CN102509112A - Number plate identification method and identification system thereof - Google Patents

Abstract

The invention belongs to the technical field of vehicle identification and particularly discloses a number plate identification method and an identification system of the number plate identification method. The method comprises the steps of number plate image positioning, character separation and character identification, wherein the character identification step comprises pre-treatment, coarse classification, Chinese character identification and number and letter identification. Through the improvement on a Chinese character identification method (through a Gabor filter and a minimum distance classifier) and a number and letter identification method (based on correlation feature selection (CFS) and a Bayes classifier), the number plate Chinese character identification rate and the number plate number and letter identification rate are greatly improved.

Description

Licence plate recognition method and recognition system thereof

Technical field

The invention belongs to the vehicle recongnition technique field, be specifically related to a kind of high-precision licence plate recognition method and recognition system thereof to China's car plate.

Background technology

China's car plate mainly contains four types of wrongly written or mispronounced character of the blue end, yellow end surplus, white gravoply, with black engraved characters The Scarlet Letter and black matrix wrongly written or mispronounced characters, and it comprises Chinese character, English and numeral totally seven characters, and wherein Chinese character is usually located at the first end of car plate, and middle five characters are generally numeral or alphabetical.

License plate recognition technology is the important component part of automatic vehicle identification technology.Thereby mainly to be the thought of utilizing each automobile all to have unique number-plate number different vehicle to have different licence plates discern and calculating vehicle for it.It through technology such as a series of Flame Image Process, pattern-recognitions, carries out the identification of the number-plate number with the vehicle image of shot by camera.Under the situation that does not influence vehicle condition, computing machine is accomplished the identification of the number-plate number automatically.

As shown in Figure 1, a complete Vehicle License Plate Recognition System comprises 3 core subsystems: license plate image positioning unit, Character segmentation unit and character recognition unit, the accuracy of each subsystem all has direct influence to next system.

The principle of work of Vehicle License Plate Recognition System is, when the vehicle passing detection zone, camera is caught vehicle image and is sent in the image pick-up card, and image pick-up card is sent the vehicle image that collects into computing machine and discerned, and detailed step is following:

1, the license plate image positioning unit filters out the license plate image zone from the vehicle image of input;

2, the Character segmentation unit is a plurality of single characters with said license plate image Region Segmentation;

3, character recognition unit carries out Classification and Identification to each character, and exports corresponding recognition result and deliver to information processing places such as Surveillance center or cash desk.

Accurate location and accurately cutting apart of characters on license plate to car plate are the prerequisites of carrying out car plate identification work.Because, English alphabet and arabic numeral are not only arranged in China's car plate, also comprised the Chinese character that stroke is numerous and diverse, the difficulty of Recognition of License Plate Characters is more much bigger than numeral and alphabetical foreign car plate are only arranged.Therefore, Recognition of License Plate Characters is especially crucial for China's vehicle identification system.

At present; Domestic main license plate character recognition method comprises based on binary picture with based on the method for gray-scale map; Specifically: because Chinese character can be lost many useful informations in the binaryzation process; Produce the fracture of unnecessary noise and stroke, nearest research all is based on gray-scale map and carries out Chinese Character Recognition work.These class methods can be divided into two kinds: based on local feature and based on global characteristics.The former basic thought is to be divided into the partial structurtes characteristic that some primitives extract Chinese character to character picture; Research emphasis is placed on the mutual relationship of analyzing between the primitive and queueing discipline, and more typical algorithm has: geology architectural feature, stroke density characteristic; The stroke distribution characteristics; The stroke characteristic, the stroke direction characteristic, or the like.The latter also claims statistical method, and these class methods are frequency domain with the Chinese character gray matrix by spatial transform often, extracts frequency domain character, Fourier transform for example, Wal conversion, Huffman conversion, Karhunen-Loeve conversion, Gabor wave filter or the like.

Recognition of License Plate Characters algorithm commonly used is varied, mainly can reduce following several kinds: based on the recognizer of assemblage characteristic, combine the Zernike square like wavelet character; Through these characteristics characters on license plate is carried out Classification and Identification; This method robustness is high, accuracy good, but owing to adopted composite character; Therefore calculated amount is big, and real-time haves much room for improvement; Based on template matches, this method is mated character and template, finds the recognition result of the highest template of similarity degree as this character, as utilizes Euclidean distance that character is mated identification; The method of SVMs, this method are through a large amount of sample training is identified character, and the classic applications of this method is to handle two types of identification problems, and robustness is not strong more for a long time when the classification number; Based on neural network method, can identification and some pre-service be combined together and carry out, the concurrent working mode to the information-distribution type memory, has robustness, but this method is hidden the number of plies, the node number not have fixing mode definite, and be not easy to restrain; Based on the method for Bayesian model, Bayes classifier is done well in the situation of dealing with complicated problem, but a key issue is to know prior distribution, and this point just in time can make full use of the prior distribution information of character on car plate.

Though they can navigate to car plate under certain scene and condition, in the interference abominable, like the weather misty rain to some external conditions; Illumination variation, background is complicated, and the car plate wearing and tearing are faded; Under the situation such as image inclination, can't locate car plate or can't separating character and correct identification etc.

Summary of the invention

In order to solve existing license plate recognition technology low problem of discrimination under the complex environment factor, the object of the present invention is to provide a kind of high-precision licence plate recognition method and corresponding Vehicle License Plate Recognition System.

For realizing the foregoing invention purpose, technical scheme that the present invention adopts is following:

A kind of licence plate recognition method comprises license plate image location, Character segmentation and character recognition, and said character recognition specifically comprises the steps:

Remove the redundant frame of the single character picture after Character segmentation is handled, and each character graphics is transformed into identical big or small pixel;

Each character picture is divided into two types, and the first kind is that Chinese character image, second type are letter and digital picture;

Wherein, the first kind utilizes the Gabor wave filter to extract the characteristics of image of each Chinese character image, and utilizes minimum distance classifier to realize classification according to its characteristics of image;

Wherein, Second type of utilization selects from the original feature vector of numeral and letter based on the characteristic selecting method (CFS) of correlativity that representative power is strong closes with the low character subset of redundancy, and utilizes Bayes classifier to close according to said character subset numeral and letter are realized classifying.

