CN106570457A - Chinese and Japanese character identification method - Google Patents

Abstract

The invention relates to a Chinese and Japanese character identification method via a Markov random field method. The Chinese and Japanese character identification method comprises the following steps: a step of size standardization, a step of characteristic point extraction and a step of matching and identifying. The step of matching and identifying comprises the following steps: characteristic points are used as station points, character type states are used as labels, the station points are compared with the labels via the Markov random field method, and an identification result can be obtained.

Description

A kind of Chinese and japanese word recognition methodss

Technical field

The present invention relates to a kind of character recognition method, especially a kind of Chinese and japanese word recognition methodss.

Background technology

Hand script Chinese input equipment character recognition is a very important technology of smart mobile phone, iPad and panel computer, and it uses Numeral conversion handwriting input character code, this is again appreciated that to provide convenient user circle by computer in turn Face.Although online handwriting Chinese Character Recognition has been present for many years, it remains the technological challenge with identical accuracy of identification, mainly It is because that hand-written character pattern often writes on a very common or cursive script.Additionally, attempting to reach higher precision, skill Art has adopted the module of complexity, so as to increase cost, process time and memory consumption.This work has been solved, and is studied Various such problems, in hand script Chinese input equipment character recognition, and propose a kind of effective and outstanding method, to improve identification Precision.

This work has been published in the 11st international litigation Document analysis and recognition, it report development markov with Airport (MRF) is to hand-written Japanese ONLINE RECOGNITION model.It is online handwriting Chinese character that Markov model (HMM) is most common method Identification.HMM model unitary and binary feature.Per unitary each input point of feature extraction, and each binary feature is reflected Neighborhood relationship between an input point of continuous adjacent.Although however, the characteristics of HMM application binary, the binary that they only merge Unitary feature (rather than considering their independence) is characterized as, its accuracy of identification is which has limited.This problem is being solved, is being carried herein The hand-written Japanese character of the ONLINE RECOGNITION of MRF models is gone out, the characteristics of the probability of effective evaluation not only unitary feature is also binary. Particularly, per unitary signature as a label, probability of each binary feature for a pair of tag, this is the model evaluation In a kind of new method that former model is compared.

The content of the invention

Binary feature is only merged into unitary feature and is identified by existing HMM modes.The invention provides one kind can be with The character recognition method of identification binary feature.The present invention mainly carries out Text region using the random field (MRF) of markov.

Many character recognition methods of the present invention, comprise the steps：

Step S10：Size-normalized, the input pattern for tracking acquisition by nib carries out size-normalized；

Step S20：Feature point extraction, extracts characteristic point on the input pattern after size-normalized；

And step S30：Matching and recognize, using markov random field method by characteristic point and existing character Class is compared, and will be correct characters with that character recognition in the immediate character type of characteristic point.

Step S30 includes,

Step S301：Using characteristic point as website；

Step S302：Using the state of character type as label；

Step S303：Website is compared with label using the method for the random field of markov, and obtains identification knot Really.

Many character recognition methods of the present invention, add weight parameter to improve identification essence in the method for Markov random field Degree.Weight parameter can be obtained using condition random field (CRF) method.

Many character recognition methods of the present invention, add weight parameter to improve identification essence in the method for Markov random field Degree.Weight parameter can be obtained using minimum classification error method.

Many character recognition methods of the present invention, the formula of energy function is

Many character recognition methods of the present invention, add weight parameter to improve identification essence in the method for Markov random field Degree.Weight parameter can be obtained using maximum matching method, and formula is

Many character recognition methods of the present invention, add weight parameter to improve identification essence in the method for Markov random field Degree.Weight parameter can be obtained using maximum matching method, and formula is

By L_NLL(λ, O)=- log P (C | O) formula is optimized to weight parameter.

