CN103065137B - A kind of license plate character recognition method - Google Patents

Abstract

The present invention proposes a kind of method of new Recognition of License Plate Characters, adopt the method for canny algorithm and bianry image combination to carry out identification character. And when identification is owing to having extracted edge and the saltus step information of character, then the saltus step of character edge pixel is done and mated with the saltus step of template character set, find the template character that matching degree is the highest, and then the recognition result of acquisition character. This character recognition based on saltus step, has solved character identification problem under various interference preferably, can effectively overcome various disturbing factors, keeps comparatively stable high discrimination. And, because the canny algorithm to standard is simplified, greatly improve the efficiency of identification, reduce the system resource taking, improve recognition speed.

Description

A kind of license plate character recognition method

Affiliated technical field:

Patent of the present invention relates to a kind of character identifying method in Vehicle License Plate Recognition System, belongs to technical field of image processing.

Background technology:

Along with the fast development of Chinese national economy, the requirement of control of traffic and road is progressively improved to intelligent transportation systemArise at the historic moment. In intelligent transportation system, after accurate positioning licence plate and Character segmentation, carry out character by Vehicle License Plate Recognition SystemIdentification just finally completes complete car plate identification, and therefore the quality of character recognition plays vital work to character identification rateWith. Therefore, as the important component part of intelligent transportation system, Vehicle License Plate Recognition System has obtained fast development, is learned both at home and abroadPerson's broad research. Vehicle License Plate Recognition System is divided into image acquisition, car plate location, Character segmentation, character recognition four parts, wherein characterIdentification is the Research Emphasis of the each enterprise of the industry. Aspect character feature extraction, need to suppress various interference, such as uneven illuminationInconsistent and the situation such as fracture or adhesion of the rotation of even, character and deformation, stroke weight, just can extract stable character spyLevy, finally complete character recognition.

In prior art, common threshold obtains bianry image, and in the time that light is inhomogeneous, discrimination is not high, and originalAlgorithm steps is more, and the system resource taking while identification is too high, and recognition speed is not high; Meanwhile, for because turn inside diameter,That shooting angle etc. cause is license plate sloped, the fracture of characters on license plate stroke, character are stained, low light according to time noise etc., all can cause knowledgeRate does not decline.

Summary of the invention:

In order to improve the discrimination of character, the present invention proposes a kind of method of new Recognition of License Plate Characters, adopt cannyThe method of algorithm and bianry image combination is carried out identification character, and the canny algorithm of standard is simplified, specific as follows:

First license plate grey level image is divided into some character zones, to the gray level image I of each character zone_charPressProcess according to following steps order:

A, to I_charAsk adaptive threshold;

B, to I_charAdopt Canny algorithm to determine character edge image I_edge；

C, to above-mentioned character edge image I_edgeCarry out filling cavity;

D, again determine above-mentioned character edge image I_edgeBorder;

E, to above-mentioned I_edgeBe normalized;

The horizontal hopping sequences S at F, calculating character edge_HoriWith vertical transition sequence S_Vert, and according to above-mentioned horizontal saltus stepSequence S_HoriWith vertical transition sequence S_VertDetermine upper number of transitions density Hop_Density_Top, lower number of transitions density Hop_Density_Bottom, left number of transitions density Hop_Density_Left, right number of transitions density Hop_Density_Right, and further trueFixed upper and lower number of transitions density ratio Hop_Density_Top/BottomWith left and right number of transitions density ratio Hop_Density_Left/Right；

G, by above-mentioned F step determine above-mentioned horizontal hopping sequences S_Hori, vertical transition sequence S_Vert, upper and lower number of transitionsDensity ratio Hop_Density_Top/BottomWith left and right number of transitions density ratio Hop_Density_Left/RightWith from Character mother plate concentrate readThe horizontal hopping sequences SM of the current character template of getting_Hori, vertical transition sequence SM_Vert, upper and lower number of transitions density ratio Hop_Density_M_Top/Bottom, left and right number of transitions density ratio Hop_Density_M_Left/RightMate one by one, find matching degreeHigh character index value, completes character recognition.

Adopt method of the present invention to process characters on license plate gray level image, owing to having extracted edge and the jumping of characterChange information, then the saltus step of character edge pixel is done and mated with the saltus step of template character set, the template word that matching degree is the highest foundSymbol, and then the recognition result of acquisition character. This character recognition based on saltus step, has solved preferably character under various interference and has knownOther problem, can effectively overcome various disturbing factors, keeps comparatively stable high discrimination. And, due to the canny to standardAlgorithm is simplified, and has greatly improved the efficiency of identification, has reduced the system resource taking, and has improved recognition speed.