In the described licence plate recognition method, each character graphics is transformed into identical big or small pixel, specifically: utilize Gaussian function that image is done a cube interpolation arithmetic, each character graphics all is transformed into 60 * 30 pixels.

In the described licence plate recognition method; After each character graphics was transformed into identical big or small pixel, this licence plate recognition method also comprised: use mean filter that each character picture is carried out smothing filtering and uses maximum variance between clusters (OTSU) that each character picture is carried out binary conversion treatment.

In the described licence plate recognition method, each character picture is divided into two types, the first kind is that Chinese character image, second type are letter and digital picture, specifically comprises:

Distinguish first last character and intermediate character, intermediate character directly is judged to be second class-letter and digital picture;

Utilize the binary tree sort device that first last character image is carried out the rough segmentation of Chinese character, letter and number once more, the first kind is a Chinese character image, and second type is letter and digital picture; If said initial character is " W ", then gets into People's Armed Police's Vehicle License Plate Recognition System and discern.

In the described licence plate recognition method, the Gabor function that said Gabor wave filter uses is:

Wherein, x '=xcos θ+ysin θ, y '=-x sin θ+ycos θ,

Wherein, λ is a wavelength, and its unit is a pixel, and its effective wavelength is more than or equal to 2, simultaneously less than 1/5th of input picture size;

Wherein, θ is the anglec of rotation, and its unit degree of being, its effective interval are [0 °, 360 °];

Wherein, φ is phase shift, its unit degree of being, and effectively span is [180 °, 180 °];

Wherein, γ is an aspect ratio;

Wherein, bw is a bandwidth.

In the described licence plate recognition method, utilize the Gabor wave filter to extract the characteristics of image of each Chinese character image, comprising:

Pending character picture and convolution kernel are carried out convolution get 60 * 30 complex matrix;

To same character picture get the textural characteristics of four direction and splice 60 * 120 real number matrix, said four direction is respectively θ=0 °, 45 °, 90 °, 135 ° direction;

Said real number matrix is transformed to 7200 dimensional feature vectors, and proper vector is reduced to 78 dimensions through PCA algorithm (Principle Component Analysis);

Said convolution kernel is made up and is got by said Gabor function, its convolution kernel as shown in the formula:

Wherein, λ=8, φ=0 °, γ=0.5, bw=1, convolution yardstick are 35 * 35.

In the described licence plate recognition method, said numeral is following with the concrete obtaining step of the original feature vector of letter:

Define one 10 * 5 structural element;

This structural element is whenever moved once at a distance from a pixel on character picture, until covering whole character picture;

During each moving, add up the white point number of pixels of this structural element overlay area and divided by the area of structural element as a characteristic, obtain the original feature vector of one 1326 dimension at last.

In the described licence plate recognition method, utilize and from the original feature vector of numeral and letter, to select the strong and low character subset of redundancy of representative power based on the characteristic selecting method (CFS) of correlativity and close, specifically:

With training sample X=(X ₁, X ₂..., X _m) ^TWith classification C=(c ₁, c ₂..., c _m) ^TSubstitution first formula and second formula are realized selecting of representative power is strong and redundancy is low characteristic according to the result of calculation of the 3rd formula;

Wherein, first formula is following:

r _fc＝corr(col(X，j)，C)

Col (X j) is the j row of sampling eigen set X, j=1, and 2 ..., n;

Wherein, second formula is following:

r _ff＝corr(col(X，i)，col(X，j))，i≠j；

Wherein, the 3rd formula is following:

$M_s=\frac{{\mathrm{kr}}_{\mathrm{fc}}}{\sqrt{k+k(k-1)r_{\mathrm{ff}}}}$

M _sBe the tolerance of characteristic set s for the correct classification of realization contribution, k is the number of set s institute containing element.

In the described licence plate recognition method, utilize Bayes classifier to close classification realized in numeral and letter according to said character subset, specifically:

In sample characteristics X to be identified and the first formula substitution, second formula, accomplish numeral and alphabetical classification according to the result of calculation of second formula;

Wherein, first formula is following:

$\mathrm{In}\det \left({\mathrm{\Σ}}_i\right)=\ln ({\mathrm{\σ}}_{i,1}^2{\mathrm{\σ}}_{i,2}^2...{\mathrm{\σ}}_{i,r}^2)$

$=2\underset{t=1}{\overset{r}{\mathrm{\Σ}}}\ln {\mathrm{\σ}}_{i,t};$

Wherein, second formula is following:

lnp(c _i|x)∝lnp(x|c _i)+lnp(c _i)

＝-0.5(nln(2π)+lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))。

∝-0.5(lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))

The corresponding Vehicle License Plate Recognition System of a kind of and aforementioned licence plate recognition method comprises license plate image positioning unit, Character segmentation unit and character recognition unit, and said character recognition unit specifically comprises following subelement:

The pre-service subelement is removed the redundant frame of the single character picture after Character segmentation is handled, and each character graphics is transformed into identical big or small pixel;

The rough sort subelement is divided into two types with each character picture, and the first kind is that Chinese character image, second type are letter and digital picture;

Chinese Character Recognition subelement, the first kind utilize the Gabor wave filter to extract the characteristics of image of each Chinese character image, and utilize minimum distance classifier to realize classification according to its characteristics of image;

Digital and alphabetical recognin unit; Second type of utilization selects from the original feature vector of numeral and letter based on the characteristic selecting method (CFS) of correlativity that representative power is strong closes with the low character subset of redundancy, and utilizes Bayes classifier to close according to said character subset numeral and letter are realized classifying.

