CN106384113A - Odd-and-even-line license plate character dividing method based on projection and template matching - Google Patents

Abstract

The invention provides an odd-and-oven-line license plate character dividing method based on projection and template matching, wherein the method relates to the field of machine vision. The odd-and-oven-line license plate character dividing method comprises the steps of performing binarization on a license plate image according to an improved blocked binarization method, and performing inverse color judgment according to proportion of pixels of which the pixel value is zero in a binarized image, thereby unifying a character background color; performing horizontal and vertical inclination correction of the license plate according to a rotation projection improved method; determining the structure type of the license plate through prior knowledge of the license plate, extracting upper boundaries and lower boundaries of character areas of an odd-line license plate and an even-line license plate according to different methods, constructing a license plate matching template and a matching degree, and performing optimal matching in a character width reference range for dividing the characters of the license plate; and finally performing fine dividing on the characters by means of a vertical projection method. The odd-and-oven-line license plate character dividing method can realize effective dividing on the characters of the odd-line license plate and the characters of the even-line license plate and furthermore settles a problem that the even-line license plate is mistakenly divided as the odd-line license plate in performing character dividing in prior art.

Description

Single-double-row registration number character dividing method based on projection and template matching

Technical field

The present invention relates to computer vision field, refer in particular to a kind of divide based on the single-double-row characters on license plate of projection and template matching Segmentation method.

Background technology

Increasing with vehicle, the phenomenon of overspeed of vehicle, overload and lost vehicles on highway is increasingly severe； Meanwhile, the charging aperture of highway, mainly using manual charge method, greatly reduces the utilization rate of highway.Automobile board According to the important symbol being automobile, it is the important evidence of automobile identification.If being capable of identify that car plate, traffic intersection will be greatly enhanced Intelligent traffic management, improve charge station work efficiency, thus improving the service efficiency of highway, and can improve reality When the efficiency that monitors.Vehicle License Plate Recognition System is mainly by License Plate, VLP correction, License Plate Character Segmentation and Recognition of License Plate Characters four Individual part.Wherein License Plate Character Segmentation is a key link of license plate recognition technology, and the accuracy of Character segmentation directly affects Final Car license recognition rate.

China's car plate is broadly divided into single file car plate and duplicate rows car plate.Car plate typically has seven characters, seven of single file car plate In same a line, in different rows, the character arrangements of different car plates increase character for the first two character of duplicate rows car plate and rear five characters The complexity of Character segmentation；Affected by actual photographed environment or weather, uneven illumination usually in license plate image, character takes off The problems such as color；The acquisition of license plate image is erected above road generally by fixation and the ccd video camera shooting of both sides obtains, The randomness of storing cycle position makes the vehicle image shooting produce tilt phenomenon, and this will have a strong impact on Character segmentation and knowledge Not.License Plate Character Segmentation common at present is primarily directed to single file car plate, often duplicate rows car plate can be mistakenly considered single file in segmentation Car plate, undesirable to the segmentation effect of duplicate rows car plate.

Towards single-double-row License Plate Character Segmentation, for the problem in existing License Plate Character Segmentation it is proposed that based on projection and The single-double-row registration number character dividing method of template matching.

Content of the invention

The technical problem to be solved in the present invention is：For this particular problem of License Plate Character Segmentation, propose one kind and be based on Projection and the single-double-row board character segmentation method of template matching.

For solving above-mentioned technical problem, the present invention adopts the technical scheme that：A kind of list based on projection and template matching Duplicate rows car plate method, comprises the steps：

Step 1) using improved piecemeal binarization method, binarization operation is carried out to license plate image, further according to binary picture As the pixel proportion that intermediate value is 0 carries out Judging Reverse Color, thus unified car plate background color；；

Step 2) car plate mixing slant correction is carried out by improved rotating projection method；

The license plate sloped correction calculation amount ratio of traditional rotating projection method is larger, and the present invention is thrown using a kind of improved rotation Image method carries out car plate mixing slant correction, first carries out horizontal tilt correction, is carrying out vertical skew correction.

The horizontal tilt correction of 2.1 car plates；

The vertical skew correction of 2.2 car plates.

Step 3) license plate structure is judged by the priori of car plate, extract car plate up-and-down boundary, construction car plate coupling mould Plate, by template matching method cutting characters on license plate, it includes：

3.1 judge license plate structure using car plate priori；

If 3.2 judge car plate for single file structure by 3.1, extract the up-and-down boundary of single file car plate, reconstruct single file car Board template, by template matching method cutting characters on license plate；

By step S3.1,3.3 judge that car plate, for duplicate rows structure car plate, is first carried out to duplicate rows structure car plate using projecting method Cutting, obtains upper part F of duplicate rows car plate₁With lower part F₂；Extract F more respectively₁And F₂Up-and-down boundary；Finally construct duplicate rows Plate template, by template matching method cutting characters on license plate.

