CN106250913A - A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis - Google Patents

Abstract

The invention discloses a kind of combining classifiers licence plate recognition method based on local canonical correlation analysis, including 1, process sample set, sample set is divided into original training sample T_o, test sample T_s, second test sample T_p, and according to sample feature extraction sample characteristics；2, utilize the sample characteristics extracted, train multiple relatively independent single classifier by parameter when changing training single classifier with kernel function；3, the method utilizing class KNN, finds the sample set similar with test sample, and utilizes the method for canonical correlation analysis to consider the relation between fractional sample feature and overall grader classification results, thus adjust each grader weight and obtain integrated model of classifiers；4, the final detection result of sample to be tested is together decided on by the grader after step 3 is integrated, it is judged that sample to be tested generic.The present invention can be self adaptive tackle different test samples and change grader weight, simultaneously can effectively improve grader classification accuracy.

Description

A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis

Technical field

The present invention relates to pattern recognition, integrated study, computer vision field, be based especially on multimedia application technical side To image processing techniques, a kind of combining classifiers licence plate recognition method based on local canonical correlation analysis.

Background technology

License plate recognition technology can effectively strengthen relevant department's supervision to road traffic at intelligent transportation field, significantly drops The use of low human and material resources, is the important component part of modern smart city development.The Car license recognition skill that present stage is main Art, such as: licence plate recognition method based on support vector machine (SVM), licence plate recognition method based on rarefaction representation (SRC) and base Licence plate recognition method etc. in fuzzy classification.Above-mentioned licence plate recognition method, can not fully play according to different sample characteristics The maximum utility of grader.Traditional licence plate recognition method is often because of the classification results excessively centralization of single classifier, list One change feature and occur classification results to have occasionality.The training of single classifier simultaneously can depend on training sample set unduly originally Body.In the case of working as lack of training samples, or training sample excessively complexity, the sorter model that training obtains can not be expired The identification of foot normal road traffic car plate and detection.

To sum up, on the basis of above-mentioned technology develops, introduce the concept of combining classifiers, by multiple relatively independent dividing Class device together decides on the classification of sample to be tested, thus improves the verification and measurement ratio of car plate.Now, the scheme of combining classifiers such as, base In the classification ensemble method of Adaboost (adaptive boosting), based on Bagging (Bootstrap Aggregating) categorizer integration method etc..AdaBoost method, is classifier optimization method based on weak point of thought.Should The grader that training obtains can be utilized the principle that its detection mistake classification learns again by method, optimizes traditional classifier classification effect Really.Bagging is a kind of algorithm selecting thought based on ballot.Have an advantage in that and any grader can be carried out extensive melting Closing, define grader and do not have special requirement, for the learning algorithm of similar neutral net, Bagging can instruct parallel Practicing manifold classification device, can shorten the time overhead of algorithm greatly, its classification results is chosen in a vote by majority.

Can be seen that above categorizer integration method does not take into account fractional sample feature and overall grader classification results Between contact, balance the fractional sample feature weight when sample classification.Thus can be because overall classifying quality is depended on unduly Training sample set, the Weak Classifier that training classifying quality is poor.Such as, AdaBoost algorithm is susceptible to the feelings of overtraining Condition, can not balance contacting between overall training sample and grader, along with its accuracy rate of increase of training iterations is anti- And can decline.And training sample is depended on unduly by Bagging method due to its grader trained, in the face of small sample problem Time, particularly in the case of the complicated various training sample of classification, its classifying quality is the most barely satisfactory, can not take into account local The effect that sample plays when classification with overall Weak Classifier.

Reply the problems referred to above, the present invention proposes a kind of combining classifiers Car license recognition side based on local canonical correlation analysis Method.The thought searching feature of present invention application canonical correlation analysis and classification results, the dependency relation of these two groups of variablees, and then really The confidence level of each single classifier fixed, thus distribute weight to each single classifier.And in order to balance fractional sample with overall Relation between classification results utilizes the method for class KNN, searches out K neighbour of sample to be tested such that it is able to utilize this K closely Adjacent analysis fractional sample feature contacts with classification results, the weight of dynamically adjustment grader that simultaneously can be self adaptive.Experiment Prove that the present invention can effectively improve the accuracy rate of Car license recognition, and the different sample set of reply has higher robustness.

