CN101354749A - Method for making dictionary, hand-written input method and apparatus - Google Patents

Abstract

The invention discloses a method for manufacturing a dictionary, a handwriting input method and equipment. The method for manufacturing the dictionary can predict a character input by strokes so as to relieve the burden of a user. The method comprises the following steps: the feature of the integral character of an integral character sample of characters and the feature of partial strokes of a partial stroke sample of the characters with the number of the strokes larger than a preset value are extracted; and a learning algorithm by a machine is used to carry out learning to the feature of the integral character and the feature of the partial strokes to generate an integral character template and/or a partial stroke template of the characters, which are used as items of the dictionary. The system has the advantages of simple structure, low hardware requirement, rapid identification speed, high identification rate, and so on, and can be realized in an embedded type system and so on.

Description

Dictionary method for making, hand-written inputting method and equipment

Technical field

The present invention relates to the literal input, be specifically related to a kind of dictionary method for making, hand-written inputting method and equipment that is applied to electronic product, the literal that the stroke prediction that can enoughly import will be imported is so that alleviate user's burden.

Background technology

For a long time, the quick input of non-alphabetic character is puzzlement people's a difficult problem such as Chinese character, because the keyboard of computing machine is applicable to the input of western language such as English, and is not suitable for the input of Chinese character.

The input of Chinese character is divided into two kinds of keyboard input and handwriting inputs usually.The keyboard input is to give certain coding for each Chinese character according to certain coding rule, discerns Chinese character by the keyboard input coding, for example various spelling input methods and five-stroke character input method.Handwriting input is discerned Chinese character by the person's handwriting that the identification people directly write, thereby reaches the purpose of input Chinese character.Because the keyboard input needs the user skillfully to use keyboard and/or keeps the various codes or the coding rule of each Chinese character firmly in mind, can't promote the use of in the people that seldom use keyboard.In addition, because the cause of dialect, a lot of people can write certain Chinese character, really can't accurately read its pronunciation, and this makes spelling input method run into the difficulty that is difficult to overcome in actual use.

Because do not need the user to carry out the study of input method before using, hand-writing input method has obtained seldom using the people's of keyboard welcome.In principle, handwriting input does not need the user to carry out any study, as long as he can write Chinese character.

Patent documentation 1 (US6028959) has disclosed a kind of method that adopts the stroke forecasting techniques to import Chinese character.According to the stroke of the Chinese character of having write, before whole Chinese character has been write, the just measurable Chinese character that goes out will write, thus improved the speed of handwriting input greatly.Particularly, the method for patent documentation 1 has adopted time-delay neural network (TDNN) to carry out the stroke prediction in conjunction with the mode of multilayer perceptron network (MLP).

Patent documentation 2 (spy opens flat 2005-25566) has disclosed a kind of method of handwriting input Chinese character, wherein the hand-written stroke centralized stores that input part is imported is in storage part, with comprising the coordinate feature in the storage part, mating, use to comprise various method for mode matching generation forecast candidates such as OCR, DP to the searching object information of measure feature and graphic feature etc. and the dictionary of creating in advance.Then, the candidate selection portion literal that selection will be imported from the candidate of prediction.Input part with the input stroke set as a whole image handle, can reduce user's burden.

But the hand-written inputting method prediction steps of patent documentation 1 is too complicated, it adopted comprise two classes totally 68 neural networks relate to many parameters of 5M, cause forecasting process very complicated.Do not need to write complete word though the method for patent documentation 2 has provided, just can want the method for input results by the coupling prediction, this invention is handled the integral body of the stroke of input as image, cause inefficiency.

Summary of the invention

In view of prior art problems, finished the present invention.The purpose of this invention is to provide a kind of dictionary creating method and apparatus, hand-written inputting method and the equipment of innovation fully, can predict the literal that to import by the stroke of writing, so that alleviate user's burden.

In a first aspect of the present invention, a kind of method of making dictionary has been proposed, comprise step: the whole word feature of the whole printed words basis of extraction literal and stroke number are greater than the part stroke feature of the part stroke sample of the literal of predetermined value; And by described whole word feature and described part stroke feature being learnt to generate the whole character matrix plate and/or the part stroke template of literal, as the project in the dictionary with machine learning algorithm.

In a second aspect of the present invention, a kind of equipment of making dictionary has been proposed, comprising: the whole word feature of the whole printed words basis of extraction literal and stroke number are greater than the device of the part stroke feature of the part stroke sample of the literal of predetermined value; And by described whole word feature and described part stroke feature being learnt to generate the whole character matrix plate and/or the part stroke template of literal, as the device of the project in the dictionary with machine learning algorithm.

In a third aspect of the present invention, a kind of hand-written inputting method has been proposed, comprise the feature to the small part handwriting of step extraction literal; And calculate described feature and the dictionary created according to the method for above-mentioned first aspect in template between distance; And will be apart from the literal of the template representative of less at least one as recognition result.

