CN102663454B - Method and device for evaluating character writing standard degree - Google Patents

Abstract

The invention relates to the field of pattern recognitions, in particular to a method and a device for evaluating a character writing standard degree. The method includes the steps of extracting primitive characteristic vectors of stroke tracks of writing characters, matching the extracted primitive characteristic vectors with character templates in a first character set to obtain a first confidence degree, implementing a secondary matching process if the first confidence degree can not meet a requirement of a preset evaluation condition, obtaining secondary characteristic vectors, matching the secondary characteristic vectors with character templates in a second character set to obtain a second confidence degree, and performing an evaluation for the writing characters based on the second confidence degree. The provided method effectively solves the problem that the evaluation for confusing characters is inaccurate in prior art, and the reasonability and accuracy of the evaluation for the writing standard degree are improved.

Description

A kind of method and apparatus of character writing standard degree evaluation and test

Technical field

The present invention relates to area of pattern recognition, particularly relate to a kind of method and apparatus of character writing standard degree evaluation and test.

Background technology

Along with the development of information interaction, computer-aided instruction is widely used.For example, aspect Chinese teaching, computer-aided instruction provides the application such as Chinese character evolution, phonetic demonstration, Chinese-character writing dynamic demonstration, yet, at but shorter mention aspect the assessment of user's Chinese-character writing standard degree.The Chinese character of a normalized written requires stroke standard, the order of strokes observed in calligraphy to meet standard conventionally, and character compact overall structure meets requirement attractive in appearance simultaneously.Because Chinese character quantity is larger, the standardization assessment of Chinese character is realized comparatively complicated, relate to the technology such as image processings, pattern-recognition, therefore how effectively for the character of user writing, carry out standard degree and assess and become a challenging problem.

In prior art, when written character being carried out to the evaluation of standard degree, often adopt written character and single standard character relatively to calculate the method that similarity is passed judgment on, the evaluation result of obtaining is thus often reliable not.At this moment, there is a kind of Chinese-character writing quality evaluating method based on degree of confidence.In this method, the degree of confidence of mating by calculating character, knows the degree of reliability that current written character is similar with standard form, and then Chinese-character writing quality is evaluated.In this method, first utilize correction second judgement function category device to identify handwritten Chinese character, obtain K candidate word, and calculate the distance of each candidate word and handwriting samples; Utilize subsequently candidate word apart from calculating degree of confidence, recycling degree of confidence is carried out Chinese-character writing quality assessment.Here, adopt the method based on handwritten Chinese character Character mother plate and candidate characters collection template score ratio to calculate degree of confidence.In general, degree of confidence is higher, and the differentiation of instructions write characters and other candidate characters is larger, and character writing is got over standard.This method is compared with traditional Chinese-character writing quality evaluating method, and the standard of evaluation mainly, based on the choosing of training sample, is write under carefully and neatly done condition at training sample, and this evaluation system is write trimness to sample good evaluating ability.

In realizing process of the present invention, inventor finds that in prior art, at least there are the following problems: in the method that prior art provides, the confidence calculations based on application Character mother plate and candidate characters collection template score ratio adopting is assessed the standard degree of written character, although can distinguish preferably the normalized written degree of most of character, yet to such as " my god; husband ", " people; enter ", " day, say " " oneself,, the sixth of the twelve Earthly Branches " etc. only has the character of nuance to but there is the problem of underestimating its normalized written degree in part.In general, the reference template of obscuring character is comparatively similar, the similarity of the proper vector of corresponding itself and input character is also comparatively approaching, under the confidence calculations based on ratio is set, the degree of confidence score calculating is often on the low side, even if inputted to the correct standard of user this character, system also easily provides the conclusion of the inadequate standard of character writing.For example, for user, input Chinese character " own ", by identification, show that candidate characters is respectively " own ", " ", " the sixth of the twelve Earthly Branches ", suppose that respectively the distance (or similarity) calculating with candidate characters is 0.9,0.8,0.7, under the confidence calculations based on ratio is set, the degree of confidence drawing is 0.375.Although the character of user's input and the similarity of standard form are very high, because the degree of confidence score drawing is on the low side, system can draw writes nonstandard conclusion.Therefore the method that, prior art provides is often not accurate enough for the evaluation of confusable character.

Summary of the invention

For solving the problems of the technologies described above, the embodiment of the present invention provides the method and apparatus of character writing standard degree evaluation and test, can effectively solve in prior art and evaluate inaccurate problem to obscuring character, has improved rationality, the accuracy of normalized written degree assessment.

On the one hand, the embodiment of the present invention provides a kind of method of character writing standard degree evaluation and test, and described method comprises:

Extract the primitive character vector of the stroke track of written character;

The described primitive character vector extracting is mated with the Character mother plate in the first character set, obtain the first degree of confidence;

When described the first degree of confidence of judgement does not meet default evaluation and test condition, carry out Secondary Match and process, obtain Second Characteristic vector, described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence; According to described the second degree of confidence, described written character is evaluated.

Preferably, the primitive character vector of the stroke track of described extraction written character comprises:

Gather and record the stroke track of written character;

Stroke track to described written character carries out pre-service;

Extract the primitive character vector of the stroke track of pretreated written character.

