CN106570458A - Recognition method for recognizing on-line handwritten Chinese and Japanese - Google Patents

Abstract

The invention provides a recognition method for recognizing the on-line handwritten Chinese and Japanese. According to the method, the structured dictionary representation and the vector quantization (VQ) are adopted to establish a compact MRF-based online character recognition method, and the method can be used for recognizing large-scale Chinese and Japanese character sets. Characters are decomposed into the free radicals of basic elements for constituting characters. By means of an MRF-based model, identical free radicals for forming different characters are shared, so that the storage space of a dictionary model is saved. In addition, the method simultaneously applies the VQ technique to compress a character recognizer. Therefore, the storage space of the dictionary model is further greatly compressed on the premise that the recognition rate is not lost at all.

Description

A kind of recognition methodss of online handwriting Chinese and japanese

Technical field

The present invention relates to character recognition technologies field, more particularly to a kind of online handwriting Chinese and japanese recognition methodss.

Background technology

In online handwriting character recognition process, when user using writing pencil on a special notebook or panel computer When writing, there are the movement of sensor identification nib and the conversion of lift pen/start to write.This data is referred to as " digital ink ", can To be considered the dynamic performance of a hand-written process.Online handwriting character recognition is that digital ink is automatically converted to calculating by one kind The technology of the code of machine editable and retrieval.Due to panel computer, electronic whiteboard, smart mobile phone and iPad etc., these possess touch-control The appearance of the electronic equipment on interactive interface and surface so that online handwriting character recognition is more widely applied.Therefore, Ren Menxu Ask hand-written character string pattern identification (having less restriction during character writing) technology, rather than independent character recognition (each word During symbol requires the grid that write is specified).Continuous hand-written character string pattern is a kind of rapid style of writing style, be result in character recognition process The difficulty of middle separating character.Just because of this reason, needs a high property that can provide natural convenient user input Can hand-written character string recognition methodss.

As Chinese, Japanese and Korean have thousands of different classes of and huge character set, existing handwriting recognition side Method is problematic in that in terms of discrimination, recognition speed and memory consumption.Currently used for identification 7097 Japanese characters based on The recognition methodss of markov random file (MRF) need 20 Mbytes of memory consumption, if MRF recognition methodss combine non-structural Improvement quadratic function (MQDF) and linguistic context process recognition methodss, due to MQDF dictionaries and the large volume of linguistic context dictionary, its need Want higher volume of internal memory (about 65 Mbytes).So, it is badly in need of the modularity handwriting recognition side of high speed low memory consumption on market Method.Within 0.5 second, memory consumption possesses actual application value in 10 Mbytes of hand-written recognition method to every page of recognition speed.

The content of the invention

It is an object of the invention to provide a kind of high performance online handwriting Chinese and japanese evaluator method, methods described has Higher discrimination and recognition speed, while having less memory consumption.

For solve above-mentioned technical problem, the invention discloses a kind of compact for being applied to Chinese and japanese large character set based on The online character identifying methods of MRF, wherein apply structuring dictionary representing and vector quantization (VQ) technology.

Handwritten text based on MRF recognizes the structure of ONLINE RECOGNITION device.Online list word is in identification process with extraction Characteristic point sequence substitute the nib landing point coordinates sequence that gathers when word is input into, can not only more efficiently express this hand-written Body word, and can reduce during Text region because of the excessive and caused computing resource consumption of characteristic point. first input Handwritten text carries out linear normalization process in the case of the horizontal and vertical constant rate for preserving former input word, changes into The word of normal size, then extraction feature point again.For every a line of the single word of input, first by the starting point on the side Characteristic point is elected as with terminating point；Secondly, if the certain point on the side in addition to having elected characteristic point as is to adjacent feature point Distance is more than a certain threshold value, then the point is also chosen as characteristic point. characteristic point is so selected always until no another characteristic point can Only elect as.

Using MRF models as ONLINE RECOGNITION device recognition function.The recognition function is input into the feature of word by comparing Point and the state of all kinds of prototypes in prototype dictionary, obtain being input into the similarity of word and all kinds of prototypes, with maximum similarity Recognition result of the word class as input handwritten text.

Application structure dictionary presentation technology reduces memory consumption.Chinese and japanese ideogram, is referred to as Kanji, tool in Japan There are system and hierarchical structure.Free radical is conventional basic semantic or phonetic unit and then is built into ideogram.Each is Sino-Japan The ideogram of text contains one or several free radicals.By a set of rational free radical, all of Chinese and japanese table can be represented Meaning word.Overall verbal model is replaced largely to reduce the quantity of model using the model of free radical.