In a fourth aspect of the present invention, a kind of handwriting input device has been proposed, comprising: the device that extracts literal to small part handwriting feature; The device of the distance between the template in the dictionary that calculates described feature and create according to the method for above-mentioned first aspect; And will be apart from the literal of the template representative of less at least one device as recognition result.

The present invention proposes complete printed words basis and part stroke printed words originally, and made recognition dictionary on this basis, by mating with this recognition dictionary, hand-written user is input characters not exclusively, automatically dope the literal candidate that to import, reduce Writer's burden.That native system has is simple in structure, hardware requirement is low, recognition speed is fast, and the discrimination advantages of higher can realize on embedded system etc.

In addition, in the manufacturing process of recognition dictionary, the feature of being extracted and the order of strokes observed in calligraphy of handwriting, connect pen and stroke number is irrelevant, thereby make the user when writing, break away from the order of strokes observed in calligraphy, connect the restriction of pen and stroke number.

In addition,, both greatly reduced the required internal memory of recognition dictionary, realized miniaturization, reduced the calculated amount in the identifying again, avoided floating-point operation, improved recognition speed, helped the realization of high speed by dimensionality reduction and quantification.

In addition, in identifying, adopted the sectional type fast matching method, filtering candidate item is step by step dwindled comparison range, under the situation that influences discrimination hardly, improve recognition speed greatly, finally ensured the realization of hand script Chinese input equipment character identification system high speed.

In addition, the tabulation of ten candidate provides a kind of friendly more operation interface in conjunction with the background prompting mode, has avoided the frequent transfer of user's sight line between input field and candidate regions, reduce user's working strength so on the one hand, improved handwriting input speed on the other hand again.

In addition, send the word mode automatically by the self-adaptation non-timed, system can write according to user's writing style and institute, adjusts the interval stand-by period between word and the word intelligently, and a kind of more humane control mode is provided, and also makes handwriting input more efficient.

Description of drawings

By below in conjunction with description of drawings the preferred embodiments of the present invention, will make above-mentioned and other purpose of the present invention, feature and advantage clearer, wherein:

Fig. 1 shows the functional block diagram according to the handwriting input device of the embodiment of the invention;

Fig. 2 is the automatic generation of part stroke printed words basis and the synoptic diagram of whole word class and part stroke word class;

Fig. 3 is the process flow diagram of describing according to the dictionary method for making of the embodiment of the invention;

Fig. 4 is the synoptic diagram that is described in the equidistant re-sampling operations of carrying out in the preprocessing process;

Fig. 5 is the synoptic diagram of the barycenter normalization carried out in preprocessing process and non-linear normalizing operation;

Fig. 6 is a synoptic diagram of describing the process of extracting the stroke direction distribution characteristics;

Fig. 7 is a synoptic diagram of describing the process of extracting the grid stroke feature;

Fig. 8 is a synoptic diagram of describing the process of extracting the peripheral direction feature;

Fig. 9 describes the synoptic diagram utilize the process that GLVQ learns;

Figure 10 is the detail flowchart according to the identifying in the hand-written inputting method of the embodiment of the invention;

Figure 11 is a synoptic diagram of describing quick matching process; And

Figure 12 is the synoptic diagram of ten candidate tabulation in conjunction with first-selected word background prompting.

Embodiment

To a preferred embodiment of the present invention will be described in detail, having omitted in the description process is unnecessary details and function for the present invention with reference to the accompanying drawings, obscures to prevent that the understanding of the present invention from causing.

Fig. 1 shows the functional block diagram according to the handwriting input device of the embodiment of the invention.As shown in Figure 1, comprise according to the handwriting input device of the embodiment of the invention: handwriting input unit 110, be used to gather user's person's handwriting, and to its digitizing, as input person's handwriting signal; Handwriting storage unit 120 is used to store the input person's handwriting signal that handwriting input unit 110 produces; Identification prediction unit 130 is used for extracting from input person's handwriting signal the feature of this signal, the feature of the some or all of stroke of certain word for example, and itself and the template in the dictionary that is pre-created mated, produce the input candidate according to matching degree; Indicative control unit 150, when the user is by handwriting input unit 110 input strokes, on hand-written display unit 160, show person's handwriting, present to the user, on the other hand, on hand-written display unit 160, show the candidate item that identification prediction unit 130 is produced, for example the candidate item of ten literal that will import the most approaching arranging according to degree of closeness; Candidate item selected cell 140, the literal that selection will be imported from ten candidate item under user's operation is shown to the user by hand-written display unit 160 then.

Describe the constructive process of the above-mentioned dictionary of mentioning below in detail, just generate process from hand-written literal sample at the template of each literal by machine learning method.