Preferably, the described described primitive character vector by extraction mates with the Character mother plate in the first character set, obtains the first degree of confidence and comprises:

Obtain the standard character template corresponding with written character;

Obtain the first character set;

Respectively the described primitive character vector extracting is mated with the Character mother plate in described standard character template, the first character set, obtain a plurality of similarity values;

According to described a plurality of similarity values, obtain the first degree of confidence.

Preferably, respectively by the described primitive character vector extracting with before Character mother plate in described standard character template, the first character set mates, described method also comprises:

Described primitive character vector is carried out to Feature Conversion, obtain First Characteristic vector;

The described respectively described primitive character vector extracting coupling with Character mother plate in described standard character template, the first character set, is:

The First Characteristic vector obtaining is mated with the Character mother plate in described standard character template, the first character set.

Preferably, described execution Secondary Match is processed, and obtains Second Characteristic vector, and described Second Characteristic vector is mated with the Character mother plate in the second character set, obtains the second degree of confidence and comprises:

Obtain the second character set;

According to described the second character set, obtain Second Characteristic transition matrix, according to described Second Characteristic transformation matrix, described primitive character vector is carried out to eigentransformation to obtain Second Characteristic vector;

Described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence.

Preferably, describedly according to the second character set, obtain Second Characteristic transition matrix and comprise:

According to the standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate determine described standard character template with described in obscure the subset that Character mother plate belongs to altogether;

Judge whether described subset is the root node of decision tree, if not, described subset characteristic of correspondence transition matrix obtained as Second Characteristic transformation matrix.

Preferably, the described basis standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate determine described standard character template with described in obscure the subset that Character mother plate belongs to altogether and comprise:

Judge standard character that written character is corresponding and corresponding with described written character obscure character and whether belong to same subset;

If not, obtain described standard character and described in obscure the upper level subset of character, repeat a determining step;

If so, using described subset as described standard character template with described in obscure the subset that Character mother plate belongs to altogether.

Preferably, described method further comprises:

When described the second degree of confidence of judgement does not meet when pre-conditioned, carry out again Secondary Match and process.

On the other hand, the embodiment of the invention also discloses a kind of character writing standard degree evaluating apparatus, described device comprises:

Feature extraction unit, for extracting the primitive character vector of the stroke track of written character;

The first matching unit, for the described primitive character vector extracting is mated with the Character mother plate of the first character set, obtains the first degree of confidence;

The first judging unit, for judging whether described the first degree of confidence meets pre-conditioned;

The second matching unit, for receiving the judged result of the first judging unit, when judged result shows that described the first degree of confidence does not meet default evaluation and test condition, execution Secondary Match is processed, obtain Second Characteristic vector, described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence;

The second evaluation unit, for evaluating described written character according to described the second degree of confidence.

Preferably, described feature extraction unit comprises:

Collecting unit, for gathering and record the stroke track of written character;

Pretreatment unit, carries out pre-service for the stroke track to described written character;

Extraction unit, for extracting the primitive character vector of the stroke track of pretreated written character.

Preferably, described the second matching unit comprises the first acquiring unit, second acquisition unit, Feature Conversion unit, the 3rd acquiring unit, wherein:

Described the first acquiring unit is used for obtaining the second character set;

Described second acquisition unit is for obtaining Second Characteristic transition matrix according to described the second character set;

Described Feature Conversion unit, for carrying out eigentransformation to obtain Second Characteristic vector according to described Second Characteristic transformation matrix to described primitive character vector;

Described the 3rd acquiring unit, for described Second Characteristic vector is mated with the Character mother plate of the second character set, obtains the second degree of confidence.

Preferably, described second acquisition unit comprises:

Determining unit, for according to the standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate determine described standard character template with described in obscure the subset that Character mother plate belongs to altogether;

The second judging unit, for judging whether described subset is the root node of decision tree, if not, obtain described subset characteristic of correspondence transition matrix as Second Characteristic transformation matrix.

The beneficial effect that the embodiment of the present invention can reach is: the method that the embodiment of the present invention provides is extracted the primitive character vector of the stroke track of written character, the described primitive character vector extracting is once mated with the Character mother plate in the first character set, obtain the first degree of confidence; When judgement does not meet when pre-conditioned according to described the first degree of confidence, carry out Secondary Match and process, obtain Second Characteristic vector, described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence; According to described the second degree of confidence, described written character is evaluated.In the method providing in the embodiment of the present invention, in the time cannot making correct evaluation to written character according to the degree of confidence that once coupling is obtained, written character is carried out to Secondary Match, in Secondary Match owing to having adopted the Second Characteristic vector of high differentiation, and according to described Second Characteristic vector and the second character set, obscure character set and mate, improved the differentiation of Model Matching, the degree of confidence assessment result of obtaining is thus often reliable, has improved accuracy and rationality that normalized written degree is evaluated.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, to the accompanying drawing of required use in embodiment or description of the Prior Art be briefly described below, apparently, the accompanying drawing the following describes is only some embodiment that record in the present invention, for those of ordinary skills, do not paying under the prerequisite of creative work, can also obtain according to these accompanying drawings other accompanying drawing.

Character writing standard degree evaluating method the first embodiment process flow diagram that Fig. 1 provides for the embodiment of the present invention;

Character writing standard degree evaluating method the second embodiment process flow diagram that Fig. 2 provides for the embodiment of the present invention;

The clustering algorithm schematic diagram based on decision tree that Fig. 3 provides for the embodiment of the present invention;

The tree construction schematic diagram of the character subset that Fig. 4 provides for the embodiment of the present invention;

The character writing standard degree evaluating apparatus schematic diagram that Fig. 5 embodiment of the present invention provides.

