CN101944174B - Identification method of characters of licence plate - Google Patents

Abstract

The invention provides an identification method of characters of a licence plate, mainly solving the problems of low identification speed and complex realization in the prior art. The identification process of the identification method comprises the following step of: individually taking out the pretreated characters of the licence plate, uniformizing, and scanning in vertical and horizontal directions; recording horizontal and vertical characteristic vectors, and respectively calculating similarity with characteristic vectors already well stored in a standard character library; finding a certain character of the standard character library, which has maximum similarity, through similarity magnitude comparison, wherein the character is a first-time identification result of characters to be identified; dividing partial similar characters in the first-time identification result into an up and down part and a left and right part, and respectively scanning the characteristic vectors of the similar characters; and then carrying out similarity calculation with a characteristic vector library of the characters, and finding a certain character of the standard character library, which has the maximum similarity, as a final identification result. The invention has high identification speed and identification accuracy rate and can be used for the real-time detection of the licence plate of a high-speed running vehicle.

Description

The recognition methods of characters on license plate

Technical field

The invention belongs to technical field of image processing, particularly relate to a kind of character identifying method, can be used for identification characters on license plate.

Background technology

In recent years since; China has accelerated paces in the construction of infrastructure; By contrast; Facilities Construction such as road management monitoring and science charging software but relatively lag behind, and to this situation, administrative authority has set about carrying out such as traffic signals adaptive control system, intelligent traffic monitoring system, GPS vehicle management and navigational system, the not development of intelligent traffic system ITS such as Auto Fare Collection Parking System.Wherein the vehicle license recognition system is in a crucial ring in a plurality of fields; Therefore, Vehicle License Plate Recognition System will have high recognition, simultaneously to the ambient lighting condition; Factor affecting such as camera site and Vehicle Speed also should have bigger robustness, and can satisfy the requirement of real-time.

Through licence plate vehicle being managed is that traffic realizes modernization and intelligentized basis; License plate recognition technology is the branch that Computerized intelligent is used; Be the important subject that computer vision and mode identification technology are used at intelligent transportation field, belong to a very important field of Flame Image Process.Its purposes is very extensive, comprises substantially:

(1) expressway tol lcollection, monitoring management.

(2) sub-district, parking lot management.

(3) urban road monitor, management violating the regulations.

(4) car plate lands, verifies.

(5) wagon flow statistics, safety management.

Along with the raising of computing power and the development of computer vision technique, the license plate automatic identification technology reaches its maturity.Present plate recognition system has reached certain discrimination, and still, under the situation at weather condition difference or night, discrimination has tangible reduction.Existing plate recognition system will reach completely, and practicability still has a long way to go.With regard to the development trend of license plate recognition technology, main target is to improve the discrimination and the robustness of recognition system, so that work under various conditions.Here said various conditions comprise; Fine, cloudy day, sleet sky different weather, daytime, night different time, the different occasions in highway limit or parking lot or the like.Another requirement to recognition system is temporal, and processing speed will reach real-time as quickly as possible.

Existing license plate character recognition method mainly is a neural network recognition method, and neural network recognition method is a new research direction of the area of pattern recognition that rose in recent years.Because characteristic conforms human visual systems' such as the high-speed parallel processing of neural network, distribution storage information groundwork principle; Have and learn by oneself habit, self-organization, fault-tolerance, highly non-linear, high robustness, function of associate memory and reasoning consciousness function etc. very by force, can realize that the pattern information that can't accomplish based on the institute of the pattern recognition theory on the theory of computation level at present deals with the work.

Traditional Vehicle License Plate Recognition System based on optics is generally all chosen the BP neural network, i.e. error back-propagating neural network is as the main method of character recognition module.It is 3 layers that the image classification device network of neural network is divided into, the totally interconnected mode of many employings between layer and the layer.Do not interconnect with not existing between one deck unit.The one-component data of each input node presentation video proper vector of network are gray-scale value, output node presentation class sequence number, and the output maximum value process is adopted in the classification judgement.

