CN105447457A - License plate character identification method based on adaptive characteristic - Google Patents

Abstract

The invention provides a license plate character identification method based on an adaptive characteristic. The method comprises the following steps of constructing a multi-scale spatial convolution template; carrying out normalization on a license plate character image size to be identified; acquiring a multi-scale spatial license plate character characteristic set; acquiring an adaptive license plate character characteristic vector; and identifying a license plate character image. In the method, based on a human visual identification character principle, through analyzing convolution images of different scale spaces and different convolution template sizes, an optimal characteristic of the character is adaptively acquired. The optimal characteristic comprises a global characteristic and a local characteristic. The license plate character can be accurately identified, robustness is high and a requirement to character image quality is low.

Description

A kind of license plate character recognition method based on self-adaptive features

Technical field

The present invention relates to technical field of image processing, specifically a kind of license plate character recognition method based on self-adaptive features.

Background technology

Recognition of License Plate Characters is the final step of Vehicle License Plate Recognition System, and be also a vital step, it directly affects recognition accuracy and the efficiency of whole system.For the identification of desirable characters on license plate, existing a lot of ripe method, can reach very high accuracy rate both at home and abroad.And the license plate image gathered in actual environment, character often have that resolution is lower, part shoals or lack, edge fog, the feature such as character inclination, make accurate identification character become very difficult, based on existing character identifying method, satisfied accuracy rate cannot be reached.Therefore, accurately identify characters on license plate, become the difficult point of domestic and international Vehicle License Plate Recognition System.

Recognition of License Plate Characters adopts three kinds of recognition methodss usually, is the method based on template matches, the method based on neural network, method based on support vector machine (SVM) respectively.Known according to pattern recognition theory, affecting the biggest factor of recognition accuracy, is not which kind of sorting technique selected, but the target signature selected, as long as character feature selection is suitable, above-mentioned three kinds of methods all can reach good classifying quality.Therefore, how to select character feature, become the key determining Recognition of License Plate Characters success or failure.

At present, conventional characters on license plate feature mainly contains following a few class:

(1) global characteristics, this category feature adopts global change to obtain the global feature of character, use orderly global feature or subset feature to carry out constitutive characteristic vector, common feature has GABOR transform characteristics, moment characteristics, projection properties, stroke density feature, HARR feature, HOG feature etc.The advantage of these features is insensitive to localized variation, and antijamming capability is strong; Its shortcoming easily ignores some important local feature, cannot distinguish similar character.

(2) local feature, this category feature is in multiple regional areas of character, calculate corresponding feature, use the orderly local feature of series connection to form final proper vector, principal character comprises local gray level histogram feature, LBP feature, threading feature, SIFT feature etc.The advantage of this category feature is that to distinguish the ability of character strong; Its shortcoming too pays close attention to the local feature of character, and often fault discrimination has the character of noise.

Summary of the invention

The object of the invention is to the shortcoming for using global characteristics or local feature recognition characters on license plate, a kind of license plate character recognition method based on self-adaptive features is provided, by paying close attention to the character convolved image under different scale space, obtain the optimal characteristics of character adaptively, improve the accuracy rate of character recognition.

Technical scheme of the present invention is:

Based on a license plate character recognition method for self-adaptive features, comprise the following steps:

(1) multiscale space convolution mask is built;

(2) by characters on license plate picture size to be identified normalization;

(3) the multiscale space convolution mask built is utilized to carry out convolution algorithm to the characters on license plate image to be identified after normalization, the assemblage characteristic collection of characters on license plate convolved image under acquisition different scale space, to connect the assemblage characteristic collection of characters on license plate convolved image under all metric spaces, obtain multiscale space characters on license plate feature set;

(4) extraction process is carried out to the multiscale space characters on license plate feature set obtained, be obtained from and adapt to characters on license plate proper vector;

(5) the self-adaptation characters on license plate proper vector input Bayes classifier will obtained, the characters on license plate proper vector calculating input belongs to the probability of a certain classification, selects classification that most probable value is corresponding as the character class in characters on license plate image to be identified.