Using VQ technique compresses Character recognizers, in MRF evaluators of the present invention, due to the one of character and free radical There are many similar orders in unit with dual nature.Therefore, it can for same method to be applied to the parameter of each average vector, every The covariance matrix of individual unitary characteristic and binary feature and the transition probability of each state, cluster argument sequence are divided into group, respectively A common argument sequence is shared by individual group.It can make compression more efficient.

The invention discloses an online Japanese OCR method, application structure dictionary presentation technology and VQ technologies, are protecting Memory headroom is have compressed while holding accuracy of identification further.Increased extensive experimental result, the number of character type is from 2965 It is individual to increase to 4218 (the conventional character set of daily life).Propose the multiple MRF models classes of construction to improve recognition performance Method.For different actual demands in the present invention, multiple MRF models are constructed.Structuring MRFs is combined with VQ technologies, is being ensured Can reduce by 16% internal memory while discrimination, but process time increases by 2 times.The combination of original MRF models and VQ2 technologies Then can be with time-consuming.Consider, structurized MRFs can more meet actual demand with reference to VQ technologies.

Description of the drawings

For the technical scheme being illustrated more clearly that in the embodiment of the present invention, below will be to making needed for embodiment description Accompanying drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the present invention, for For those of ordinary skill in the art, on the premise of not paying creative work, can be obtaining other according to these accompanying drawings Accompanying drawing.

Fig. 1 is the flow chart of online handwriting Chinese and japanese evaluator method in the present invention

Fig. 2 is characterized an extraction process schematic diagram.

Fig. 3 is MRF labelling schematic diagrams.

Fig. 4 is linear chain MRF schematic diagrams.

Fig. 5 is to create an interface to build the schematic diagram of structurized MRF.

Fig. 6 is free radical pattern normalization schematic diagram.

Fig. 7 is by VQ focuses parameters schematic diagrams.

Statistics of database of the Fig. 8 for character pattern.

Fig. 9 is the comparison of discrimination, memory consumption and the recognition speed of three kinds of models.

Figure 10 is the comparison of discrimination, memory consumption and the recognition speed of four kinds of models.

Specific embodiment

With reference to specific embodiment, the present invention is expanded on further.It should be understood that these embodiments are merely to illustrate the present invention Rather than limit the scope of the present invention.

The present invention provides a kind of high performance online handwriting Chinese and japanese evaluator method, as shown in figure 1, in accordance with the following steps Realize：

Step S1：Character recognition, is preserving the former horizontal and vertical ratio for being input into word the handwritten text of input first Linear normalization process is carried out in the case of rate is constant, the word of normal size is changed into, then extraction feature point again.For defeated Enter every a line of word, elect the starting point and ending point on the side as characteristic point first；Secondly, if except on the side The certain point elected as outside characteristic point is more than a certain threshold value to the distance of adjacent feature point, then the point is also chosen as characteristic point. so Selection characteristic point is till no another characteristic point is optional always；

Step S2：Using MRF models as ONLINE RECOGNITION device recognition function.By compare input word characteristic point with The state of all kinds of prototypes in prototype dictionary, obtains being input into the similarity of word and all kinds of prototypes, the word with maximum similarity Recognition result of the class as input handwritten text, obtains the other MRF evaluators of character level；

Step S3：Structuring MRF evaluators are built, further, is created an interface to build structurized MRF, is obtained To unitary characteristic mean vector：The average coordinates of the characteristic point of training pattern.The all strokes of each character are connected to a stroke To obtain stroke number independence.In characteristic point, a mouse click right button cuts the character pattern of characteristic point.One left mouse button point Hit, two free radicals will be connected in characteristic point.Character model can be divided into by several parts by the method.By decomposing at normalization Reason is registered in each section in free radical dictionary, while identical free radical can be classified as same class.

Step S4：All character MRF model decompositions are registered, makes training pattern and character MRF synchronously decompose free radical Level.Average bounding box is obtained from the bounding box of free radical training pattern.The free radical for decomposing out in kinds of characters may There are different size and locations, by these free radical normalization, then with normalized free radical model training free radical MRF moulds Type.By the average bounding box of normalization average value vector sum covariance matrix, we arrange free radical MRF models for character mould Type.