Detailed description of the invention:

Method of the present invention is first license plate grey level image to be carried out to Character segmentation, and license plate image is divided into several charactersRegion, to the gray level image I of each character zone_charRepeat following operation, finally obtain each character in predetermined template wordAccord with concentrated index value Index_m, complete the identification to all characters with this. Wherein, predetermined masterplate character set can be according to needThe characters on license plate of identifying be the situations such as numeral, letter, Chinese character, font self-defined (for example, can comprise each province be called for short, 26English alphabet capital and small letter, numeral, can also set other words or other language as required). Specifically to gray level image I_charTreatment step as follows:

1. couple I_charAsk adaptive threshold T

2. couple I_charAdopt Canny algorithm to obtain character edge image I_edge

1) adopt the template of 3*3, ask that current pixel point P is upper and lower, the gradient G of left and right, two clinodiagonals₁、G₂、G₃、G₄，The corresponding operator of each gradient direction is respectively H₁、H₂、H₃、H₄：

$\begin{array}{cccc}{H}_1=\left[\begin{array}{ccc}0& -1& 0\\ 0& 2& 0\\ 0& -1& 0\end{array}\right]& H_2=\left[\begin{array}{ccc}0& 0& 0\\ -1& 2& -1\\ 0& 0& 0\end{array}\right]& H_3=\left[\begin{array}{ccc}-1& 0& 0\\ 0& 2& 0\\ 0& 0& -1\end{array}\right]& H_4=\left[\begin{array}{ccc}0& 0& -1\\ 0& 2& 0\\ -1& 0& 0\end{array}\right]\end{array}$

2) find maximum gradient G_max＝MAX(G₁，G₂，G₃，G₄), and with mode mark P point below:

if((G_max≥MIN_EDGE_GRADIENT)&&(Gray(P)≥T))

{

P is marginal point, at I_edgeIn be labeled as 1;

else

{

P is non-marginal point, at I_edgeIn be labeled as 0;

According to G_max＝MAX(G₁，G₂，G₃，G₄) find the greatest gradient of current pixel point P, and judge current pixel point PWhether meet the gray value that greatest gradient is greater than predetermined minimum marginal value and current pixel point simultaneously and whether be greater than above-mentioned self adaptationThreshold value, if it is mark is judged to be marginal point, I_edgeBe labeled as 1; If not being judged to be not to be marginal point, I_edgeMarkBe 0.

The gray value that wherein Gray (P) is ordered for P, MIN_EDGE_GRADIENT is defaulted as 15, can be according to captured imageThe difference of type is got different values.

3) get back to step 1), repeat this process, until travel through completely, finally obtain edge image I_edge。

3. pair character edge image I_edgeFilling cavity, the impact of character saltus step being calculated with cancellation. Be specially

1) at I_edgeIn, obtain current pixel point P_ij；

2) if Gray is (P_ij)==0, with P_ijCentered by get the window of 3 × 3, pixel value forms square in windowBattle array W_3×3, definition operator $H=\left[\begin{array}{ccc}1& 1& 1\\ 1& 0& 1\\ 1& 1& 1\end{array}\right],$ Ask the scalar product C of the two_ij＝W_3×3H, works as C_ij≥T_c(T_c=6), justI_edgeMiddle by P_ijBe set to marginal point.

3) get back to step 1, repeat this process, until travel through complete.

4. recalculate the border of character edge image, the object of doing is like this impact in order to eliminate noise at the boundary

1) look for I_edgeUpper marginal position index Index_Top

Method: row traversal from top to bottom, the Num until count in row edge_Hori≥T_num，T_num＝1；

2) look for I_edgeLower limb location index Index_Bottom

Method: row traversal from top to bottom, the Num until count in row edge_Hori≥T_num，T_num＝1；

3) look for I_edgeLeft hand edge location index Index_Left

Method: from left to right row traversal, until column border points N um_Vert≥T_num，T_num＝1；

4) look for I_edgeRight hand edge location index Index_Right

Method: row traversal from right to left, until column border points N um_Vert≥T_num，T_num＝1。

5. couple I_edgeDo normalized, to reduce the requirement to Character segmentation precision

1)Index_Top、Index_Bottom、Index_Left、Index_RightFour determined images in border areI_{divided_edge}, by I_{divided_edge}Do size normalization, its wide height is equated with the wide height of Character mother plate, obtain I_{normal_edge}；

2) record I_{normal_edge}Height H eight, width is Width.