Many character recognition methods of the present invention, add weight parameter to improve identification essence in the method for Markov random field Degree.Weight parameter can be obtained using minimum classification error method, and formula is

L_MCE(λ, 0)=σ (max (fractions_{It is incorrect})-fraction_Correctly)

σ (x)=(1+e^-x)^-1

Website is compared with label by many character recognition methods of the present invention using the method for the random field of markov Compared with, and obtain recognition result.Wherein, by Viterbi algorithm or Baum-Welch algorithms training the random field of markov Parameter.

Many character recognition methods of the present invention, step S10：It is size-normalized, the input pattern for obtaining is tracked by nib Carry out size-normalized.Input pattern retention level vertical scale is transformed into into standard size.

Many character recognition methods of the present invention, step S20：Feature point extraction, the input pattern after size-normalized Upper employing Ramner methods extract characteristic point.

Description of the drawings

Fig. 1 is the block diagram of the character recognition method for representing the present invention, shows that the present invention has three key steps in block diagram, its Also include three steps in middle third step S30.

Fig. 2 be represent feature of present invention point extract schematic diagram, right side represent input pattern, left side represents existing character Class.

Fig. 3 is the schematic diagram for representing straight chain MRF of the present invention, and circle represents a character type, and every kind of character type has three kinds Transition form.

Specific embodiment

Specific embodiment one

Below in conjunction with the accompanying drawings, the specific embodiment of the present invention is specifically described.

The present invention provides a kind of quickly many character recognition methods, as shown in figure 1, comprising the steps：

Step S10：It is size-normalized；

Step S20：Feature point extraction；

And step S30：Matching and identification.

The size-normalized of step S10 is carried out first, and we will track the input pattern for obtaining by nib and carry out size Standardization.Such as it is that input pattern retention level vertical scale is transformed into into standard size.

The feature point extraction of step S20 is subsequently carried out, on the input pattern after size-normalized, characteristic point is extracted.This In can for example adopt Ramner methods carry out the extraction of characteristic point.

Finally by step S30, characteristic point is compared with existing character type, will be with the immediate character of characteristic point That character recognition of apoplexy due to endogenous wind is correct characters.

In step S30, the characteristic point that we are matched with the state of MRF models and each character type, and obtain each word The similarity of symbol class.Then, we select the character type with maximum comparability as recognition result.

Concrete step includes.

Step S301：Using characteristic point as website；

Step S302：Using the state of character type as label；

Step S303：Website is compared with label, and obtains recognition result.

It is further elaborated with as follows：

Step S301：Using characteristic point as website.Website S={ s are formed for example₁,s₂,s₃,…,s_I}。

Step S302：Using the state of character type as label.Label L={ l are formed for example₁,l₂,l₃,…,l_J}。

Step S303：Website is compared with label, using the character with maximum comparability more afterwards as identification knot Really.

If Fig. 2, so-called identification are exactly that website is compared with label, the value of label is carried out into assignment to website subsequently, Such as F={ s₁=l₁,s₂=l₁,s₃=l₃,…,s₉=l₈,s₁₀=l₈, wherein F is thus referred to as a kind of configuration, represents that a kind of S is arrived The matching of L.

Hereinafter, it is described in detail with regard to step S303.

First, the characteristic vector of characteristic point forms an observation collection O.

In statistic algorithm or Bayesian algorithms, it is based on maximum a posteriori probability by the decision-making that Character recognizer is made (MAP) concept：

Wherein, P (C) is a prior probability of the given patterns of character type C, and P (O | C) it is observation collection O under the conditions of character type C Likelihood function, P (O) is the probability of observation collection O, and P (C | O) is the probability of the character type C input patterns under the conditions of observation collection O

Last decision-making uses equation below：

Wherein C* is approximate character type.

If P (C) be set to it is constant, and MAP approximately to become maximum likelihood (ML) approximate.So formula (2) just becomes how Approximate P (O | C).P (O | C) can be expressed as by we

Here, F={ s₁=l_i,s₂=l_j,…,s_I=l_k|l_i,l_j,l_k∈ L } it is from the website S to character type C being input into In label L matching

P (O, F | C)=P (F | C) P (O | F, C) (4)

The value of Direct calculation formulas (3) is relatively difficult.For HMM typically has two kinds of solutions：Baum-Welch Algorithm and Viterbi algorithm.

