CN100545858C - Based on the vehicle license plate extraction method in the complex background of wavelet transformation - Google Patents

Abstract

Provided by the inventionly provide a kind of based on the vehicle license plate extraction method in the complex background of wavelet transformation, it is to adopt coarse positioning and accurately locate to combine, the way of employing wavelet analysis is extracted the high-frequency information in the license plate area in accurately locating, and accurately carrying out the normalized of brightness and color behind the location, but finally obtained a car plate that includes identification number with unified background and number color.Adopt vehicle license plate extraction method of the present invention, the high-frequency information of extraction license plate area that not only can be as much as possible, thus accurately locate car plate, the environmental adaptation degree of method is improved greatly; And also improved the computing velocity of method in conjunction with the method for coarse positioning and fine positioning, satisfied the needs of real-time.

Description

Based on the vehicle license plate extraction method in the complex background of wavelet transformation

Technical field

The invention belongs to the electronic information processing technology field, particularly the vehicle license plate extraction method in the complex background in the license plate recognition technology.

Background technology

Intelligent transportation system is the forward position research topic of present world traffic and transport field; developed country proposes and has carried out a series of projects; its core is at serious day by day transport need and environmental protection pressure; adopt infotech, the communication technology, computer technology, control technology etc. that conventional traffic transportation system is carried out deep transformation; to improve service efficiency, the security of system of system resource; reduce the consumption and the environmental pollution of resource, China is no exception in this regard.License plate recognition technology in the intelligent transportation system is mainly realized the driving vehicle licence plate is discerned automatically, thereby finish automatic charge, unmanned parking management, the traffic control of critical junction and the tracking of vehicles peccancy or the like function, save manpower, fund with this, improved the efficient of traffic administration simultaneously.Nowadays, along with the raising of computing power and the development of image processing techniques, Vehicle License Plate Recognition System reaches its maturity.See document for details: Da-Shan Gao, Jie Zhou, Car license plates detection from complexscene, Signal Processing Proceedings, 5th International Conference, 21-25Aug.2000 and document: Shyang-Lih Chang, Li-Shien Chen, YunChung Chung, Sei-Wan Chen, Automatic license platerecognition, Intelligent Transportation Systems, IEEE Transactions on, March 2004 is described.

In Vehicle License Plate Recognition System, it is the technology of a key that car plate extracts, and its purpose is exactly to be partitioned into license plate image from the digital picture of a width of cloth complex background, requires to have high recognition and stronger environmental suitability.But because the complicacy of car plate background and the diversity of vehicle license plate characteristic, up to now, the complete general license plate intelligently localization method of neither one still.Most of localization methods are confined to the solution of certain side problem, as the interference of the inclination of licence plate, illumination, The noise etc., from practical application bigger distance are arranged still.Therefore, how existing achievement in research is combined, consider the ability to work of existing equipment simultaneously, make our Vehicle License Plate Recognition System have good performance and recognition speed is the direction of our current research.See document Zhi-Bin Huang for details, Yan-Feng Guo, Classifier fusion-based vehicle license plate detectionalgorithm, Machine Learning and Cybernetics, 2003International Conference on, 2-5Nov.2003 and document Yoshimori, S.Mitsukura, Y.Fukumi, M.Akamatsu, N.Khosal, R.License plate detection system in rainydays, Computational Intelligence in Robotics and Automation., 16-20July 2003.

And based on the vehicle license plate extraction method of wavelet analysis, might address this problem, because the high-frequency information of license plate area in entire image is obvious and concentrated, thereby the while wavelet analysis can effectively extract and outstanding these high-frequency informations extract license plate image accurately.

Wherein, so-called wavelet transformation refers to the small-sized ripple that is called small echo based on some, has the frequency of variation and limited duration.Complex background refers to then that in containing the image of license plate image having only a very little part is license plate area, and other parts are background.And complexity is meant that background contains various images and textural characteristics, some feature even similar with license plate area.Last car plate refers to the number that is used to identify testing vehicle register that each car plate is installed, and production standardization generally hangs on the front end or the rear portion of vehicle.

The method that present normally used car plate extracts has:

1. the color property based on car plate carries out vehicle license plate extraction method.It is to distinguish license plate area and background area by other regional rainbow features that are different from of extracting car plate, thereby extracts car plate.Advantage is the chromatic information that has made full use of image; Shortcoming is that speed is slower, may lose efficacy for the profuse image of those colors.See document Zhu Wei-gang for details; Hou Guo-jiang; Jia Xing; A study of locating vehicle license plate based on colorfeature and mathematical morphology.Signal Processing, 26-30Aug.2002.

2. use the Hough conversion and carry out the method that car plate extracts.Thereby it is to search the car plate position by the straight line that extracts the car plate frame.Its shortcoming is that speed is slower, and locatees inaccurately, and it is many interference region to occur.See document K.M.Kim for details, B.J.Lee, K.Lyou.The automatic coefficient and Hough transform.Journal of Control, Automatic and System Engineering.3 (5): 511-519,1997.

3. based on the vehicle license plate extraction method of morphological operator.It is to strengthen by the edge feature of operations such as use expansion, corrosion to car plate, and then extracts car plate.Its shortcoming is can find out considerable interference region to the abundant relatively zone of those edge features, thereby influences the effect that car plate extracts.See document M.Shridhar for details, J.W.Miller.Recognition of license plate image, In Processings of International Conference on DocumentAnalysis and Recognition, 17-20,1999.