The described license plate character recognition method based on self-adaptive features, in step (2), described by characters on license plate picture size to be identified normalization, realized by following bilinear interpolation formula:

$\left\{\begin{array}{c}f\left(x,y\right)=s_{11}*f\left(x_1,y_1\right)+s_{21}*f\left(x_2,y_1\right)+s_{12}*f\left(x_1,y_2\right)+s_{22}*f\left(x_2,y_2\right)\\ s_{11}=\frac{\left(x_2-x\right)\left(y_2-y\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{21}\frac{\left(x-x_1\right)\left(y_2-y\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{12}=\frac{\left(x_2-x\right)\left(y-y_1\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{22}=\frac{\left(x-x_1\right)\left(y-y_1\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\end{array}\right.$

Wherein, (x, y) represents the pixel coordinate of gray-scale value f (x, y) to be asked, (x ₁, y ₁), (x ₂, y ₁), (x ₁, y ₂), (x ₂, y ₂) represent four known ash angle value f (x that distance pixel coordinate (x, y) is nearest respectively ₁, y ₁), f (x ₂, y ₁), f (x ₁, y ₂), f (x ₂, y ₂) pixel coordinate.

The described license plate character recognition method based on self-adaptive features, in step (3), the assemblage characteristic collection of characters on license plate convolved image under described acquisition different scale space, specifically comprises the following steps:

(31) certain metric space is chosen as current scale space;

(32) following formula is adopted, the characters on license plate convolved image of selected convolution mask size under obtaining current scale space:

$g(x,y)=\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}f(x_i+dx,y_j+dy)*G(x_i,y_j)$

$G\left(x_i,y_i\right)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{\frac{{(x_i-\frac{m}{2})}^2+{(y_i-\frac{n}{2})}^2}{2{\mathrm{\σ}}^2}}$

Wherein, m represents the width of convolution mask, and n represents the height of convolution mask, and σ represents Gaussian distribution standard deviation, G (x _i, y _j) represent (x on convolution mask _i, y _j) weighted value corresponding to place's pixel, the position (x, y) on the characters on license plate image to be identified after normalization of the center that dx, dy represent convolution mask relative to the side-play amount in the image upper left corner, f (x _i+ dx, y _j+ dy) represent (x on the characters on license plate image to be identified after normalization _i+ dx, y _j+ dy) gray-scale value of place's pixel, g (x, y) represents the gray-scale value of (x, y) place pixel on characters on license plate convolved image;

(33) the characters on license plate convolved image of the selected convolution mask size obtained is divided into a series of overlapped subregion block;

(34) feature set of every sub regions block is obtained;

(35) to connect the feature set of all subregion blocks, obtain the feature set of the characters on license plate convolved image of selected convolution mask size;

(36) convert convolution mask size, repeat step (32) ~ (35), until the feature set of the characters on license plate convolved image of all convolution mask sizes under obtaining current scale space;

(37) to connect the feature set of the characters on license plate convolved image of all convolution mask sizes under current scale space, obtain the assemblage characteristic collection of characters on license plate convolved image under current scale space;

(38) change of scale space, repeats step (31) ~ (37), until obtain the assemblage characteristic collection of characters on license plate convolved image under different scale space.

The described license plate character recognition method based on self-adaptive features, in step (4), the described multiscale space characters on license plate feature set to obtaining carries out extraction process, specifically adopts classical linear discriminant analysis method, reduces the dimension of described multiscale space characters on license plate feature set.

The described license plate character recognition method based on self-adaptive features, in step (5), the described characters on license plate proper vector calculating input belongs to the probability of a certain classification, is realized by following formula:

$pi\left(X\right)=\frac{n_i}{2\mathrm{\π}N{\left|Cov\right|}^{0.5}}{e}^{\[-0.5*\left(X-U\right){\mathrm{Cov}}^{-1}{\left(X-U\right)}^T\]}$

Wherein, X represents the characters on license plate proper vector of input, n _irepresent the training sample number of the i-th class characters on license plate, N represents the number of all training samples, and U represents the corresponding average of each dimension element of a certain class characters on license plate proper vector, and Cov represents the covariance matrix of each dimension element of a certain class characters on license plate proper vector.

The described license plate character recognition method based on self-adaptive features, in step (33), the described characters on license plate convolved image by the selected convolution mask size obtained is divided into a series of overlapped subregion block, specifically comprises the following steps:

(331) size dimension of selected square area, is that half size dimension slides on described characters on license plate convolved image according to step-length, obtains the subregion block that this size dimension is corresponding;

(332) adjust the size dimension of square area, repeat step (331).