Embodiment

The embodiment of the present invention provides the method and apparatus of character writing standard degree evaluation and test, can effectively solve in prior art and evaluate inaccurate problem to obscuring character, has improved rationality, the accuracy of normalized written degree assessment.

In order to make those skilled in the art person understand better the technical scheme in the present invention, below in conjunction with the accompanying drawing in the embodiment of the present invention, technical scheme in the embodiment of the present invention is clearly and completely described, obviously, described embodiment is only the present invention's part embodiment, rather than whole embodiment.Embodiment based in the present invention, those of ordinary skills, not making the every other embodiment obtaining under creative work prerequisite, should belong to the scope of protection of the invention.

Referring to Fig. 1, be the method first embodiment process flow diagram of character writing standard degree evaluation and test provided by the invention, described method comprises:

S101, the primitive character vector of the stroke track of extraction written character.

In the method providing in the embodiment of the present invention, can the chosen in advance current Chinese character of wanting exercise of user, and write corresponding character in default writing in region, to set up the character to be assessed of user writing and the corresponding relation of standard character.Certainly, also can not comprise the step of selection, directly provide and write region, gather the stroke track of the character of user writing.System is a series of two-dimensional coordinate point range P by the stroke track record of the character collecting _i(x _i, y _i).

The two-dimensional coordinate point column signal of original person's handwriting is easily subject to the interference of various noises, and has bulk redundancy information, and directly according to it, carrying out the assessment of normalized written degree will cause the decline of operand and assessment accuracy.Therefore, first the method that the embodiment of the present invention provides extracts the proper vector with high sign power, for example eight direction characters or DEF feature (Directional Element Feature, direction element characteristic) etc. from original stroke track.By the extraction of primitive character vector, the person's handwriting of original T sampled point sequence of the written character of user's input is characterized by a D dimensional feature vector.

During specific implementation, first system adjusts to character script predefined size, carries out size normalization processing.Preferably, can also carry out pre-service to the character script after adjusting, such as obtaining clean character track by preconditioning techniques such as non-linear regular, level and smooth, resamplings.The two-dimensional space that system forms from the character trajectory coordinates of original time domain is subsequently considered each sampling point position of person's handwriting, forms and describes character pixel distribution two dimensional image.Finally, on this two dimensional image, it is carried out to subregion, in each individual region, extract the feature of describing this area pixel regularity of distribution, and form accordingly the pixel distribution law characteristic vector of describing whole two dimensional image.Concrete, the eigenvector of extraction can be eight direction characters or DEF feature etc., the present invention does not limit this.

To extract from all directions, to being characterized as example, describe below.Wherein, eight direction characters are for weighing the regularity of distribution of the projection components of character picture in eight directions of systemic presupposition.Concrete, before mention, system is transformed into two dimensional image by the character trajectory coordinates of original time domain, then two dimensional image is carried out to subregion, at this moment, for each the character sampled point in each the independent subregion obtaining, calculate respectively itself and the projected size of section in 8 directions that a last or rear sampled point forms, subsequently the component in eight directions in this subregion is passed through to a new octuple feature of the method formation such as accumulative total.Finally by the octuple merging features of each subregion, or form new eigenvector by statistical methods such as Gauss fit.In embodiments of the present invention, character script is used as to a two dimensional image and extracts eigenvector, consider emphatically its characteristic of spatial distribution.Describe for example, system by the person's handwriting of time sequencing according to the regular image to N*N size space of the position distribution of person's handwriting sampled point, subsequently the image of this N*N is divided into 8*8 little subregion, the projection of all sampled points of considering respectively each subregion on eight different directions, and accumulative total obtains the eigenvector of one 8 dimension.Last each subregion provides 8 dimensional features, and 8*8 subregion will provide the 8*8*8 feature of dimension.Optionally, system also can continue to the feature of 8*8*8 to be carried out to aftertreatment, or fits method to give prominence to the differentiation between character by Gauss.

S102, mates the described primitive character vector extracting with the Character mother plate in the first character set, obtain the first degree of confidence.

S102A, obtains the first character set.

In first embodiment of the invention, the first character set is the character model collection of system intialization, stores all character reference templates that system is supported.Concrete, described character model collection is that system is by training and obtain on the magnanimity training sample gathering, for simulating the features such as stroke, the order of strokes observed in calligraphy and font architecture of character.Concrete, in order to improve running efficiency of system, in embodiments of the present invention, each character adopts the model structure of single mode plate, and utilizes its corresponding training sample to train its model parameter.Wherein the single mode Slab parameter of i character is designated as:

${\mathrm{\μ}}_i=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{X}_j}{N}---\left(1\right)$

Wherein, N is the training sample number that this character is corresponding, X _jit is the eigenvector of j training sample of this character.

S102B, calculates similarity.

Concrete, when obtaining the first degree of confidence, be that the described primitive character vector extracting is mated to calculating similarity value with the Character mother plate in the first character set, according to described similarity value, obtain the first degree of confidence.Wherein, under the setting of single mode Slab, input character eigenvector X and i character model M _ibetween similarity be calculated as shown in formula (2):

$s(X,{\mathrm{\μ}}_i)=1/\sqrt{{(X-{\mathrm{\μ}}_i)}^T(X-{\mathrm{\μ}}_i)}---\left(2\right)$

S102C, calculates the first degree of confidence.