Step S5：Cluster different size and each free radical of position.Each free radical MRF is from kinds of characters class In free radical pattern, thus size position is different.Different sizes brings different variances, if by difference big free radical MRF merges and may cause relatively low recognition performance.Various sizes of each free radical of aggregation is different groups, separates free radical and is Multiple classifications.Optimal stage size quantity.

Step S6：Dictionary is compressed by VQ technologies, in MRF evaluators, there is numerous identical unitary and binary Feature.For unitary and binary feature, each mean vector has two param elements, and each covariance matrix has four ginsengs Number element each state has three transition probabilities to have three param elements.By the mean vector of unitary and binary feature and association side Difference matrix parameter element is combined as one.Three transition probabilities of each state have three param elements to be set to another group Close.It is different groups by parameter combination cluster, the group of shared identical parameters collection is set to into a group center.By storing group index and center Parameter, can effectively be compressed to evaluator.

Specific embodiment is：

1) set up a compact online MRF evaluator

First the handwritten text of input is carried out in the case of the horizontal and vertical constant rate for preserving former input word Linear normalization process, changes into the word of normal size, then extraction feature point again.For each of the single word of input Side, elects the starting point and ending point on the side as characteristic point first；Secondly, if on the side in addition to having elected characteristic point as Certain point is more than a certain threshold value to the distance of adjacent feature point, then the point is also chosen as characteristic point. characteristic point is so selected always Till no another characteristic point is optional.As shown in Figure 2, as the round dot in a, b is characterized a little.Using MRF model conducts The recognition function of ONLINE RECOGNITION device.Characteristic point is set from an input mould set S={ s₁,s₂,s₃,…,s_I}.Characteristic point is from one Individual input pattern set S={ s₁,s₂,s₃,…,s_I}.The recognition function is by comparing the characteristic point for being input into word and prototype dictionary In all kinds of prototypes state, obtain being input into the similarity of word and all kinds of prototypes, the word class with maximum similarity is used as defeated Enter the recognition result of handwritten text.As shown in Figure 3.

Neighborhood system is to be chosen continuously to close on characteristic point according to sequential write, is arranged

Odd number set:C1={ s₁,s₂,s₃,…,s₁₀,…}

Pairing set:C2={ { s₁,s₂},{s₂,s₃},{s₃,s₄},{s₄,s₅},……,{s₉,s₁₀},…}

A linear chain MRF is defined to each character type, as shown in figure 4, each label has a state, each state There are three conversion.

In MAP frameworks, maximum posterior probability is equal to the energy function of minimum to condition conversion probability.

Energy function

Wherein:One class distributes label s_i

:Unitary feature, including X and Y coordinates point

:Binary features (dx with dual element:X coordinate of s_i-X coordinate of s_i-1,dy: Y coordinate of s_i-Y coordinate of s_i-1)

Gaussian function is used for evaluation condition probability

State transition probability：

Each character type has a straight chain MRF, and we are using the input pattern of Viterbi search's matching characteristic point with MRF's The state of each character type, is path and least energy that each character type finds matching.

2) structuring MRF evaluators are built

The other MRF evaluators of character level are created first.Further, create an interface structurized to build MRF, obtains unitary characteristic mean vector：The average coordinates of the characteristic point of training pattern.Each little red squares：Feature Point coordinates.The HCCI combustion of character (baby) as shown in Figure 5.Each little red rectangle shows a characteristic point.

The all strokes of each character are connected to a stroke to obtain stroke number independence.In characteristic point, a mouse click is right Key cuts the character pattern of characteristic point.One left mouse button is clicked on, and will connect two free radicals in characteristic point.Can by the method Character model is divided into into several parts.Each section is registered in free radical dictionary by decomposing normalized, while can be with Identical free radical is classified as into same class.

All character MRF model decompositions are registered, training pattern and character MRF are synchronously decomposed free radical water by us It is flat.Average bounding box is obtained from the bounding box of free radical training pattern.The free radical for decomposing out in kinds of characters may have Different size and location, by these free radical normalization.By average bounding box by the freedom of the appearance of identical characters apoplexy due to endogenous wind Base normalization.As shown in fig. 6, orange represent input free radical pattern.After free radical pattern is cut out, returned by average bounding box One changes, then with normalized free radical model training free radical MRF models.By normalization average value vector sum covariance matrix Average bounding box, we arrange free radical MRF models for character model.