Summary of the invention

It is an object of the invention to play each single classifier maximum utility when the different sample set of reply, it is provided that A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis.Research fractional sample feature and classification further Contact between device classification results, thus it is effectively improved Car license recognition accuracy rate and dynamic programming.

In order to solve above-mentioned technical problem, the concrete technical scheme that the present invention uses is as follows:

A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis, comprises the steps:

Step one: process set of data samples, data set is divided into training sample T_o, test sample T_s, second test sample T_p, and according to the HOG feature of sample feature extraction sample；

Step 2: on the basis of step one, utilizes the training sample T extracted_oFeature, by changing the single classification of training Parameter during device and kernel function train multiple relatively independent single classifier；

Step 3: the method utilizing class KNN, finds and test sample T_sSimilar sample, and utilize canonical correlation to divide The method of analysis considers the relation between fractional sample feature and overall grader classification results, thus adjusts each grader weight Obtain integrated model of classifiers；

Step 4: the final detection result of sample to be tested is together decided on by the grader after step 3 is integrated, it is judged that Sample to be tested generic.

Further, implementing of described step one includes: described extraction sample characteristics includes three parts: Part I It is training sample T_oHOG feature extraction, for training multiple different and relatively independent grader；Part II is to utilize Class KNN method find with test sample T_sK close second test sample T_pFeature extraction, for each single classification Device detects its sample accuracy rate thus realizes the dynamic adjustment to single classifier confidence level；Part III is test sample T_sSpy Levy extraction, for detecting the classification results that sample is final.

Further, implementing of described step 2 includes: the grader of employing is support vector machines, for obtaining Multiple different graders, when carrying out classifier training, use and adjust punishment correlation coefficient C and gamma correlation coefficient G, and And close in different parameter group and to take different kernels, to form multiple different grader.

Further, implementing of described step 3 comprises the steps:

A. utilize the method for class KNN from original training sample T_rIn to each test sample T_sSelect several second tests Sample, and then the single classifier obtained by training carried out test to these several second test samples and obtains classification results probability Value DT_p(x)；And by probit DT obtained_pX () calculates each grader test accuracy rate AT for sample X_p, computing formula As follows:

AT_p ⁱ=DT_p(X_C=j) (i=1.2.3 ... L)

Wherein d_LCX () represents that n-th test sample X (N≤K) records at l-th grader is the probit of classification C, DT_p(X_C=j) represent the probability that classification C is sample concrete class j recording sample X, the detection to sample X of the most each single classifier Accuracy rate；

By the AT obtained_p ⁱIt is worth to as K second test sample test result values Y in L grader_j ⁱ:

${Y_j}^i=\left(\begin{array}{ccc}{y}_1^1& ...& y_j^i\end{array}\right)=\left(\begin{array}{ccc}{{\mathrm{AT}}_p}^1& ...& {{\mathrm{AT}}_p}^i\end{array}\right)$

WhereinRepresent that jth (j≤K) individual sample is in the correct probit of each single classifier classification results；

B. the feature vector, X of K sample is constructed_K ⁱ, and utilize the method for canonical correlation analysis to calculate fractional sample feature X_nWith L grader classification accuracy Y_c ⁱBetween overall correlation matrix R_xy, computing formula is as follows:

X_K ⁱ=(b₁ … b_q)

Wherein b_qRepresent the feature in sample X q dimension, i=1.2.3 ... K；M₁₁Represent variable X_KThe covariance matrix of self COV(X,X)；M₁₂Represent COV (X, Y)；M₂₁Represent COV (Y, X)；M₂₂Represent COV (Y, Y)；C_q+k,q+kRepresent between each variable Dependency relation；