According to the present invention, the handwriting samples of each literal is divided into two classes.One class is called whole word class, and the sample that belongs to such possesses whole strokes, is a complete word.Reasons such as " even pens " can not require that the sample that belongs in such has identical stroke number when considering character calligraph.The another kind of part stroke word class that then is called, the sample stroke disappearance in such is the word of not write.Some literal, especially single character, for example " people ", " soil " etc., stroke number own seldom, generating portion stroke word class no longer just.On the other hand, a predetermined value STH can be set to the stroke of literal, for example the STH value is 4.Stroke number only has whole word class sample smaller or equal to 4 literal, and other literal has two kinds of handwriting samples, promptly whole word class sample and part stroke word class sample.Fig. 2 is the automatic generation of part stroke printed words basis and the synoptic diagram of whole word class and part stroke word class.

As shown in Figure 2, write literal, and write down its person's handwriting by different users.The complete person's handwriting of a literal is called aforesaid ' whole word class sample '.By whole word class sample is begun stroke removal one by one from last stroke, obtain aforesaid ' part stroke word class sample '.Row ' ' shown in the left side of Fig. 2 are exactly the whole word class sample of this literal, and the person's handwriting shown in the right side of Fig. 2 is exactly the part stroke word class sample of this literal, and they are to form by remove the part stroke from the whole word class sample in left side.As mentioned above, the stroke number of part stroke word class sample is more than or equal to a predetermined value, and for example 4.

Fig. 3 is the process flow diagram of describing according to the dictionary method for making of the embodiment of the invention.As shown in Figure 3, at step S110, gather each literal, the person's handwriting that produces when promptly being write by the user is as the sample set that generates dictionary.At step S120, the sample storage of each literal of collection is the whole word class sample of this literal.

Then, at step S130, judge that whether number of strokes at each literal is greater than predetermined value STH.If stroke number, then means this literal greater than STH and can have whole word class sample and part stroke word class sample.Like this, at step S140, from the whole word class sample generating portion stroke word class sample of this literal, and at step S150, the part stroke word class sample of this literal and this literal and whole word class sample are stored explicitly.

If the stroke number of this literal, then means this literal smaller or equal to STH and only has whole word class sample that flow process forwards step S160 to.

At step S160, carry out pre-service at the whole word class sample and/or the part stroke word class sample of each literal, operations such as for example equidistant level and smooth, barycenter normalization and non-linear normalizing are so that make the feature of this sample become regular.

Equidistant smooth operation is that the sampled point to handwriting samples resamples, and makes it at interval evenly, and Fig. 4 is the synoptic diagram that is described in the equidistant re-sampling operations of carrying out in the preprocessing process.

As shown in Figure 4, in order to reduce the operand of smoothing processing, reduce the number of sampled point significantly by equidistant resampling.The original input coordinate sequence of (A) expression of Fig. 4, the result of sampling when it is grade.Therefore, the local sampled point slow in writing speed is very intensive, even has many apart to be zero sampled point, to be called rest point.

Can delete all rest points by equidistant resampling, and reduce the number of sampled point significantly.The equidistant process that resamples is as follows:

Make the starting point P of stroke ₀Be resample points, calculate P ₁And P ₀Distance.If distance is then left out P less than the resampling interval T ₁, continue to calculate P ₂And P ₀Distance.Otherwise, keep P ₁As resample points, continue to calculate P ₂And P ₁Distance, by that analogy.

Because original sample point is very intensive, the resampling interval T is generally much larger than the interval of original sample point.Therefore, needn't consider that the interval of original sample point needs to replenish the situation of sampled point greater than T.(B) of Fig. 4 shows the result of equidistant resampling.As can be seen from the figure, need the data volume of processing to reduce greatly.

Barycenter normalization operation is that sample size is regular to predetermined size, and (W/2 H/2) overlaps at the center of for example W * H, and barycenter and external square frame.The normalized purpose of barycenter is to adjust position, the size of input characters, makes it consistent with mode standard in the recognition dictionary, and the geometric center of the boundary rectangle frame of this literal is overlapped with the barycenter of literal.Fig. 5 (A) shows original handwriting samples, and (B) of Fig. 5 shows the synoptic diagram of the barycenter normalization process of carrying out in preprocessing process.The height of bidding quasi-mode is H, and wide is W.The height of input characters is h, and wide is w.The computing formula of barycenter is:

$\left\{\begin{array}{c}{x}_0=\underset{k=1}{\overset{n-1}{\mathrm{\Σ}}}(\frac{x_k+x_{k+1}}{2}\·\sqrt{{(x_k-x_{k+1})}^2+{(y_k-y_{k+1})}^2})\\ y_0=\underset{k=1}{\overset{n-1}{\mathrm{\Σ}}}(\frac{y_k+y_{k+1}}{2}\·\sqrt{{(x_k-x_{k+1})}^2+{(y_k-y_{k+1})}^2})\end{array}\right....\left(1\right)$

The normalized computing formula of barycenter is:

$x^{\&prime;}=\left\{\begin{array}{cc}x\&CenterDot;\frac{W}{2}/x_0& \mathrm{ifx}<x_0\\ \frac{W}{2}+(x-x_0)\&CenterDot;\frac{W}{2}/(w-x_0)& \mathrm{ifx}>x_0\end{array}\right....\left(2\right)$

$y^{\&prime;}=\left\{\begin{array}{cc}y\&CenterDot;\frac{H}{2}/y_0& \mathrm{ify}<y_0\\ \frac{H}{2}+(y-y_0)\&CenterDot;\frac{H}{2}/(h-y_0)& \mathrm{ify}\&GreaterEqual;y_0\end{array}\right....\left(3\right)$

But barycenter normalization operation can only be corrected centroid motion, and it is powerless to the problem out of proportion of the partial structurtes of input characters.