Above-mentioned three kinds of methods all possessed a common characteristic when car plate extracted: these methods all are at a particular environment, and the high-frequency information that extracts is limited.In case environmental change, it extracts accuracy rate, and bigger fluctuation will take place, and the performance of whole Vehicle License Plate Recognition System is descended.Therefore lack all-environment adaptability, the method robustness is bad.Our said environment refers to the environment that the car plate original image is gathered herein, different environment, such as weather, constantly, vehicle background of living in and vehicle displacement situation or the like, the image of collection sizable difference often occurs at aspects such as resolution and textural characteristics.High-frequency information refers to the local detail feature of image, and the local detail feature of license plate area is more obvious with respect to other zones.

Summary of the invention

Task of the present invention provides a kind of based on the vehicle license plate extraction method in the complex background of wavelet transformation, adopt method of the present invention, can improve the versatility of vehicle license plate extraction method widely, described complex background comprises fine day, rainy day, greasy weather, car plate level, license plate sloped, stationary vehicle, vehicle movement, daytime, night or the like situation.

The present invention ground content for convenience of description, at first make a term definition at this:

1. car plate: number plate that is used to identify testing vehicle register of installing in each car, include 7 characters in number plate, unified production standard is arranged, generally hang on the front end or the rear portion of vehicle, but may be different the standard of the car plate of the vehicle of different purposes.

2. gray level image: only comprised monochrome information in the image and without any the image of colouring information.

3. gradation conversion: a kind of coloured image is converted into the means of gray level image, utilize formula f (i, j)=0.114*I (i, j, 1)+0.587*I (i, j, 2)+0.299*I (i, j, 3) can be the width of cloth greyscale image transitions gray level image, the line position of i presentation video wherein, the column position of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of i in the gray level image after the expression conversion, I (i, j, 1), I (i, j, 2) and I (i, j, 3) represent the R of the pixel of the capable j row of i in the coloured image respectively, G, the value of B component.

4. gray-scale map stretches: the image to those intensity profile are relatively concentrated, and carry out functional transformation its uniform gray level is distributed within 0～255 this scope, increase picture contrast with this, this functional transformation formula is $f\left(x\right)=\tanh \left(\frac{c}{\mathrm{\σ}}x\right),$ Wherein x represents the gray-scale value of input, and the gray-scale value behind f (x) representation transformation, tanh are two tans.σ represents the mean square deviation of gray level image, and c represents a constant, is taken as 2.4 herein.

5. VG (vertical gradient): for the vertical detail information of outstanding original image, extract the edge, deduct the gray-scale value of previous pixel, obtain the VG (vertical gradient) value with back pixel gray-scale value of each row.Computing formula is g _V(i, j)=| f (i, j+1)-f (i, j) |, the line position of i presentation video wherein, the column position of j presentation video, f (i, the j) gray-scale value of the pixel of the capable j row of expression i, f (i, the j+1) gray-scale value of the pixel of the capable j+1 row of expression i, g _V(i, j) the corresponding VG (vertical gradient) value of expression.

6. gradient horizontal projection: a kind of gradient of image and gray scale value in the two-dimensional space is transformed into method in the one-dimensional space after by a functional transformation, this functional transformation is $T_H\left(i\right)=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{g}_V(i,j),$ G wherein _V(i, j) the VG (vertical gradient) value of the capable j row of expression i, T _H(i) be the capable projection value of i.

7. Gaussian filter: a kind of famous smoothing filter, carry out Filtering Processing by a filter function transfer pair curve, can make curve more level and smooth, filter function is $T_H^{\&prime;}\left(i\right)=\frac{1}{k}\{T_H\left(i\right)+\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}[T_H(i-j)h(j,\mathrm{\&sigma;})+T_H(i+j)h(j,\mathrm{\&sigma;})\left]\right\}.$ T ' wherein _H(i) be filtered gradient horizontal projection value, the variation range of i is from 1 to n, the height of n representative image; W has represented the width size of smooth region, gets 8 herein; $h(j,\mathrm{\&sigma;})=\exp \left(\frac{-{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20061002079600201.png,C20061002079600202.png"\; state="\{\{state\}\}">j</figure-callout>}}^2}{2{\mathrm{\&sigma;}}^2}\right)$ Be Gaussian function, $k=2\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}h(j,\mathrm{\&sigma;})+1,$ σ represents the mean square deviation of gray level image; T _H(i) i gradient horizontal projection value of expression, T _H(i-j) expression (i-j) individual gradient horizontal projection value T _H(i+j) expression (i+j) individual gradient horizontal projection value.

8. crest: a kind of feature of the value on the curve; Curve values at this place is all bigger than previous curve values that is close to and next-door neighbour's a back curve values.

9. trough: a kind of feature of the value on the curve; Curve values at this place is all littler than previous curve values that is close to and next-door neighbour's a back curve values.

10. crest integrated value: the mathematical measure of the area in the zone under curve covers in a kind of coordinates computed axle representation; Specially refer to two areas between trough herein.

11. gradient horizontal projection curved scanning: a kind ofly search crest value or the bigger method of crest integrated value on the curve, concrete grammar is, seek crest from the starting point of curve, calculate this crest integrated value after finding crest, note the position of the bigger crest of the position of big crest value and crest integrated value simultaneously, so search terminal point up to curve.

12. license plate candidate area: some have the zone of some feature in piece image, these features are, at first, it one be rectangle or parallelogram with certain size size, to such an extent as to this rectangle or parallelogram can not occupy whole vehicle image too greatly, to such an extent as to can not too little naked eyes invisible, generally be rectangle about high 140mm or parallelogram zone about a long 440mm; Secondly, must there be character in the zone in the rectangle; At last, this rectangle is arranged in the middle and lower part of a vehicle image.