C. the dependency relation matrix R of sample characteristics and grader classification accuracy is obtained by step B_xy, then apply this step In Zhou, the 3rd formula is asked and is worked as variable X_KWith Y_cIt is in canonical variable coefficient a during maximal correlation relation_i ^T, b_i ^T, obtain X_kWith Y_cIt Between maximum linear correlation combiner U_q、V_p(p≤q, p, q=1.2.3 ... .) it is i.e. required canonical variable, computing formula is as follows:

U_p=a^ΤX_k ^q

V_p=b^ΤY_c ^p

${\mathrm{max\ρ}}_{xy}=corr(u,v)=\frac{\mathrm{cov}(u,v)}{{\mathrm{\σ}}_u{\mathrm{\σ}}_v}=\frac{a^T{M}_{12}b}{\sqrt{a^T{M}_{11}a}\sqrt{b^T{M}_{22}b}}$

Subject to:

a^TM₁₁A=1

b^TM₂₂B=1

D. obtained between the feature of K the neighbour for some test sample and grader classification results by step C The grader linear combination V that big dependency relation obtains_p, simultaneously according to the coefficient b of the canonical variable obtained_i ^TDistribution single classifier power Weight, and then the approximation of grader linear combination and sample characteristics maximum correlation can be calculated, thus obtain different grader Confidence level ￡ (X, V when sample X is predicted by combination_p), computing formula is as follows:

Wherein D_ijRepresent feature and the grader linear combination V of test sample X_pDependency relation.

Further, implementing of described step 4 includes:

Determine final test sample T_sThe classification of middle sample to be tested X, needs to detect its K relevant neighbour and tests the category Relevant approximation Ψ of sample_j(x), and process sample to be tested X, obtain its relation D with sample characteristics (x, y), Jin Erji Calculate approximation MT of sample to be tested X_s ⁿ(x), computing formula is as follows:

${{\mathrm{MT}}_s}^n\left(x\right)=\frac{1}{pq}*{\Σ}_{ij}D\left(d_{ij}\right)$

Wherein MT_s ⁿ(n=1.2.3 ...) expression sample to be tested X is belonging respectively to approximation during the n-th class；

Finally by the difference calculated between test sample approximation and such sample class approximation, this difference minimum is i.e. Being the final classification of sample to be tested, computing formula is as follows:

$Fin-test\left(x\right)=\mathrm{argmin}\frac{1+\left|\right|{\mathrm{\ψ}}_j-{{\mathrm{MT}}_s}^n\left(x\right)|{|}_2^2}{{\Σ}_j(1+{\mathrm{\ψ}}_j-{{\mathrm{MT}}_s}^j(x\left)\right)}.$

Beneficial effects of the present invention:

1, the present invention has evaded tradition and has asked based on testing result occasionality present in single grader licence plate recognition method Topic.Utilize the thought of integrated study, multiple grader codetermination results are analyzed, finally the label of sample to be tested are entered Row decision-making.This method can be effectively improved classification accuracy.

2, the present invention and existing categorizer integration method have bigger difference.Traditional classifier Integrated Solution, not Have and take into account contacting between local and entirety, thus can cause training the Weak Classifier obtained to depend on training sample set unduly originally Body.When sample size is little, or during training sample set complexity, training to the effect to single classifier also can weaken.Simultaneously When combining classifiers optimizes, the problem optimized also can be there is so that Detection accuracy declines on the contrary.And side used by the present invention Method utilizes the thought of class KNN to find K neighbour of sample such that it is able between balance fractional sample feature and whole detection result Contact, each test sample can be adjusted dynamically the multi-categorizer confidence level when classification, thus efficient solution simultaneously Certainly the problems referred to above.

3, the scheme of tradition Multi-classifers integrated, is to consider that the optimization of multi-categorizer is integrated, do not consider sample with Contact between grader.The thought of present invention application canonical correlation analysis (CCA), have studied sample characteristics and classifier result Between canonical correlation relation, by its study understanding the multi-categorizer sensitivity to sample characteristics, it is thus possible to find Multi-categorizer weight when going out more can be suitable for certain class pattern detection adjusts parameter.