The BP network model has been realized the imagination of multitiered network study.When input pattern of given network, it delivers to a latent layer unit by the input layer unit, after a latent layer unit successively handled, delivers to output layer unit again.Handle the back by output layer unit and produce an output mode, so be called propagated forward.If the output response has error with the desired output pattern, and does not meet the demands, that just imports the error back-propagating into, is about to error amount and successively transmits rearward along connecting path, and revise each layer connection weights.

The BP e-learning is typically to have the tutor to learn.Training set comprises M sample, to p training sample (p=1,2 ..., M), the actual O that is output as of unit j _Pj, its i input also is i the neuronic O that is output as _Pi, then:

$u_{\mathrm{Pj}}=\underset{i=0}{\overset{N}{\mathrm{\Σ}}}{W}_{\mathrm{ji}}{O}_{\mathrm{pi}}---\left(1\right)$

Select for use the S type function as output function, that is: in the BP algorithm mostly

$O_{\mathrm{pi}}=f\left(u_{\mathrm{Pj}}\right)=\frac{1}{1+\exp (-u_{\mathrm{Pj}})}---\left(2\right)$

The define grid error function is:

$E=\underset{p}{\mathrm{\Σ}}{E}_p---\left(3\right)$

$E_p=\frac{1}{2}\underset{j}{\mathrm{\Σ}}{(d_{\mathrm{ji}}-O_{\mathrm{pi}})}^2---\left(4\right)$

In the formula, d _PjExpression is to p training sample, the desired output of unit j.The purpose of training network is to find one group of weight, makes the error function minimization.

Utilize the quick descent method of gradient, weights are changed along the negative gradient direction of error function.If the variable quantity of weights is designated as Δ W _Ij, that is:

$\mathrm{\Δ}{W}_{\mathrm{ij}}=-\frac{\∂E_P}{\∂W_{\mathrm{ji}}}---\left(5\right)$

Order $-\frac{\∂E_P}{\∂U_{\mathrm{Ji}}}={\mathrm{\δ}}_{\mathrm{Pj}},$ Then:

$\frac{\∂E_P}{\∂W_{\mathrm{ji}}}=\frac{{\∂E}_P}{\∂E_{\mathrm{pi}}}\frac{{\∂U}_{\mathrm{pj}}}{\∂W_{\mathrm{ji}}}=\frac{\∂E_P}{\∂E_{\mathrm{pi}}}{O}_{\mathrm{pj}}=-{\mathrm{\δ}}_{\mathrm{pj}}{O}_{\mathrm{pj}}---\left(6\right)$

Obtain:

ΔW _ji＝ηδ _pjO _pj，η＞0 (7)

η is the study factor in the formula.

Generally, BP Learning Algorithms step is described below:

(1) variable and parameter are set, comprising training sample, weight matrix, learning rate;

(2) initialization, small random non-vanishing vector of given each weight matrix;

(3) input random sample;

(4) to the input sample, every layer of neuronic input signal of forward calculation BP network and output signal;

(5) by the error of reality output and desired output ball.Judge whether to meet the demands, changeed for the 8th step if satisfy; Not satisfying changeed for the 6th step;

(6) judge whether to have reached maximum iteration time,, changeed for the 8th step if arrive, otherwise every layer of neuronic partial gradient of backwards calculation;

(7) get weights according to each matrix of partial gradient correction:

W _ji(t+1)＝W _ji(t)+ηδ _pjO _pj (8)

Error term δ _PjTwo kinds of situation are arranged:

(8) judging whether to have learnt all samples, is then to finish, otherwise forwards for the 3rd step to;

Usually in order being that study factor η value is enough beaten, to be unlikely to again to produce vibration, in weights correction formula (8), to add a state of affairs item, that is:

W _ji(t+1)＝W _ji(t)+ηδ _pjO _pj+a[W _ji(t)-W _ji(t+1)] (10)

0＜a＜1 wherein.

