable, so the recursive algorithm in 1D cases can be applied for their calculation. Fig. 6 shows the Fourier transform amplitude of a bidimensional GHMs kernels of different orders. We use GHMs to efficiently recognition

 Fig. 6. Frequency behavior of 2D base functions of GHMs (orders: (0,1), (1, 0), (0, 3) and (3, 0))

recognition.

defined similarly

Where *Gtv* (, , )

the character plate image.

polynomial of order (p,q) , with

A neural network (NN) is an artificial network model, which emulates the cerebral nerve network in the brain. Characters are recognized using a supervised back propagation neural network (BPNN) classifier (see Fig.7). A BPNN is trained by adjusting the weights of the connections between the nodes of the different layers before the BPNN can be used. The input patterns are fed into the input layer and the error between the expected output and the actual output are propagated backwards through the network such that the weights can be adjusted to minimize the errors. This training procedure is repeated until the error is sufficiently small. When learning of neural network complete, we can used that for recognize character of license plate. A major advantage of BPNN is that a trained network is capable of classifying unknown pattern with little computational effort. In this paper, we use a three-layered BPNN architecture. Fig. 7 shows the three-layered BP neural network architecture.

#### **4.2 Input of BPNN with GHMs features**

The Gaussian-Hermite moments feature obtained from license plate character as the input vector of BPNN. We use GHMs feature extraction as our method for character recognition. Since the base functions of GHMs are much more smoothed and thus less sensitive to noise, GHMs could facilitate the recognition of character in noisy image sequences. This method distinguishes characters by their unique features.

The license plate character images is characterized by a vector of Gaussian-Hermite moment Mp+q (in our experiment, = 0.6). The order of used moments are of 0th-20th (N = 20).

The character images (including references and noisy images) is then characterized by 231 2D moments of orders (0,0),(0,1),…,(0.20), (1,0), (1,1) ,…, (1,19),…,(20,0).

The number of units in BPNN is shown in Table 1. The three-layered BPNN that contains 231 input units, 120 hidden units. Learning rate was 0.05, and the number of learning cycles was 5000, an error value less than 1.0\*10-7, then learning stopped. It has 34 types of characters, including 24 character and 10 numbers, the maximum sample number for training and testing is 200.

The number of units in BPNN

Application of Gaussian-Hermite Moments in License 93

of training samples would improve recognition accuracy. As a whole, these results are

(Pan et al., 2005) 86.7 89.2 87.4 GHMs feature 98.6 97.4 97.8

Although significant progress has been made in the last decade, there is still work to be done, as a robust LP recognition system should effectively work for a variety of environmental illumination, plate types/conditions, as well as acquisition parameters. Moreover, most LPR systems focus on the processing of images with one vehicle.

In addition, assuming that LP regions are detectable even in very low resolution, an open topic for future research is the readability improvement of LP text using image processing techniques. Research for improving degraded plates has lately been directed to superresolution methods for video sequences or to blurred plate images with promising

Though the new method proposed in this paper is still in its stage of a prototype, it has already shown its potential for various implementations. This method can also be used for

The work was supported by the National Natural Science Foundation of China (No. 60965001) and The Guizhou Key Laboratory of Pattern Recognition and Intelligent System.

Jia, W.; Zhang, H. & He, X. (2007). Region-based license plate detection. *Journal of Network and Computer Applications*, Vol. 30, 2007, pp. 1324-1333, ISSN: 1084-8045 Rosenfeld, A. (1969). Picture Processing by Computer. *Academic Press*, New York, 1969,

Huang, Y.P.; Chang, T.W. Chen, Y.R & Sandnes, F.E. (2008). A back propagation based real-

Shen, J. (1997). Orthogonal Gaussian-Hermite Moments for Image Characterization. *Proc.* 

*Artificial Intelligence*, vol. 22, no. 2, pp.233-251, 2008, ISSN: 0218-0014 Christos-Nikolaos, E.; Loannis, E. & Loannis, D. (2008). License Plate Recogniton From Still

*Transportation Systems*, Vol.9, no.3, 2008. ISSN: 1524-9050

time license plate recognition system. *International Journal of Pattern Recognition and* 

Images and Video Sequences: A Survey. *IEEE Transaction on Intelligent* 

*Intelligent Robots and Computer Vision XVI: Algorithms Techniques, Active Vision, and* 

Nevertheless, input images may contain more than one vehicle or motorcycles.

detecting the moving objects. Our research will be carried on following this track.