The non-linear normalizing operation can make the interval of sample stroke even.(C) of Fig. 5 shows the synoptic diagram of non-linear normalizing operation.Dynamic scaling is then adopted in non-linear normalization in a literal, according to the different scaling of factors such as stroke distribution density degree decision of the different parts of input characters.The non-linear normalizing method that is based on dot density that present embodiment adopts is divided cloth density according to the calculating pen reciprocal of the projection of stroke pixel, dynamically adjusts scaling with this.The local suitably amplification that the stroke distribution density is high, low density place suitably dwindles.The stroke of adjusted like this literal distributes will be tending towards even.Non-linear normalization can more effectively reduce the literal deformation extent, and the font difference that weakens different writing styles generations can improve the discrimination of hand-written regular literal effectively with discrete.

Make f (i, j) expression one hand-written literal bianry image, i=1,2 ..., I, j=1,2 ..., J.

H (i) and V (j) are respectively f (i, j) projection function in the horizontal and vertical directions.

$H\left(i\right)=\underset{j=1}{\overset{J}{\mathrm{\&Sigma;}}}f(i,j)$

$V\left(j\right)=\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}f(i,j)$

Make g (s, t) be f (i, the j) result behind the non-linear normalizing, s=1,2 ..., S, t=1,2 ..., T, its computing formula is as follows:

$s=\underset{k=1}{\overset{i}{\mathrm{\&Sigma;}}}H\left(k\right)\&times;\frac{S}{\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}H\left(k\right)}...\left(4\right)$

$t=\underset{l=1}{\overset{j}{\mathrm{\&Sigma;}}}V\left(l\right)\&times;\frac{T}{\underset{l=1}{\overset{J}{\mathrm{\&Sigma;}}}V\left(l\right)}...\left(5\right)$

After pre-treatment step,, from sample, extract the M dimensional feature: can comprise stroke direction distribution characteristics M at step S170 ₁Dimension, grid stroke feature M ₂Dimension, peripheral direction feature M ₃The dimension stroke direction, and with other irrelevant features of stroke number, the order of strokes observed in calligraphy.

Fig. 6 is a synoptic diagram of describing the process of extracting the stroke direction distribution characteristics.Shown in Fig. 6 (A), the grid that earlier literal is divided into n * n, calculate each section in each grid to the projection of 8 directions, the projected length of each grid on 8 directions shown in Fig. 6 (B) promptly constituted the stroke direction distribution characteristics of this grid.

Fig. 7 is a synoptic diagram of describing the process of extracting the grid stroke feature.The grid that earlier literal is divided into n * n calculates the shared area of pen section in each grid, promptly stroke point count with, just be the grid stroke feature of this grid.

Make bianry image

Then

The grid stroke feature of representing this grid.

Fig. 8 is a synoptic diagram of describing the process of extracting the peripheral direction feature.Earlier literal is vertically carried out the n five equilibrium.The length that from left to right searches along horizontal scanning line till half of first stain or literal width is called the left side search length, the average left side search length of the horizontal scanning line of each piecemeal such as grade is the horizontal contour feature in left side of these piecemeals divided by half of literal width, shown in Fig. 8 (A).Similarly, average right side search length can be calculated the horizontal contour feature in right side divided by the literal width.Equally, shown in Fig. 8 (B), literal along horizontal direction n five equilibrium, is calculated upper vertical contour feature and downside vertically profiling feature.

Three kinds of features are connected into following M dimension associating proper vector successively

X＝[x ₁，…x _M1，x _M1+1，…，x _M1+M2，x _M1+M2+1，…，x _M1+M2+M3]

M=M wherein ₁+ M ₂+ M ₃

With Fig. 2 is example, can obtain the complete word characteristic set { X of " " word _{Whole 1}..., X _{Whole 5}And the partial words characteristic set { X of " " word _{Portion 1}..., X _{Portion 20}}

At step S180, dimensionality reduction and quantification treatment: adopt the KL conversion, the dimension of proper vector is reduced to the N dimension from the M dimension, again the proper vector behind KL transformation matrix and the dimensionality reduction is quantized, represent its element with WORD (16bits) type and BYTE (8bits) type variable respectively.

The KL conversion is chosen N feature in can selecting from the M dimensional vector, former vectorial dimensionality reduction is become a N dimensional vector (N＜M), and keeping as much as possible in the vector originally to the classification Useful Information.