13. car plate coarse positioning computing: the simple operation of a kind of rough estimation car plate position, algorithm is: according to the position of the bigger crest value of the curve that provides or bigger crest integrated value search for this crest next-door neighbour about two wave trough position, the coordinate figure correspondence of left side trough be the position of the coboundary of car plate in a vehicle image, the position, coboundary of representing car plate herein with top_boundary, the right trough the coordinate figure correspondence be the position of the lower boundary of car plate in a vehicle image, represent the lower boundary position of car plate herein with bot_boundary.After the up-and-down boundary of car plate was determined, the rough position of car plate just can have been determined.Owing to be the computing of car plate coarse positioning, herein so the situation of a plurality of license plate candidate areas may occur.

14. fast wavelet transform (Mallat): be a kind of efficient calculation that realizes wavelet transform (DWT), the relation of the practicality between adjacent yardstick DWT coefficient has been found in this conversion.Concrete conversion principle is seen accompanying drawing 1.

15. normalization: being distributed in the interior functional value of different range by the mathematical operation of mathematical operation ruleization to a specified scope, this mathematical operation is $n(i,j)=\frac{f(i,j)-\min \left(f\right)}{\max \left(f\right)}*255,$ F (i wherein, j) the original functional value of expression, n (i, j) functional value after the expression normalization, minimum function value among min (f) the expression original function f, the maximal function value in the function f is willing in max (f) expression, and 255 expressions normalize to original functional value between 0 to 255, and 255 can change other numerical value into.

16. the corresponding sum operation of gradient: a kind of the Grad of same position is carried out the mathematical operation of addition, specially finger carries out the VG (vertical gradient) of three high fdrequency component LH, HL among the wavelet analysis result and HH the computing of addition herein; Calculating formula is: Dif_I (i, j)=| LH (i, j+1)-LH (i, j) |+| HL (i, j+1)-HL (i, j) |+| HH (i, j+1)-HH (i, j) |, LH (i, j) the Normalized Grey Level value of the capable j row of the i of expression LH subgraph wherein, LH (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression LH subgraph, HL (i, j) the Normalized Grey Level value of the capable j row of the i of expression HL subgraph, HL (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression HL subgraph, HH (i, j) the Normalized Grey Level value of the capable j row of the i of expression HH subgraph, HH (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression HH subgraph, Dif_I (i, j) the small echo subgraph difference summation of the capable j row of expression i.

17. car plate accurately cuts: specially refer to utilize the car plate position l_top that provides, l_bot, the computing that l_left and l_right accurately cut herein.This algorithm is, at first, defines a full matrix mask, and total columns of matrix mask is (l_right-l_left), and total line number of matrix is (l_bot-l_top).; Next matrix mask and license plate candidate area carry out point multiplication operation; Once more, fixed full null matrix la, its total columns is (l_right-l_left), its total line number is (l_bot-l_top).; At last the value of non-zero among the result of point multiplication operation is composed to matrix la, finally just obtained a license plate area la accurately.

18. gray scale upset: to the conversion that gray-scale map carries out, it is just passable to deduct each gray-scale value with 255, deceives to be converted into whitely, is converted into black in vain.

19. car plate color normalization: be meant the car plate of different background and character color is all become the license plate image of black matrix wrongly written or mispronounced character by a mathematical operation, its mathematical principle is, at first, and white pixel number and black picture element number in the statistics license plate area; Secondly compare black picture element number and white pixel number; Once more, if the white pixel number is most, then carry out the gray scale upset, if the black picture element number is majority then is left intact.

20. brightness strengthens: a kind of image processing means, be used for improving the black white contrast of piece image, its mathematical principle is: $R^{\&prime;}(i,j)=\left\{\begin{array}{ccc}R(i,j)& \mathrm{if}& R(i,j)>\mathrm{thres}\\ R(i,j)\&times;0.4& \mathrm{if}& R(i,j)\&le;\mathrm{thres}\end{array}\right.,$ Wherein ((thres is a threshold values to R ' to R for i, the j) gray-scale value of the car plate of the capable j row of i after the expression conversion, the minimum gradation value in bigger 20% pixel of the board gradation of image value of generally picking up the car for i, the j) gray-scale value of the capable j row of expression license plate image i.

According to of the present invention a kind of based on the vehicle license plate extraction method in the complex background of wavelet transformation, it comprises the following step:

Step 1. is installed on the appropriate location in highway crossing or parking lot with camera head, carries out image acquisition after in vehicle enters image pickup scope, obtains containing the original image of license plate image;

Resulting original image carries out gradation conversion in the step 2. pair step 1, obtains the gray level image that a width of cloth comprises car plate.Concrete grammar for adopt formula f (i, j)=0.114*I (i, j, 1)+0.587*I (i, j, 2)+0.299*I (i, j, 3) changes, the line position of i presentation video wherein, the column position of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of i in the gray level image after the expression conversion, * be the multiplying symbol, I (i, j, 1), I (i, j, 2) and I (i, j, 3) represent the R of the pixel of the capable j row of i in the coloured image respectively, G, the value of B component;

Resulting gray level image carries out the gray-scale map stretching in the step 3. pair step 2, and concrete grammar is for adopting formula $f\left(x\right)=\tanh \left(\frac{c}{\mathrm{\&sigma;}}x\right)$ Stretch, wherein x represents the gray-scale value of input, and the gray-scale value behind f (x) representation transformation, tanh are two tans; σ represents the mean square deviation of gray level image, and c represents a constant, is taken as 2.4 herein; Obtain a gray-scale value after the grey level stretching and be uniformly distributed in gray level image between 0～255;

Resulting gray level image carries out VG (vertical gradient) calculating in the step 4. pair step 3, obtains a vertical shade of gray figure who includes the vehicle gray level image of car plate; Concrete grammar is for adopting formula g _V(i, j)=| f (i, j+1)-f (i, j) | carry out VG (vertical gradient) and calculate, the line position of i presentation video wherein, the column position of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of expression i, f (i, the j+1) gray-scale value of the pixel of the capable j+1 row of expression i, g _V(i, j) the capable j of expression i row the VG (vertical gradient) value;