But neural network recognition method also has its weak point as the higher character identifying method of a kind of recognition accuracy:

1. the software of neural network character recognition is realized very complicated.Because the learning functionality of neural network self; Comprise training sample; Calculate weight matrix etc., the mathematical operation of number of complex will be passed through in the centre, realizes and is not easy with software programming; Such as the involved character picture size final to be identified of BP nerve net all is 20 * 35 pixels; So nearly 700 of the input feature vector values of each character of neural network input end, multiplying each other with its transposed matrix with these 700 eigenwerts obtains one 700 * 700 matrix again again, and then calculates the eigenwert of this 700 * 700 matrix; Only the realization of this step has just needed bigger computational complexity, and this step is the sub-fraction of the training process of neural network.Neural network recognition method must constantly be revised weights through a large amount of study in addition, and the result that just can make each input obtain expecting also has no small complexity in realization.

2. because the process more complicated of neural network identification character; Processing speed to Recognition of License Plate Characters has caused influence; The neural network identification character needs the long time, when being integrated into neural network recognition method after car plate detects in the recognition system, because its real-time is bad; Influence the efficient of Recognition of License Plate Characters, cause the mistake identification of whole car plate detection system and leak identification.

Summary of the invention

The object of the invention is to overcome the deficiency of above-mentioned application process, has proposed a kind of license plate character recognition method that is easy to realize, to improve the efficient and the accuracy rate of Recognition of License Plate Characters.

For realizing above-mentioned purpose, license plate character recognition method of the present invention comprises the steps:

1. choose characters on license plate, numeral, letter, each one of Chinese character, and it is carried out normalization respectively;

2. the character after the normalization is carried out the scanning of vertical and horizontal direction respectively; Write down the sequence of its scanning black-white transition times; Be feature value vector VTD and HTD; And these proper vectors are made the proper vector library storage of standard, select some to cause the close character of mistake identification to set up close character set easily;

3. car plate to be identified is carried out binaryzation and Character segmentation, and the single character that takes out carries out normalization;

4. the single character after the normalization is carried out the scanning of vertical and horizontal direction respectively, write down vertical features value vector VTD and horizontal properties vector HTD respectively;

5. through following similarity formula, distinguish the vertical features vector sum horizontal properties vector of calculating character and the vertical similarity and the horizontal similarity in proper vector storehouse;

$\mathrm{Similarity}(X_m,X_n)=\frac{\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{mj}}{x}_{\mathrm{nj}}}{\sqrt{\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{mj}}^2\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{nj}}^2}}$

In the formula, X _mAnd X _nBe respectively the proper vector of character m and character n, J is the dimension of proper vector, x _MjJ eigenwert of expression character m proper vector, x _NjJ eigenwert of expression character n proper vector, Similarity (X _m, X _n) similarity of expression character m and character n;

6. calculate the corresponding horizontal similarity of each character and the weighted sum of vertical similarity, obtain one group of final similarity:

Simi _mn＝αSimi _VTD(X _m，X _n)+βSimi _HTD(X _m，X _n)

In the formula, Simi _VTD(X _m, X _n) be the similarity of vertical direction, Simi _HTD(X _m, X _n) be the similarity of horizontal direction, α and β are respectively VTD and the shared weight of HTD, and in application, choosing α is 0.4, and β is 0.6, Simi _MnWeighted sum for horizontal similarity and vertical similarity;

7. the similarity that obtains is carried out the size comparison, find out maximum similarity, and take out the pairing character result of maximum similarity;

8. will take out the pairing character result of maximum similarity compares with the close character set that causes mistake identification easily; If the character result is not in close character set; Then end of identification is discerned if the character result in close character set, then carries out the secondary of step 9;

9. character is divided into up and down or its proper vector of left and right sides two parts scanning, getting wherein, a part recomputates similarity as the Partial Feature vector; The similarity that obtains is carried out the size comparison; Find out maximum similarity, and take out the pairing character result of maximum similarity, obtain recognition result.