Accompanying drawing explanation

Fig. 1 is Multi-classifers integrated model based on CCA.

Fig. 2 is Multi-classifers integrated strategic process figure based on CCA.

Fig. 3 is car plate data set experimental result picture.

Fig. 4 is data set Madelon experimental result picture.

Detailed description of the invention

Below in conjunction with the accompanying drawings, technical scheme is described in further detail.

Fig. 1 describes the general thought of the present invention.The method that the present invention proposes comprises the following steps:

Step one: process set of data samples, data set is divided into original training sample, test sample, second test sample, Training sample, and according to sample feature extraction sample characteristics.Specific as follows:

In the embodiment of licence plate recognition method of the present invention, set of data samples S is existing car plate sample.Sample set includes Original training sample T_rWith test sample T_s.Wherein original training sample T_rIn be divided into again training sample T_oWith second test sample T_p (i.e. by the method for class KNN from original training sample T_rIn pick out with test sample T_sThe most close sample), i.e. S=T_r+ T_s=T_o+T_p+T_s.In car plate data set, altogether comprising 2100 sample object, wherein original training sample has T_r=1550, Test sample has T_s=550, second test sample T_p=9.

Described sample characteristics extracts and includes three parts: all data sets of this patent are all to extract histograms of oriented gradients (Histogram of Oriented Gradient, HOG) feature.Part I is training sample T_oHOG feature extraction, its Purpose is to train multiple different and relatively independent grader；Part II be utilize that class KNN method finds with test Sample T_sK close second test sample T_pFeature extraction, this part sample is mainly used for detecting each single classifier Its sample accuracy rate thus realize the dynamic adjustment to single classifier confidence level；Part III is test sample T_sFeature carry Take, for detecting the classification results that sample is final.

Step 2: on the basis of step one, utilizes the training sample T extracted_oFeature, by changing the single classification of training Parameter during device and kernel function train multiple relatively independent single classifier.Specific as follows:

The grader used is support vector machines, for obtaining multiple different grader, is carrying out classifier training Time, use and adjust punishment correlation coefficient C and gamma correlation coefficient G, and close in different parameter group and take different kernels Function, forms multiple different grader with this；Described kernel function includes: RBF, Liner, Polynomial, Sigmoid. Concrete as shown in table 1.

More than 1 sorter model of table

Step 3: the method utilizing class KNN, finds and test sample T_sSimilar sample, and utilize canonical correlation to divide The method of analysis considers the relation between fractional sample feature and overall grader classification results, thus adjusts each grader weight Obtain integrated model of classifiers.As in figure 2 it is shown, be implemented as follows:

A. utilize the method for class KNN from original training sample T_rIn to each test sample T_sSelect T_p=9 second tests Sample.And then the single classifier obtained by training is carried out test to these 9 second test samples and obtains classification results probit DT_p(x).And by probit DT obtained_pX () calculates each grader test accuracy rate AT for sample X_p.Computing formula is such as Under:

AT_p ⁱ=DT_p(X_C=j) (i=1.2.3 ... L) (2)

Wherein d_LCX () represents that n-th test sample X (N≤K) records at l-th grader is the probit of classification C. DT_p(X_C=j) represent the probability that classification C is sample concrete class j recording sample X, the detection to sample X of the most each single classifier Accuracy rate；K represents the number of second test sample.

By the AT obtained_p ⁱIt is worth to as K second test sample test result values Y in L grader_j ⁱ。

${Y_j}^i=\left(\begin{array}{ccc}{y}_1^1& ...& y_j^i\end{array}\right)=\left(\begin{array}{ccc}{{\mathrm{AT}}_p}^1& ...& {{\mathrm{AT}}_p}^i\end{array}\right)---\left(3\right)$

WhereinRepresent that jth (j≤K) individual sample is in the correct probit of each single classifier classification results.