The described license plate character recognition method based on self-adaptive features, in step (34), the feature set of the every sub regions block of described acquisition, specifically comprises the following steps:

(341) adopt following formula, obtain the grey level histogram of subregion block:

$hist\left(i\right)=\left\{\begin{array}{cc}hist\left(i\right)+1& 10*\left(i-1\right)\&le;g\left(x,y\right)<10*i\\ hist\left(i\right)& 10*i\&le;g\left(x,y\right)org\left(x,y\right)<10*\left(i-1\right)\end{array}\right.$

Wherein, hist (i) represents the i-th dimension grey level histogram of subregion block, and i is integer and 1≤i≤26, and g (x, y) represents the gray-scale value of (x, y) place pixel in subregion block.

(342) following formula is adopted, the grey level histogram normalization by the subregion block of acquisition:

$\left\{\begin{array}{c}{\mathrm{hist}}^{\&prime;}\left(i\right)=\frac{hist\left(i\right)}{rs}\\ rs=\frac{1}{N}\underset{I=1}{\overset{N}{\mathrm{\&Sigma;}}}hist\left(i\right)\end{array}\right.$

Wherein, N represents the dimension of subregion block grey level histogram, and hist ' (i) represents the grey level histogram of the subregion block after normalization;

(343) using the feature set of the grey level histogram after normalization as subregion block.

Beneficial effect of the present invention is:

As shown from the above technical solution, the present invention is based on the principle of human vision identification character, by analyzing the convolved image of different scale space, different convolution mask size, obtain the optimal characteristics of character adaptively, this optimal characteristics contains global characteristics and local feature, more accurately can identify characters on license plate, robustness is higher, simultaneously lower for the quality requirements of character picture.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is the process flow diagram of the character feature collection under acquisition single scale space;

Convolution mask weighted value distribution plan when Fig. 3 is Gaussian distribution standard deviation is 1.2, template size is 5*5;

Character convolved image when Fig. 4 is Gaussian distribution standard deviation is 1.2, template size is 5*5, Fig. 4 (a) for former figure, Fig. 4 (b) be image after convolution;

Fig. 5 divides character convolved image subregion block schematic diagram, and wherein, b0, b1, b2, b3, b4, c0, c1, c2 represent a sub regions block respectively.

Embodiment

The present invention is further illustrated below in conjunction with the drawings and specific embodiments.

As shown in Figure 1, a kind of license plate character recognition method based on self-adaptive features, comprises the following steps:

S1, structure multiscale space convolution mask, the computing of Gaussian convolution template is the effective ways realizing multiscale space conversion, dimensional Gaussian convolution mask function, as shown in formula (1), adopts different Gaussian distribution standard deviations, can obtain the convolution mask in different scale space.

$G\left(x,y\right)=\frac{1}{2{\mathrm{\&pi;\&sigma;}}^2}{e}^{\frac{{\left(x-\frac{m}{2}\right)}^2+{\left(y-\frac{n}{2}\right)}^2}{2{\mathrm{\&sigma;}}^2}}---\left(1\right)$

Wherein, (x, y) represent the coordinate of certain pixel on convolution mask, G (x, y) represents (x, y) weighted value that place's pixel is corresponding, m represents the width of convolution mask, and n represents the height of convolution mask, and σ represents Gaussian distribution standard deviation, σ value is larger, and the image space yardstick of acquisition is larger.

Convolution mask weighted value distributed effect when Gaussian distribution standard deviation is 1.2, convolution mask is of a size of 5*5 as shown in Figure 3.

S2, the normalization of character picture size, mainly according to bilinear interpolation theoretical formula (2), the character picture of different size, transform to identical size, eliminates the impact that size difference is brought.

$\left\{\begin{array}{c}f\left(x,y\right)=s_{11}*f\left(x_1,y_1\right)+s_{21}*f\left(x_2,y_1\right)+s_{12}*f\left(x_1,y_2\right)+s_{22}*f\left(x_2,y_2\right)\\ s_{11}=\frac{\left(x_2-x\right)\left(y_2-y\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{21}\frac{\left(x-x_1\right)\left(y_2-y\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{12}=\frac{\left(x_2-x\right)\left(y-y_1\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{22}=\frac{\left(x-x_1\right)\left(y-y_1\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\end{array}\right.---\left(2\right)$

Wherein, (x, y) represents the pixel coordinate of gray-scale value f (x, y) to be asked, (x ₁, y ₁), (x ₂, y ₁), (x ₁, y ₂), (x ₂, y ₂) represent four known ash angle value f (x that distance pixel coordinate (x, y) is nearest respectively ₁, y ₁), f (x ₂, y ₁), f (x ₁, y ₂), f (x ₂, pixel coordinate y).

