CN101256624B

CN101256624B - Method and system for establishing HMM topological structure being suitable for recognizing hand-written East Asia character

Info

Publication number: CN101256624B
Application number: CN200710085282A
Authority: CN
Inventors: 韩石; 邹宇; 常明; 刘鹏; 吴义坚; 马磊; 宋謌评; 张冬梅; 王坚
Original assignee: Microsoft Corp
Current assignee: Microsoft Technology Licensing LLC
Priority date: 2007-02-28
Filing date: 2007-02-28
Publication date: 2012-10-10
Anticipated expiration: 2027-02-28
Also published as: WO2008104130A1; CN101256624A

Abstract

The invention, aiming at characteristics of writing East Asia characters, provides a design proposal of topology structure of HMM model suitable for identifying writing East Asia characters and a method for establishing the HMM model of the topology structure; with the sufficient view to many strokes, diverse forms of stroke orders, complicated structures, various writing styles and uncertain connection between the strokes in the writing East Asia characters, the invention introduces a corner state, provides a plurality of paths and parallel state in the HMM topology structure to solve the above problem, besides, data quantity is reduced by the operations of convergence and combination, thus lessening the complexity of the operation.

Description

Foundation is applicable to the method and system of the HMM topological structure of identification hand-written East Asia character

Technical field

The present invention relates to Character Recognition, more particularly, relate to the recognition technology of hand-written East Asia character.

Background technology

Handwriting input identification is an important research direction in the Computer Applied Technology always.Different with the image of static state or literal identification is that online handwriting input identification is a time stochastic process, rather than the object of a static state.Therefore, in the online handwriting recognition field, recessive markov (HMM) model often is used.

Brief account HMM model once at first, the HMM model is a hidden markov models, is a kind of of Markov model.Markov model can be used for predicting that the probability of certain state appears in following a certain incident, and this probability will be only based on this incident current states.Markov model shows as a finite-state automata, can change mutually between the state, and, each time will be from a state exchange to next state (possibly be other state, also possibly be this state itself).

Markov model can be divided into two kinds, a kind of dominance Markov model that is referred to as, and in the dominance Markov model, the conversion sequence between the state is known.

Another kind is called hidden markov models (HMM), and wherein the conversion sequence between the state is unknown, and what known only is the transition probability between the state.Therefore, the HMM model can be defined as and have following characteristic:

1) be a finite-state automata, can change mutually between the state, and, each time will be from a state exchange to next state (possibly be other state, also possibly be this state itself);

2) transfer between the state is determined by one group of transition probability, and the probability of occurrence of one group of observed events (observation sequence) is by the transition probability decision relevant with state.

Such as, the explanation HMM model of giving one example.A beverage dispenser being arranged, two kinds of beverages are provided, is respectively can happy tea.This beverage dispenser can have two states; The state of " preference is laughable " and the state of " preference tea "; When dropping into one piece of coin and obtain your drink, the beverage that vending machine can be sold according to its residing state decision, promptly; Cola will be sold when being in the state of " preference laughable ", and tea will be sold when being in the state of " preference tea ".This vending machine can be changed between two states after each the sale, and the probability of conversion is following:

When being in " preference is laughable " state, the probability that is transformed into " preference tea " state is 30%, and the probability that remains on " preference is laughable " state is 70%;

The probability that when starting from the state of " preference tea ", is transformed into " preference is laughable " state is 50%, and the probability that remains on " preference tea " state also is 50%.

For above-mentioned beverage dispenser,, can set up the HMM model for above-mentioned beverage dispenser if hope to confirm the probability of a certain specific beverage vending utensil sequence.This model can be with reference to shown in the figure 1b, comprising two states: " preference is laughable " state and " preference tea " state, according to the top state transition probability of listing, confirm the probability of each bar transfer path of transition, shown in Fig. 1 b.When needs are confirmed some specific sale sequences, can be through calculating the probability that all path of The above results can occur, and, just can obtain the probability of occurrence of this a kind of specific sale sequence with they summations.

The HMM topological structure has reflected the transfer sequence annexation between each state in the HMM model just.

The application of HMM model and HMM topological structure relates to three types of main problems:

1) when providing a HMM model, confirms the probability that a certain specific " observed events " (corresponding state transitions order) occurs.