B. the feature vector, X of K sample is constructed_K ⁱ, and utilize the method for canonical correlation analysis (CCA) to calculate fractional sample Feature X_nWith L grader classification accuracy Y_c ⁱBetween overall correlation matrix R_xy.Computing formula is as follows:

X_K ⁱ=(b₁ … b_q) (4)

Wherein b_qRepresent the feature in sample X q dimension, i=1.2.3 ... K.M₁₁Represent variable X_KThe covariance matrix of self COV(X,X)；M₁₂Represent COV (X, Y)；M₂₁Represent COV (Y, X)；M₂₂Represent COV (Y, Y)；C_q+k,q+kRepresent between each variable Dependency relation.

C. the dependency relation matrix R of sample characteristics and grader classification accuracy is obtained by step B_xy, then apply formula (8) ask and work as variable X_kWith Y_cIt is in canonical variable coefficient a during maximal correlation relation_i ^T, b_i ^T, obtain X_kWith Y_cBetween maximum linear Correlation combiner U_q、V_p(p≤q, p, q=1.2.3 ... .) it is i.e. required canonical variable, computing formula is as follows:

U_p=a^ΤX_k ^q (6)

V_p=b^ΤY_c ^p (7)

${\mathrm{max\ρ}}_{xy}=corr(u,v)=\frac{\mathrm{cov}(u,v)}{{\mathrm{\σ}}_u{\mathrm{\σ}}_v}=\frac{a^T{M}_{12}b}{\sqrt{a^T{M}_{11}a}\sqrt{b^T{M}_{22}b}}---\left(8\right)$

Subject to:

a^TM₁₁A=1

b^TM₂₂B=1

D. obtained between the feature of K the neighbour for some test sample and grader classification results by step C The grader linear combination V that big dependency relation obtains_p, simultaneously according to the coefficient b of the canonical variable obtained_i ^TDistribution single classifier power Weight, and then the approximation of grader linear combination and sample characteristics maximum correlation can be calculated, thus obtain different grader Confidence level ￡ (X, V when sample X is predicted by combination_p), computing formula is as follows:

Wherein D_ijRepresent feature and the grader linear combination V of test sample X_pDependency relation.

Step 4: the final detection result of sample to be tested is together decided on by the grader after step 3 is integrated, it is judged that Sample to be tested generic.Specific as follows:

Determine final test sample T_sThe classification of middle sample to be tested X, needs to detect its K relevant neighbour and tests the category Relevant approximation Ψ of sample_j(x), and process sample to be tested X, obtain its relation D with sample characteristics (x, y), Jin Erji Calculate approximation MT of sample to be tested X_s ⁿ(x), computing formula is as follows:

${{\mathrm{MT}}_s}^n\left(x\right)=\frac{1}{pq}*{\Σ}_{ij}D\left(d_{ij}\right)---\left(12\right)$

Wherein MT_s ⁿ(n=1.2.3 ...) expression sample to be tested X is belonging respectively to approximation during the n-th class.

Finally by the difference calculated between test sample approximation and such sample class approximation, this difference minimum is i.e. It it is the final classification of sample to be tested.Computing formula is as follows:

$Fin-test\left(x\right)=\mathrm{argmin}\frac{1+\left|\right|{\mathrm{\ψ}}_j-{{\mathrm{MT}}_s}^n\left(x\right)|{|}_2^2}{{\Σ}_j(1+{\mathrm{\ψ}}_j-{{\mathrm{MT}}_s}^j(x\left)\right)}---\left(13\right)$

Experiment conclusion

Below from several groups of experimental datas, verify effectiveness of the invention.Make to evade the occasionality of experiment simultaneously The error become, provides the test result of one group of data set of increasing income here, and data set Madelon data set is from NIPS 2003Feature Selection Challenge data base obtains gained.The following is the test for different pieces of information collection Result.