At first, supposing needs K character of identification now, and the capacity of promptly supposing this recognition dictionary is a K character, owing to produced part stroke word class, (K≤W≤2K), the prior probability of all kinds of appearance is P (ω to total W classification _i), i=1,2 ..., W.With X _iExpression is from the M dimensional vector of i class, then the autocorrelation matrix R of i class cluster _iFor

$R_i=E\left\{X_i{X}_i^T\right\}...\left(6\right)$

The autocorrelation matrix R of mixed distribution is

$R=E\left\{X{X}^T\right\}=\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)R_i=\underset{i=1}{\overset{W}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)E\left\{X_i{X}_i^T\right\}...\left(7\right)$

Be that R is the statistical average of all kinds of autocorrelation matrixes.

Secondly, obtain eigenvectors matrix Φ and the eigenvalue matrix Λ of R,

$\mathrm{\&Lambda;}=\left[\begin{array}{cccccc}{\mathrm{\&lambda;}}_1& & & & & 0\\ & {\mathrm{\&lambda;}}_2& & & & \\ & & .& & & \\ & & & .& & \\ & & & & .& \\ 0& & & & & {\mathrm{\&lambda;}}_M\end{array}\right]...\left(8\right)$

Φ＝[Φ ₁Φ ₂…Φ _M]…(9)

And require eigenwert to arrange descendingly:

λ ₁≥λ ₂≥…≥λ _M…(10)

Difference character pair vector Φ ₁, Φ ₂..., Φ _M

The 3rd, get top n proper vector Φ _i(i=1,2 ..., N), constitute transformation matrix A

$A={\left[\begin{array}{c}{\mathrm{\&Phi;}}_1^T\\ {\mathrm{\&Phi;}}_2^T\\ .\\ .\\ .\\ {\mathrm{\&Phi;}}_N^T\end{array}\right]}_{N\&times;M}...\left(11\right)$

Get conversion Y=AX again, Y is the N dimensional vector.

With Fig. 2 is the characteristic set of example explanation through obtaining behind the dimensionality reduction, and the complete word characteristic set of " " word is { Y _{Whole 1}..., Y _{Whole 5}, each Y wherein _i=[y ₁, y ₂Y _N], the partial words characteristic set for " " word is { Y equally _{Portion 1}..., Y _{Portion 20}, each Y wherein _i=[y ₁, y ₂Y _N].

Then, matrix A and vectorial Y are quantized:

The quantification of transformation matrix A

Transformation matrix A is a floating-point matrix,

$A=\left[\begin{array}{cccccc}{a}_{11}& a_{12}& .& .& .& a_{1M}\\ a_{21}& a_{22}& .& .& .& a_{2M}\\ .& .& & & & .\\ .& .& & & & .\\ .& .& & & & .\\ a_{N1}& a_{N2}& .& .& .& a_{\mathrm{NM}}\end{array}\right]...\left(12\right)$

Make Q=max (| a _Ij|), K=32767, then quantitative formula is as follows:

a′ _ij＝round(a _ij×K/Q)…(13)

Wherein, round () is the round function.Like this, a ' _IjCan represent with a WORD (16bits) type variable.

The quantification of new feature vector Y

New feature behind KL conversion dimensionality reduction vector Y, its quantification manner and front similar, different is Y this as fixed point vector, quantize afterwards each element of Y and represent with BYTE (8bits) type variable:

$Y_{N\&times;1}=A_{\mathrm{NxM}}{X}_{M\&times;1}=\left[\begin{array}{c}{y}_1\\ y_2\\ .\\ .\\ .\\ y_N\end{array}\right]...\left(14\right)$

Template in the recognition dictionary can produce by the whole word class sample characteristics vector sum part stroke class sample characteristics vector after dimensionality reduction and the quantification is carried out the training of GLVQ algorithm respectively, is the proper vector T that whole word class sample and part stroke printed words originally generate a N dimension respectively _iAs template, T _i=[t ₁, t ₂T _N].With Fig. 2 is example, and 5 complete word characteristic sets generate a complete character matrix plate through the training back, and 20 part stroke word characteristic sets generate a partial words class template through the training back.

At step S190, GLVQ (the Generalized Learning VectorQuantization) algorithm after utilization improves is learnt the handwriting samples collection after dimensionality reduction and quantification treatment and is trained, and finally is that each classification generates a template.