Described close character set comprises: A and 4; P and 9; B and 8; D and O; T and L; 7 and 1.

Described single character after the normalization being carried out vertical scanning, is from top to bottom, writes down the number of times that each row white pixel and black picture element replace.

Described single character after the normalization being carried out horizontal scanning, is from left to right, writes down the number of times that each row white pixel and black picture element replace.

The present invention has following advantage with respect to neural network recognition method:

1) the present invention calculates through simple mathematical owing to only scan the level and the vertical features vector of character, just can obtain recognition result, thereby recognition speed is fast, and real-time is good.

2) the present invention is owing to have the secondary recognition function, and the close character of ability effective recognition can solve other recognition methodss preferably and discern the not high problem of close character accuracy rate, thereby can reach higher recognition accuracy.

Description of drawings

Fig. 1 is a recognition methods process flow diagram of the present invention;

Fig. 2 is pre-service and the proper vector leaching process of the present invention to character " 4 ";

Fig. 3 is that the present invention is to character " 1 " and character " 7 " normalization figure;

Fig. 4 is secondary identification realization figure of the present invention;

Fig. 5 is that the present invention is to complete car plate identification instance graph;

Fig. 6 is selected 20 car plates of experiment;

Embodiment

With reference to Fig. 1, embodiment of the present invention is following:

Step 1, the pre-service of character to be identified.

After obtaining character to be identified, need treat identification character and carry out binary conversion treatment, and, carry out normalization and handle according to the level of character and the projection on the vertical direction.As shown in Figure 2, the character to be identified " 4 " that reads in is a coloured image, and this character " 4 " is treated to binary image earlier, obtains the character picture " 4 " after the normalization according to projection again.

Step 2 is extracted proper vector.

The characteristic of each character object comprises two proper vectors, i.e. vertical features vector VTD and horizontal properties vector HTD.As shown in Figure 2; The method that this method is extracted characteristics of image is: with row or one-row pixels is unit, picture is scanned the number of times that record white pixel and black picture element replace; Column scan with character " 4 " is an example; From top to bottom, Yi Lie Baihei is alternately for once, so eigenwert is designated as 1.Can see obviously that through image white pixel and black picture element alternate frequency are twice the most for a long time, and minimum for once.Picture is pressed leu time scanning, can obtain the series of features value, this stack features value is designated as VTD; In like manner, with character object " 4 " from left to right, by line scanning; Obtain a stack features value; Be designated as HTD, VTD and HTD carry out the needed proper vector of character recognition exactly, have write down the characteristics of image of target character their complete and accurates.In the method; The size of setting character to be identified is 20 * 35 pixels; So it is 35 dimensions that scanning obtains proper vector HTD, VTD is 20 dimensions, and is as shown in Figure 2; The horizontal properties vector HTD of character " 4 " is: 1111111111111112222222222111111111, and vertical features vector VTD is: 11111222222222222211.

Step 3 is set up the proper vector storehouse.

Extract the proper vector of one group of standard character, set up the proper vector storehouse.Character in the general-utility car licence plate that China adopts comprises 31 Chinese characters, 24 capitalization English letters and 10 arabic numeral, totally 65 characters altogether.Choose characters on license plate clearly; Each character only selects one; After the character of choosing is handled according to foregoing preprocess method, scan the horizontal properties vector HTD and the vertical features vector VTD of character respectively, can obtain 130 proper vectors altogether; These vectors are corresponding one by one and store according to the sequence of character, just generated the proper vector storehouse of standard character.

Step 4 is set up close character set.