For the ease of calculating, the best match that we only consider.Therefore using in the case of Viterbi algorithm, obtain,

P (O | C) ≈ P (O, F_best|C) (5)

Here,

Therefore, the problem of identification is translated into, the P (F for obtaining_best|C)P(O|_Fbest, C) value and best match.

Calculating probability P (F | C) it is relatively difficult, because the interaction between variable is of overall importance.For ease of Process, MRFs is by assuming that label only depends on adjacent bus station to constrain the relation of interdependence between label.This can be with neck Domain system is being indicated.

Then neighborhood system Ni represents website S_iIt must is fulfilled forOne label and adjacent mark Sign and issue raw interaction.According to neighborhood system, a c is defined as into the subset of all common website S.

According to Hammersley-Clifford theorems set up between markov random file and gibbs random field etc. Formula,

Here,

It is referred to as priori energy function.Wherein V^F _c(F | C) it is referred to as the priori group potential function of a c.

It is normalization factor, is referred to as partition function.

Then it is contemplated that P (O | F, C), can equally obtain

And it is referred to as global likelihood energy function,

And

Wherein V^O _c(O | F, C) it is referred to as likelihood group potential function.

For the sake of simplicity, we only consider single site group c1={ si } and dual station point group c2={ si, sj }.By formula (8) and Formula (11), we obtain

Wherein l_siIt is the label in class c to Si assignment, O_siFor slave site s_iThe unitary characteristic vector of extraction, O_Si,SjIt is from s_i And s_jCombination in the binary feature vector that extracts.

Likelihood group's gesture is to describe the demographic information with regard to giving label observation collection, and priori group gesture is encoded with regard to neighbouring mark The prior information of label.

Under MAP frameworks, the posterior probability in equation (2) is maximized equivalent to the energy function in the formula of minimum (12)

Our groups of definition are as follows：

Single site group:c₁={ s₁,s₂,s₃,…,s₁₀,…}

Dual station point group:c₂={ { s₁,s₂},{s₂,s₃},{s₃,s₄},{s₄,s₅},……,{s₉,s₁₀},…}

Neighborhood system is that the characteristic point of the continuous adjacent according to sequential write is set up.As shown in figure 3, we are each character Class defines a straight chain MRF.Here, each label has a state, and each state has 3 kinds of conversions.

Energy function is expressed as：

Wherein, I is the quantity of characteristic point.

We can derive likelihood group gesture from the negative logarithm of conditional probability.

Wherein,It is set to 1.

We are the straight chain MRF using Fig. 3, therefore status condition probability can be used to derive priori energy function

Last energy function can be expressed as：

When energy function is less, then mean that the similarity being input between pattern and character type C is higher.

Each character type has a straight chain MRF, and the characteristic point that system searches for match input pattern using Viterbi With the MRF states of each character type, all pass through the energy of the minimum of formula (16) to find matching finally for there is no character type Path.

Unitary characteristic vector O_siIncluding s_iX, Y-coordinate.Binary feature vector O_sisi-1With two element (dx：s_iX Coordinate-s_i-1X-coordinate, dy：s_iY-coordinate-s_i-1Y-coordinate), an or element (tan-1 (dy/dx)).

Gaussian function can be used to approximatelyWith

In order to train the MRF of each character type, we first, by training pattern character type in any character pattern spy Levy and a little initialized as the state of MRF, subsequently, for each single state, the unitary characteristic vector of each characteristic point is made For the average of Gaussian function, subsequently, for each double state, each the binary feature vector between two adjacent feature points is made For the meansigma methodss of Gaussian function, then, it is 1 that we initialize the variance of those Gaussian functions and state transition probabilities.

Subsequently, we trained by Viterbi algorithm or Baum-Welch algorithms MRF parameter (Gaussian function and The average and variance of state transition probabilities).Our repetition trainings are until obtaining optimized parameter.