S3, acquisition multiscale space character feature collection, mainly obtain under different scale space, the assemblage characteristic collection of character convolved image, concrete steps are as follows:

Character feature collection under S31, acquisition single scale space, mainly for some metric spaces, obtain the assemblage characteristic collection of character convolved image, as shown in Figure 2, concrete steps are as follows:

The character convolved image of S311, the selected convolution mask size of acquisition, mainly for specific metric space, the Gaussian convolution template of selected some convolution mask sizes, according to convolution algorithm formula (3), obtain corresponding character convolved image, effect as shown in Figure 4:

$g(x,y)=\underset{i=1}{\overset{m}{\&Sigma;}}\underset{j=1}{\overset{n}{\&Sigma;}}f(x_i+dx,y_j+dy)*G(x_i,y_j)---\left(3\right)$

Wherein, G (x _i, y _j) represent (x on convolution mask _i, y _j) weighted value corresponding to place's pixel, m represents the width of convolution mask, and n represents the height of convolution mask, and dx, dy represent the side-play amount of the position (x, y) of the center of convolution mask on character picture relative to the character picture upper left corner, f (x _i+ dx, y _j+ dy) represent (x on character picture _i+ dx, y _j+ dy) gray-scale value of place's pixel, g (x, y) represents the gray-scale value of (x, y) place pixel on character convolved image.

S312, division character convolved image subregion block, mainly character convolved image is divided into a series of overlapped zonule block, concrete steps are as follows:

The size dimension of S3121, selected square area, as shown in Figure 5, is that half size dimension slides according to step-length, obtains corresponding character convolved image subregion block;

The size dimension of S3122, adjustment square area, repeated execution of steps S3121, obtains all possible character convolved image subregion block.

The feature set of S313, acquisition character convolved image subregion block, concrete steps are as follows:

S3131, for each subregion block, obtain corresponding subregion block grey level histogram according to formula (4):

$hist\left(i\right)=\left\{\begin{array}{cc}hist\left(i\right)+1& 10*\left(i-1\right)\&le;g\left(x,y\right)<10*i\\ hist\left(i\right)& 10*i\&le;g\left(x,y\right)org\left(x,y\right)<10*\left(i-1\right)\end{array}\right.---\left(4\right)$

Wherein, hist (i) represents the i-th dimension grey level histogram of subregion block, and i is integer and 1≤i≤26, and g (x, y) represents the gray-scale value of (x, y) place pixel in subregion block.

S3132, according to formula (5), carry out the normalization of subregion block grey level histogram, eliminate the impact of local anomaly gray-scale value:

$\left\{\begin{array}{c}{\mathrm{hist}}^{\&prime;}\left(i\right)=\frac{hist\left(i\right)}{rs}\\ rs=\frac{1}{N}\underset{I=1}{\overset{N}{\mathrm{\&Sigma;}}}hist\left(i\right)\end{array}\right.---\left(5\right)$

Wherein, N represents the dimension of subregion block grey level histogram, and hist ' (i) represents the feature set of the subregion block grey level histogram after normalization and subregion block.

The feature set of S314, the character convolved image obtained under selected convolution mask size, method is the feature set of all subregion blocks of connecting.

S315, whether process whole convolution mask size, if so, enter step S316, otherwise conversion convolution mask size, continues to perform step S311 to step S315.

The assemblage characteristic collection of character convolved image under S316, acquisition single scale space, method is the feature set of the character convolved image under all convolution mask sizes of series connection;

S317, according to formula (6), the assemblage characteristic collection of character convolved image under single scale space is selected in normalization, eliminates the impact that light inequality is brought:

${\mathrm{allhist}}^{\&prime;}\left(i\right)=\frac{allhist\left(i\right)}{\max \left\{allhist\left(i\right)\right\}}---\left(6\right)$

Wherein, allhist (i) represents grey level histogram and the assemblage characteristic collection of character convolved image under selected single scale space, max{} represents and gets maximal value, and allhist ' (i) represents the grey level histogram of character convolved image under the selected single scale space after normalization.

S32, change of scale space, repeat step S31, the assemblage characteristic collection of character convolved image under acquisition different scale space.

Under S33, all metric spaces of connecting, the assemblage characteristic collection of character convolved image, namely obtains multiscale space character feature collection.