In Vehicle License Plate Recognition System, the identification of some close characters is difficult points, need set up a close character set in the method, is used for the secondary identification of back, and the close character in the character set all is to occur in pairs.These close character set comprise: A and 4; P and 9; B and 8; D and O; T and L; 7 and 1.

Step 5 scans the proper vector of character to be identified.

After obtaining a character to be identified, carry out pre-service according to method noted earlier, extract the horizontal properties vector HTD and the vertical features vector VTD of character to be identified.

Step 6 is calculated similarity.

After vertical features vector VTD that obtains character to be identified and horizontal properties vector HTD, need to calculate the proper vector of character to be identified and the similarity in proper vector storehouse.

The similarity of two vectors is defined as: their scalar product is divided by their length, and its similarity formula is specially:

$\mathrm{Similarity}(X_m,X_n)=\frac{\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{mj}}{x}_{\mathrm{nj}}}{\sqrt{\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{mj}}^2\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{nj}}^2}}---\left(1\right)$

In the formula, X _mAnd X _nBe respectively the proper vector of character m and character n, J is the dimension of proper vector, x _MjJ eigenwert of expression character m proper vector, x _NjJ eigenwert of expression character n proper vector, Similarity (X _m, X _n) characteristic similarity of expression character m and character n.

Similarity to the otherness of eigenvector embody very obvious, disturb on these two performances at the identification and the noise resistance of close character, be significantly improved.Present Vehicle License Plate Recognition System is nearly all discerned respectively single character in the car plate; The proper vector of using in the character recognition of the present invention comprises vertical features vector VTD and horizontal properties vector HTD; Their dimension is respectively 20 and 35; Explain that these two quantity of information that proper vector comprised are different, the dimension of HTD is more than the dimension of VTD.According to VTD and the shared weight of HTD, similarity weighted calculation formula does

Simi _mn＝αSimi _VTD(X _m，X _n)+βSimi _HTD(X _m，X _n) (2)

Simi wherein _VTD(X _m, X _n) expression vertical similarity component that the vectorial VTD substitution formula of the vertical features of character m and character n (1) is calculated, Simi _HTD(X _m, X _n) expression horizontal similarity component that the vectorial HTD substitution formula of the horizontal properties of character m and character n (1) is calculated.In the formula, α and β are respectively vertical features vector VTD and the shared weight of horizontal properties vector HTD, and in application, choosing α is 0.4, and β is 0.6, Simi _MnWeighted sum for horizontal similarity and vertical similarity;

When binary image had obvious noise, the normalization effect of character was undesirable, and the character fracture is serious.But because horizontal properties vector HTD weights VTD weights vectorial than vertical features are big in the weighting formula, so noise do not influence the identification of character, and visible introducing weights α and β are rational.

Step 7, the homotaxy degree.

With the weighted sum Simi that obtains _MnSort by size, the pairing character of maximum similarity is recognition result.

Step 8 is judged recognition result.

The character result who identifies is compared with the close character set that causes mistake identification easily, if the character result not in close character set, end of identification then is if the character result in close character set, then carries out secondary identification;

Step 9, secondary identification.

Secondary identification is on the result's that formerly discerns the basis, carries out finer identification, in order to distinguish the very approaching character of those results.Shown in character among Fig. 4 " A ", character " A " recognition result for the first time is character " 4 ".Through judging that recognition result character " 4 " belongs to similar character set, need carry out secondary identification.Examine this two characters, find that easily though their the first half are very similar, the latter half otherness is bigger, can the proper vector HTD of character to be identified be divided, HTD is the vector of 35 dimensions, and preceding 20 dimensions are designated as HTD ₁, back 15 dimensions are designated as HTD ₂, what they were represented respectively is the horizontal properties vector of target character the first half and the latter half.Same, the horizontal properties of character " A " and character " 4 " vector HTD also can do identical division in the standard character proper vector storehouse.So just can calculate a final horizontal similarity:

Simi′ _HTD＝Simi _HTD1+Simi _HTD2 (4)

Simi _HTD1Be the horizontal similarity of the first half of character, Simi _HTD2Being the horizontal similarity of the latter half of character, is example with character " A " and character " 4 ", though the horizontal similarity Simi of their the first half _HTD1Size is about the same, but the horizontal similarity Simi of their the latter half _HTD2Obvious difference is but arranged.Therefore, according to final horizontal similarity Simi ' _HTDDraw the maximum character of similarity, this character is exactly final recognition result.And similar character D and O also need same secondary recognition methods.