Specific embodiment two

In step S303, we can will add weight parameter in formula (16) (λ=λ 1, λ 2, λ are 3), special to adjust unitary The binary feature value of value indicative and state transition probability：

We can be optimized to weighting parameters based on condition random field.

Different weighting parameters may apply to different character types.

The state transfer that we can also adjust more parameters, the average and variance of such as Gaussian function, or MRF is general Rate.But, so need more training.

As an example, we have adjusted three of unitary eigenvalue, binary feature value and state transition probability value it is common Weighting parameters.

According to conditional random field models, the posterior probability of character type C is given by：

Wherein, F_CIt is a matching of a character type C.

We set P (C) be it is constant, therefore E (C)=- log P (C) be also it is constant,

Therefore, posterior probability is：

We can also pass through the stochastic gradient descent algorithm of following negative log-likelihood (NLL) loss function to vectorial λ It is optimized

L_NLL(λ, O)=- log P (C | O) (21)

Wherein C is correct characters class O.

Specific embodiment three

In step S303, we can will add weight parameter in formula (16) (λ=λ 1, λ 2, λ are 3), special to adjust unitary The binary feature value of value indicative and state transition probability：

We can be optimized to weighting parameters based on minimum classification error.

Different weighting parameters may apply to different character types.

The state transfer that we can also adjust more parameters, the average and variance of such as Gaussian function, or MRF is general Rate.But, so need more training.

As an example, we have adjusted three of unitary eigenvalue, binary feature value and state transition probability value it is common Weighting parameters.

We can also pass through the stochastic gradient descent algorithm of following negative log-likelihood (NLL) loss function to vectorial λ It is optimized

We can also be optimized by stochastic gradient descent algorithm vector λ by minimum classification error standard.Pass through Minimize immediate character type and correct character type between fraction gap drawing optimal solution：

L_MCE(λ, 0)=σ (maX (fractions_{It is incorrect})-fraction_Correctly)

σ (x)=(1+e^-x)^-1

Fraction_CorrectlyThe fraction of=correct characters class

Fraction is not_CorrectlyThe fraction (22) of=incorrect character type

Wherein, being input into pattern and the fraction of character type is：

Each character type has a MRF.To each character type, weight parameter and system are searched by using Viterbi The state of rope MRF is matching the characteristic point of input pattern, and is each character by E minimum in formula (18) (λ, O, F | C) Class finds smallest match path.

By the following examples, the effect that is embodied as of the present invention is illustrated.

Embodiment one

The MRFs methods and conventional HMMs methods in the specific embodiment of the invention one be we used to handwritten pattern It is identified.

Test altogether 1000 class of kanji and Chinese character, hiragana (subset of assumed name) 46 classes and 26 english lowercases Alphabetical classification.We use Viterbi algorithm and Baum-Welch Algorithm for Training models.

MRFs and HMM identical recognition time and frequency of training.

As shown result in following table.

The average character recognition time, it is 2.9 milliseconds when using function (x, y, dx, dy), when (x, y, dir) is used is 2.7 milliseconds, and be 2.2 milliseconds when (x, y) is used.

It is 16s to the average workout times of Viterbi algorithm iteration, and Baum-Welch algorithms uses about 51s.

As a result show have than conventional HMMs methods using the MRFs methods used in the specific embodiment of the invention one There is higher identification accuracy.

Embodiment two

Using the condition random field in specific embodiment two, to weighting parameters, (λ 1, λ 2,3) λ is optimized, as a result for we (λ 1, λ 2, λ are 3)=(0.28,0.48,0.94) for acquisition.

See the table below using the recognition result after weighting parameters.

Weighting parameters estimate that maximum matching method contrasts simple MRF models than minimum classification error method and significantly improves The accuracy rate of character recognition.For weighting parameters estimate that maximum matching method is more accurate than minimum classification error method.

Claims (8)

1. a kind of Chinese and japanese word recognition methodss, many character recognition methods of the present invention, it is characterised in that comprise the steps：

Step S10：Size-normalized, the input pattern for tracking acquisition by nib carries out size-normalized；

Step S20：Feature point extraction, extracts characteristic point on the input pattern after size-normalized；

And step S30：Characteristic point is entered with existing character type by matching and identification using the method for the random field of markov Row compares, and will be correct characters with that character recognition in the immediate character type of characteristic point.