3)

6. the horizontal and vertical hopping sequences S at calculating character edge_HoriAnd S_Vert, and obtain upper and lower number of transitions density Hop_Density_Top、Hop_Density_Bottom, left and right number of transitions density Hop_Density_Left、Hop_Density_Right, jump up and downParameter density ratio Hop_Density_Top/BottomWith left and right number of transitions density ratio Hop_Density_Left/Right

1) obtain horizontal hopping sequences S_Hori

Method: row traversal I from top to bottom_{normal_edge}All row, for the capable R of j_jIf there is some P_i(P_i-1＝0，P_i＝1), R_jOn number of transitions Hop_jAdd 1 (Hop_jInitial value is 0), travel through so all row (R₁，R₂，..，R_j，..，R_Height)，Obtain horizontal hopping sequences S_Hori＝{Hop₁，Hop₂，..，Hop_j，...，Hop_Height}；

2) obtain vertical transition sequence S_Vert

Method: from left to right row traversal I_{normal_edge}All row, for j row C_jIf there is some P_i(P_i-1＝0，P_i＝1), C_jOn number of transitions Hop_jAdd 1 (Hop_jInitial value is 0), travel through so all row (C₁，C₂，..，C_j，..，C_Width)，Obtain vertical transition sequence S_Vert＝{Hop₁，Hop₂，..，Hop_j，...，Hop_Width}；

3) obtain upper and lower number of transitions density Hop_pensity_Top、Hop_Density_Bottom

(k default value is Height/2, Hop_j∈S_Hori)

$Hop\_{\mathrm{Density}}_{Bottom}=\frac{1}{h-k}\underset{j=k+1}{\overset{h}{\Σ}}{\mathrm{Hop}}_j$

(k default value is Height/2, h=Height, Hop_j∈S_Hori)

4) obtain left and right number of transitions density Hop_Density_Left、Hop_Density_Right

(k gets default value Width/2, Hop_j∈S_Vert)

$Hop\_{\mathrm{Density}}_{Right}=\frac{1}{h-k}\underset{j=k+1}{\overset{h}{\Σ}}{\mathrm{Hop}}_j$

(k gets default value Width/2, h=Width, Hop_j∈S_Vert)

5) upper and lower number of transitions density ratio:

Hop_Density_Top/Bottom＝Hop_Density_Top/Hop_Density_Bottom

Left and right number of transitions density ratio:

Hop_Density_Left/Right＝Hop_Density_Left/Hop_Density_Rigth

7, all characters of Character mother plate collection do and mate one by one, find the Character mother plate index value that matching degree is the highest

1) concentrate and obtain current character template Tamplate at Character mother plate_i(1≤i≤n, n is template set largest indexValue), obtain its horizontal hopping sequences SM_Hori, vertical transition sequence SM_Vert, upper and lower number of transitions density ratio Hop_Density_M_Top/Bottom, left and right number of transitions density ratio Hop_Density_M_Left/Right；

2)S_HoriWith SM_HoriDo and mate, obtain matching degree M₁；

3)S_VertWith SM_VertDo and mate, obtain matching degree M₂；

4)Hop_Density_Top/BottomWith Hop_Density_M_Top/BottomDo and mate, obtain mating M₃；

5)Hop_Density_Left/RightWith Hop_Density_M_Left/RightDo and mate, obtain matching degree M₄；

6) calculate total matching degree Mt_i：Mt_i＝w₁M₁+w₂M₂+w₃M₃+w₄M₄

Wherein, w₁、w₂、w₃、w₄Be respectively M₁、M₂、M₃、M₄Weight;

7) get back to step 1, repeat this process, until travel through complete;

8) get Index_mFor { Mt₁，Mt₂，…，Mt_i，…，Mt_nIn peaked index value, Index_mBe final identificationResult. So far, this character recognition is complete.