For similar character such as P and 9, B and 8 need divide their vertical features vector VTD, and VTD is 20 dimensions, can be divided into preceding 10 dimensions and be designated as VTD ₁, back 10 dimensions are designated as VTD ₂, what they were represented respectively is the vertical features vector of target character left-half and right half part, the vertical features of respective symbols vector VTD also does identical division in the standard character proper vector storehouse, obtains new vertical similarity with following similarity formula.

Simi′ _VTD＝Simi _VTD1+Simi _VTD2 (5)

Simi _VTD1Be the vertical similarity of left-half of character, Simi _VTD2Be the vertical similarity of character right half part, the vertical similarity Simi of their right half part _VTD2Size is about the same, but the horizontal similarity Simi of their left-half _VTD1Distinguish very big.According to final vertical similarity Simi ' _VTDDraw the maximum character of similarity, this character is exactly final recognition result.

For similar character T and L, 7 and 1, because division methods is carried out secondary identification about dividing perhaps about their employings, all can not reach ideal effect.Under such kind situation, need judge differentiation with the area size of part character, shown in Figure 4 is character " 1 " and character " 7 " process pre-service and character normalization design sketch afterwards.Can see that from Fig. 3 through after the normalization, from the white portion area of the latter half, character " 1 " as criterion, is distinguished identification to two characters much larger than character " 7 ".

Experimental result and data:

As shown in Figure 5, read in a complete car plate, through binaryzation, Character segmentation is discerned with the present invention after handling, obtain alphabet recognition result.For the objectivity of accomplishing to contrast, under same computer hardware condition, choose 20 car plate photos; As shown in Figure 6, in these car plate photos, comprise some clearly, inclination; Fuzzy car plate carries out the pre-service of car plate with method of the present invention, that is to say; Except the character recognition module difference, all conditions are the same in the experimentation, and obtaining experimental result like this is accurately.Experiment Main Conclusions data are respectively recognition accuracy and identification required time.

Experimental data: first character of car plate is a Chinese character, does not do statistics, and remaining numeral and alphabetic character add up to 120, and the present invention discerns correct 108, and accuracy rate is 90%, and neural network is discerned correct 112, and accuracy rate is 93.3%.Discern used averaging time, this method is 4.39 milliseconds, and neural network is 91.38 milliseconds, and is as shown in table 1.

The recognition accuracy of table 1 the present invention and neural network and recognition time contrast

Picture number	The present invention discerns correct number of characters	Recognition time unit of the present invention: millisecond	Neural network is discerned correct number of characters	Neural network recognition time unit: millisecond
					1	5	4.33	5	70.67
2	5	4.42	5	65.76
					3	6	4.38	6	80.28

4	5	4.40	5	74.89
					5	6	4.49	6	97.52
6	5	4.39	6	95.58
					7	6	4.33	6	86.45
8	5	4.31	5	92.71
					9	6	4.32	6	96.27
10	5	4.35	6	98.38
					11	5	4.38	6	94.16
12	6	4.64	5	96.50
					13	6	4.40	5	93.43
14	6	4.36	6	96.12
					15	6	4.36	6	98.45
16	5	4.63	5	97.91
					17	5	4.33	5	97.38
18	5	4.35	6	98.15
					19	5	4.37	6	99.03
20	6	4.31	6	98.12

Can find out from table 1, more lower slightly than traditional neural networks method on accuracy rate based on the character identifying method of proper vector scanning, still, see that from the speed of identification the former will be far superior to the latter.That is to say that aspect the real-time of method the method that the present invention proposes has very big superiority.In addition, also find in the experimentation, in the target license plate of various conditions; The identification of correct recognition rata, the especially numerical character of most of character, two methods are equal; Just in some close words; Because the distinctive learning process of neural net method, its performance is comparatively good, has caused final recognition accuracy slightly higher than this method.