Resulting vertical shade of gray figure carries out the gradient horizontal projection in the step 5. pair step 4, obtains a coarse gradient horizontal projection curve; The computing formula of gradient horizontal projection is $T_H\left(i\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{g}_V(i,j),$ G wherein _V(i, j) the VG (vertical gradient) value of the capable j row of expression i, T _H(i) be the capable projection value of i;

Resulting coarse gradient horizontal projection curve carries out gaussian filtering in the step 6. pair step 5, obtains a level and smooth gradient horizontal projection curve.Filtering method is for adopting formula $T_H^{\&prime;}\left(i\right)=\frac{1}{k}\{T_H\left(i\right)+\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}[T_H(i-j)h(j,\mathrm{\&sigma;})+T_H(i+j)h(j,\mathrm{\&sigma;})\left]\right\}$ Carry out filtering.T ' wherein _H(i) be filtered gradient horizontal projection value, the variation range of i is from 1 to n, the height of n representative image; W has represented the width size of smooth region, gets 8 herein; $h(j,\mathrm{\&sigma;})=\exp \left(\frac{-{\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20061002079600201.png,C20061002079600202.png"\; state="\{\{state\}\}">j</figure-callout>}}^2}{2{\mathrm{\&sigma;}}^2}\right)$ Be Gaussian function, $k=2\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}h(j,\mathrm{\&sigma;})+1,$ σ represents the mean square deviation of gray level image; T _H(i) i gradient horizontal projection value of expression, T _H(i-j) expression (i-j) individual gradient horizontal projection value T _H(i+j) expression (i+j) individual gradient horizontal projection value;

Resulting level and smooth gradient horizontal projection curve carries out gradient horizontal projection curved scanning in the step 7. pair step 6, obtains the position coordinates of the bigger crest of bigger crest value or crest integrated value in level and smooth gradient horizontal projection curve; Concrete grammar is: seek crest from the starting point of curve, calculate this crest integrated value after finding crest, note the position coordinates of the bigger crest of the position of big crest value and crest integrated value simultaneously, so search the terminal point up to curve;

Step 8. utilizes that resulting crest location coordinate carries out the computing of car plate coarse positioning in the step 7, obtains the position coordinates of one or more license plate candidate areas in including the vehicle image of car plate.Concrete grammar is: the position coordinates of the crest that the bigger crest value of the curve that provides according to step 7 or crest integrated value are bigger search for this crest of stating next-door neighbour about the position coordinates of two troughs, the position coordinates correspondence of left side trough be the position coordinates of the coboundary of car plate in including the vehicle image of car plate, the coboundary position coordinates of representing car plate herein with top_boundary, the right trough the coordinate correspondence be the position coordinates of the lower boundary of car plate in including the vehicle image of car plate, represent the lower boundary position coordinates of car plate herein with bot_boundary;

Step 9. is utilized in the step 8 resulting positional information to carry out the car plate rough lumber and is cut processing, obtains one or more license plate candidate areas.Concrete grammar is, at first, the definition element is zero matrix mask, and the line number of mask is consistent with the line number and the columns of the vehicle gray level image matrix that includes car plate respectively with columns; Secondly, all be changed to 1 being positioned at top_boundary elements capable and bot_boundary all row in the ranks among the matrix mask, wherein top_boundary represents the coordinate figure of the coboundary of car plate in including the vehicle gray level image of car plate, and bot_boundary represents the coordinate figure of the lower boundary of car plate in vehicle image; Then, the element with same coordinate value in matrix mask and the vehicle gray level image matrix is multiplied each other, after multiplying each other, the element in the vehicle gray level image matrix in the non-license plate candidate area is zero, and the element value in the license plate candidate area remains unchanged; At last, the definition element is zero matrix I_cut, the line number of matrix I_cut is bot_boundary-top_boundary, columns and vehicle gray level image matrix column number equate, nonzero element in the matrix of consequence of matrix mask and vehicle gray level image matrix multiple is composed to I_cut, and matrix I_cut just represents license plate candidate area;

Resulting license plate candidate area carries out wavelet analysis in the step 10. pair step 9, obtains subimage LL, LH, HL and the HH of four license plate candidate areas.Concrete grammar is, adopts the haar small echo, utilizes the quick wavelet decomposition algorithm of Mallat that candidate's license plate area is carried out wavelet decomposition, and wherein, the quick wavelet decomposition process flow diagram of Mallat is seen accompanying drawing 1;

Subimage LH, the HL of resulting back three license plate candidate areas and HH carry out VG (vertical gradient) calculating in the step 11. pair step 10, obtain three VG (vertical gradient) images;

The gray-scale value of resulting three VG (vertical gradient) images carries out normalization in the step 12. pair step 11, and concrete grammar is for adopting formula $n(i,j)=\frac{f(i,j)-\min \left(f\right)}{\max \left(f\right)}*255$ Carry out normalization, f (i wherein, j) the original functional value of expression, n (i, the j) functional value after the expression normalization, the minimum function value among min (f) the expression original function f, the maximal function value in the function f is willing in max (f) expression, 255 expressions normalize to original functional value between 0 to 255, and 255 can change other numerical value into, obtain three normalized VG (vertical gradient) images after the normalized;

Resulting three normalized VG (vertical gradient) images carry out the corresponding sum operation of gradient in the step 13. pair step 12, obtain the Wavelet image difference summation of car plate.Concrete grammar is for adopting formula Dif_I (i, j)=| LH (i, j+1)-LH (i, j) |+| HL (i, j+1)-HL (i, j) |+| HH (i, j+1)-HH (i, j) | carry out the corresponding sum operation of gradient, wherein LH (i, j) the Normalized Grey Level value of the capable j row of the i of expression LH subgraph, LH (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression LH subgraph, HL (i, j) the Normalized Grey Level value of the capable j row of the i of expression HL subgraph, HL (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression HL subgraph, HH (i, j) the Normalized Grey Level value of the capable j row of the i of expression HH subgraph, HH (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression HH subgraph, Dif_I (i, j) the small echo subgraph difference summation of the car plate of the capable j row of expression i;