Step 4) fine segmentation of line character is entered using upright projection method, optimize Character segmentation result further.

As the improvement further of technical solution of the present invention, described step 1) in, using improved piecemeal binarization method Binarization operation is carried out to license plate image, from wicket as minimum processing unit, by wicket with have formed objects 8 field windows merge into a big window, simultaneously with horizontal properties line replace whole big window in pixel transported Calculate, the pixel proportion being 0 further according to pixel value in binary image carries out Judging Reverse Color, thus unified car plate background face Color.

As the improvement further of technical solution of the present invention, in the horizontal tilt correction of described step 2.1 car plate, first to figure As carrying out horizontal difference, in the range of -30 °～30 °, increase by 2 degree successively and level side is carried out to image by linear interpolation To rotation, obtain bianry image using method in S1, then bianry image is carried out with floor projection and obtain floor projection value array p [m], wherein m are the line numbers of license plate image, obtain 4 maximum projection value sums and are designated as S₁, formula is as follows：

$S_1=\underset{k=1}{\overset{4}{\Σ}}p\[h\left(k\right)\]$

Wherein, h (k) is the position of four maximum projection values of array p [m],

Search out S₁Maximum when stop the rotation, then now the anglec of rotation is horizontal tilt angle, by corresponding Rotation obtains the image of level correction；

In the vertical skew correction of described step 2.2 car plate, using the method in S1 to the car plate after horizontal tilt correction Image carries out binarization operation, then carries out upright projection computing to binary image, sets projection the lowest point according to projection histogram Height threshold is t；Carry out vertical rotary more successively and obtain array q [n] using upright projection, wherein n is the row of license plate image Number, statistics upright projection array q [n] intermediate value is less than the array value sum of t, is designated as S₂, that is,：

$S_2=\underset{k=1}{\overset{n}{\&Sigma;}}q\&lsqb;k\&rsqb;<t$

In formula, k is that the value in upright projection array is less than the columns residing for t license plate image；

Work as S₂Reach maximum to stop the rotation, the anglec of rotation now is vertical tilt angle, carries out phase according to this angle The vertical tilt to eliminate car plate should be rotated.

As the improvement further of technical solution of the present invention, in described step 3.2, car is judged according to the priori of car plate Board structure type, extracts the license plate area after correction for single file structure car plate first with Sobel operator and horizontal projection method Up-and-down boundary, obtains the height H of character zone, then is mated by constructing car plate matching template and template matching degree, passes through Carry out best match to divide characters on license plate in reference character width.

As the improvement further of technical solution of the present invention, in described step 3.3, for the duplicate rows structure car judged Board, carries out horizontal cutting first with sciagraphy to duplicate rows structure car plate, obtains upper part F of car plate₁With lower part F₂；Divide again Indescribably take F₁And F₂Character up-and-down boundary, obtain F₁Character height H₁And F₂Character height H₂；Finally construct corresponding car Board template and matching template degree, by carrying out best match come a point characters on license plate in reference character width.

As the improvement further of technical solution of the present invention, described step 4) in, the step of the fine segmentation of described character As follows：

(1) first chinese character subdivision is cut：From the beginning of its right margin, about boundary, adjacent certain distance is hung down on the right Deliver directly shadow, the left side starting point of projective transformation be chosen as character right margin turn left character duration 1/3 at, the right of projective transformation is eventually Click the left margin being selected as second character, put from left to right and put from right to left after determining, scan for the lowest point from left to right, the lowest point determines New right margin for first character；The new left margin of first chinese character, upright projection are determined using same method Put at the 1/3 of character duration of turning left for the original left margin of character from left to right, the right terminal of upright projection is that character is original Left margin turn right character duration 1/3 at, equally from left to right scan for the lowest point, the lowest point is defined as the new of first character Left margin；

The subdivision of (2) second characters is cut：For the determination method of the new left margin of second character, in second character Originally carry out upright projection in adjacent certain distance about left margin, the left side starting point of upright projection is chosen as the right side of first character Border, the right starting point of upright projection be chosen as second character left margin turn right character duration 1/3 at, sweep from left to right Retouch searching the lowest point, the lowest point is the new left margin of second character；For the determination method of the new right margin of second character, Carry out upright projection in adjacent certain distance, the left side starting point of upright projection is chosen as character about the original right margin of two characters Right margin turn left character duration 1/3 at, the right terminal be chosen as right margin turn right character duration 1/3 at, same from a left side The lowest point is selected to determine the new right margin of second character to by scanning；

New right boundary is determined using the upright projection that same method carries out remaining character.