Obtain the method for the first degree of confidence concrete can select the method based on posterior probability simulation, the method based on training sample score distribution simulation etc., the present invention does not limit this.

Such as system can be calculated the first degree of confidence with respect to the mode of the posterior probability of standard character template by calculating described written character:

$p\left(M_c\right|X)\≈\frac{p\left(X\right|M_c)}{\underset{j\∈S}{\mathrm{\Σ}}p\left(X\right|M_j)}\≈\frac{s(X,{\mathrm{\μ}}_c)}{\underset{j\∈S}{\mathrm{\Σ}}s(X,{\mathrm{\μ}}_j)}---\left(3\right)$

P (M wherein _c| X)) represent that described written character feature X belongs to standard character template M _cprobability.P (X|M _c) represent described written character feature X and standard character template M _csimilarity.S set is the first character set, p (X|M _j) be Character mother plate M in described written character feature X and the first character set _jsimilarity.

Obviously posterior probability is larger, and the first recognition result more can distinguish over other recognition results, and its identification is more reliable, otherwise more unreliable.

S103, when described the first degree of confidence of judgement does not meet default evaluation and test condition, carries out Secondary Match and processes, and obtains Second Characteristic vector, and described Second Characteristic vector is mated with the Character mother plate in the second character set, obtains the second degree of confidence.

Default condition can be the threshold interval of systemic presupposition, for example, if system can arrange degree of confidence score when being less than predetermined threshold value T, judges that it does not meet default condition, carries out Secondary Match.Further, in order to improve the efficiency of Secondary Match, if system can also arrange degree of confidence score being greater than T1, be less than T2 this when interval, carry out Secondary Match.Here, threshold value T, T1, T2 are system default parameters on training sample analysis foundation, for simulating the not high interval of degree of confidence causing because character is similar.

When described the first degree of confidence of judgement does not meet when pre-conditioned, carry out Secondary Match and process.Here, the second character set is specially and obscures character set.In embodiments of the present invention, the second character set is directly known according to a matching result.When being greater than systemic presupposition value such as the difference when N and N+1 optimal identification result similarity, we are using character corresponding to top n recognition result as the second character set.In first embodiment of the invention, while obtaining Second Characteristic vector, the Second Characteristic transition matrix that first finds the second character set to be associated by traditional decision-tree.According to the Second Characteristic transformation matrix obtaining, primitive character vector is carried out to eigentransformation to obtain Second Characteristic vector, described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence.

S104, evaluates described written character according to described the second degree of confidence.

If the second degree of confidence is greater than default thresholding, normalized written is described, otherwise lack of standardization.

In first embodiment of the invention, in the time cannot making correct evaluation to written character according to the degree of confidence that once coupling is obtained, written character is carried out to Secondary Match, in Secondary Match owing to having obtained the Second Characteristic vector of high differentiation, and according to described Second Characteristic vector and the second character set, obscure character set and mate, improved the differentiation of Model Matching, the degree of confidence assessment result of obtaining is thus often reliable, has improved accuracy and rationality that normalized written degree is evaluated.

Referring to Fig. 2, it is the method second embodiment process flow diagram of character writing standard degree evaluation and test provided by the invention.

S201, gathers and records the stroke track of written character.

In method provided by the invention, can the chosen in advance current Chinese character of wanting exercise of user, and write corresponding character in default writing in region, to set up the character to be assessed of user writing and the corresponding relation of standard character.The stroke track that gathers the character of user writing is a series of two-dimensional coordinate point range P by the stroke track record of the character collecting _i(x _i, y _i).

S202, carries out pre-service to the stroke track of described written character.

By the written character stroke track gathering is carried out to pre-service, can reduce the random signals such as the burr in person's handwriting, wild point, reduce noise jamming.

Concrete, the stroke track of described written character is carried out to pre-service and can comprise following any one or more step:

(1) the stroke track of described written character is carried out to open country point Transformatin.

(2) the stroke track of described written character is carried out to size normalization processing.Concrete, the stroke trajectory map of the written character collecting is to default size, concrete, can be mapped to the size identical with character in Character mother plate.

(3) the stroke track of described written character is carried out to pixel equilibrium treatment.

(4) the stroke track of described written character is carried out to smoothing processing.

(5) the stroke track of described written character being carried out to stroke resampling processes.

S203, extracts the original feature vector of the stroke track of pretreated written character.

Because original two-dimensional coordinate point column signal is easily subject to the interference of various noises, and have bulk redundancy information, directly according to it, carrying out the assessment of normalized written degree will cause the decline of operand and assessment accuracy.Therefore, first the method that the embodiment of the present invention provides extracts the proper vector with high sign power, such as eight direction characters or DEF feature etc. from original stroke track.By the extraction of primitive character vector, the person's handwriting of original T sampled point sequence of the written character of user's input is characterized by a D dimensional feature vector.Concrete implementation method is identical with the first embodiment, does not repeat them here.

S204, mates the described primitive character vector extracting with the Character mother plate in the first character set, obtain the first degree of confidence.

Step S204 can realize by following steps:

S204A, obtains the standard character template corresponding with written character.

As above-mentioned, be different from hand-written discrimination system and cannot judge the markup information of input character before system identification, in hand-written evaluating system, user's character that often chosen in advance need to be learnt is also write in default region, because the corresponding correct markup information of its written character can obtain in advance.To this, in second embodiment of the invention, first obtain the reference template that input character is corresponding, and for the similarity of subsequent calculations current input character person's handwriting and its standard form.