Step 14. is utilized the Wavelet image difference summation of license plate candidate area in the step 13, carries out car plate and accurately locatees, and obtains the exact position of car plate in vehicle image, i.e. four of car plate the coordinate figures of border in the vehicle image matrix.Concrete grammar is, from left to right, element value forms the zone of a bigger connected region in the searching Wavelet image difference summation matrix from top to bottom than big and these bigger element values, if the area of this connected region is in the certain value scope, think that then described connected region is exactly a car plate, note the coordinate figure on four borders up and down of this connected region, the coordinate figure on these four borders is exactly the coordinate figure of four borders in vehicle gray level image matrix of car plate, and be designated as top, bot, left and right, wherein top represents the abscissa value of coboundary in vehicle gray level image matrix of car plate, bot represents the abscissa value of lower boundary in vehicle gray level image matrix of car plate, left represents the ordinate value of left margin in vehicle gray level image matrix of car plate, and right represents the ordinate value of right margin in vehicle gray level image matrix of car plate;

Step 15. is utilized in the step 14 resulting car plate exact position to carry out car plate and is accurately cut, but is included the license plate image of identification number accurately.Concrete grammar is, at first, the definition element is one matrix A 1, and total columns of matrix A 1 is (right-left), and total line number of matrix A 1 is (bot-top).; Secondly the element correspondence with same coordinate value in matrix A 1 and the license plate candidate area matrix is multiplied each other; Then, define an element and be zero matrix A 0, its total columns is (right-left), and its total line number is (bot-top).; At last the nonzero element value in matrix A 1 and the license plate candidate area matrix multiple matrix of consequence is composed in the matrix A 0, but obtained a license plate image matrix A 0 that includes identification number accurately;

Resulting license plate image accurately carries out car plate color normalized in the step 16. pair step 15, but obtains a license plate image with black background and white number that includes identification number; Concrete grammar is at first, to add up white pixel number and black picture element number in the license plate area; Secondly compare black picture element number and white pixel number; Then, if the white pixel number is most, then carry out the gray scale upset, if the black picture element number is majority then is left intact;

In the step 17. pair step 16 resulting normalization license plate area carries out brightness and strengthens, but obtains the license plate image that includes identification number that a brightness has strengthened.Concrete grammar utilizes formula $R^{\&prime;}(i,j)=\left\{\begin{array}{ccc}R(i,j)& \mathrm{if}& R(i,j)>\mathrm{thres}\\ R(i,j)\&times;0.4& \mathrm{if}& R(i,j)\&le;\mathrm{thres}\end{array}\right.$ Strengthen, wherein R (i, j) gray-scale value of the capable j row of expression license plate image i, R ' (i, j) gray-scale value of the car plate of the capable j row of the i of expression after the conversion, thres is a threshold values, the minimum gradation value in bigger 20% pixel of the board gradation of image value of generally picking up the car.

By above step, we have just extracted the license plate image after strengthening from the original image that contains car plate.

Need to prove:

1., then no longer do gradation conversion and handle if the original vehicle image that includes car plate that collects in the step 2 is a gray level image.

2. the purpose of carrying out grey level stretching in the step 3 is in order to strengthen the contrast of the vehicle gray level image that includes car plate.

3. the calculating of the VG (vertical gradient) in the step 4 is because the pixel value in the license plate area changes soon and be concentrated, and more obviously concentrates on the VG (vertical gradient) aspect.

4. the gradient horizontal projection in the step 5 is because license plate area one is positioned to have in the zonule of quite high gradient density value, so we at first carry out the horizontal candidate region of horizontal projection location car plate.

5. not all crest that satisfies the crest condition is all adopted when searching for crest in the step 7, just think crest when only those values being satisfied the certain value scope greater than the area that certain value and crest covered, crest greater than the certain value scope that area satisfied that certain value and crest covered be the empirical value of testing gained.

6. in the step 8 in the computing of car plate coarse positioning, about search and the crest that satisfies condition next-door neighbour, only those troughs that satisfy the certain value scope are thought true trough during two troughs, because thereby the little fluctuating on the curve has caused the existence at false wave peak, the certain value scope that true trough satisfied is the empirical value that obtains by experiment.In addition, resulting in the coarse positioning computing is car plate rough position in including the vehicle image of car plate, owing to be the coarse positioning computing of car plate in this step, so the position coordinates of a plurality of license plate candidate areas might occur.

7. carry out among the result of wavelet analysis in the step 10, LL represents the low frequency subgraph picture among the wavelet decomposition result, it be one of former input license plate candidate area approximate, its size is 1/4th of a former input license plate candidate area, LH represents the high frequency subimage in vertical direction among the wavelet decomposition result, it has given prominence in the license plate candidate area high-frequency information in vertical direction, such as in the license plate candidate area horizontal linear, its size is 1/4th of a former input license plate candidate area, HL represents the high frequency subgraph in the horizontal direction among the wavelet decomposition result, it has given prominence in the license plate candidate area high ordinary mail breath in the horizontal direction, such as the vertical line in the license plate candidate area, its size is 1/4th of a former input license plate candidate area, HH represents the high frequency subimage on diagonal among the wavelet decomposition result, it has given prominence to the high-frequency information on diagonal in the former input license plate candidate area, as the oblique line in the license plate candidate area, its size is 1/4th of a former input license plate candidate area.