As the improvement further of technical solution of the present invention, described construction car plate matching template includes：According to single file structure The character depth-width ratio of car plate, the width of character is W ≈ H/2；In order to License Plate Character Segmentation is carried out with best match principle, set w For character reference width, in the range fromW is increased with a pixel every time, for each word Symbol width, the overall movement from left to right of template is mated, and the width between the second character and the 3rd character is 34/ 45w, the width between other characters is 4/15w；

Described template matching degree is：

The pixel total quantity of template character zone is：N_s=7wH

The pixel total quantity of interval region is：D_s=34/45wH+4/3wH

Wherein, N_u(u=1 ..., 7) it is character zone pixel quantity；D_u(u=1 ..., 6) it is inter-character space region picture Vegetarian refreshments quantity；w_eIt is the weight coefficient more than zero, scalable inter-character space width and character duration ratio；Template is mated with car plate Must be better, matching formula character portionIt is worth less, inter-character space partAlso less, then Match is also less；

For the template of different size character duration, translate template in license plate image horizontal direction, by smallest match degree Carry out license plate image Character segmentation.

As the improvement further of technical solution of the present invention, described construction plate template is：

F₁Middle character duration W₁≈4/3H₁, F₂Middle character duration W₂≈13/22H₂；

Set w₁For F₁Character reference width, in the range fromSet w₂For F₂Character reference Width, in the range fromw₁And w₂Increased with a pixel, for each character width every time Degree, the overall movement from left to right of template is mated, and F₁Character pitch be 3/4w₁, F₂Character pitch be 3/13w₂；

Described template matching degree is：

Part F on car plate₁Template matching degree Match₁With part F under car plate₂Matching degree Match₂Following institute respectively Show：

${\mathrm{Match}}_1=w_{e1}(1-\frac{1}{N_{s1}}\underset{u=1}{\overset{2}{\&Sigma;}}{N}_u^{\&prime;})+\frac{1}{D_{s1}}{D}_1^{\&prime;}$

${\mathrm{Match}}_2=w_{e2}(1-\frac{1}{N_{s2}}\underset{u=3}{\overset{7}{\&Sigma;}}{N}_u^{\&prime;})+\frac{1}{D_{s2}}\underset{u=2}{\overset{5}{\&Sigma;}}{D}_u^{\&prime;}$

Wherein, N_s1=2w₁H₁

D_s1=3/4w₁H₁

N_s2=5w₂H₂

D_s2=12/13w₂H₂

In formula, N'_uFor character zone pixel quantity；D₁' it is part inter-character space area pixel point quantity on car plate, D'_u For part inter-character space area pixel point quantity under car plate；N_s1And D_s1It is respectively part F on duplicate rows car plate₁Character zone picture The total quantity of the pixel in the total quantity of vegetarian refreshments and inter-character space region；N_s2And D_s2It is respectively part F under duplicate rows car plate₂Character The total quantity of the pixel in the total quantity of the pixel in region and inter-character space region；w_e1w_e2It is greater than zero weight coefficient, can Adjust inter-character space width and character duration ratio；

Template is mated better with car plate, matching formula character portionWith It is worth less, inter-character space partWithAlso less, then Match₁And Match₂Also less, for different size The template of character duration, translates template in license plate image horizontal direction, carries out license plate image Character segmentation by smallest match degree.

Compared with prior art, the invention has the advantages that：

1st, binarization operation is carried out to license plate image using improved piecemeal binarization method, from wicket as minimum Processing unit, by middle wide mouth with there are 8 field windows of formed objects merge into a big window, simultaneously special with level Levying line replaces the pixel in whole window to enter row operation, decreases amount of calculation with this, also to uneven illumination, the car such as low contrast Board image has preferable binaryzation effect.

2nd, carry out the horizontal and vertical tilt correction of license plate image using rotation projecting method, in rotation projection process not Need all of angle is calculated, only need to stop the rotation when finding the maximum of S1 or S2, using rotation now Angle is carrying out horizontal tilt correction or vertical skew correction.

3rd, license plate structure type is judged according to the priori of car plate, for single file structure car plate first with Sobel operator Extract the up-and-down boundary of the license plate area after correction with horizontal projection method, obtain the high H of character zone, then by constructing car plate Join template and template matching degree is mated, by carrying out best match in reference character width come a point characters on license plate；For The duplicate rows structure car plate judged, carries out horizontal cutting first with sciagraphy to duplicate rows structure car plate, obtains the top of car plate Divide F₁With lower part F₂；Extract F more respectively₁And F₂Character up-and-down boundary, obtain F₁Character height H₁And F₂Character height H₂；Finally construct corresponding plate template and matching template degree, by carrying out best match in reference character width come a point car Board character.

4th, carry out the structure of Character segmentation is entered with the fine segmentation of line character using upright projection method, improve character and divide The precision cut.

Brief description

Fig. 1 is the implementation case algorithm overall flow figure.

Algorithm single-double-row License Plate Character Segmentation result figure described in Fig. 2 embodiment.