3) cluster each free radical of different size and position.

Free radical patterns of each free radical () MRF from kinds of characters apoplexy due to endogenous wind, thus size position is different.Different Size brings different variances, if difference big free radical MRF is merged may cause relatively low recognition performance.Thus application Assembling various sizes of each free radical, separations free radical is multiple classifications to cluster analysis method, optimal stage size quantity.

4) by VQ technique compresses

As shown in fig. 7, in MRF evaluators, there is numerous identical unitary and binary feature.It is special for unitary and binary Levy, each mean vector there are two param elements, each covariance matrix there are four param elements each states there are three Transition probability has three param elements.Using the mean vector and covariance matrix param elements of unitary and binary feature as one Individual combination.Represent that three param elements of three transition probabilities of each state are set to another combination.Parameter combination is clustered For different groups, the group of shared identical parameters collection is set to into a group center.Store group index and Center Parameter.Evaluator is pressed Contracting.Combinations thereof mode is set to one group compared to by mean vector, covariance matrix and state transition probability, to evaluator Compression effectiveness it is more preferable.

5) training test

Recognize to evaluate compact evaluator, character type identification by 2965 chinese characters (1 grade of Japanese Industrial Standards) Device is trained by online Japanese hand-written data storehouse Nakayosi.Performance test uses online Japanese hand-written data storehouse Kuchibue. The statistics of database data of character model are shown in Fig. 8.

Compare the performance of three kinds of models, be respectively original MRF models, structuring MRF models (Str-MRF1) and this Bright model Str-MRF2,.Running environment is in the CPU w5590 double-core 12GB of Intel's Xeon (R)=3.36GHz Deposit.Comparing result is as shown in figure 9, wherein speed represents the average time of one character of identification.Although Str-MRF1 compression storages Space is very big, but largely reduces the accuracy rate of character recognition.Str-MRF2 possesses preferably identification than Str-MRF1 Precision, also have compressed memory space while recognition speed is ensured.

6) compared for the performance of evaluator after VQ technique compresses.

Original MRF：Using the state transition probability and covariance matrix of the data storage of 4 bytes each parameter, And using 2 bytes data storage average vector each parameter.

VQ1：Parameter set is divided into into 255 groups, and each group index is saved as the data of 1 byte.

VQ2：Parameter set is divided into into some groups, and each group index is saved as the data of 2 bytes.

VQ2+Str-MRF2：Dictionary dictionary is further compressed to structure using vector quantization method.

Comparing result as shown in Figure 10, although VQ2 consumes more memory headrooms than VQ1, has reached more preferable character The accuracy rate of identification.The combination of Str-MRF2 and VQ2 is higher than the not only accuracy of identification of VQ1, and memory headroom is also less, comprehensively examines Consider, can more meet actual demand.

It is more than presently preferred embodiments of the present invention, all changes made according to technical solution of the present invention, produced function are made During with scope without departing from technical solution of the present invention, protection scope of the present invention is belonged to.

Claims (6)

1. a kind of online handwriting Chinese and japanese recognition methodss, it is characterised in that realize in accordance with the following steps：

Step S1：Character recognition, is preserving the former horizontal and vertical ratio for being input into word not the handwritten text of input first Linear normalization process is carried out in the case of change, the word of normal size is changed into, then extraction feature point again.For input text Every a line of word, elects the starting point and ending point on the side as characteristic point first；Secondly, if except having elected as on the side Certain point outside characteristic point is more than a certain threshold value to the distance of adjacent feature point, then the point is also chosen as characteristic point. so always Selection characteristic point is till no another characteristic point is optional；

Step S2：Using MRF models as ONLINE RECOGNITION device recognition function.By the characteristic point and prototype that compare input word The state of all kinds of prototypes in dictionary, obtains being input into the similarity of word and all kinds of prototypes, and the word class with maximum similarity is made To be input into the recognition result of handwritten text, the other MRF evaluators of character level are obtained；

Step S3：Structuring MRF evaluators are built, further, is created an interface to build structurized MRF, is obtained one First characteristic mean vector：The average coordinates of the characteristic point of training pattern.The all strokes of each character are connected to a stroke to obtain To stroke number independence.In characteristic point, a mouse click right button cuts character (character) pattern of characteristic point.One mouse Left button is clicked on, and will connect two free radicals (radical) in characteristic point.Character model can be divided into by several parts by the method. Each section is registered in free radical dictionary by decomposing normalized, while can be classified as identical free radical same Class；