Character Uppercase Number

satisfactory enough for the character recognition process.

(%)

Correctness Rate

proposed feature

Input of BPNN

**6. Conclusion and future work** 

Table 2. Experimental results

results.

**7. Acknowledgment** 

ISBN: ISBN: 0-12-597350-0

**8. References** 


Table 1.

We use Network to denote the number and character network. According to the feature of Network, 6 neurons are set. So the output vector is <sup>123456</sup> *O oooooo* [ , , , , , ,], the expectation output of network is {0,0,0,0,0,0}, {0,0,0,0,0,1}, {0,0,0,0,1,0}, {0,0,0,0,1,1},…, {0,1,1,1,1,1}, {1,0,0,0,0,0}, {1,0,0,0,0,1} corresponds number and character separately.

The number of hidden neurons relates to the input and output unit directly. If the hidden neurons are too few, the network possibly cannot train well, local minimum are more and no robust, it could not distinguish the sample which had not seen before, and the fault tolerance is bad. The increase hidden neurons maybe improve the match precision between the network and the training set, but it will causes the study time too long again, the error is also uncertain the best. The choice of hidden neurons is given according to the empirical formula usually:

$$m = \sqrt{n\_1 n\_0} + \delta \left(1 \le \delta \le 10\right) \tag{20}$$

n1, n0 are the number of inputs and the output respectively, so the neural network structure is shown as Fig.7.

Fig. 7. Three layers BPNN Structure

#### **5. Experimental result**

In this section, preliminary experiments for testing the feasibility and robustness of our method are conducted by applying the above-mentioned procedure. One is license plate images which are directly captured by a mobile surveillance system. There are 200 training and testing sample images to be processed using our algorithm. Table 2 shows the results using this set of images.

The testing results using the images in the first category are very encouraging. The average recognition accuracy is 97.93%. Among 200 testing images, only few very poorly focused samples have failed for the number of character training samples is not many. The increase


of training samples would improve recognition accuracy. As a whole, these results are satisfactory enough for the character recognition process.

Table 2. Experimental results

92 Recent Advances in Document Recognition and Understanding

We use Network to denote the number and character network. According to the feature of Network, 6 neurons are set. So the output vector is <sup>123456</sup> *O oooooo* [ , , , , , ,], the expectation output of network is {0,0,0,0,0,0}, {0,0,0,0,0,1}, {0,0,0,0,1,0}, {0,0,0,0,1,1},…, {0,1,1,1,1,1},

The number of hidden neurons relates to the input and output unit directly. If the hidden neurons are too few, the network possibly cannot train well, local minimum are more and no robust, it could not distinguish the sample which had not seen before, and the fault tolerance is bad. The increase hidden neurons maybe improve the match precision between the network and the training set, but it will causes the study time too long again, the error is also uncertain the best. The choice of hidden neurons is given according to the empirical

> 

n1, n0 are the number of inputs and the output respectively, so the neural network structure

**...**

**...**

In this section, preliminary experiments for testing the feasibility and robustness of our method are conducted by applying the above-mentioned procedure. One is license plate images which are directly captured by a mobile surveillance system. There are 200 training and testing sample images to be processed using our algorithm. Table 2 shows the results

Input layer Hidden layer Output layer

The testing results using the images in the first category are very encouraging. The average recognition accuracy is 97.93%. Among 200 testing images, only few very poorly focused samples have failed for the number of character training samples is not many. The increase

 

1 0 *n nn* 1 10 (20)

*O* 

{1,0,0,0,0,0}, {1,0,0,0,0,1} corresponds number and character separately.

**..**

*M*

Input layer 231 Hidden layer 120 Output layer 6

Table 1.

formula usually:

is shown as Fig.7.

Fig. 7. Three layers BPNN Structure

**5. Experimental result** 

using this set of images.