S4, be obtained from adapt to character feature vector, mainly step S3 obtain multiscale space character feature collection basis on, automatically choose the feature with strong distinction, as the final proper vector of character picture.Concrete grammar adopts classical linear discriminant analysis method, reduces the dimension of multiscale space character feature collection, obtains the character feature vector with stronger distinction.

S5, identification character image, mainly the self-adaptation character feature vector input Bayes classifier obtained, the character feature vector obtaining input according to formula (7) belongs to the probability of a certain classification, selects classification that most probable value is corresponding as the classification of current character.

$pi\left(X\right)=\frac{n_i}{2\mathrm{\&pi;}N{\left|Cov\right|}^{0.5}}{e}^{\&lsqb;-0.5*\left(X-U\right){\mathrm{Cou}}^{-1}{\left(X-U\right)}^T\&rsqb;}---\left(7\right)$

Wherein, X represents the character feature vector of input, n _irepresent the training sample number of the i-th class character, N represents the number of all training samples, U, Cov are through Bayes classifier training and obtain, represent corresponding average and the covariance matrix of each dimension element of a certain class character feature vector respectively, training Bayes classifier carried out before whole identifying.

The above embodiment is only be described the preferred embodiment of the present invention; not scope of the present invention is limited; under not departing from the present invention and designing the prerequisite of spirit; the various distortion that those of ordinary skill in the art make technical scheme of the present invention and improvement, all should fall in protection domain that claims of the present invention determine.

Claims (7)

1. based on a license plate character recognition method for self-adaptive features, it is characterized in that, comprise the following steps:

(1) multiscale space convolution mask is built;

(2) by characters on license plate picture size to be identified normalization;

(3) the multiscale space convolution mask built is utilized to carry out convolution algorithm to the characters on license plate image to be identified after normalization, the assemblage characteristic collection of characters on license plate convolved image under acquisition different scale space, to connect the assemblage characteristic collection of characters on license plate convolved image under all metric spaces, obtain multiscale space characters on license plate feature set;

(4) extraction process is carried out to the multiscale space characters on license plate feature set obtained, be obtained from and adapt to characters on license plate proper vector;

(5) the self-adaptation characters on license plate proper vector input Bayes classifier will obtained, the characters on license plate proper vector calculating input belongs to the probability of a certain classification, selects classification that most probable value is corresponding as the character class in characters on license plate image to be identified.

2. the license plate character recognition method based on self-adaptive features according to claim 1, is characterized in that, in step (2), described by characters on license plate picture size to be identified normalization, is realized by following bilinear interpolation formula:

$\left\{\begin{array}{c}f\left(x,y\right)=s_{11}*f\left(x_1,y_1\right)+s_{21}*f\left(x_2,y_1\right)+s_{12}*f\left(x_1,y_2\right)+s_{22}*f\left(x_2,y_2\right)\\ s_{11}=\frac{\left(x_2-x\right)\left(y_2-y\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{21}=\frac{\left(x-x_1\right)\left(y_2-y\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{12}=\frac{\left(x_2-x\right)\left(y-y_1\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\\ s_{22}=\frac{\left(x-x_1\right)\left(y-y_1\right)}{\left(x_2-x_1\right)\left(y_2-y_1\right)}\end{array}\right.$

Wherein, (x, y) represents the pixel coordinate of gray-scale value f (x, y) to be asked, (x ₁, y ₁), (x ₂, y ₁), (x ₁, y ₂), (x ₂, y ₂) represent four known ash angle value f (x that distance pixel coordinate (x, y) is nearest respectively ₁, y ₁), f (x ₂, y ₁), f (x ₁, y ₂), f (x ₂, y ₂) pixel coordinate.

3. the license plate character recognition method based on self-adaptive features according to claim 1, is characterized in that, in step (3), the assemblage characteristic collection of characters on license plate convolved image under described acquisition different scale space, specifically comprises the following steps:

(31) certain metric space is chosen as current scale space;

(32) following formula is adopted, the characters on license plate convolved image of selected convolution mask size under obtaining current scale space:

$g(x,y)=\underset{i=1}{\overset{m}{\&Sigma;}}\underset{j=1}{\overset{n}{\&Sigma;}}f\left(x_i+dx,y_j+dy\right)*G\left(x_i,y_j\right)$

$G\left(x_i,y_i\right)=\frac{1}{2{\mathrm{\&pi;\&sigma;}}^2}{e}^{\frac{{\left(x_i-\frac{m}{2}\right)}^2+{\left(y_i-\frac{n}{2}\right)}^2}{2{\mathrm{\&sigma;}}^2}}$