If X is a handwritten word sample, P _iAnd P _jBe two templates in the recognition dictionary, and X and P _iBelong to a class, X and P _jDo not belong to a class, P ' _iAnd P ' _jBe the new template after upgrading, its training formula is as follows:

$P_i^{\&prime;}=P_i+\mathrm{\&alpha;}{l}_k(1-l_k)\frac{D_j}{{(D_i+D_j)}^2}(X-P_i)...\left(15\right)$

$P_j^{\&prime;}=P_j-\mathrm{\&alpha;}{l}_k(1-l_k)\frac{D_i}{{(D_i+D_j)}^2}(X-P_j)...\left(16\right)$

Wherein, α is a learning rate,

D _i＝||P _l-P _i||…(17)

D _j＝||P _l-P _j||…(18)

${\mathrm{\&mu;}}_k=\frac{D_i-D_j}{D_i+D_j}...\left(19\right)$

$l_k=l_k\left({\mathrm{\&mu;}}_k\right)=\frac{1}{1+e^{-{\mathrm{\&mu;}}_k}}...\left(20\right)$

D _iAnd D _jBe the matching distance of two patterns, can adopt Euclidean distance, mahalanobis distance and other various distance measures.P _iAnd P _jInitial value can be made as its geometric center of class separately.Take turns circulation, P through this _iClose with X, and P _jBe pushed far.All classifications are repeated this process, reach certain level until the discrimination of training set sample.

Fig. 9 describes the synoptic diagram utilize the process that GLVQ learns.In Fig. 9, comprised the whole word class and the part stroke word class of " woods " word, and the whole word class of " wood " word.Because the sample of whole word class is compared with part stroke word class sample, the differences between samples that belongs to such is less relatively, therefore the feature of the whole word class sample write of different people is comparatively approaching, distribution in the space is comparatively concentrated, shown in the broken circle of Fig. 9, and the differences between samples in the part stroke word class sample of ' woods ' word is bigger, so the distribution of part stroke word class sample in the space that different people is write comparatively disperse, shown in the dotted ellipse among Fig. 9.As can be seen from Figure 9, when ' woods ' word had just begun to write, when being write as ' wood ', ' wood ' of its feature and complete word class was approaching; And work as ' woods ' word when having write soon, its feature is with approaching as ' woods ' word of complete word class.Therefore, the sample characteristics of part stroke word class distributes and comparatively disperses, and the sample characteristics of whole word class distributes then comparatively concentrated.In view of the above, we have done improvement to the GLVQ of standard.If X is for belonging to whole word class, learning rate is α ₁If X is for belonging to part stroke word class, learning rate then adopts α ₂It satisfies α ₁＞α ₂, in actual applications, get α ₁=4 α ₂Like this can accelerating convergence.

Among Fig. 9, black circle is represented all kinds of cluster centres, and it is produced by the study of GLVQ algorithm.In when identification, calculate the matching distance of input sample and each cluster centre, apart from reckling as recognition result.

Like this,, represent complete word class and part stroke word class respectively, finally finish the making of recognition dictionary at step S200 for each literal generates one or two template.

Though above with the GLVQ arthmetic statement automatic learning process, this only is for purposes of illustration, but not wants to limit the present invention.The present invention also can adopt other machine learning algorithm, SOM (Self-Organizing Maps) network for example, LVQ (Learning VectorQuantization) algorithm, and the modified LVQ1 of LVQ algorithm, LVQ2 and LVQ3 etc.

For the recognition dictionary that can discern the band forecast function of K character, it comprises W template, and it satisfies K≤W≤2W.Making the proper vector dimension behind the dimensionality reduction is N, and then recognition dictionary is that W N dimensional feature vector is connected in series the one-dimension array that forms from beginning to end:

Dictionary＝[T ₁…T _W]′＝[t ₁₁…t _1N…t _W1…t _WN]′…(21)

Characteristics of the present invention are miniaturizations, and for the single mode plate dictionary of the no forecast function that can discern K character, its template number is K, and the proper vector dimension behind the dimensionality reduction is N, and then the size of recognition dictionary is KN (bytes).Make K=7000, N=64, then dictionary size is 448000bytes.

And for the dictionary of creating according to the method for present embodiment that can discern K character equally, then its template number is W (K≤W≤2K).This moment, the size of recognition dictionary was WN (bytes).Therefore, under the worst case, the data volume of the recognition dictionary that the method for present embodiment is created can increase to original 2 times.But because the data volume of original dictionary is less, so the data volume of new dictionary is also little.Still with K=7000, N=64 is an example, and the data volume maximum of new dictionary is no more than 896000bytes.

Below in conjunction with Figure 10 process according to the hand-written inputting method of the embodiment of the invention is described.Figure 10 is the detail flowchart according to the identifying in the hand-written inputting method of the embodiment of the invention.

As shown in figure 10, at step S310, the user is by the stroke of handwriting input unit 110 certain literal of input.Then, at step S320, store by handwriting storage unit 120, and be presented on the hand-written display unit 160 by indicative control unit 150.

Be stored in the operation of the stroke execution in step S330～S360 in identification prediction unit 130 in the handwriting storage unit 120.

In view of step S330 basic identical to operation and the step S160 in the above-mentioned dictionary creating method of S350 to the operation of S180, so no longer these steps are elaborated here.