Specific embodiment

, in conjunction with accompanying drawing, the present invention is described in further details, the present invention is a kind of based on projection taking license plate image as a example With the single-double-row registration number character dividing method of template matching, the method specifically includes following steps：

S1 carries out binarization operation using improved piecemeal binarization method to license plate image, further according in binary image It is worth the pixel proportion for 0 and carry out Judging Reverse Color, thus unified car plate background color；

(1) improved piecemeal binarization operation；

The present invention carries out binarization operation using improving piecemeal Binarization methods to license plate image, this improvement piecemeal binaryzation Using fritter window as minimal processing unit, the translation by covering, overlapping is processed the binaryzation threshold between window and window algorithm Value associates, it is to avoid the unified goal binaryzation being caused due to the difference of threshold value between block and block in direct piecemeal binaryzation The discontinuous phenomenon of result, in window, threshold value is obtained by the line statistics based on horizontal four split windows, reduces amount of calculation, place The license plate image of reason uneven illumination and low contrast also can obtain preferable binaryzation effect.Piecemeal schematic diagram is as shown in Fig. 2 vacation If desired for the binary-state threshold trying to achieve middle window 5, first it is merged into one with 8 field windows with same size size Individual big window, and calculate big window threshold value as the binary-state threshold of window 5, the wherein size of wicket is 1/4H, and H is car plate The height of image；Select three horizontal properties lines when calculating threshold value replace the pixel in whole big window to enter row operation simultaneously, Amount of calculation is effectively reduced with this.Taking obtain the binary-state threshold of Fig. 2 center window 5 as a example, it is improved that the present invention adopts Piecemeal Binarization methods step is as follows：

1) three horizontal linear l that big window is carried out horizontal 4 deciles are chosen₁、l₂And l₃；

2) calculate l respectively₁、l₂And l₃Article three, on straight line pixel average m₁、m₂And m₃, and standard deviation v₁、v₂And v₃；

3) then required threshold valueWherein N₁、N₂And N₃It is respectively three The points of pixel on bar horizontal line, according to threshold value T as window 5 binary-state threshold.

(2) Judging Reverse Color is carried out according to the pixel proportion that pixel value in binary image in (1) of S1 is 0, from And unification car plate background color.

Whole region ratio shared by background in license plate image is more than the ratio of whole region shared by target character, if binaryzation In image pixel value be 255 point proportion be more than 47%, then explanation background then need to carry out inverse process for light tone region, Otherwise, illustrate that background does not need to carry out inverse process for dark areas.Carrying out the unified inverse process formula of background color is：

G (x, y)=255-f (x, y) (1)

In formula, g (x, y) is the image after inverse conversion process, and f (x, y) carries out the binary image before inverse conversion.

S2 carries out car plate mixing slant correction by improved rotating projection method, and it includes：

The license plate sloped correction calculation amount ratio of traditional rotating projection method is larger, and the present invention is thrown using a kind of improved rotation Image method carries out car plate mixing slant correction, first carries out horizontal tilt correction, is carrying out vertical skew correction.

The horizontal tilt correction of S2.1 car plate；

In horizontal tilt trimming process, first horizontal difference is carried out to image, in the range of -300～300, increase by 2 successively Degree carries out horizontal direction rotation by linear interpolation to image, obtains bianry image using method in S1, then to bianry image Carry out floor projection and obtain floor projection value array p [m], wherein m is the line number of license plate image, obtain 4 maximum projection values Sum is designated as S₁, formula is as follows：

$S_1=\underset{k=1}{\overset{4}{\&Sigma;}}p\&lsqb;h\left(k\right)\&rsqb;---\left(2\right)$

Wherein, h (k) is the position of four maximum projection values of array p [m].

Search out S₁Maximum when stop the rotation, then now the anglec of rotation is horizontal tilt angle, by corresponding Rotation obtains the image of level correction.

The vertical skew correction of S2.2 car plate.

After horizontal tilt corrects, then carry out vertical skew correction.Vertical skew correction is mainly due to car plate figure The dislocation of same a line pixel of picture causes, and equally vertically can be inclined using horizontal tilt bearing calibration in similar S2.1 Tiltedly correct.Using the method in S1, binarization operation is carried out to the license plate image after horizontal tilt correction, then to binary image Carry out upright projection computing, projection the lowest point height threshold is set as t according to projection histogram.Carry out vertical rotary profit more successively Obtain array q [n] with upright projection, wherein n is the columns of license plate image, statistics upright projection array q [n] intermediate value is less than t's Array value sum, is designated as S₂, that is,：

$S_2=\underset{k=1}{\overset{n}{\&Sigma;}}q\&lsqb;k\&rsqb;<t---\left(3\right)$

In formula, k is that the value in upright projection array is less than the columns residing for t license plate image.

Work as S₂Reach maximum to stop the rotation, the anglec of rotation now is vertical tilt angle, carries out phase according to this angle The vertical tilt to eliminate car plate should be rotated.