#### **6. Conclusion and future work**

Although significant progress has been made in the last decade, there is still work to be done, as a robust LP recognition system should effectively work for a variety of environmental illumination, plate types/conditions, as well as acquisition parameters. Moreover, most LPR systems focus on the processing of images with one vehicle. Nevertheless, input images may contain more than one vehicle or motorcycles.

In addition, assuming that LP regions are detectable even in very low resolution, an open topic for future research is the readability improvement of LP text using image processing techniques. Research for improving degraded plates has lately been directed to superresolution methods for video sequences or to blurred plate images with promising results.

Though the new method proposed in this paper is still in its stage of a prototype, it has already shown its potential for various implementations. This method can also be used for detecting the moving objects. Our research will be carried on following this track.

#### **7. Acknowledgment**

The work was supported by the National Natural Science Foundation of China (No. 60965001) and The Guizhou Key Laboratory of Pattern Recognition and Intelligent System.

#### **8. References**


*Materials Handling*, pp.224-233, ISBN: 0-8194-2640-7, Pittsburgh, USA, 15-17, Oct. 1997


94 Recent Advances in Document Recognition and Understanding

Shen, J.; Shen, W. & Shen, D. (2000). On Geometric and orthogonal moments. *International* 

Wu, Y. & Shen, J. (2004). Moving object detection using orthogonal Gaussian-Hermite

Wang, L.; Dai, M. & Geng, G.H. (2004). Fingerprint Image Segmentation by Energy of

Comput.Sci., pp.414-423, 2004. ISBN: 978-3-540-24029-7, Springer-Verlag. Wang, L. & Dai, M. (2007). Application of New Type of Singular Points in Fingerprint

Fernandez-Garcia, N.L. & Medina-Carnicer, R. (2004). Characterization of empirical

Ma, X.; Pan, R. & Wang, L. (2009). A Method Based on Orientation Field for Skew Correction

Pan, X.; Ye, X. & Zhang S.Y. (2005). A hybrid method for robust car plate character

1997

ISSN: 0218-0014

ISBN: 9780819452115

47, ISSN:0167-8655

972, ISSN: 1062-922X

Nov. 2009

*Materials Handling*, pp.224-233, ISBN: 0-8194-2640-7, Pittsburgh, USA, 15-17, Oct.

*Journal of Pattern Recognition and Artificial Intelligence*, Vol.14 (7), 2000, pp.875-894,

moments. *Visual communication and image processing*, Vol.5308, 2004, pp.841-849,

Gaussian-Hermite Moments. *Sinobiometrics 2004*, vol. 3338 of Lecture Notes

Classification. *Pattern Recognition Letter*, vol.28,2007, pp.1640-1650, ISSN:0167-8655

discrepancy evaluation measures, *Pattern Recognition Letters*, Vol.25(1), 2004, pp.35-

of License Plate. *2009 Asia-Pacific Conference on Computational Intelligence and Industrial Applications*, pp. 308-311, ISBN: 978-1-4244-4606-3, Wuhan, China, 28-29

recognition. *Engineering Applications of Artificial Intelligence*, Vol.18, 2005, pp. 963-

### *Edited by Minoru Mori*

In the field of document recognition and understanding, whereas scanned paper documents were previously the only recognition target, various new media such as camera-captured documents, videos, and natural scene images have recently started to attract attention because of the growth of the Internet/WWW and the rapid adoption of low-priced digital cameras/videos. The keys to the breakthrough include character detection from complex backgrounds, discrimination of characters from noncharacters, modern or ancient unique font recognition, fast retrieval technique from large-scaled scanned documents, multi-lingual OCR, and unconstrained handwriting recognition. This book aims to present recent advances, applications, and new ideas that are relevant to document recognition and understanding, from technical topics such as image processing, feature extraction or classification, to new applications like camera-based recognition or character-based natural scene analysis. The goal of this book is to provide a new trend and a reference source for academic research and for professionals working in the document recognition and understanding field

Photo by Evgen\_Prozhyrko / iStock

Recent Advances in Document Recognition and Understanding

Recent Advances in

Document Recognition and

Understanding

*Edited by Minoru Mori*