8. during handle the car plate exact position in the step 14, only big and these the bigger element values of element value can form a zone of satisfying the connected region of certain areal extent and just think license plate area in Wavelet image difference summation matrix, certain areal extent herein is the empirical value that obtains by experiment, if the value of some element is big but do not go into connected region then it is not a car plate in the matrix, if the connected region that some bigger values forms is less than certain areal extent then be not car plate, because the area of car plate in a car has a certain size, although when taking vehicle image because the angle of inclination of camera head, variation that distance is far and near and being uneven of road surface and caused the variation of car plate area, but it changes always within the specific limits, can be too not little, also not too large, utilize this information, just can get rid of those element values more greatly but the point that does not form connected region is the possibility of car plate, also got rid of simultaneously those element values big and formed connected region but the area of its connected region less than or be the possibility of car plate greater than the zone of certain areal extent, only find those element values to be only license plate area than the zone of area in the certain value the scope big and connected region that their form, if finally do not find the zone of satisfying certain areal extent yet, then think and do not contain car plate in the captured vehicle image, through this step processing, simultaneously those are thought car plate in coarse positioning is handled but in fact be not that get rid of in the zone of car plate, thereby obtain four coordinate figures of border in vehicle image of real car plate.

9. in the step 17, carry out purpose that brightness strengthens and be making cutting apart and discerning more or less freely of the follow-up number-plate number.

The present invention carries out car plate accurately and extracts by the high-frequency information in the wavelet analysis extraction license plate area; It is to adopt coarse positioning and accurately locate to combine, in accurately locating, adopt the way of wavelet analysis, and accurately carrying out the normalized of brightness and color behind the location, but finally obtained a car plate that includes identification number with unified background and number color.Adopt vehicle license plate extraction method of the present invention, the high-frequency information of extraction license plate area that not only can be as much as possible, thus accurately locate car plate, the environmental adaptation degree of method is improved greatly; And also improved the computing velocity of method in conjunction with the method for coarse positioning and fine positioning, satisfied the needs of real-time.

Innovation part of the present invention is:

1. (LH, HL HH) extract, and by outstanding these high-frequency informations of vertical difference summation Dif_I, just can find out license plate image accurately by the Special matrix that scans these high-frequency informations formations then to utilize wavelet analysis that high-frequency information is carried out in the candidate region.

2. propose car plate coarse positioning and the accurate method that combines of location, so that the speed of raising method, satisfied the needs of Vehicle License Plate Recognition System real-time.At first utilize the horizontal projection of the vertical gradient map of original image to extract a few license plate candidate area roughly, and then these candidate regions are used the minutia (being high-frequency characteristic) that wavelet analysis methods extract license plate image, and outstanding this feature, and then utilize the high-frequency characteristic bianry image of gained to carry out the accurate location of car plate.

Since last car plate position by the decision of high-frequency characteristic bianry image, and these high-frequency informations differ from a pixel at most with original image, therefore the accuracy of extraction is quite high, is convenient to subsequent treatment.

Description of drawings

Fig. 1 is the process flow diagram that fast wavelet transform decomposes;

Wherein (i, j) the original input picture of expression specially refers to license plate candidate area, h herein to f _ψ(-n) expression low-pass filter, so the haar small echo of Cai Yonging herein is h _ψThe concrete parameter of (-n) is set to [0.5,0.5], i.e. h _ψ(-n) comprises two parameters, and first parameter and second parameter are 0.5;

The expression Hi-pass filter, concrete parameter is set to [0.5,0.5], promptly Comprise two parameters, first parameter is that-0.5, second parameter is 0.5;

Expression is to carrying out promptly getting data every data on column direction every a sampling on the column direction of image data matrix through filtered image;

Expression is to carrying out promptly getting data every data on line direction every two sampling on the line direction of image data matrix through filtered image array; LL represents through the image on the low frequency direction after the wavelet decomposition, it be to original input image f (i, j) one approximate, its size is 1/4th of an original input image; LH represents that it has given prominence to the high-frequency information of original input image in vertical direction, and has kept the gradual information on the horizontal direction through the high frequency imaging on the vertical direction after the wavelet decomposition; HL represents that it has given prominence to original input image high-frequency information in the horizontal direction, and has kept the gradual information on the vertical direction through the high frequency imaging on the horizontal direction after the wavelet decomposition; HH represents that it has given prominence to the high-frequency information on the diagonal in the original image through the high frequency imaging on the diagonal after the wavelet decomposition.

Fig. 2 is the original image synoptic diagram that contains car plate;

Wherein mark the windshield of 1 expression vehicle, the bonnet of mark 2 expression vehicles, car lights are installed in mark 3 expressions, hang car plate and and the zone of bumper bar, mark 4 expression car plates, mark 5 expression wheels are in mark 4 represented car plates, capitalization A, B, C, D, E, F and G represent first of car plate respectively, second, the 3rd, the 4th, the 5th, the 6th and the 7th number, wherein the spacing of second number and the 3rd number is slightly larger.

Fig. 3 is the filtered horizontal projection curve of VG (vertical gradient);

The value of the numeric representation horizontal projection curve on the ordinate wherein, the coordinate figure of the numeric representation horizontal projection curve on the horizontal ordinate.

Fig. 4 is the license plate image synoptic diagram that finally obtains;

Wherein, A, B, C, D, E, F and G represent first of car plate respectively, second, the 3rd, the 4th, the 5th, the 6th and the 7th number.

Fig. 5 is the process flow diagram of the inventive method.