What Fig. 3 described is this method experimental result in car plate data set.Car plate data set we have found that in phase Under same kernel, experiment is had a great impact by different C (loss parameter), G (gamma parameter) parameters.Single classifier is minimum 41.82% can be reached, 97.82% can be reached.And after Multiple Classifier Fusion, that a part of grader that classification accuracy is low Allocated weight is close to 0.And this experiment institute extracting method can reach 98.32%.

What Fig. 4 described is the experimental result at Madelon data set, and the experiment during we are extracted under different parameters is tied Fruit research, finds that between relatively independent single classifier, sample classification is accurate when using multinomial kernel (Polynomial) Rate can reach 67.5%.And apply classification accuracy rate proposed by the invention can reach 75.28%.

To sum up, a kind of based on local canonical correlation analysis the combining classifiers licence plate recognition method that the present invention proposes can Effectively promote Car license recognition accuracy rate.Analyze experimental data and find in the case of single grader classifying quality is preferable, this reality Testing institute's extracting method classifying quality and promote limited, maximum can improve 1%.The experimental result of similar Madelon data set finds Sample classification accuracy rate is had by the institute of the present invention extracting method in the case of gap is relatively big between single classifier classification results originally Obvious raising, the classification accuracy after Multiple Classifier Fusion improves nearly 8%.Classification based on CCA of the present invention Device integrated approach can effectively improve the accuracy rate of picture classification in complex samples, and has higher robustness.

The a series of detailed description of those listed above is only for the feasibility embodiment of the present invention specifically Bright, they also are not used to limit the scope of the invention, all equivalent implementations made without departing from skill of the present invention spirit Or change should be included within the scope of the present invention.

Claims (5)

1. a combining classifiers licence plate recognition method based on local canonical correlation analysis, it is characterised in that include walking as follows Rapid:

Step one: process set of data samples, data set is divided into training sample T_o, test sample T_s, second test sample T_p, and And according to the HOG feature of sample feature extraction sample；

Step 2: on the basis of step one, utilizes the training sample T extracted_oFeature, by change training single classifier time Parameter and kernel function train multiple relatively independent single classifier；

Step 3: the method utilizing class KNN, finds and test sample T_sSimilar sample, and utilize the side of canonical correlation analysis Method considers the relation between fractional sample feature and overall grader classification results, thus adjusts each grader weight and divided Class device integrated model；

Step 4: the final detection result of sample to be tested is together decided on by the grader after step 3 is integrated, it is judged that to be measured Sample generic.

A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis the most according to claim 1, its Being characterised by, implementing of described step one includes: described extraction sample characteristics includes three parts: Part I is training sample This T_oHOG feature extraction, for training multiple different and relatively independent grader；Part II is to utilize class KNN side Method find with test sample T_sK close second test sample T_pFeature extraction, for each single classifier is detected it Sample accuracy rate thus realize the dynamic adjustment to single classifier confidence level；Part III is test sample T_sFeature extraction, For detecting the classification results that sample is final.

A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis the most according to claim 1, its Being characterised by, implementing of described step 2 includes: the grader of employing is support vector machines, for obtaining multiple difference Grader, when carrying out classifier training, use and adjust punishment correlation coefficient C and gamma correlation coefficient G, and in difference Parameter group close and take different kernels, to form multiple different grader.

A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis the most according to claim 1, its Being characterised by, implementing of described step 3 comprises the steps:

A. utilize the method for class KNN from original training sample T_rIn to each test sample T_sSelect several second test samples, And then the single classifier obtained by training is carried out test to these several second test samples and obtains classification results probit DT_p (x)；And by probit DT obtained_pX () calculates each grader test accuracy rate AT for sample X_p, computing formula is as follows:

AT_p ⁱ=DT_p(X_C=j) (i=1.2.3 ... L)

Wherein d_LCX () represents that n-th test sample X (N≤K) records at l-th grader is the probit of classification C, DT_p (X_C=j) representing the probability that classification C is sample concrete class j recording sample X, the most each single classifier is accurate to the detection of sample X Really rate；