The present invention is through Chinese characters recognition method (Gabor wave filter+minimum distance classifier) and the numeral and the improvement of alphabetical recognition methods (based on characteristic selecting method (the CFS)+Bayes classifier of correlativity), makes the present invention can reach 98.125% car plate Chinese Character Recognition rate and 98.58% car plate numeral and alphabetical discrimination.

Description of drawings

The picture that this description of drawings provided is used for auxiliary to further understanding of the present invention, constitutes the application's a part, does not constitute to improper qualification of the present invention, in the accompanying drawings:

Fig. 1 is the frame diagram of Vehicle License Plate Recognition System;

Fig. 2 is the schematic flow diagram of the embodiment of the invention;

Fig. 3 is the car plate positioning flow of the embodiment of the invention;

Fig. 4 is the pack framework flow process of the embodiment of the invention;

Fig. 5 is the further detail flowchart of character recognition unit of the present invention.

Embodiment

To combine accompanying drawing and practical implementation method to specify the present invention below, be used for explaining the present invention in schematic enforcement of the present invention and explanation, but not as to qualification of the present invention.

Embodiment 1:

As illustrated in fig. 1 and 2, the invention discloses a kind of licence plate recognition method, its general thought still comprises following three steps:

S1: license plate image location;

S2: Character segmentation;

S3: character recognition.

One, license plate image location

The present invention can use the license plate locating method based on mathematical morphology, and utilizes the colouring information of car plate to come car plate is carried out assist location.This localization method has been taken all factors into consideration gray scale and the edge and the colouring information of license plate image, realization flow such as Fig. 3, and it is following specifically to locate implementation procedure:

1) license plate image has been carried out pre-service,, improved the quality of image, strengthened image-region like methods such as greyscale transformation, denoising, enhancing contrast ratio, rim detection, binaryzations.

For ease of handling; The coloured image of at first video camera being captured is transformed to gray level image, and the average gray of statistical picture top four/part then is if average gray is [20; 70] between; Show dark images, it is done histogram equalization to improve the contrast of image, remake medium filtering; The picture that is generally evening less than 20 need not done histogram equalization.Utilize two prewitt vertical operator to come the vertical edge in the detected image, can detect the vertical edge of car plate simultaneously, do binaryzation again.Employed prewitt operator is in this programme:

$\mathrm{per}1=\left(\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right),$ $\mathrm{<figure-callout\; id="2"\; label="pre"\; filenames="HDA0000104749110000011.png,HDA0000104749110000012.png"\; state="\{\{state\}\}">pre</figure-callout>}2=\left(\begin{array}{ccc}-1& 0& 1\\ -1& 0& 1\\ -1& 0& 1\end{array}\right)$

2) utilize morphologic closed operation and opening operation scheduling algorithm that image is processed, the candidate region of outstanding car plate.

Use closed operation to connect license plate area, the size of structural element is 36 * 10; For removing some independently noises and break off linking of license plate area and non-license plate area, to do opening operation again and handle, the size of structural element is 28 * 4.

3) utilize the saltus step of car plate geometric configuration and picture pixel to come accurately to filter out license plate area from a plurality of car plate candidate regions.

Utilize the textural characteristics of profile (the ratio of width to height etc.) and the character of car plate to come from image, to choose the candidate region of car plate; The ratio of width to height of car plate selection interval is [2.0,7.5] in this programme, and the interval range of candidate's car plate district area is [4300; 35000], to the coarse positioning of car plate.

4) utilize the Hough transformation scheduling algorithm to come the car plate that tilts is done angularity correction, and accurately orient car plate.

Because adopting the problem of figure angle, video camera cause license plate image to tilt.At first utilize Hough transformation to search the horizontal tilt angle of license plate area, and license plate image is carried out level correction, promptly detect the car plate edge line, calculate inclination angle theta, then car plate is done geometric transformation to accomplish the rectification of level angle through Hough transformation.License plate image to after the level angle rectification carries out binary conversion treatment; Utilize the character texture (discontinuous point number) on the car plate to come the accurately up-and-down boundary of location car plate then; This process is to ask for the row that a maximum is interrupted to count; Utilize this maximum to be interrupted to count and the ratio of counting that is interrupted of certain row decides up-and-down boundary, ratio range is [0.38,0.40] in this programme.

The vertical bank angle of utilizing rotation vertical projection method to search car plate; And license plate image carried out vertical correction; Basic thought is to do projection for the binaryzation car plate that vertical direction tilts; The non-zero region of its projection result is maximum, is that unit vertically rotates with 0.1 ° in this programme, converges on the non-zero region smallest point place of projection result.After vertical correction is accomplished, can accurately determine the border, the left and right sides of car plate according to the projection of binaryzation car plate.

5) if above morphology algorithm can not effectively be oriented car plate, just adopt subsequent use localization method.

Utilize the car plate colouring information to locate.Localization method based on the car plate color characteristic mainly is to convert the RGB coloured image into the HSV coloured image to handle; Because the Euclidean distance and the color distance of point-to-point transmission are nonlinear relationship in the RGB three primary colors model commonly used; The colour that is not easy to carry out image is cut apart; And the HSV model can reflect perception and the distinguishing ability of people to color preferably, and this model receives illumination effect not quite easily, therefore; This alternative mean is transformed into the HSV space to the car plate coloured image, utilizes the colouring information of car plate to cut apart car plate.

Two, Character segmentation

Cutting apart of characters on license plate; What present embodiment adopted is vertical projection method; Effective object is the binaryzation license plate image after accurately locating, and be black matrix wrongly written or mispronounced character (need carry out the color upset for yellow end surplus, white gravoply, with black engraved characters, white background The Scarlet Letter car plate) so need to guarantee image to be split, and the license plate image after the binaryzation is done vertical projection; Search for the lowest point of drop shadow curve then, classify partitioning boundary as with place, the lowest point pixel.