Embodiment:

Adopt method of the present invention, at first in the porch of highway, charge station and other any correct positions adopt the original image of camera head automatic shooting vehicle; Secondly the vehicle original image that photographs is input to adopt in the software of method of the present invention with the Matlab language compilation as former data and handles; But finally obtain the license plate image that a width of cloth includes identification number.Experiment adopts 353 colored vehicle images of taking as former data altogether on the spot, coarse positioning goes out 348, and locating accuracy is 98.58%, accurately orients 346, locating accuracy is 98.02%, only needs 50ms but locate the license plate image that a width of cloth includes identification number.

In sum, but utilize method of the present invention from the vehicle original image that is provided, to orient the license plate image that includes identification number fast and accurately.

Claims (1)

1, a kind of based on the vehicle license plate extraction method in the complex background of wavelet transformation, it comprises the following step:

Step 1. is installed on the appropriate location in highway crossing or parking lot with camera head, carries out image acquisition after in vehicle enters image pickup scope, obtains containing the original image of license plate image;

Resulting original image carries out gradation conversion in the step 2. pair step 1, obtains the gray level image that a width of cloth comprises car plate; Concrete grammar for adopt formula f (i, j)=0.114*I (i, j, 1)+0.587*I (i, j, 2)+0.299*I (i, j, 3) changes, the line position of i presentation video wherein, the column position of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of i in the gray level image after the expression conversion, * be the multiplying symbol, I (i, j, 1), I (i, j, 2) and I (i, j, 3) represent the R of the pixel of the capable j row of i in the coloured image respectively, G, the value of B component;

Resulting gray level image carries out the gray-scale map stretching in the step 3. pair step 2, and concrete grammar is for adopting formula $f\left(x\right)=\tanh \left(\frac{c}{\mathrm{\&sigma;}}x\right)$ Stretch, wherein x represents the gray-scale value of input, and the gray-scale value behind f (x) representation transformation, tanh are two tans; σ represents the mean square deviation of gray level image, and c represents a constant, is taken as 2.4 herein; Obtain a gray-scale value after the grey level stretching and be uniformly distributed in gray level image between 0～255, the image array after this grey level stretching is designated as A;

Gray level image in the step 4. pair step 3 after the resulting grey level stretching carries out VG (vertical gradient) and calculates, and obtains a vertical shade of gray figure who includes the vehicle gray level image of car plate; Concrete grammar is for adopting formula g _V(i, j)=| f (i, j+1)-f (i, j) | carry out VG (vertical gradient) and calculate, the line position of i presentation video wherein, the column position of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of expression i, f (i, the j+1) gray-scale value of the pixel of the capable j+1 row of expression i, g _V(i, j) the capable j of expression i row the VG (vertical gradient) value;

Resulting vertical shade of gray figure carries out the gradient horizontal projection in the step 5. pair step 4, obtains a coarse gradient horizontal projection curve; The computing formula of gradient horizontal projection is $T_H\left(i\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{g}_V(i,j),$ G wherein _V(i, j) the VG (vertical gradient) value of the capable j row of expression i, T _H(i) be the capable projection value of i;

Resulting coarse gradient horizontal projection curve carries out gaussian filtering in the step 6. pair step 5, obtains a level and smooth gradient horizontal projection curve; Filtering method is for adopting formula $T_H^{\&prime;}\left(i\right)=\frac{1}{k}\{T_H\left(i\right)+\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}[T_H(i-j)h(j,\mathrm{\&sigma;})+T_H(i+j)h(j,\mathrm{\&sigma;})\left]\right\}$ Carry out filtering; T ' wherein _H(i) be filtered gradient horizontal projection value, the variation range of i is from 1 to n, the height of n representative image; W has represented the width size of smooth region, gets 8 herein; $h(j,\mathrm{\&sigma;})=\exp \left(\frac{{-j}^2}{2{\mathrm{\&sigma;}}^2}\right)$ Be Gaussian function, $k=2\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}h(j,\mathrm{\&sigma;})+1,$ σ represents the mean square deviation of gray level image; T _H(i) i gradient horizontal projection value of expression, T _H(i-j) expression (i-j) individual gradient horizontal projection value, T _H(i+j) expression (i+j) individual gradient horizontal projection value;

Resulting level and smooth gradient horizontal projection curve carries out gradient horizontal projection curved scanning in the step 7. pair step 6, obtains crest satisfies the certain value scope greater than certain value and crest integrated value the position coordinates of crest in level and smooth gradient horizontal projection curve; Concrete grammar is: seek crest from the starting point of curve, calculate this crest integrated value after finding crest, note simultaneously greater than the position of the crest value of certain value and the position coordinates that the crest integrated value satisfies the crest of certain value scope, so search terminal point up to curve;

Step 8. utilizes that resulting crest location coordinate carries out the computing of car plate coarse positioning in the step 7, obtains the position coordinates of one or more license plate candidate areas in including the vehicle image of car plate; Concrete grammar is: the position coordinates that the crest of the curve that provides according to step 7 satisfies the crest of certain value scope greater than certain value and crest integrated value search for this crest of stating next-door neighbour about the position coordinates of two troughs, the position coordinates correspondence of left side trough be the position coordinates of the coboundary of car plate in including the vehicle image of car plate, the coboundary position coordinates of representing car plate herein with top_boundary, the right trough the coordinate correspondence be the position coordinates of the lower boundary of car plate in including the vehicle image of car plate, represent the lower boundary position coordinates of car plate herein with bot_boundary;

Step 9. is utilized in the step 8 resulting positional information to carry out the car plate rough lumber and is cut processing, obtains one or more license plate candidate areas; Concrete grammar is, at first, the definition element is zero matrix mask, and the line number of mask is consistent with the line number and the columns of matrix A respectively with columns; Secondly, all be changed to 1 being positioned at top_boundary elements capable and bot_boundary all row in the ranks among the matrix mask; Then, the element with same coordinate value in matrix mask and the matrix A is multiplied each other, after multiplying each other, the element in the matrix A in the non-license plate candidate area is zero, and the element value in the license plate candidate area remains unchanged; At last, the definition element is zero matrix I_cut, and the line number of matrix I_cut is bot_boundary-top_boundary, and the columns of columns and matrix A equates, nonzero element in matrix mask and the matrix A multiplied result matrix is composed to I_cut, and matrix I_cut just represents license plate candidate area;