By the AT obtained_p ⁱIt is worth to as K second test sample test result values Y in L grader_j ⁱ:

${Y_j}^i=\left(\begin{array}{ccc}{y}_1^1& ...& y_j^i\end{array}\right)=\left(\begin{array}{ccc}{{\mathrm{AT}}_p}^1& ...& {{\mathrm{AT}}_p}^i\end{array}\right)$

Wherein Y_j ⁱ(i=1 ... L) represents that jth (j≤K) individual sample is in the correct probit of each single classifier classification results；

B. the feature vector, X of K sample is constructed_K ⁱ, and utilize the method for canonical correlation analysis to calculate fractional sample feature X_nWith L Individual grader classification accuracy Y_c ⁱBetween overall correlation matrix R_xy, computing formula is as follows:

X_K ⁱ=(b₁ … b_q)

Wherein b_qRepresent the feature in sample X q dimension, i=1.2.3 ... K；M₁₁Represent variable X_KThe covariance matrix COV of self (X,X)；M₁₂Represent COV (X, Y)；M₂₁Represent COV (Y, X)；M₂₂Represent COV (Y, Y)；C_q+k,q+kRepresent between each variable Dependency relation；

C. the dependency relation matrix R of sample characteristics and grader classification accuracy is obtained by step B_xy, then apply in this step 3rd formula is asked and is worked as variable X_KWith Y_cIt is in canonical variable coefficient a during maximal correlation relation_i ^T, b_i ^T, obtain X_kWith Y_cBetween Big linear correlation combination U_q、V_p(p≤q, p, q=1.2.3 ... .) it is i.e. required canonical variable, computing formula is as follows:

U_p=a^TX_k ^q

V_p=b^TY_c ^p

${\mathrm{max\ρ}}_{xy}=corr(u,v)=\frac{\mathrm{cov}(u,v)}{{\mathrm{\σ}}_u{\mathrm{\σ}}_v}=\frac{a^T{M}_{12}b}{\sqrt{a^T{M}_{11}a}\sqrt{b^T{M}_{22}b}}$

Subject to:

a^TM₁₁A=1

b^TM₂₂B=1

D. maximum phase between the feature of K the neighbour for some test sample and grader classification results is obtained by step C The grader linear combination V that pass relation obtains_p, simultaneously according to the coefficient b of the canonical variable obtained_i ^TDistribution single classifier weight, And then the approximation of grader linear combination and sample characteristics maximum correlation can be calculated, thus obtain different classifiers combination Confidence level when sample X is predictedComputing formula is as follows:

Wherein D_ijRepresent feature and the grader linear combination V of test sample X_pDependency relation.

A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis the most according to claim 1, its Being characterised by, implementing of described step 4 includes:

Determine final test sample T_sThe classification of middle sample to be tested X, needs to detect its K relevant neighbour to category test sample Relevant approximation Ψ_jX (), and process sample to be tested X, (x y), and then calculates and treats to obtain its relation D with sample characteristics Approximation MT of this X of test sample_s ⁿ(x), computing formula is as follows:

${{\mathrm{MT}}_s}^n\left(x\right)=\frac{1}{pq}*{\mathrm{\Σ}}_{ij}D\left(d_{ij}\right)$

Wherein MT_s ⁿ(n=1.2.3 ...) expression sample to be tested X is belonging respectively to approximation during the n-th class；

Finally by the difference calculated between test sample approximation and such sample class approximation, this difference minimum is i.e. treated This final classification of test sample, computing formula is as follows:

$Fin-test\left(x\right)=\mathrm{argmin}\frac{1+\left|\right|{\mathrm{\ψ}}_j-{{\mathrm{MT}}_s}^n\left(x\right)|{|}_2^2}{{\mathrm{\Σ}}_j(1+{\mathrm{\ψ}}_j-{{\mathrm{MT}}_s}^j(x\left)\right)}.$