1) segmentation candidates point breaks off

Reasons old owing to the part car plate or each side such as pollution, weather environment, image capture device or binaryzation poor effect cause characters on license plate that connection phenomenon is in various degree arranged, so before cutting apart, according to setting threshold possible cut-point is broken off earlier.Threshold value is set to 2 in the present embodiment, and the point that promptly is lower than this threshold value in the drop shadow curve is changed to 0.

2) border, the character left and right sides confirms

Judge according to threshold conditions such as wide, the height ratio of character and areas whether current segmentation result is a significant character, thereby confirm the deletion and the reservation on border, the left and right sides.In the present embodiment, the ratio of width to height span is [0.33,0.8], and wherein to account for the ratio range of this character area be [0.2,0.75] to the Chinese character pixel.

3) single character obtains

According to the border, effective left and right sides that the last step is confirmed, from the positioning result image, can take out single character picture successively, get the first seven character in this programme, be non-significant character image if there is remaining character boundary then to think.For People's Armed Police's car plate, then individual processing.

Three, character recognition

Recognition of License Plate Characters is the part of most critical of the present invention.Characters on license plate is actually the print hand writing that is attached on the car plate, and its recognition technology needs application mode identification, technology such as neural network and OCR (Optical Character Recognition).Can identification characters on license plate that correct be not only the problem of printed Chinese character identification, also need consider the influence of environment.Because the performance of video camera, illumination condition, vehicle movement all can make the character in the licence plate serious bluring occur, crooked, damaged or stain disturbs, and has therefore increased the difficulty of Recognition of License Plate Characters.The character recognition of car plate is subdivided into the car plate Chinese Character Recognition again, with car plate numeral and letter identification two aspect contents.The licence plate recognition method of our research and development can access the car plate numeral and alphabetical discrimination of 98.125% car plate Chinese Character Recognition rate and 98.58%.

Concrete identification step is as shown in Figure 2:

S301: discern pre-service, remove the redundant frame of the single character picture after Character segmentation is handled, and each character graphics is transformed into identical big or small pixel;

S302: Chinese character, numeral and letter are thick to be distinguished, and each character picture is divided into two types, and the first kind is that Chinese character image, second type are letter and digital picture;

S303: Chinese Character Recognition, utilize the Gabor wave filter to extract the characteristics of image of each Chinese character image for the first kind, and utilize minimum distance classifier to realize classification according to its characteristics of image;

S304: numeral and letter identification; Select from the original feature vector of numeral and letter based on the characteristic selecting method (CFS) of correlativity for second type of utilization that representative power is strong closes with the low character subset of redundancy, and utilize Bayes classifier to close numeral and letter are realized classifying according to said character subset.

In the present embodiment, Chinese character, numeral and alphabetical identification are handled respectively, mainly be to consider that they have feature extraction algorithm and concrete classifier design different of resolving ability most, but overall frame model are consistent.After obtaining primitive character, the method that adopts characteristic to select is carried out dimensionality reduction, and the characteristic selection method has three main branches, filter method (filter), pack (wrapper) and embedding inlay technique (embed) under the framework of classification.Filter method is set up the degree of correlation of data through analyzing correlativity between the data characteristics, removes low relevant characteristic, thereby obtains the character subset of a high degree of correlation.The advantage of this method is fast succinct, even if but to high dimensional data also fast processing, in addition, it is independent of disaggregated model.The major defect of this method is it based on univariate analysis, thereby has ignored the relevance between the characteristic, and the prediction effect that causes classifying is bad.Although there is the researcher to utilize multivariable analysis to overcome this shortcoming, select still fuzzyyer for the variable number purpose.Pack is attempted that characteristic is selected process and is in the same place with sorting algorithm " parcel "; Through on character subset, selecting the searching direction; Set up a plurality of character subsets and close, seek optimum classifying quality, but the dimension that closes along with character subset increases; The capacity that character subset closes can sharply increase, and increases the difficulty of seeking the character subset optimal direction.Solution commonly used comprises that determinacy is selected with randomness and selects.The major advantage of the method has been to consider the relevance of characteristic set and sorting algorithm, and the correlativity between the characteristic.

Yet; The completion that is closely linked of the selection of the feature extraction of character picture and sorter in the present embodiment; Promptly adopt pack; Its realization flow is as shown in Figure 4, comprises search characteristics subclass, classifiers combination and combined effect assessment, its objective is and guarantees that the characteristic and the employed sorter that extract are best of breeds.

The extraction algorithm of present embodiment Chinese character characteristics of image adopts Gabor, and Digital Character Image Feature Extraction algorithm use is based on the characteristic selecting method (CFS) of correlativity, and it realizes that idiographic flow is as shown in Figure 5:

1), identification pre-service:

1.1, at first, secondary splitting-removal character is separated the redundant frame of the single character picture after handling:

Single character picture by being partitioned into passes to identification module, at first will do secondary splitting, guarantees that each character does not have redundant frame.

1.2, then, normalization handles-each character picture is transformed into identical big or small pixel:

Utilize Gaussian function that image is done a cube interpolation arithmetic, single character picture is carried out normalization, in this programme, be to normalize to 60 * 30 pixel sizes.

1.3, last, the mean filter denoising:

Owing in the middle of interpolation arithmetic carries out normalized process, can produce some noises, to carry out The disposal of gentle filter and binary conversion treatment after the normalization.

Smothing filtering mainly is filtering because some edge noises of generation when the character picture interpolation amplified, and the smoothing filter that uses in the present embodiment is a mean filter, and its filtering speed is fast, satisfactory for result.

Binary conversion treatment is used maximum variance between clusters (OTSU), is big law again, can adaptive definite binaryzation process in the threshold value of needs.

2) Chinese character, numeral and alphabetical thick differentiation:

2.1, distinguish first last character and intermediate character, intermediate character directly is judged to be second class-letter and digital picture;

2.2, utilize the binary class device that first last character image is carried out the rough segmentation of Chinese character, letter and number once more, the first kind is a Chinese character image, second type is the letter and number image.