At step S360, carry out the feature of input stroke and the quick coupling of the template in the dictionary.Figure 11 is a synoptic diagram of describing quick matching process.Be located at step S350 and adopt the N dimensional feature vector that generates after the KL conversion, each is tieed up contained quantity of information and successively decreases successively.Therefore, the N dimensional feature vector can be divided into the d section, every section comprises N respectively ₁, N ₂..., N _dIndividual element, and satisfy following formula

N＝N ₁+N ₂+…+N _d…(22)

In practice, the segmentation majority meets following condition

N ₁≤N ₂≤…≤N _d…(23)

First round screening: choose first section N ₁Individual element participates in coupling, calculates the characteristic distance between each template in sample to be identified and the recognition dictionary, and threshold value TH1 is set, and the template that keeps greater than threshold value continues to participate in the next round screening.In practice, threshold value is made as the intermediate value of all distances.Adopt distance measure identical when training herein with GLVQ.

Second takes turns screening: choose second section N ₂Individual element also participates in coupling, i.e. (N ₁+ N ₂) individual element, calculate the characteristic distance between the remaining template of sample to be identified and last round of screening, threshold value TH2 is set, the template that keeps greater than threshold value continues to participate in the next round screening.

……

Last takes turns screening: choose last remaining N _dIndividual element participates in coupling, and promptly whole N elements calculate the characteristic distance between the remaining template of sample to be identified and last round of screening, and minimum 10 are final TOP10 recognition result.

At step S370, upgrade original recognition result, above-mentioned TOP10 recognition result is presented on the hand-written display unit 160, select for the user.

At step S380,, then finished the input of this literal if the user has selected certain candidate item among the TOP10 by candidate item selected cell 140.Otherwise flow process forwards step S310 to, repeats above-mentioned operating process.

Figure 12 is the synoptic diagram of ten candidate tabulation in conjunction with first-selected word background prompting.As shown in figure 12, inputting interface roughly is divided into three zones, the zone in left side is used for showing the TOP10 candidate item, select for the user, middle zone is to write the district, shown user's stroke of input in real time on it, and on the background of the stroke of importing, shown current recognition result, as the prompting in the input process.If the user certain word has been write when a part of and just have been identified this word, then this word need not have been write and just finish input operation.On the right side of inputting interface is the instruction area, wherein is provided with a plurality of function keys, and the confession user carries out other operations such as editor in input process.

In addition, in identifying, can set user's ' lift pen ' and finish after the input, identification prediction unit 130 is after waiting for the appropriate time, automatically first candidate item is sent to indicative control unit 150, be presented in the candidate item viewing area of hand-written display unit 160.Whether identification prediction unit 130 is that the single character factor is adjusted this wait input time according to the literal of user's writing style and input.

In addition, in the superincumbent description, only generated a template at the part stroke word class sample of certain literal, as shown in Figure 2.But also can generate two or more part stroke word class templates, further improve predictive ability the part stroke word class sample evidence stroke number of this literal with different priorities.

So far invention has been described in conjunction with the preferred embodiments.Should be appreciated that those skilled in the art can carry out various other change, replacement and interpolations under the situation that does not break away from the spirit and scope of the present invention.Therefore, scope of the present invention is not limited to above-mentioned specific embodiment, and should be limited by claims.

Claims (36)

1. method of making dictionary comprises step:

The whole word feature of the whole printed words basis of extraction literal and stroke number are greater than the part stroke feature of the part stroke sample of the literal of predetermined value; And

By described whole word feature and described part stroke feature being learnt to generate the whole character matrix plate and/or the part stroke template of literal, as the project in the dictionary with machine learning algorithm.

2. the method for claim 1, wherein said whole word feature and part stroke feature all are the M dimensions, and wherein M is a natural number, and described extraction step comprises:

The whole word feature of described M dimension is dropped to the N dimension, and wherein N is that natural number and M are greater than N.

3. the method for claim 1, wherein said whole word feature and described part stroke feature are and stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy.

4. method according to claim 1 also comprises step:

Before extraction step, described whole printed words this and described part stroke sample are equidistantly resampled, make that the distance between the sampled point of stroke of sample is equal substantially.

5. method according to claim 1 also comprises step:

Before extraction step, the size of adjusted size for being scheduled to described whole printed words basis and described part stroke sample makes the barycenter of sample overlap with the center of the rectangle of preliminary dimension.

6. method according to claim 1 also comprises step:

Before extraction step, divide cloth density according to the calculating pen reciprocal of the projection of stroke pixel, dynamically adjust the scaling of sample.

7. method according to claim 1, wherein said part stroke sample generates from described whole word stroke sample.

8. method according to claim 3, wherein said and stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy comprise following one of at least: stroke direction distribution characteristics, grid stroke feature and peripheral direction feature.

9. method according to claim 2 also comprises step:

Each amount of element of N dimensional feature is turned to integer.

10. method according to claim 2 wherein drops to the N dimension by the KL conversion with described M dimensional feature.

11. method according to claim 1, wherein said machine learning algorithm comprise GLVQ algorithm, SOM network, LVQ algorithm, LVQ1 algorithm, LVQ2 algorithm and LVQ3 algorithm one of at least.