S204B, obtains the first character set.

Different from the first embodiment, in order to reduce the operand of system, in second embodiment of the invention, the first character set is the obscure character set relevant to written character.When calculating similarity, can directly calculate the distortion of obscuring character subset relevant to input character.This obscures character subset can be by system in advance according to the settings such as degree of obscuring of the similarity of character and character recognition on training sample.

S204C, mates the described primitive character vector extracting respectively with the Character mother plate in described standard character template, the first character set, obtain a plurality of similarity values.

Optionally, by primitive character vector with before Character mother plate in described standard character template, the first character set mates, can also comprise: described primitive character vector is carried out to Feature Conversion, obtain First Characteristic vector.By the primitive character vector X of input person's handwriting is carried out to eigentransformation, to improve the differentiation between kinds of characters and to reduce system operand.For example, by LDA (Linear Discriminant Analysis, linear differentiation analyzed) scheduling algorithm, obtain eigentransformation matrix W, wherein, W is the matrix of a D * M, and meets M < D.By eigentransformation, by primitive character X, from the eigenvector transform of a D dimension, be the proper vector of a M dimension:

X‘＝W ^TX (4)

Like this, after eigentransformation, with respect to primitive character vector, improved the differentiation of coupling.

When coupling, be that the First Characteristic vector after conversion is mated with the Character mother plate in described standard character template, the first character set, calculating similarity value.Input character eigenvector X and i character model M _ibetween similarity be calculated as shown in formula (5):

$s(X,{\mathrm{\&mu;}}_i)=1/\sqrt{{(X-{\mathrm{\&mu;}}_i)}^T(X-{\mathrm{\&mu;}}_i)}---\left(5\right)$

From above-mentioned formula, can find out, between described written character feature and Character mother plate, distance is larger, and its probability that belongs to Character mother plate is less.

S204D, obtains the first degree of confidence according to described a plurality of similarity values.

Such as system can be calculated the first degree of confidence with respect to the mode of the posterior probability of standard character template by calculating described written character:

$p\left(M_c\right|X)\&ap;\frac{p\left(X\right|M_c)}{\underset{j\&Element;S}{\mathrm{\&Sigma;}}p\left(X\right|M_j)}\&ap;\frac{s(X,{\mathrm{\&mu;}}_c)}{\underset{j\&Element;S}{\mathrm{\&Sigma;}}s(X,{\mathrm{\&mu;}}_j)}---\left(6\right)$

P (M wherein _c| X)) represent that described written character feature X belongs to standard character template M _cprobability.P (X|M _c) represent described written character feature X and standard character template M _csimilarity.P (X|M _j) be Character mother plate M in described written character feature X and the first character set _jsimilarity.

Obviously posterior probability is larger, and the first recognition result more can distinguish over other recognition results, and its identification is more reliable, otherwise more unreliable.

S205, judges that the first degree of confidence is whether in (T1, T2) interval, if enter step S206, if not, enter step S211.

Here, threshold value T1, T2 are system default parameters on training sample analysis foundation, for simulating the not high interval of degree of confidence causing because character is similar.

S206, obtains the second character set.

In second embodiment of the invention, be to obtain based on decision Tree algorithms Second Characteristic transformation matrix that the second character set is associated.Concrete, the second character set is for obscuring character set.In embodiments of the present invention, the second character set is directly according to a matching result, to obtain, when being greater than systemic presupposition value such as the difference when N and N+1 optimal identification result similarity, we are using character corresponding to top n recognition result as the second character set.

S207, obtains Second Characteristic transformation matrix according to described the second character set.

In embodiments of the present invention, use decision Tree algorithms to obtain the Second Characteristic transition matrix that the second character set is associated.Because the character that the second character set comprises may disperse in tree, so need to upwards obtain one by one covering all Second Characteristic transformation matrixs of obscuring character by recursive fashion.

The clustering algorithm schematic diagram based on decision tree providing for the embodiment of the present invention referring to Fig. 3.To the present invention is based on the character subset building process of decision tree structure, be introduced below.

The embodiment of the present invention adopts clustering algorithm from bottom to top, and the character set of system support is clustered into multistage subset structure, idiographic flow as shown in Figure 3:

S301, initialization.Current system subset number M is set and equals character number, the leaf node that tree structure is set is single character.

S302, calculates the similarity of every two subclasses.

Concrete can obtain subclass similarity by calculating the variance of two subclass sample datas.Suppose subclasses C _iin comprise N _iindividual sample, subclasses C _jin comprise N _jindividual sample, subclasses C _iand subclasses C _jsample variance after merging is calculated as:

${\mathrm{\&Sigma;}}_{\mathrm{ij}}=\frac{1}{N_i}\underset{S_k\&Element;C_i}{\mathrm{\&Sigma;}}\sqrt{{(S_k-{\mathrm{\&mu;}}_{\mathrm{ij}})}^T(S_k-{\mathrm{\&mu;}}_{\mathrm{ij}})}+\frac{1}{N_j}\underset{S_m\&Element;C_j}{\mathrm{\&Sigma;}}\sqrt{{(S_m-{\mathrm{\&mu;}}_{\mathrm{ij}})}^T(S_m-{\mathrm{\&mu;}}_{\mathrm{ij}})}---\left(7\right)$

Here S _krepresent subclasses C _iin k sample, S _mrepresent subclasses C _jin m sample, μ _ijrepresent subclasses C _iand subclasses C _jthe average of all samples.Sample variance is less, and subclass similarity is larger.