In this step, Chinese character, letter and number will be handled respectively, but can't judge the position of Chinese character according to priori accurately for different car plates; This programme hypothesis Chinese character only might appear on the position of first last character, is first last character so distinguish earlier, directly calls letter and digit recognition for non-first last character; To do the thick differentiation (binary tree sort device) of Chinese character, letter and number earlier for first last character; Want individual processing for People's Armed Police's car plate in addition, promptly be:, judge that it is " W " if initial character is a letter and number; If " W " then gets into People's Armed Police's Vehicle License Plate Recognition System and discerns.

3), Chinese Character Recognition:

3.1, Feature Extraction:

Use the Gabor wave filter to do the feature extraction of car plate Chinese character image in the present embodiment.Adopt the method for Gabor wave filter mainly to be based on stroke width and the orientation-sensitive of Gabor wave filter to Chinese character.To having the character picture of different stroke widths and direction, filtered is widely different.The Gabor wave filter is to inferior quality simultaneously, and the character picture of low resolution has certain robustness.

Wherein, the Gabor bank of filters that present embodiment adopts, the concrete design as follows:

The Gabor bank of filters is actually the two-dimensional Gabor function by one group of parameter control, so core technology is Determination of Parameters, in this programme, the Gabor function of use is

Wherein, x '=xcos θ+ysin θ;

Wherein, y '=-x sin θ+ycos θ;

Wherein, λ is a wavelength, and unit is a pixel, and effective wavelength requires more than or equal to 2, is less than 1/5th of input picture size simultaneously, and this embodiment is selected λ=8;

Wherein, θ is the anglec of rotation, unit degree of being, and effectively interval is [0 a °, 360 °], owing to be the textural characteristics that extracts four direction, so θ=0 °, 45 °, 90 °, 135 °;

Wherein, φ is phase shift, unit degree of being, and effectively span is [180 °, 180 °], present embodiment is selected, φ=0 °;

Wherein, γ is an aspect ratio, has shown the shape of the supporting ellipse of Gabor function, and present embodiment is selected γ=0.5;

Wherein, σ can not direct representation, can only be derived by bandwidth, selects bandwidth bw=1 in this example, and its derivation formula does

$\mathrm{\&sigma;}=\frac{\mathrm{\&lambda;}}{\mathrm{\&pi;}}\sqrt{\frac{1}{2}\mathrm{<figure-callout\; id="2"\; label="log"\; filenames="HDA0000104749110000011.png,HDA0000104749110000012.png"\; state="\{\{state\}\}">log</figure-callout>}2}\frac{2^{\mathrm{bw}}+1}{2^{\mathrm{bw}}-1}$

Chinese character to be identified is handled through the Gabor bank of filters and is obtained corresponding unity and coherence in writing image.Through multiple dimensioned, after multidirectional Gabor filtering, we extract the primitive character of the wave filter output coefficient on a plurality of sub-planes as the car plate Chinese character.

3.1.1, pending character picture and convolution kernel carried out convolution get 60 * 30 complex matrix;

3.1.2, to same character picture get the textural characteristics of four direction and splice 60 * 120 real number matrix, said four direction is respectively θ=0 °, 45 °, 90 °, 135 ° direction;

Said convolution kernel is made up and is got by aforementioned Gabor function, its convolution kernel as shown in the formula:

Wherein, λ=8, φ=0, γ=0.5, bw=1, convolution yardstick are 35 * 35.

In this step pending character picture examined therewith and carry out convolution; Pixel value outside the edge is chosen as the pixel value of nearest neighbor point; Convolution results is one 60 * 30 a complex matrix (onesize with former character picture); So will ask mould and do normalization this complex matrix, the method for normalizing that in this programme, adopts be the greatest member value of each element of matrix divided by this matrix.Then,, for a character picture, extract the textural characteristics of four direction (θ=0 °, 45 °, 90 °, 135 °), and filtered is stitched together, obtain one 60 * 120 real number matrix through adjustment θ value.

3.1.3, principal component PCA (Principle Component Analysis) dimensionality reduction

At first, above-mentioned real number matrix is transformed to 7200 dimensional feature vectors, through the PCA algorithm proper vector is reduced to 78 dimensions then.

In this step; The primitive character matrixing that each Chinese character image filtering is obtained is one 7200 dimensional feature vector; Select a template vector to form a pattern matrix for each character type; Calculate its covariance, find the pairing proper vector of big eigenwert to constitute the resolving ability projecting direction.Through projection finally realizes the characteristic dimensionality reduction in that this side up with original feature vector.

3.2, Chinese character classification

In this step, the characteristics of image of selecting for use minimum distance classifier will go up step extraction is realized classification.It is following that detailed sorter is selected principle:

Pack is adopted in the selection of Chinese character sorter in the present embodiment; Promptly gather based on primitive character; Sorter in conjunction with classical comprises nearest neighbor classifier, minimum distance classifier, SVMs (SVM), decision tree, Bayes classifier, hidden Markov model; C4.5 does the classification forecast analysis, through the degree of accuracy, time complexity etc. of relatively identification, chooses the combination of optimum characteristic set+sorter.Through a large amount of experiments, we get minimum distance classifier is optimal selection.

Primitive character has the strongest recognition capability master composition through obtaining one group behind the dimensionality reduction; The basic thought of confirming still to be based on pack of major component dimension; Promptly under selected minimum distance classifier, the quantity of information of screening major component is since 50%, and every increase by 5% makes up a character subset; Through comparing discrimination and time complexity, the selected major component of present embodiment is 78 dimensions at last.This main composition matees as the characteristic and the template of each Chinese character image, and immediate that template Chinese character of matching result promptly is judged to be the affiliated classification of this Chinese character to be identified.