S3 judges license plate structure by the priori of car plate, extracts car plate up-and-down boundary, constructs car plate matching template, leads to Cross template matching method cutting characters on license plate, it includes：

S3.1 judges license plate structure using car plate priori；

Judge that license plate structure judges that car plate is single file car according to the ratio that car plate priori is car plate length and width size Board or duplicate rows car plate.

S3.2 judges car plate for single file structure by step S3.1, then extract the up-and-down boundary of single file car plate, reconstruct single file Car plate matching template, by template matching method cutting characters on license plate.

If judging, car plate is single file structure, and single file License Plate Character Segmentation step is：

(1) up-and-down boundary extracts：Using hanging down of the license plate area after correcting in Sobel vertical direction operator detecting step S2 Straight edge, prominent characters on license plate region vertical edge feature, distinguish the impact such as car plate frame, then carry out floor projection and obtain Projection histogram, determines the up-and-down boundary of characters on license plate, it is highly designated as H by analyzing projection histogram.

(2) construct car plate matching template：Character duration, gap width equal proportion numerical value are also held in stability range substantially Interior, therefore construct a similar car plate template, character duration within certain zone of reasonableness, width increase from small to large, Other character pitches and character height are similarly scaled, and for each width, template integrally moves from left to right, with This substantially can outline characters on license plate.By draw in (1) of step S3.2 character up-and-down boundary height be H, according to single file The character depth-width ratio of structure car plate, the width of character is W ≈ H/2, and that is, the height of character is about the twice of character duration.In order to License Plate Character Segmentation is carried out with best match principle, sets w as character reference width, in the range fromW is every Secondary increased with a pixel, for each character duration, the overall movement from left to right of template is mated, and the Width between two characters and the 3rd character is 34/45w, and the width between other characters is 4/15w.

Obtain best match according to the template matching rule in following (3).

(3) template matching rule：Template moves on license plate image, and preferable position should be that square frame outlines character, word Accord with all of pixel all in square frame.To by step S1, binary conversion treatment is carried out to license plate image, then do upright projection, obtain Drop shadow curve, upright projection computing formula is as shown in (4).

$vp\&lsqb;j\&rsqb;=\underset{i=0}{\overset{H-1}{\&Sigma;}}value(i,j)---\left(4\right)$

Wherein, the upright projection when columns is for j for vp [j] the license plate binary image；Value (i, j) is pixel (i, j) The grey scale pixel value that position goes out；H is image character height；Seven character zones are followed successively by a₁,a₂,a₃,a₄,a₅,a₆,a₇；Character Between gap be d₁,d₂,d₃,d₄,d₅,d₆；Inter-character space is contour with character zone.

Character pixels point quantity in first character region and first character gap area pixel quantity are respectively such as Shown in formula (5) and (6).

$N_1=\underset{j=a_{11}}{\overset{a_{12}}{\&Sigma;}}\frac{vp\&lsqb;j\&rsqb;}{p}---\left(5\right)$

D₁=(d₁₂-d₁₁)·H (6)

Wherein, a₁₁And a₁₂It is respectively template character zone and character zone a to be matched during template matching₁Overlap Partial transverse axis beginning and end；d₁₁And d₁₂It is respectively template character gap area and word to be matched during template matching Symbol gap area d₁Lap transverse axis beginning and end；P=255 is bianry image pixel value normalization coefficient；

In the same manner, N can be tried to achieve according to formula (5)₂～N₇；

Gap area pixel quantity D can be tried to achieve in the same manner according to formula (6)₂～D₆；

The pixel total quantity of template character zone and interval region is as shown in formula (7) and (8) respectively.

N_s=7wH (7)

D_s=34/45wH+4/3wH (8)

The matching degree of template matching is：

Wherein, N_u(u=1 ..., 7) it is character zone pixel quantity；D_u(u=1 ..., 6) it is inter-character space region picture Vegetarian refreshments quantity；w_eIt is the weight coefficient more than zero, scalable character and effect played in template matching degree for the gap, the present invention Take w_e=12/45, i.e. inter-character space width and character duration ratio.Template is mated better with car plate, matching formula character portion PointIt is worth less, inter-character space partAlso less, then Match is also less.For different size The template of character duration, translates template in license plate image horizontal direction, carries out license plate image Character segmentation by smallest match degree.

By step S3.1, S3.3 judges that car plate, for duplicate rows structure car plate, is first carried out to duplicate rows structure car plate using projecting method Cutting, obtains upper part F of duplicate rows car plate₁With lower part F₂；Extract F more respectively₁And F₂Up-and-down boundary；Finally construct duplicate rows Plate template, by template matching method cutting characters on license plate.