Step S4：All character MRF model decompositions are registered, makes training pattern and character MRF synchronously decompose Free Radical Level. Average bounding box (mean bounding box) is obtained from the bounding box of free radical training pattern.Decomposite in kinds of characters The free radical for coming may have different size and locations, by these free radical normalization, then with normalized free radical model Training free radical MRF models.By the average bounding box of normalization average value vector sum covariance matrix, we arrange free radical MRF models are character model；

Step S5：Cluster different size and each free radical of position.Each free radical MRF is from kinds of characters apoplexy due to endogenous wind Free radical pattern, thus size position is different.Various sizes of each free radical of aggregation is different groups, and it is many to separate free radical Individual classification.Optimal stage size quantity；

Step S6：Dictionary is compressed by VQ technologies, in MRF evaluators, there is numerous identical unitary and binary feature. For unitary and binary feature, each mean vector has two param elements, and each covariance matrix has four parameter units Plain each state has three transition probabilities to have three param elements.By unitary and the mean vector and covariance square of binary feature Battle array param elements are combined as one.Represent that three param elements of three transition probabilities of each state are set to another group Close.It is different groups by parameter combination cluster, the group of shared identical parameters collection is set to into a group center.By storing group index and center Parameter, is compressed to evaluator.

2. a kind of online handwriting Chinese and japanese recognition methodss according to claim 1, it is characterised in that in the step S1 word In symbol identification process, using Ramner methods.

3. a kind of online handwriting Chinese and japanese recognition methodss according to claim 1, it is characterised in that step S2 builds The method of MRF evaluators is：Characteristic point is set from an input mould set S={ s₁,s₂,s₃,…,s_I}.Characteristic point is defeated from one Enter set of modes S={ s₁,s₂,s₃,…,s_I}.The recognition function is each with prototype dictionary by the characteristic point for comparing input word The state of class prototype, obtains being input into the similarity of word and all kinds of prototypes, and the word class with maximum similarity is used as input handss Write the recognition result of body word.

Neighborhood system is to be chosen continuously to close on characteristic point according to sequential write, is arranged

Odd number set:C1={ s₁,s₂,s₃,…,s₁₀,…}

Pairing set:C2={ { s₁,s₂},{s₂,s₃},{s₃,s₄},{s₄,s₅},……,{s₉,s₁₀},…}

A linear chain MRF is defined to each character type, each label has a state, and each state there are three conversion.

In MAP frameworks, maximum posterior probability is equal to the energy function of minimum to condition conversion probability.

Energy function

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

Wherein:One class distributes label s_i

:Unitary feature, including X and Y coordinates point

:Binary features (dx with dual element:X coordinate of s_i-X coordinate of s_i-1,dy:Y coordinate of s_i-Y coordinate of s_i-1)

Gaussian function is used for evaluation condition probability

$P\left(O_{s_i}\right|l_{s_i},C)P\left(O_{s_i{s}_{i-1}}\right|l_{s_i},l_{s_{i-1}},C)$

State transition probability：

$P\left(l_{s_i}\right|l_{s_{i-1}},C)$

$P\left(l_{s_i}\right|l_{s_{i-1}},C)=\frac{Numberoftransitionsfrom{l}_{s_i-1}to{l}_{s_i}}{Numberofsitesassigned{l}_{s_i-1}}$

$P\left(l_{s_1}\right|l_{s_{i-1}},C)=\frac{Numberof{s}_{1}assigned{l}_{s_1}}{Numberof{s}_1}---\left(2\right)$

Each character type has a straight chain MRF, and we use each of the input pattern and MRF of Viterbi search's matching characteristic point The state of character type, is path and least energy that each character type finds matching.

4. a kind of online handwriting Chinese and japanese recognition methodss according to claim 1, it is characterised in that in step S5, During each free radical of cluster different size and position, assemble various sizes of each free radical using equal cluster analysis method For different groups, separation free radical is multiple classifications, optimal stage size quantity.

5. a kind of online handwriting Chinese and japanese recognition methodss according to claim 1, it is characterised in that character type evaluator leads to Cross online Japanese hand-written data storehouse Nakayosi training.

6. a kind of online handwriting Chinese and japanese recognition methodss according to claim 1, it is characterised in that performance test using Line Japanese hand-written data storehouse Kuchibue.