4), letter and number identification

4.1, the extraction of original feature vector

Binaryzation character picture after the secondary splitting is 60 * 30 pixel sizes of standard, and the original feature vector obtaining step that this step is concrete is following:

4.1.1, defined one 10 * 5 structural element;

4.1.2, this structural element is whenever moved once at a distance from a pixel on character picture, until covering whole character picture;

4.1.3, when moving at every turn, the white point number of pixels of statistical framework element overlay area and divided by the area of structural element as a characteristic, obtain the original feature vector of one 1326 dimension at last.

4.2, utilize and from above-mentioned original feature vector, to select the strong and low character subset of redundancy of representative power based on the characteristic selecting method (CFS) of correlativity and close, specifically:

With training sample X=(X ₁, X ₂..., X _m) ^TWith classification C=(c ₁, c ₂..., c _m) ^TSubstitution first formula and second formula are realized selecting of representative power is strong and redundancy is low characteristic according to the result of calculation of the 3rd formula;

Wherein, first formula is following:

r _fc＝corr(col(X，j)，C)

Col (X j) is the j row of sampling eigen set X, j=1, and 2 ..., n;

Wherein, second formula is following:

r _ff＝corr(col(X，i)，col(X，j))，i≠j；

Wherein, the 3rd formula is following:

$M_s=\frac{{\mathrm{kr}}_{\mathrm{fc}}}{\sqrt{k+k(k-1)r_{\mathrm{ff}}}}$

M _sBe the tolerance of characteristic set s for the correct classification of realization contribution, k is the number of set s institute containing element.

This step principle remains and adopts pack to carry out the optimum combination screening of character subset and sorter.Based on original feature vector, at first use the different character selection method to choose the alternative set of character subset that the representative ability is strong, redundancy is low.The method that the construction feature subclass is selected for use has relief, CFS (based on the characteristic selecting method of correlativity), Chi (side's of card distribution), Mutual information (interactive information method), Information gain (information is obtained method) etc.The character subset that obtains through this characteristic selection method closes; In conjunction with nearest neighbor classifier, minimum distance classifier, SVMs (SVM), decision tree, Bayes classifier, hidden Markov model; C4.5, BP neural networks etc. are through the different combinations of sorter with character subset; Based on the strategy of alternate analysis, select optimum sorter+character subset selection method: Bayes classifier+CFS (based on the characteristic selecting method of correlativity).

CFS is a didactic signature search algorithm based on correlativity, and this algorithm hypothesis is beneficial to the characteristic set and the classification height correlation of classification, but uncorrelated between each component of characteristic.CFS selects characteristic through the correlativity of tolerance between the variable, and this method do not rely on any specific data conversion, mainly is to be applied to the supervised classification problem.This algorithm acts on the primitive character space, that is to say through the select low dimensional characteristics subclass of CFS algorithm to be represented rather than transformation space by primitive character.Its mechanism is similar to a wave filter, can not cause high time complexity because of an algorithm iteration.The CFS hypothesis is characterized in that condition independently for concrete class, as long as there are not the dependence of height in characteristic and non-target class, CFS can pick out the characteristic relevant with target class efficiently, and removes incoherent and redundant characteristic.

The ultimate principle of CFS is if the changing of the value of characteristic system along with the difference of classification; Think that then these characteristics and classification are relevant; Perhaps can make prediction if in other words be exactly that characteristic is relevant with class, think that then these characteristics are useful to classifying class.The algorithm for estimating of CFS relativity measurement is following:

Suppose that the augmented matrix that sample characteristics set and its corresponding class are formed does,

A＝(X，C)

Wherein the sample characteristics set is:

$X={(X_1,{\mathrm{<figure-callout\; id="2"\; label="X"\; filenames="HDA0000104749110000011.png,HDA0000104749110000012.png"\; state="\{\{state\}\}">X</figure-callout>}}_2,...,X_m)}^T=\left(\begin{array}{cccc}{x}_{11},& x_{12},& ...,& x_{1n}\\ x_{21},& x_{22},& ...,& x_{2n}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ x_{m1},& {\mathrm{<figure-callout\; id="2"\; label="x\; m"\; filenames="HDA0000104749110000011.png,HDA0000104749110000012.png"\; state="\{\{state\}\}">x</figure-callout>}}_m,& ...,& x_{\mathrm{mn}}\end{array}\right)$

Classification is:

C＝(c ₁，c ₂，…，c _m) ^T

Then the correlativity of sample characteristics-classification is:

r _fc＝corr(col(X，j)，C)

Wherein, col (X j) is the j row of sampling eigen set X, j=1, and 2 ..., n.

Correlativity between each component of sample characteristics does,

r _ff＝corr(col(X，i)，col(X，j))，i≠j

The character subset valuation functions of CFS as shown in the formula:

$M_s=\frac{{\mathrm{kr}}_{\mathrm{fc}}}{\sqrt{k+k(k-1)r_{\mathrm{ff}}}}$

Wherein, M _sBe the tolerance of characteristic set s for the correct classification of realization contribution, k is the number of set s institute containing element.Through calculating M _sAssess each characteristic component, then according to M _sSize each characteristic component is sorted.The parcel strategy uses a kind of Algorithm for Reduction to accomplish selecting of characteristic through the advantage of estimating each character subset.Under the strategy of pack, each character subset is combined with various sorters, confirm the characteristic component that helps classifying most at last.