Step S30 includes,

Step S301：Using characteristic point as website；

Step S302：Using the state of character type as label；

Step S303：Website is compared with label using the method for the random field of markov, and obtains recognition result.

In step S303, weight parameter in the method for described Markov random field, is added to improve accuracy of identification, weight Parameter can utilize condition random field or minimum classification error method to obtain.

2. many character recognition methods as claimed in claim 1, it is characterised in that the formula of the method for Markov random field It is:

$\begin{array}{c}E(O,F|C)=E\left(O\right|F,C)+E\left(F\right|C)\\ =\underset{i=1}{\overset{I}{\Σ}}\[-\log P\left(O_{s_i}\right|l_{s_i},C)-\log P\left(O_{s_i{s}_{i-1}}\right|l_{s_i},l_{s_{i-1}},C)-\log P\left(l_{s_i}\right|l_{s_{i-1}},C)\]\end{array}.$

3. many character recognition methods as claimed in claim 1, it is characterised in that power is added in the method for Markov random field Weight parameter improves accuracy of identification.Weight parameter can be obtained using maximum matching method, and formula is

$\begin{array}{c}P\left(C\right|O)=\frac{\underset{F_c}{\Σ}\exp (-E(\mathrm{\λ},O,F_c,C\left)\right)}{\underset{C_i}{\Σ}\underset{F_{c_i}}{\Σ}\exp (-E(\mathrm{\λ},O,F_{c_i},C_i\left)\right)}\\ =\frac{\underset{F_C}{\Σ}\exp (-E(\mathrm{\λ},O,F_c|C)-E(C\left)\right)}{\underset{C_i}{\Σ}\underset{F_{C_i}}{\Σ}\exp (-E(\mathrm{\λ},O,F_{c_i},C_i)-E(C_i\left)\right)}\\ =\frac{\underset{F_C}{\Σ}\exp (-E(\mathrm{\λ},O,F_c|C\left)\right)\exp (-E(C\left)\right)}{\underset{C_i}{\Σ}\underset{F_{C_i}}{\Σ}\exp (-E(\mathrm{\λ},O,F_{c_i}|C_i\left)\right)\exp (-E(C_i\left)\right)}\end{array}.$

4. many character recognition methods as claimed in claim 1, it is characterised in that power is added in the method for Markov random field Weight parameter improves accuracy of identification.Weight parameter can be obtained using maximum matching method, and formula is

$P\left(C\right|O)=\frac{\underset{F_c}{\Σ}\exp (-E(\mathrm{\λ},O,F_c|C\left)\right)}{\underset{C_i}{\Σ}\underset{F_{c_i}}{\Σ}\exp (-E(\mathrm{\λ},O,F_{c_i}|C_i\left)\right)}$

By L_NLL(λ, O)=- log P (C | O) formula is optimized to weight parameter.

5. many character recognition methods as claimed in claim 1, it is characterised in that power is added in the method for Markov random field Weight parameter improves accuracy of identification.Weight parameter can be obtained using minimum classification error method, and formula is

L_MCE(λ, O)=σ (max (fractions_{It is incorrect})-fraction_Correctly)

σ (x)=(1+e^-x)^-1

6. many character recognition methods as claimed in claim 1, it is characterised in that will using the method for the random field of markov Website is compared with label, and obtains recognition result.Wherein, by Viterbi algorithm or Baum-Welch algorithms training The random field parameters of markov.

7. many character recognition methods as claimed in claim 1, it is characterised in that step S10：It is size-normalized, by nib The input pattern that tracking is obtained carries out size-normalized.Input pattern retention level vertical scale is transformed into into standard size.

8. many character recognition methods as claimed in claim 1, it is characterised in that step S20：Feature point extraction, advises from size Characteristic point is extracted using Ramner methods on input pattern after generalized.