Certainly the measure of subclass similarity is not only confined to sample variance, can also simply by calculating the distance of the sample average of two subclasses, obtain subclass similarity, concrete:

${\mathrm{\&Sigma;}}_{\mathrm{ij}}=\sqrt{{({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)}^T({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_j)}---\left(8\right)$

Or

${\mathrm{\&Sigma;}}_{\mathrm{ij}}=\frac{1}{2}\sqrt{{({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_{\mathrm{ij}})}^T({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_{\mathrm{ij}})}+\frac{1}{2}\sqrt{{({\mathrm{\&mu;}}_j-{\mathrm{\&mu;}}_{\mathrm{ij}})}^T({\mathrm{\&mu;}}_j-{\mathrm{\&mu;}}_{\mathrm{ij}})}---\left(9\right)$

μ _i, μ _jit is respectively subclasses C _iand subclasses C _jsample average.Similarly, sample average distance is less, and subclass similarity is larger.

Above-mentioned account form is only preferred embodiment of the present invention, here similarity calculating method is not limited.

S303, two subclasses selecting to have maximum similarity are the subclass of new higher level.

S304, judges that whether current subclass number equals 1, if proceed to S305, otherwise proceeds to S302.

S305, finishes.

Especially, the subset node to each structure, system all utilizes corresponding training sample to train corresponding eigentransformation matrix, for improving the differentiation of kinds of characters in subset.The root node of special this decision tree is the set of all characters of system support, and its characteristic of correspondence transformation matrix is the global change's matrix adopting in a template matches.Fig. 4 has shown the tree construction schematic diagram of character subset.

S207A, according to the standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate settle the standard Character mother plate with obscure the subset that Character mother plate belongs to altogether.

Concrete, judge standard character that written character is corresponding and corresponding with described written character obscure character and whether belong to same subset; If not, obtain described standard character and described in obscure the upper level subset of character, repeat a determining step; If so, using described subset as standard character template with obscure the subset that Character mother plate belongs to altogether.

Here, obscure character set and be like this and determine: mention above, in step S204D, can calculate written character with respect to the similarity of the first recognition result and described written character the difference with respect to the second recognition result similarity.Similarly, can calculate the difference between the similarity of optimal identification result of N and N+1, when the difference of judgement N and N+1 recognition result similarity is greater than systemic presupposition value, we are using character corresponding to top n recognition result as obscuring character set.

S207B, judges that whether described subset is the root node of decision tree, if so, enters step S207C; If not, enter S207D.

S207C, finishes Secondary Match.

S207D, obtains Second Characteristic transition matrix according to described subset.

In embodiments of the present invention, the subset node to each structure, system all utilizes corresponding training sample to train corresponding eigentransformation matrix, for improving the differentiation of kinds of characters in subset.

S208, carries out eigentransformation to obtain Second Characteristic vector according to described Second Characteristic transformation matrix to described primitive character vector.

The concrete eigentransformation matrix W of utilizing, is mapped to new feature space by primitive character vector X, obtains Second Characteristic vector:

Y＝W ^TX (10)

S209, mates described Second Characteristic vector with the Character mother plate in the second character set, obtain similarity.

Concrete, be that Second Characteristic vector corresponding to written character mated between two with the character of obscuring of obscuring in character set, obtain similarity.

S210, obtains the second degree of confidence.

Same, the described character calculating according to Secondary Match and standard form and the similarity of obscuring character corresponding templates are calculated the second degree of confidence.Concrete, the mode that can adopt the described character of calculating to belong to the posterior probability of standard form is obtained the second degree of confidence.

Such as system can be calculated the second degree of confidence with respect to the mode of the posterior probability of standard character template by calculating described written character:

$p\left(M_c\right|Y)\&ap;\frac{p\left(Y\right|M_c)}{\underset{j\&Element;S^{\&prime;}}{\mathrm{\&Sigma;}}p\left(Y\right|M_j)}\&ap;\frac{s(Y,{\mathrm{\&mu;}}_c)}{\underset{j\&Element;S^{\&prime;}}{\mathrm{\&Sigma;}}s(Y,{\mathrm{\&mu;}}_j)}---\left(11\right)$

P (M wherein _c| Y) represent that described written character characteristic Y belongs to standard character template M _cprobability.P (Y|M _c) represent described written character characteristic Y and standard character template M _csimilarity.S set ' be the second character set, p (Y|M _j) be Character mother plate M in described written character characteristic Y and the second character set _jsimilarity.Similarly, s (Y, μ _i) be defined as eigenvector Y and character model μ _ibetween similarity, by formula (12), calculated:

$s(Y,{\mathrm{\&mu;}}_i)=1/\sqrt{{(Y-W^T{\mathrm{\&mu;}}_i)}^T(Y-W^T{\mathrm{\&mu;}}_i)}---\left(12\right)$

Wherein W is that system is according to the Second Characteristic transition matrix of described subset Dynamic Acquisition.Obviously posterior probability is larger, and the first recognition result more can distinguish over other recognition results, and its identification is more reliable, otherwise more unreliable.

S211, evaluates written character according to the degree of confidence of obtaining.