4.3, letter and number classification

4.1, utilize Bayes classifier to close classification realized in numeral and letter according to said character subset, specifically:

In sample characteristics X to be identified and the first formula substitution, second formula, accomplish numeral and alphabetical classification according to the result of calculation of second formula;

Wherein, first formula is following:

$\mathrm{In}\det \left({\mathrm{\&Sigma;}}_i\right)=\ln ({\mathrm{\&sigma;}}_{i,1}^2{\mathrm{\&sigma;}}_{i,2}^2...{\mathrm{\&sigma;}}_{i,r}^2)$

$=2\underset{t=1}{\overset{r}{\mathrm{\&Sigma;}}}\ln {\mathrm{\&sigma;}}_{i,t};$

Wherein, second formula is following:

lnp(c _i|x)∝lnp(x|c _i)+lnp(c _i)

＝-0.5(nln(2π)+lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))。

∝-0.5(lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))

The principle explanation is as follows in detail:

The character subset that this step makes up selected sample storehouse after through CFS (based on the characteristic selecting method of correlativity) screening characteristic is as the training sample set of Bayes classifier; Because the possibility that each character of each character position occurs on the car plate is identical; So each character type of hypothesis is equiprobable, promptly

p(c _i)＝1/34，i＝1，2，…，34

Simultaneously, suppose appearance Normal Distribution, promptly for each each character picture of character type

$f\left(x\right)=\frac{1}{\sqrt{{\left(2\mathrm{\&pi;}\right)}^n\det \left({\mathrm{\&Sigma;}}_i\right)}}{e}^{-\frac{{\left({x-u}_i\right)}^T{\mathrm{\&Sigma;}}_i^{-1}\left({x-u}_i\right)}{2}}$

Wherein, x, u _i∈ R ⁿ, x is that the original feature vector of character picture is screened gained character subset, u through CFS _iBe the expectation vector of i class sample, ∑ _iBeing the real symmetric positive definite square formations in n rank, is the covariance matrix of i class sample, n=82 in this programme.

Know by Bayesian formula,

$p\left(c_i\right|x)=\frac{p\left(x\right|c_i)p\left(c_i\right)}{p\left(x\right)}$

Wherein, c _iType of being mark.According to the Bayes principle, the following formula two ends are taken the logarithm:

lnp(c _i|x)＝lnp(x|c _i)+lnp(c _i)-lnp(x)

Make posterior probability lnp (c _i| x) maximum i class is exactly the target class of sample to be identified, and ln p (x) is a constant for all classifications, thus can omit, so there is second formula to set up:

lnp(c _i|x)∝lnp(x|c _i)+lnp(c _i)

＝-0.5(nln(2π)+lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))

∝-0.5(lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))

Wherein, u _i, ∑ _iEstimated to draw by training sample, specific algorithm is following:

$u_i=\frac{1}{m}\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{x}_{i,j}$

M is the number of samples of i class, x _{I, j}Be j training sample of i class.To X _i-u _iBe QR and decompose, get Q _iR _i=X _i-U _i

∑ _i＝(X _i-U _i) ^T(X _i-U _i)

＝(Q _iR _i) ^T(Q _iR _i)

＝R _i ^TQ _i ^TQ _iR _i

＝R _i ^TR _i

Wherein, X _i∈ R ^{M * n}Be the training sample matrix of i class, U _iBy the matrix that such expectation vector is formed, Q _i∈ R ^{M * m}Be a unitary matrix, R _i∈ R ^{M * n}It is a upper triangular matrix.

Then, to R _iBe SVD (svd),

R _i＝S _iV _iD _i

Wherein, S _iBe m rank orthogonal matrix, D _iBe n rank orthogonal matrix,

V _i＝diag(σ _i，1，σ _i，2，…，σ _i，r，0，…，0)

Wherein, σ _{I, t}Be R _iSingular value, r=r (R _i).The formula so win:

$\mathrm{In}\det \left({\mathrm{\&Sigma;}}_i\right)=\ln ({\mathrm{\&sigma;}}_{i,1}^2{\mathrm{\&sigma;}}_{i,2}^2...{\mathrm{\&sigma;}}_{i,r}^2)$

$=2\underset{t=1}{\overset{r}{\mathrm{\&Sigma;}}}\ln {\mathrm{\&sigma;}}_{i,t}$

In sample characteristics x to be identified and the first formula substitution formula, second formula, can judge the size of posterior probability through the result who calculates the right side formula, and finally accomplish digital, alphabetical classification.

Embodiment 2:

Present embodiment discloses the corresponding Vehicle License Plate Recognition System of a kind of and embodiment 1, comprises license plate image positioning unit, Character segmentation unit and character recognition unit, and said character recognition unit specifically comprises following subelement:

The pre-service subelement is removed the redundant frame of the single character picture after Character segmentation is handled, and each character graphics is transformed into identical big or small pixel;

The rough sort subelement is divided into two types with each character picture, and the first kind is that Chinese character image, second type are letter and digital picture;

Chinese Character Recognition subelement, the first kind utilize the Gabor wave filter to extract the characteristics of image of each Chinese character image, and utilize minimum distance classifier to realize classification according to its characteristics of image;

Digital and alphabetical recognin unit; Second type of utilization selects from the original feature vector of numeral and letter based on the characteristic selecting method (CFS) of correlativity that representative power is strong closes with the low character subset of redundancy, and utilizes Bayes classifier to close according to said character subset numeral and letter are realized classifying.

More than the technical scheme that the embodiment of the invention provided has been carried out detailed introduction; Used concrete example among this paper the principle and the embodiment of the embodiment of the invention are set forth, the explanation of above embodiment only is applicable to the principle that helps to understand the embodiment of the invention; Simultaneously, for one of ordinary skill in the art, according to the embodiment of the invention, the part that on embodiment and range of application, all can change, in sum, this description should not be construed as limitation of the present invention.

Claims (10)

1. a licence plate recognition method comprises license plate image location, Character segmentation and character recognition, it is characterized in that, said character recognition specifically comprises the steps:

Remove the redundant frame of the single character picture after Character segmentation is handled, and each character graphics is transformed into identical big or small pixel;

Each character picture is divided into two types, and the first kind is that Chinese character image, second type are letter and digital picture;

Wherein, the first kind utilizes the Gabor wave filter to extract the characteristics of image of each Chinese character image, and utilizes minimum distance classifier to realize classification according to its characteristics of image;

Wherein, Second type of utilization selects from the original feature vector of numeral and letter based on the characteristic selecting method (CFS) of correlativity that representative power is strong closes with the low character subset of redundancy, and utilizes Bayes classifier to close according to said character subset numeral and letter are realized classifying.