Claims (5)

1. a license plate character recognition method, is divided into some character zones by license plate grey level image, to each character zoneGray level image I_charOrder is processed in accordance with the following steps:

A, to I_charAsk adaptive threshold;

B, to I_charAdopt Canny algorithm to determine character edge image I_edge；

C, to above-mentioned character edge image I_edgeCarry out filling cavity;

D, again determine above-mentioned character edge image I_edgeBorder;

E, to above-mentioned I_edgeBe normalized;

The horizontal hopping sequences S at F, calculating character edge_HoriWith vertical transition sequence S_Vert, and according to above-mentioned horizontal hopping sequencesS_HoriWith vertical transition sequence S_VertDetermine upper number of transitions density Hop_Density_Top, lower number of transitions density Hop_Density_Bottom, left number of transitions density Hop_Density_Left, right number of transitions density Hop_Density_Right, and further trueFixed upper and lower number of transitions density ratio Hop_Density_Top/BottomWith left and right number of transitions density ratio Hop_Density_Left/Right；

G, by above-mentioned F step determine above-mentioned horizontal hopping sequences S_Hori, vertical transition sequence S_Vert, upper and lower number of transitions densityCompare Hop_Density_Top/BottomWith left and right number of transitions density ratio Hop_Density_Left/RightWith from Character mother plate concentrate readThe horizontal hopping sequences SM of current character template_Hori, vertical transition sequence SM_Vert, upper and lower number of transitions density ratio Hop_Density_M_Top/Bottom, left and right number of transitions density ratio Hop_Density_M_Left/RightMate one by one, find matching degreeHigh character index value, completes character recognition.

2. license plate character recognition method as claimed in claim 1, is further included in and in above-mentioned steps B, adopts canny algorithmObtain character edge image, be specially:

B1, adopt the template of 3*3, ask that current pixel point P is upper and lower, the gradient G of left and right, two clinodiagonals₁、G₂、G₃、G₄, each ladderThe corresponding operator of degree direction is respectively H₁、H₂、H₃、H₄：

According to G_max＝MAX(G₁，G₂，G₃，G₄) find the greatest gradient of current pixel point P, and judge that whether current pixel point P is sameIn time, meets the gray value that greatest gradient is greater than predetermined minimum marginal value and P and whether is greater than above-mentioned adaptive threshold, if it is markNote is judged to be marginal point, I_edgeBe labeled as 1; If not being judged to be not to be marginal point, I_edgeBe labeled as 0;

B2, get back to step B1, repeat this process, until travel through completely, finally obtain edge image I_edge。

3. license plate character recognition method as claimed in claim 2, further comprises in above-mentioned steps C character edge imageI_edgeThe method of filling cavity is:

C1, at I_edgeIn, obtain current pixel point P_ij；

If C2 Gray is (P_ij)==0, with P_ijCentered by get the window of 3 × 3, pixel value structure in window

Become matrix W_3×3, definition operatorAsk the scalar product C of the two_ij＝W_3×3H, works as C_ij>=6, just at I_edgeMiddle by P_ijBe set to marginal point; The gray value that wherein Gray (P) is ordered for P;

C3, get back to step C1, repeat this process, until travel through complete.

4. license plate character recognition method as claimed in claim 3, is further included in above-mentioned steps D and again determines above-mentioned wordSymbol edge image I_edgeThe method on border be:

D1, row traversal to be to determine I from top to bottom_edgeUpper marginal position index Index_Top；

D2, row traversal to be to determine I from top to bottom_edgeLower limb location index Index_Bottom；

D3, from left to right row traversal to be to determine I_edgeLeft hand edge location index Index_Left；

D4, row traversal to be to determine I from right to left_edgeRight hand edge location index Index_Right。

5. want the license plate character recognition method as described in one of 1-4 as right, further comprise in above-mentioned steps G and Character mother plate collectionThe method of coupling is as follows:

1) concentrate and obtain current character template Tamplate at Character mother plate_i, obtain its horizontal hopping sequences SM_Hori, vertical transitionSequence SM_Vert, upper and lower number of transitions density ratio Hop_Density_M_Top/Bottom, left and right number of transitions density ratio Hop_Density_M_Left/Right；

2)S_HoriWith SM_HoriDo and mate, obtain matching degree M₁；

3)SV_ertWith SM_VertDo and mate, obtain matching degree M₂；

4)Hop_Density_Top/BottomWith Hop_Density_M_Top/BottomDo and mate, obtain mating M₃；

5)Hop_Density_Left/RightWith Hop_Density_M_Left/RightDo and mate, obtain matching degree M₄；

6) calculate total matching degree Mt_i：Mt_i＝w₁M₁+w₂M₂+w₃M₃+w₄M₄

Wherein, w₁、w₂、w₃、w₄Be respectively M₁、M₂、M₃、M₄Weight, its weight numerical value can be self-defined according to the difference of concrete image;

7) get back to step 1), repeat this process, until travel through complete;

8) get Index_mFor { Mt₁，Mt₂，…，Mt_i，…，Mt_nIn peaked index value, Index_mBe final recognition result.