Resulting license plate candidate area carries out wavelet analysis in the step 10. pair step 9, obtains subimage LL, LH, HL and the HH of four license plate candidate areas; Wherein LL represents the low frequency subgraph picture among the wavelet decomposition result, LH represents the high frequency subimage in vertical direction among the wavelet decomposition result, HL represents the high frequency subgraph in the horizontal direction among the wavelet decomposition result, and HH represents the high frequency subimage on diagonal among the wavelet decomposition result; Concrete grammar is that employing haar small echo utilizes the quick wavelet decomposition algorithm of Mallat that candidate's license plate area is carried out wavelet decomposition;

Subimage LH, the HL of resulting back three license plate candidate areas and HH carry out VG (vertical gradient) calculating in the step 11. pair step 10, obtain three VG (vertical gradient) images;

The gray-scale value of resulting three VG (vertical gradient) images carries out normalization in the step 12. pair step 11, and concrete grammar is for adopting formula $n(i,j)=\frac{f^{\&prime;}(i,j)-\min \left(f^{\&prime;}\right)}{\max \left(f^{\&prime;}\right)}*255$ Carry out normalization, f ' (i wherein, j) functional value before the expression normalization, n (i, j) functional value of expression after the normalization, function f before min (f ') the expression normalization ' in the minimum function value, function f before max (f ') the expression normalization ' in the maximal function value, multiply by 255 expressions the functional value before the normalization is normalized between 0 to 255, obtain three normalized VG (vertical gradient) images after the normalized;

Resulting three normalized VG (vertical gradient) images carry out the corresponding sum operation of gradient in the step 13. pair step 12, obtain the Wavelet image difference summation of license plate candidate area; Concrete grammar is for adopting formula Dif_I (i, j)=| LH (i, j+1)-LH (i, j) |+| HL (i, j+1)-HL (i, j) |+| HH (i, j+1)-HH (i, j) | carry out the corresponding sum operation of gradient, wherein LH (i, j) the Normalized Grey Level value of the capable j row of the i of expression LH subgraph, LH (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression LH subgraph, HL (i, j) the Normalized Grey Level value of the capable j row of the i of expression HL subgraph, HL (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression HL subgraph, HH (i, j) the Normalized Grey Level value of the capable j row of the i of expression HH subgraph, HH (i, j+1) the Normalized Grey Level value of the capable j+1 row of the i of expression HH subgraph, Dif_I (i, j) the Wavelet image difference summation of the license plate candidate area of the capable j row of expression i;

Step 14. is utilized the Wavelet image difference summation of license plate candidate area in the step 13, carries out car plate and accurately locatees, and obtains the exact position of car plate in vehicle image, i.e. four of car plate the coordinate figures of border in matrix A; Concrete grammar is, from left to right, element value forms the zone of a bigger connected region in the searching Wavelet image difference summation matrix from top to bottom than big and these bigger element values, if the area of this connected region is in the certain value scope, think that then described connected region is exactly a car plate, note the coordinate figure on four borders up and down of this connected region, the coordinate figure on these four borders is exactly four coordinate figures of border in matrix A of car plate, and be designated as top, bot, left and right, wherein top represents the abscissa value of coboundary in matrix A of car plate, bot represents the abscissa value of lower boundary in matrix A of car plate, and left represents the ordinate value of left margin in matrix A of car plate, and right represents the ordinate value of right margin in matrix A of car plate;

Step 15. is utilized in the step 14 resulting car plate exact position to carry out car plate and is accurately cut, but is included the license plate image of identification number accurately; Concrete grammar is, at first, the definition element is 1 matrix A 1, and total columns of matrix A 1 is (right-left), and total line number of matrix A 1 is (bot-top); Secondly the element correspondence with same coordinate value in matrix A 1 and the license plate candidate area matrix is multiplied each other; Then, define an element and be zero matrix A 0, its total columns is (right-left), and its total line number is (bot-top); At last the nonzero element value in matrix A 1 and the license plate candidate area matrix multiple matrix of consequence is composed in the matrix A 0, but obtained a license plate image matrix A 0 that includes identification number accurately;

Resulting license plate image accurately carries out car plate color normalized in the step 16. pair step 15, but obtains a license plate image with black background and white number that includes identification number; Concrete grammar is at first, to add up white pixel number and black picture element number in the license plate area; Secondly compare black picture element number and white pixel number; Then, if the white pixel number is most, then carry out the gray scale upset, if the black picture element number is majority then is left intact;

But the resulting license plate image with black background and white number that includes identification number carries out brightness and strengthens in the step 17. pair step 16, but obtains the license plate image that includes identification number that a brightness has strengthened; Concrete grammar utilizes formula $R^{\&prime;}(i,j)=\left\{\begin{array}{cc}R(i,j)& \mathrm{ifR}(i,j)>\mathrm{thres}\\ R(i,j)\&times;0.4& \mathrm{ifR}(i,j)\&le;\mathrm{thres}\end{array}\right.$ Strengthen, R (i wherein, j) but expression includes the gray-scale value of the capable j row of the license plate image i with black background and white number of identification number, R ' (i, j) gray-scale value of the car plate of the capable j row of i after the expression conversion, thres is a threshold values, but gets the minimum gradation value in 20% bigger pixel of the license plate image gray-scale value with black background and white number that includes identification number;

By above step, just from the original image that contains car plate, extracted the license plate image after strengthening.

Images