2. licence plate recognition method according to claim 1 is characterized in that, each character graphics is transformed into identical big or small pixel, specifically:

Utilize Gaussian function that image is done a cube interpolation arithmetic, each character graphics all is transformed into 60 * 30 pixels.

3. licence plate recognition method according to claim 1 is characterized in that, after each character graphics was transformed into identical big or small pixel, this licence plate recognition method also comprised:

Use mean filter that each character picture is carried out smothing filtering and uses maximum variance between clusters (OTSU) that each character picture is carried out binary conversion treatment.

4. licence plate recognition method according to claim 1 is characterized in that, each character picture is divided into two types, and the first kind is that Chinese character image, second type are letter and digital picture, specifically comprises:

Distinguish first last character and intermediate character, intermediate character directly is judged to be second class-letter and digital picture;

Utilize the binary tree sort device that first last character image is carried out the rough segmentation of Chinese character, letter and number once more, the first kind is a Chinese character image, and second type is letter and digital picture;

If said initial character is " W ", then gets into People's Armed Police's Vehicle License Plate Recognition System and discern.

5. licence plate recognition method according to claim 1 is characterized in that, the Gabor function that said Gabor wave filter uses is:

Wherein, x '=xcos θ+ysin θ, y '=-x sin θ+ycos θ,

Wherein, λ is a wavelength, and its unit is a pixel, and its effective wavelength is more than or equal to 2, simultaneously less than 1/5th of input picture size;

Wherein, θ is the anglec of rotation, and its unit degree of being, its effective interval are [0 °, 360 °];

Wherein, φ is phase shift, its unit degree of being, and effectively span is [180 °, 180 °];

Wherein, γ is an aspect ratio;

Wherein, bw is a bandwidth.

6. licence plate recognition method according to claim 5 is characterized in that, utilizes the Gabor wave filter to extract the characteristics of image of each Chinese character image, comprising:

Pending character picture and convolution kernel are carried out convolution get 60 * 30 complex matrix;

To same character picture get the textural characteristics of four direction and splice 60 * 120 real number matrix, said four direction is respectively θ=0 °, 45 °, 90 °, 135 ° direction;

Said real number matrix is transformed to 7200 dimensional feature vectors, and proper vector is reduced to 78 dimensions through PCA algorithm (Principle Component Analysis);

Said convolution kernel is made up and is got by said Gabor function, its convolution kernel as shown in the formula:

Wherein, λ=8, φ=0 °, γ=0.5, bw=1, convolution yardstick are 35 * 35.

7. licence plate recognition method according to claim 1 is characterized in that, said numeral is following with the concrete obtaining step of the original feature vector of letter:

Define one 10 * 5 structural element;

This structural element is whenever moved once at a distance from a pixel on character picture, until covering whole character picture;

During each moving, add up the white point number of pixels of this structural element overlay area and divided by the area of structural element as a characteristic, obtain the original feature vector of one 1326 dimension at last.

8. licence plate recognition method according to claim 1 is characterized in that, utilizes from the original feature vector of numeral and letter, to select the strong and low character subset of redundancy of representative power based on the characteristic selecting method (CFS) of correlativity and close, specifically:

With training sample X=(X ₁, X ₂..., X _m) ^TWith classification C=(c ₁, c ₂..., c _m) ^TSubstitution first formula and second formula are realized selecting of representative power is strong and redundancy is low characteristic according to the result of calculation of the 3rd formula;

Wherein, first formula is following:

r _fc＝corr(col(X，j)，C)

Col (X j) is the j row of sampling eigen set X, j=1, and 2 ..., n;

Wherein, second formula is following:

r _ff＝corr(col(X，i)，col(X，j))，i≠j；

Wherein, the 3rd formula is following:

$M_s=\frac{{\mathrm{kr}}_{\mathrm{fc}}}{\sqrt{k+k(k-1)r_{\mathrm{ff}}}}$

M _sBe the tolerance of characteristic set s for the correct classification of realization contribution, k is the number of set s institute containing element.

9. licence plate recognition method according to claim 1 is characterized in that, utilizes Bayes classifier to close classification realized in numeral and letter according to said character subset, specifically:

In sample characteristics X to be identified and the first formula substitution, second formula, accomplish numeral and alphabetical classification according to the result of calculation of second formula;

Wherein, first formula is following:

$\mathrm{In}\det \left({\mathrm{\&Sigma;}}_i\right)=\ln ({\mathrm{\&sigma;}}_{i,1}^2{\mathrm{\&sigma;}}_{i,2}^2...{\mathrm{\&sigma;}}_{i,r}^2)$

$=2\underset{t=1}{\overset{r}{\mathrm{\&Sigma;}}}\ln {\mathrm{\&sigma;}}_{i,t};$

Wherein, second formula is following:

lnp(c _i|x)∝lnp(x|c _i)+lnp(c _i)

＝-0.5(nln(2π)+lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))。

∝-0.5(lndet(∑ _i)+(x-u _i) ^T∑ _i ^-1(x-u _i))

10. a Vehicle License Plate Recognition System corresponding with claim 1 comprises license plate image positioning unit, Character segmentation unit and character recognition unit, it is characterized in that said character recognition unit specifically comprises following subelement:

The pre-service subelement is removed the redundant frame of the single character picture after Character segmentation is handled, and each character graphics is transformed into identical big or small pixel;

The rough sort subelement is divided into two types with each character picture, and the first kind is that Chinese character image, second type are letter and digital picture;

Chinese Character Recognition subelement, the first kind utilize the Gabor wave filter to extract the characteristics of image of each Chinese character image, and utilize minimum distance classifier to realize classification according to its characteristics of image;

Digital and alphabetical recognin unit; Second type of utilization selects from the original feature vector of numeral and letter based on the characteristic selecting method (CFS) of correlativity that representative power is strong closes with the low character subset of redundancy, and utilizes Bayes classifier to close according to said character subset numeral and letter are realized classifying.

Images