For example, when described degree of confidence is greater than systemic presupposition threshold value, such as 0.7, can judge that described character meets normalized written requirement.In general, predetermined threshold value is larger, higher to user writing code requirement.Further, system can also arrange a plurality of threshold values, so that finer grading system to be provided.Such as when described degree of confidence is greater than system the first predetermined threshold value, such as 0.8, can judge that described character meets normalized written requirement.When if described degree of confidence is less than systemic presupposition threshold value first but is greater than the second predetermined threshold value, such as 0.6 o'clock, can judge that described character meets normalized written requirement substantially.

Optionally, after step S210, a determining step be can also comprise, when described the second degree of confidence of judgement does not meet when pre-conditioned, S206 and step afterwards repeated.That is to say, the degree of confidence of obtaining when Secondary Match does not still meet when pre-conditioned, and the present invention can carry out three couplings, four couplings ... until that the degree of confidence of obtaining meets is pre-conditioned.

In second embodiment of the invention, a kind of normalized written degree evaluation method of estimating based on multistage degree of confidence has been proposed.Concrete, first this system adopts overall general-purpose algorithm once to judge the standard degree of written character, by comparing and improved the objectivity of system evaluation with multiword symbol more.To the character that is difficult to judge, adopt Secondary Match mode to improve the rationality of normalized written degree assessment subsequently.When carrying out Secondary Match, a kind of property distinguished of the height based on decision tree algorithm has been proposed, the differentiation property improvement of obscuring character by raising is to obscuring the rationality of character writing standard degree assessment, and concrete is by obtaining written character and obscuring the differentiation that the eigentransformation matrix of the corresponding subset of character improves Model Matching.Finally, the present invention proposes a kind of subset based on tree structure and build and searching algorithm, reduced on the one hand the model parameter of system, improved the utilization factor of eigentransformation matrix, by improving search speed, improved system responses efficiency on the other hand.

Referring to Fig. 5, the device schematic diagram of the character writing standard degree evaluation and test providing for the embodiment of the present invention.Described device comprises:

Feature extraction unit 501, for extracting the primitive character vector of the stroke track of written character.

The first matching unit 502, for the described primitive character vector extracting is mated with the Character mother plate of the first character set, obtains the first degree of confidence;

The first judging unit 503, for judging whether described the first degree of confidence meets default evaluation and test condition;

The second matching unit 504, for receiving the judged result of the first judging unit, when described judged result shows that described the first degree of confidence does not meet default evaluation and test condition, carries out Secondary Match and processes, and obtains Second Characteristic vector; Described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence.

The second evaluation unit 505, for evaluating described written character according to described the second degree of confidence.

Wherein, described feature extraction unit 501 comprises:

Collecting unit, for gathering and record the stroke track of written character.

Pretreatment unit, carries out pre-service for the stroke track to described written character.

Extraction unit, for extracting the primitive character vector of the stroke track of pretreated written character.

Concrete, described pretreatment unit comprises following any one or more subelement:

Wild point is removed subelement, for the stroke track of described written character is carried out to open country point Transformatin.

Normalization subelement, for carrying out size normalization processing by the stroke track of described written character.

Balanced subelement, for carrying out pixel equilibrium treatment by the stroke track of described written character.

Level and smooth subelement, for to carrying out smoothing processing by the stroke track of described written character.

Resampling subelement, processes for the stroke track of described written character being carried out to stroke resampling.

Concrete, described the second matching unit, comprising:

The first acquiring unit, for obtaining the second character set.

Second acquisition unit is for obtaining Second Characteristic transition matrix according to described the second character set.

Feature Conversion unit is for carrying out eigentransformation to obtain Second Characteristic vector according to described Second Characteristic transformation matrix to described primitive character vector.

The 3rd acquiring unit mates described Second Characteristic vector with the Character mother plate in the second character set, obtain the second degree of confidence.

Wherein, the Second Characteristic transformation matrix that described second acquisition unit is also associated for obtain the second character set based on decision Tree algorithms.

Concrete, described second acquisition unit comprises:

Determining unit, for according to the standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate settle the standard Character mother plate with obscure the subset that Character mother plate belongs to altogether.

The second judging unit, for judging whether described subset is the root node of decision tree, if not, obtain described subset characteristic of correspondence transformation matrix.

Described device also comprises:

The first evaluation unit, when meeting default evaluation and test condition when the first degree of confidence, evaluates written character according to the first degree of confidence.

It should be noted that, in this article, relational terms such as the first and second grades is only used for an entity or operation to separate with another entity or operational zone, and not necessarily requires or imply and between these entities or operation, have the relation of any this reality or sequentially.And, term " comprises ", " comprising " or its any other variant are intended to contain comprising of nonexcludability, thereby the process, method, article or the equipment that make to comprise a series of key elements not only comprise those key elements, but also comprise other key elements of clearly not listing, or be also included as the intrinsic key element of this process, method, article or equipment.The in the situation that of more restrictions not, the key element being limited by statement " comprising ... ", and be not precluded within process, method, article or the equipment that comprises described key element and also have other identical element.

The present invention can describe in the general context of the computer executable instructions of being carried out by computing machine, for example program module.Usually, program module comprises the routine carrying out particular task or realize particular abstract data type, program, object, assembly, data structure etc.Also can in distributed computing environment, put into practice the present invention, in these distributed computing environment, by the teleprocessing equipment being connected by communication network, be executed the task.In distributed computing environment, program module can be arranged in the local and remote computer-readable storage medium that comprises memory device.