If judging, car plate is duplicate rows structure, and duplicate rows License Plate Character Segmentation step is as follows：

(1) the horizontal cutting of duplicate rows car plate：Carry out two-value using the method in S1 using to the license plate image after slant correction Change operation, then floor projection is carried out from top to bottom to binary image, deduct the projection value of lastrow with next line projection value, and And recording its difference, then the maximum position of difference is exactly duplicate rows car plate two-part demarcation line up and down, and the top minute mark of car plate is F₁, bottom minute mark is F₂.

(2) up-and-down boundary extracts：Method upper part F to car plate respectively using (1) in step S3.2₁With lower part F₂ Carry out character boundary extraction, obtain Partial Height under the height of part and car plate on car plate, be designated as H respectively₁And H₂.

(3) construct plate template：It is respectively directed to upper part F of car plate₁With lower part F₂, using in (2) in step S3.2 Principle is constructing F₁And F₂Matching template.For F₁, according to the priori of duplicate rows car plate, by F₁Up-and-down boundary height Height H₁Can get F₁Middle character duration W₁≈4/3H₁.For F₂, according to the priori of duplicate rows car plate, by F₂Up-and-down boundary Height H₂Can get F₂Middle character duration W₂≈13/22H₂.Similar to single file characters on license plate coupling, in order to former with best match Then carry out License Plate Character Segmentation, set w₁For F₁Character reference width, in the range fromSet w₂For F₂ Character reference width, in the range fromw₁And w₂Increased with a pixel, for every every time One character duration, the overall movement from left to right of template is mated, and F₁Character pitch be 3/4w₁, F₂Character pitch For 3/13w₂.

Template matching rule according to (3) in step S3.3 obtains best match.

(3) template matching rule：According to determining of the pixel quantity in the character zone of (3) single file car plate in step S3.2 The definition of the inter-character space area pixel point quantity of justice and single file car plate, the pixel number in the character zone of duplicate rows car plate in the same manner Amount N₁'～N'₇With pixel quantity D in the inter-character space region of duplicate rows car plate₁'～D₅'₅.Using similar with (3) in step S3.2 Principle carry out template matching, then part F on car plate₁Template matching degree Match₁With part F under car plate₂Matching degree Match₂As shown in formula (10) and (11) respectively.

${\mathrm{Match}}_1=w_{e1}(1-\frac{1}{N_{s1}}\underset{u=1}{\overset{2}{\&Sigma;}}{N}_u^{\&prime;})+\frac{1}{D_{s1}}{D}_1^{\&prime;}---\left(10\right)$

${\mathrm{Match}}_2=w_{e2}(1-\frac{1}{N_{s2}}\underset{u=3}{\overset{7}{\&Sigma;}}{N}_u^{\&prime;})+\frac{1}{D_{s2}}\underset{u=2}{\overset{5}{\&Sigma;}}{D}_u^{\&prime;}---\left(11\right)$

Wherein, N_s1=2w₁H₁(12)

D_s1=3/4w₁H₁(13)

N_s2=5w₂H₂(14)

D_s2=12/13w₂H₂(15)

In formula, N'_u(u=1 ..., 7) it is character zone pixel quantity；D₁' it is part inter-character space region picture on car plate Vegetarian refreshments quantity；D'_u(u=1 ..., 4) it is part inter-character space area pixel point quantity under car plate；N_s1And N_s2Definition and step N in S3.2_sDefinition identical, D_s1And D_s2Definition and step S3.2 in D_sDefinition identical.N_s1And D_s1It is respectively duplicate rows car plate Upper part F₁The total quantity of the pixel of character zone and inter-character space region pixel total quantity；N_s2And D_s2It is respectively Part F under duplicate rows car plate₂The total quantity of the pixel of character zone and inter-character space region pixel total quantity；w_e1w_e2 Weight coefficient more than zero, scalable character and effect played in template matching degree for the gap, the present invention takes w_e1=3/4, w_e2 =3/13 is inter-character space width and character duration ratio.Template is mated better with car plate, matching formula character portionWithIt is worth less, inter-character space partWithAlso less, then Match₁And Match₂Also less.For the template of different size character duration, translate template in license plate image horizontal direction, lead to Cross smallest match degree and carry out license plate image Character segmentation.

S4 enters the fine segmentation of line character using upright projection method, optimizes Character segmentation result further.

Step S3 can get relatively accurate marginal position, may there are the feelings that character edge fails completely to outline Condition, needs to segment further and cuts, and mainly carries out upright projection by binary image, a left side for the character block splitting in S3 The a range of the lowest point of right margin is found out, is defined as the new right boundary of character.The step of character fine segmentation is：

(1) first chinese character subdivision is cut：From the beginning of its right margin, about boundary, adjacent certain distance is hung down on the right Deliver directly shadow, the left side starting point of projective transformation be chosen as character right margin turn left character duration 1/3 at, the right of projective transformation is eventually Click the left margin being selected as second character, put from left to right and put from right to left after determining, scan for the lowest point from left to right, the lowest point determines New right margin for first character；The new left margin of first chinese character, upright projection are determined using same method Put at the 1/3 of character duration of turning left for the original left margin of character from left to right, the right terminal of upright projection is that character is original Left margin turn right character duration 1/3 at, equally from left to right scan for the lowest point, the lowest point is defined as the new of first character Left margin.