Claims (5)

1. a license plate character recognition method comprises the steps:

(1) choose characters on license plate, numeral, letter, each one of Chinese character, and it is carried out normalization respectively;

(2) character after the normalization is carried out respectively vertically and the scanning of horizontal direction; Write down the sequence of its scanning black-white transition times; Be feature value vector VTD and HTD; And these feature value vectors are made the feature value vector library storage of standard, select some to cause the close character of mistake identification to set up close character set easily;

(3) car plate to be identified is carried out binaryzation and Character segmentation, and take out single character and carry out normalization;

(4) the single character after the normalization is carried out respectively vertically and the scanning of horizontal direction, write down vertical features value vector VTD and horizontal properties value vector HTD respectively;

(5) through following similarity formula, the vertical features value vector sum horizontal properties value vector of difference calculating character and the vertical similarity and the horizontal similarity in feature value vector storehouse;

$\mathrm{Similarity}(X_m,X_n)=\frac{\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{mj}}{x}_{\mathrm{nj}}}{\sqrt{\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{mj}}^2\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{x}_{\mathrm{nj}}^2}}$

In the formula, X _mAnd X _nBe respectively the feature value vector of character m and character n, J is the dimension of feature value vector, x _MjJ eigenwert of expression character m feature value vector, x _NjJ eigenwert of expression character n feature value vector, Similarity (X _m, X _n) similarity of expression character m and character n;

(6) calculate the corresponding horizontal similarity of each character and the weighted sum of vertical similarity, obtain one group of final similarity:

Simi _mn＝αSimi _VTD(X _m，X _n)+βSimi _HTD(X _m，X _n)

In the formula, Simi _VTD(X _m, X _n) be the similarity of vertical direction, Simi _HTD(X _m, X _n) be the similarity of horizontal direction, α and β are respectively VTD and the shared weight of HTD, and in application, choosing α is 0.4, and β is 0.6, Simi _MnWeighted sum for horizontal similarity and vertical similarity;

(7) similarity that obtains is carried out the size comparison, find out maximum similarity, and take out the pairing character result of maximum similarity;

(8) will take out the pairing character result of maximum similarity compares with the close character set that causes mistake identification easily; If the character result is not in close character set; Then end of identification is discerned if the character result in close character set, then carries out the secondary of step (9);

(9) character is divided into up and down perhaps its feature value vector of left and right sides two parts scanning; It is vectorial as partial feature value to get the two parts that are divided into; Recomputate similarity, the similarity that obtains is carried out the size comparison, find out maximum similarity; And take out the pairing character result of maximum similarity, obtain recognition result.

2. recognition methods according to claim 1, wherein the described close character set of step (2) comprises A and 4, P and 9, B and 8, D and O, T and L and 7 and 1.

3. recognition methods according to claim 1; Wherein step (9) is described is divided into character up and down or its feature value vector of left and right sides two parts scanning; Be that two parts carried out horizontal properties value vector scan about the A in the close character set and 4, D and 0 were divided into, P and 9, B and 8 be divided into left and right sides two parts carry out the scanning of vertical features component.

4. recognition methods according to claim 1, wherein step (4) is described carries out vertical scanning to the single character after the normalization, is from top to bottom, writes down the number of times that each row white pixel and black picture element replace.

5. recognition methods according to claim 1, wherein step (4) is described carries out horizontal scanning to the single character after the normalization, is from left to right, writes down the number of times that each row white pixel and black picture element replace.

Images