The above is only the specific embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (8)

1. a method for character writing standard degree evaluation and test, is characterized in that, described method comprises:

Extract the primitive character vector of the stroke track of written character;

The described primitive character vector extracting is once mated with the Character mother plate in the first character set, obtain the first degree of confidence;

When described the first degree of confidence of judgement does not meet default evaluation and test condition, execution Secondary Match is processed: obtain the second character set, according to described the second character set, based on decision Tree algorithms, obtain Second Characteristic transformation matrix, according to described Second Characteristic transformation matrix, described primitive character vector is carried out to eigentransformation to obtain Second Characteristic vector, described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence; Wherein, described the second character set, for obscuring character set, directly obtains according to a matching result;

According to described the second degree of confidence, described written character is evaluated, when described the second degree of confidence is greater than predetermined threshold value, determined described written character normalized written; When described the second degree of confidence is not more than predetermined threshold value, definite described written character is write lack of standardization;

Wherein, according to described the second character set, based on decision Tree algorithms, obtaining Second Characteristic transformation matrix comprises:

According to the standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate determine described standard character template with described in obscure the subset that Character mother plate belongs to altogether;

Judge whether described subset is the root node of decision tree, if not, described subset characteristic of correspondence transformation matrix obtained as Second Characteristic transformation matrix.

2. method according to claim 1, is characterized in that, the primitive character vector of the stroke track of described extraction written character comprises:

Gather and record the stroke track of written character;

Stroke track to described written character carries out pre-service;

Extract the primitive character vector of the stroke track of pretreated written character.

3. method according to claim 1, is characterized in that, the described described primitive character vector by extraction once mates with the Character mother plate in the first character set, obtains the first degree of confidence and comprises:

Obtain the standard character template corresponding with written character;

Obtain the first character set;

Respectively the described primitive character vector extracting is mated with the Character mother plate in described standard character template, the first character set, obtain a plurality of similarity values;

According to described a plurality of similarity values, obtain the first degree of confidence.

4. method according to claim 3, is characterized in that, the described respectively described primitive character vector extracting coupling with Character mother plate in described standard character template, the first character set is:

Described primitive character vector is carried out to Feature Conversion, obtain First Characteristic vector;

The First Characteristic vector obtaining is mated with the Character mother plate in described standard character template, the first character set.

5. method according to claim 1, it is characterized in that, the standard character template that described basis is corresponding with described written character and with described written character corresponding obscure Character mother plate determine described standard character template with described in obscure the subset that Character mother plate belongs to altogether and comprise:

Judge standard character that written character is corresponding and corresponding with described written character obscure character and whether belong to same subset;

If not, obtain described standard character and described in obscure the upper level subset of character, repeat a determining step;

If so, using described subset as described standard character template with described in obscure the subset that Character mother plate belongs to altogether.

6. method according to claim 1, is characterized in that, described method further comprises:

When described the second degree of confidence of judgement does not meet when pre-conditioned, carry out again Secondary Match and process.

7. a character writing standard degree evaluating apparatus, is characterized in that, described device comprises:

Feature extraction unit, for extracting the primitive character vector of the stroke track of written character;

The first matching unit, for the described primitive character vector extracting is once mated with the Character mother plate of the first character set, obtains the first degree of confidence;

The first judging unit, for judging whether described the first degree of confidence meets default evaluation and test condition;

The second matching unit, for receiving the judged result of the first judging unit, when judged result shows that described the first degree of confidence does not meet default evaluation and test condition, execution Secondary Match is processed: obtain Second Characteristic vector, described Second Characteristic vector is mated with the Character mother plate in the second character set, obtain the second degree of confidence;

The second evaluation unit, for according to described the second degree of confidence, described written character being evaluated, when described the second degree of confidence is greater than predetermined threshold value, determines described written character normalized written; When described the second degree of confidence is not more than predetermined threshold value, definite described written character is write lack of standardization;

Wherein, described the second matching unit comprises the first acquiring unit, second acquisition unit, eigentransformation unit, the 3rd acquiring unit, wherein:

Described the first acquiring unit is used for obtaining the second character set; Described the second character set, for obscuring character set, directly obtains according to a matching result;

Described second acquisition unit is for obtaining Second Characteristic transformation matrix according to described the second character set based on decision Tree algorithms;

Described eigentransformation unit, for carrying out eigentransformation to obtain Second Characteristic vector according to described Second Characteristic transformation matrix to described primitive character vector;

Described the 3rd acquiring unit, for described Second Characteristic vector is mated with the Character mother plate of the second character set, obtains the second degree of confidence;

Wherein, described second acquisition unit comprises:

Determining unit, for according to the standard character template corresponding with described written character and with described written character corresponding obscure Character mother plate determine described standard character template with described in obscure the subset that Character mother plate belongs to altogether;

The second judging unit, for judging whether described subset is the root node of decision tree, if not, obtain described subset characteristic of correspondence transformation matrix as Second Characteristic transformation matrix.

8. device according to claim 7, is characterized in that, described feature extraction unit comprises:

Collecting unit, for gathering and record the stroke track of written character;

Pretreatment unit, carries out pre-service for the stroke track to described written character;

Extraction unit, for extracting the primitive character vector of the stroke track of pretreated written character.