12. method according to claim 11, wherein said machine learning algorithm at whole this learning rate that is adopted of printed words greater than the learning rate that is adopted at described part stroke sample.

13. method according to claim 12, wherein said machine learning algorithm are four times of the learning rate that adopted at described part stroke sample at whole this learning rate that is adopted of printed words.

14. an equipment of making dictionary comprises:

The whole word feature of the whole printed words basis of extraction literal and stroke number are greater than the device of the part stroke feature of the part stroke sample of the literal of predetermined value; And

By described whole word feature and described part stroke feature being learnt to generate the whole character matrix plate and/or the part stroke template of literal, as the device of the project in the dictionary with machine learning algorithm.

15. equipment as claimed in claim 14, wherein said whole word feature and part stroke feature all are the M dimensions, wherein M is a natural number, and described extraction element comprises:

The whole word feature of described M dimension is dropped to the device of N dimension, and wherein N is that natural number and M are greater than N.

16. equipment as claimed in claim 14, wherein said whole word feature and described part stroke feature are with stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy.

17. a hand-written inputting method comprises step:

Extract the feature to the small part handwriting of literal; And

Distance between the template in the dictionary that calculates described feature and create according to the method for claim 1; And

Will be apart from the literal of the template representative of less at least one as recognition result.

18. hand-written inputting method according to claim 17, wherein said feature are M dimensions, M is a natural number, and described hand-written inputting method also comprises: described M dimensional feature is dropped to the N dimensional feature, and wherein N is that natural number and M are greater than N; And the N dimensional feature vector that will represent described N dimensional feature is divided into a plurality of sections, wherein carries out described calculation procedure piecemeal.

19. hand-written inputting method as claimed in claim 17, wherein said feature are with stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy.

20. hand-written inputting method according to claim 17 also comprises step:

Before extraction step, equidistantly resample to the small part handwriting to described, make to the distance between the sampled point of the stroke of small part handwriting to equate substantially.

21. hand-written inputting method according to claim 17 also comprises step:

Before extraction step,, make described barycenter overlap with the center of the rectangle of preliminary dimension to the small part handwriting with the described size of adjusted size for being scheduled to the small part handwriting.

22. hand-written inputting method according to claim 17 also comprises step:

Before extraction step, divide cloth density according to the calculating pen reciprocal of the projection of stroke pixel, dynamically adjust described scaling to the small part handwriting.

23. hand-written inputting method according to claim 19, wherein said and stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy comprise following one of at least: stroke direction distribution characteristics, grid stroke feature and peripheral direction feature.

24. hand-written inputting method according to claim 18 also comprises step:

Each amount of element of N dimensional feature is turned to integer.

25. hand-written inputting method according to claim 18 wherein drops to the N dimension by the KL conversion with described M dimensional feature.

26. hand-written inputting method according to claim 17 also comprises step:

With described to the small part handwriting recognition result of range of a signal minimum synchronously.

27. a handwriting input device comprises:

Extract the device to small part handwriting feature of literal, and

The device of the distance between the template in the dictionary that calculates described feature and create according to the method for claim 1; And

Will be apart from the literal of the template representative of less at least one device as recognition result.

28. handwriting input device as claimed in claim 27, wherein said feature are M dimensions, M is a natural number, and described handwriting input device also comprises: described M dimensional feature is dropped to the device of N dimensional feature, and wherein N is that natural number and M are greater than N; And will represent that the N dimensional feature vector of described N dimensional feature is divided into a plurality of sections device, wherein carry out described calculating piecemeal.

29. handwriting input device as claimed in claim 27, wherein said feature are with stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy.

30. handwriting input device according to claim 27 also comprises:

Before extracting feature, equidistantly resample to the small part handwriting to described, make the device that distance between the sampled point of described stroke to the small part handwriting equates substantially.

31. handwriting input device according to claim 27 also comprises:

Before extracting feature, described adjusted size to the small part handwriting is predetermined size, make the described device that overlaps with the center of the rectangle of preliminary dimension to the barycenter of small part handwriting.

32. handwriting input device according to claim 27 also comprises:

Before extracting feature, divide cloth density according to the calculating pen reciprocal of the projection of stroke pixel, dynamically adjust the device of described scaling to the small part handwriting samples.

33. handwriting input device according to claim 29, wherein said and stroke number, connect pen and the irrelevant feature of the order of strokes observed in calligraphy comprise following one of at least: stroke direction distribution characteristics, grid stroke feature and peripheral direction feature.

34. handwriting input device according to claim 28 also comprises:

Each amount of element of N dimensional feature is turned to the device of integer.

35. handwriting input device according to claim 28 wherein drops to the N dimension by the KL conversion with described M dimensional feature.

36. handwriting input device according to claim 27 also comprises:

With described to the small part handwriting device of the recognition result of range of a signal minimum synchronously.

Images