The subdivision of (2) second characters is cut：For the determination method of the new left margin of second character, in second character Originally carry out upright projection in adjacent certain distance about left margin, the left side starting point of upright projection is chosen as the right side of first character Border, the right starting point of upright projection be chosen as second character left margin turn right character duration 1/3 at, sweep from left to right Retouch searching the lowest point, the lowest point is the new left margin of second character；For the determination method of the new right margin of second character, Carry out upright projection in adjacent certain distance, the left side starting point of upright projection is chosen as character about the original right margin of two characters Right margin turn left character duration 1/3 at, the right terminal be chosen as right margin turn right character duration 1/3 at, same from a left side The lowest point is selected to determine the new right margin of second character to by scanning.

New right boundary is determined using the upright projection that same method carries out remaining character.

The method proposing in the present invention actually can embed FPGA and realize, and apply to Car license recognition and video monitoring system In.Above example only plays the effect explaining technical solution of the present invention, and protection domain of the presently claimed invention is not limited to System and specific implementation step is realized described in above-described embodiment.Therefore, only to formula specific in above-described embodiment and algorithm Simply replaced, but still consistent with the method for the invention technical scheme of its flesh and blood, all should belong to the present invention's Protection domain.

Claims (8)

1. a kind of single-double-row registration number character dividing method based on projection and template matching, is characterized in that, comprise the steps：

Step 1) using improved piecemeal binarization method, binarization operation is carried out to license plate image, further according in binary image It is worth the pixel proportion for 0 and carry out Judging Reverse Color, thus unified car plate background color；

Step 2) car plate mixing slant correction is carried out by improved rotating projection method, it includes：

The horizontal tilt correction of 2.1 car plates；

The vertical skew correction of 2.2 car plates；

Step 3) license plate structure is judged by the priori of car plate, extract car plate up-and-down boundary, construct car plate matching template, lead to Cross template matching method cutting characters on license plate, it includes：

3.1 judge license plate structure using car plate priori；

If 3.2 steps S3.1 judge car plate for single file structure, extract the up-and-down boundary of single file car plate, reconstruct single file car plate Join template, by template matching method cutting characters on license plate；

If 3.3 steps S3.1 judge that car plate, for duplicate rows structure car plate, first carries out cutting using projecting method to duplicate rows structure car plate, Obtain upper part F of duplicate rows car plate₁With lower part F₂；Extract F more respectively₁And F₂Up-and-down boundary；Finally construct duplicate rows car plate mould Plate, by template matching method cutting characters on license plate；

Step 4) fine segmentation of line character is entered using upright projection method, optimize Character segmentation result further.

2. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 1, described Step 1) in, binarization operation is carried out to license plate image using improved piecemeal binarization method, from wicket as minimum Processing unit, by wicket with there are 8 field windows of formed objects merge into a big window, simultaneously with horizontal properties Line replaces the pixel in whole big window to enter row operation, the pixel institute accounting being 0 further according to pixel value in binary image Example carries out Judging Reverse Color, thus unified car plate background color.

3. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 1, described Step 2) in,

In the horizontal tilt correction of described step 2.1 car plate, first horizontal difference is carried out to image, in the range of -30 °～30 °, Increase by 2 degree successively and horizontal direction rotation is carried out to image by linear interpolation, obtain bianry image using method in S1, more right Bianry image carries out floor projection and obtains floor projection value array p [m], and wherein m is the line number of license plate image, obtains maximum 4 Individual projection value sum is designated as S₁, formula is as follows：

Wherein, h (k) is the position of four maximum projection values of array p [m],

Search out S₁Maximum when stop the rotation, then now the anglec of rotation is horizontal tilt angle, by rotating accordingly Image to level correction；

In the vertical skew correction of described step 2.2 car plate, using the method in S1 to the license plate image after horizontal tilt correction Carry out binarization operation, then upright projection computing is carried out to binary image, projection the lowest point height is set according to projection histogram Threshold value is t；Carry out vertical rotary more successively and obtain array q [n] using upright projection, wherein n is the columns of license plate image, system Meter upright projection array q [n] intermediate value is less than the array value sum of t, is designated as S₂, that is,：

In formula, k is that the value in upright projection array is less than the columns residing for t license plate image；

Work as S₂Reach maximum to stop the rotation, the anglec of rotation now is vertical tilt angle, is accordingly rotated according to this angle To eliminate the vertical tilt of car plate.

4. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 1, described In step 3.2, license plate structure type is judged according to the priori of car plate, for single file structure car plate first with Sobel operator Extract the up-and-down boundary of the license plate area after correction with horizontal projection method, obtain the height H of character zone, then by constructing car plate Matching template and template matching degree are mated, by carrying out best match in reference character width come a point characters on license plate.

5. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 4, described In step 3.3, for the duplicate rows structure car plate judged, first with sciagraphy, horizontal cutting is carried out to duplicate rows structure car plate, Obtain upper part F of car plate₁With lower part F₂；Extract F more respectively₁And F₂Character up-and-down boundary, obtain F₁Character height H₁ And F₂Character height H₂；Finally construct corresponding plate template and matching template degree, by carrying out in reference character width Good coupling is dividing characters on license plate.

6. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 1, described Step 4) in, the step of the fine segmentation of described character is as follows：

(1) first chinese character subdivision is cut：From the beginning of its right margin, about boundary, adjacent certain distance is vertically thrown on the right Shadow, the left side starting point of projective transformation be chosen as character right margin turn left character duration 1/3 at, the right terminal choosing of projective transformation It is selected as the left margin of second character, from left to right point and after point determines from right to left, scan for the lowest point from left to right, the lowest point is defined as the The new right margin of one character；The new left margin of first chinese character, a left side for upright projection are determined using same method Starting point be the original left margin of character turn left character duration 1/3 at, the right terminal of upright projection is the original left side of character Boundary turn right character duration 1/3 at, equally from left to right scan for the lowest point, the lowest point is defined as the new left side of first character Boundary；

The subdivision of (2) second characters is cut：For the determination method of the new left margin of second character, original in second character Carry out upright projection in adjacent certain distance, the left side starting point of upright projection is chosen as the right of first character about left margin Boundary, the right starting point of upright projection be chosen as second character left margin turn right character duration 1/3 at, scan from left to right Find the lowest point, the lowest point is the new left margin of second character；For the determination method of the new right margin of second character, second Carry out upright projection, the left side starting point of upright projection is chosen as the character right side in adjacent certain distance about the original right margin of individual character Border turn left character duration 1/3 at, the right terminal be chosen as right margin turn right character duration 1/3 at, same from a left side to The lowest point is selected to determine the new right margin of second character by scanning；

New right boundary is determined using the upright projection that same method carries out remaining character.

7. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 4, described Construction car plate matching template includes：According to the character depth-width ratio of single file structure car plate, the width of character is W ≈ H/2；In order to Good matching principle carries out License Plate Character Segmentation, sets w as character reference width, in the range fromW every time with One pixel is increased, and for each character duration, the overall movement from left to right of template is mated, and the second word Width between symbol and the 3rd character is 34/45w, and the width between other characters is 4/15w；

Described template matching degree is：

The pixel total quantity of template character zone is：N_s=7wH

The pixel total quantity of interval region is：D_s=34/45wH+4/3wH

Wherein, N_u(u=1 ..., 7) it is character zone pixel quantity；D_u(u=1 ..., 6) it is inter-character space area pixel point Quantity；w_eIt is the weight coefficient more than zero, scalable inter-character space width and character duration ratio；Template is mated with car plate and must be got over Good, matching formula character portionIt is worth less, inter-character space partAlso less, then Match also gets over Little；

For the template of different size character duration, translate template in license plate image horizontal direction, carried out by smallest match degree License plate image Character segmentation.

8. a kind of single-double-row registration number character dividing method based on projection and template matching according to claim 7, described Constructing plate template is：

F₁Middle character duration W₁≈4/3H₁, F₂Middle character duration W₂≈13/22H₂；

Set w₁For F₁Character reference width, in the range fromSet w₂For F₂Character reference width, In the range fromw₁And w₂Increased with a pixel, for each character duration, mould every time The overall movement from left to right of plate is mated, and F₁Character pitch be 3/4w₁, F₂Character pitch be 3/13w₂；

Described template matching degree is：

Part F on car plate₁Template matching degree Match₁With part F under car plate₂Matching degree Match₂As follows respectively：

Wherein, N_s1=2w₁H₁

D_s1=3/4w₁H₁

N_s2=5w₂H₂

D_s2=12/13w₂H₂

In formula, N'_uFor character zone pixel quantity；D′₁For part inter-character space area pixel point quantity, D' on car plate_uFor car Part inter-character space area pixel point quantity under board；N_s1And D_s1It is respectively part F on duplicate rows car plate₁Character zone pixel Total quantity and inter-character space region pixel total quantity；N_s2And D_s2It is respectively part F under duplicate rows car plate₂Character zone The total quantity of pixel and inter-character space region pixel total quantity；w_e1w_e2It is greater than zero weight coefficient, scalable Inter-character space width and character duration ratio；

Template is mated better with car plate, matching formula character portionWithValue is got over Little, inter-character space partWithAlso less, then Match₁And Match₂Also less, for different size character The template of width, translates template in license plate image horizontal direction, carries out license plate image Character segmentation by smallest match degree.