Wherein, m represents the width of convolution mask, and n represents the height of convolution mask, and σ represents Gaussian distribution standard deviation, G (x _i, y _j) represent (x on convolution mask _i, y _j) weighted value corresponding to place's pixel, the position (x, y) on the characters on license plate image to be identified after normalization of the center that dx, dy represent convolution mask relative to the side-play amount in the image upper left corner, f (x _i+ dx, y _j+ dy) represent (x on the characters on license plate image to be identified after normalization _i+ dx, y _j+ dy) gray-scale value of place's pixel, g (x, y) represents the gray-scale value of (x, y) place pixel on characters on license plate convolved image;

(33) the characters on license plate convolved image of the selected convolution mask size obtained is divided into a series of overlapped subregion block;

(34) feature set of every sub regions block is obtained;

(35) to connect the feature set of all subregion blocks, obtain the feature set of the characters on license plate convolved image of selected convolution mask size;

(36) convert convolution mask size, repeat step (32) ~ (35), until the feature set of the characters on license plate convolved image of all convolution mask sizes under obtaining current scale space;

(37) to connect the feature set of the characters on license plate convolved image of all convolution mask sizes under current scale space, obtain the assemblage characteristic collection of characters on license plate convolved image under current scale space;

(38) change of scale space, repeats step (31) ~ (37), until obtain the assemblage characteristic collection of characters on license plate convolved image under different scale space.

4. the license plate character recognition method based on self-adaptive features according to claim 1, it is characterized in that, in step (4), the described multiscale space characters on license plate feature set to obtaining carries out extraction process, specifically adopt classical linear discriminant analysis method, reduce the dimension of described multiscale space characters on license plate feature set.

5. the license plate character recognition method based on self-adaptive features according to claim 1, is characterized in that, in step (5), the described characters on license plate proper vector calculating input belongs to the probability of a certain classification, is realized by following formula:

$pi\left(X\right)=\frac{n_i}{2\mathrm{\&pi;}N|Cov{|}^{0.5}}{e}^{\&lsqb;-0.5*\left(X-U\right){\mathrm{Cov}}^{-1}{\left(X-U\right)}^T\&rsqb;}$

Wherein, X represents the characters on license plate proper vector of input, n _irepresent the training sample number of the i-th class characters on license plate, N represents the number of all training samples, and U represents the corresponding average of each dimension element of a certain class characters on license plate proper vector, and Cov represents the covariance matrix of each dimension element of a certain class characters on license plate proper vector.

6. the license plate character recognition method based on self-adaptive features according to claim 3, it is characterized in that, in step (33), the described characters on license plate convolved image by the selected convolution mask size obtained is divided into a series of overlapped subregion block, specifically comprises the following steps:

(331) size dimension of selected square area, is that half size dimension slides on described characters on license plate convolved image according to step-length, obtains the subregion block that this size dimension is corresponding;

(332) adjust the size dimension of square area, repeat step (331).

7. the license plate character recognition method based on self-adaptive features according to claim 3, is characterized in that, in step (34), the feature set of the every sub regions block of described acquisition, specifically comprises the following steps:

(341) adopt following formula, obtain the grey level histogram of subregion block:

$hist\left(i\right)=\left\{\begin{array}{cc}hist\left(i\right)+1& 10*\left(i-1\right)\&le;g\left(x,y\right)<10*i\\ hist\left(i\right)& 10*i\&le;g\left(x,y\right)org\left(x,y\right)<10*\left(i-1\right)\end{array}\right.$

Wherein, hist (i) represents the i-th dimension grey level histogram of subregion block, and i is integer and 1≤i≤26, and g (x, y) represents the gray-scale value of (x, y) place pixel in subregion block.

(342) following formula is adopted, the grey level histogram normalization by the subregion block of acquisition:

$\left\{\begin{array}{c}{\mathrm{hist}}^{\&prime;}\left(i\right)=\frac{hist\left(i\right)}{rs}\\ rs=\frac{1}{N}\underset{I=1}{\overset{N}{\&Sigma;}}hist\left(i\right)\end{array}\right.$

Wherein, N represents the dimension of subregion block grey level histogram, and hist ' (i) represents the grey level histogram of the subregion block after normalization;

(343) using the feature set of the grey level histogram after normalization as subregion block.