2) when providing specific " observed events " of a HMM model and, select a state transitions order, this state transitions order can be described this specific " observed events " best.

3) when providing specific " observed events " and one group of possible HMM model space, confirm that best HMM model describes this " observed events ".

The These characteristics of HMM model makes it when solving following problem, be particularly useful: the probability of occurrence of potential incident can influence observed events.A typical application is in the identification field, and especially speech recognition comprises speech recognition technology and handwriting recognition technology.The HMM model is a kind of trainable model; Through the lot of data training, can obtain one group of HMM model parameter that is suitable for expression " observed events " most, afterwards; The HMM model that adopts this from the data training, to obtain; Just can confirm the matching degree of another one " observed events " and this model,, just can confirm which model " observed events " most probable of a unknown belongs to according to matching degree.Like this, just reached the purpose of identification.

At present, though developed the recognition technology of the multiple HMM of utilization model, they are applied to western language mostly, and for orient characters, especially the literal of East Asia Region use such as Chinese character, but can not carry out effective recognition.This mainly is because western language, for example English and East Asia character, and for example Chinese character is caused in the structural difference of font.

For western language, each word all is made up of letter, and single letter configuration is simple; Usually all being one can accomplish; The problem that does not have variable order of strokes, simultaneously, the similarity degree between letter and the letter is relatively low; Except individual other letter, most of letters have own obvious characteristics.These characteristics have all been brought many facilities to handwriting recognition.Therefore, exploitation at present be that the HMM model of main identifying object generally all has a left side to right HMM topological structure with western language, through the basic characteristics that just can describe general letter word of starting point and terminating point that limit the HMM topological structure.

But East Asia character has visibly different characteristics, is example with the Chinese character:

1) stroke is many, in hand-written process, has the order of strokes different problems;

2) complex structure, it is more to make that font changes, and is embodied in hand-written aspect, has multiple handwriting style exactly;

3) have the problem that connects between stroke, because Chinese character stroke is many, complex structure is added individual hand-written custom, can cause the connection between the stroke to have a lot of uncertainties;

4) data volume is big, each word of Chinese character all be one independently individual, rather than picture English that works further degree split into a limited number of units the letter, therefore, for Chinese character, the data volume of handwriting recognition model is that ten minutes is huge.

Through top analysis, just can see, because there are a lot of evident difference in East Asia character and western language on the font characteristics; Just caused being directed against the HMM model of western language character exploitation at present and not being suitable for East Asia character, especially, present HMM model can not solve the problem of order of strokes, writing style effectively; Simultaneously; For the connection between the East Asia character stroke, also can't solution be provided well, therefore; Caused at present aspect hand-written East Asia character identification, still not having a kind of good recognition technology.

Summary of the invention

The object of the invention aims to provide the method and system that a kind of structure is used to discern the HMM topological structure of hand-written East Asia character, with the characteristics to hand-written East Asia character hand-written East Asia character is carried out effective recognition.

According to an aspect of the present invention; Many to the hand-written East Asia character stroke; Connection between stroke excessively concerns complicated characteristics; Design the method that a kind of foundation is applicable to the HMM topological structure of identification hand-written East Asia character, this design provides a description hand-written East Asia character and continues the persistent state of stroke and the corner state of describing corner between the hand-written East Asia character stroke in the HMM topological structure of being set up.Through introducing corner state, make whole HMM topological structure reflect the characteristics of hand-written East Asia character well to corner between stroke.

In a realization of the present invention, the HMM topological structure is to be designed to the right HMM topological structure of left-hand, since an initial state, finishes to a final state; Persistent state in the HMM topological structure can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting; And persistent state and corner state alternately exist successively.

According to a further aspect in the invention, various to the hand-written East Asia character order of strokes, the characteristics that writing style is various.A kind of multipath HMM topological structure is provided, in the multiple order of strokes of the corresponding hand-written East Asia character of each paths in this multipath HMM topological structure one; Perhaps in the multiple handwriting style of the corresponding hand-written East Asia character of each paths in the multipath HMM topological structure.Be integrated in the HMM topological structure through the path that will reflect multiple order of strokes or handwriting style, just can solve the various and various problem of handwriting style of hand-written East Asia character order of strokes well.

In a realization of the present invention, each paths in the multipath HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; In this multipath HMM topological structure, provide a description the corner state that hand-written East Asia character continues the persistent state of stroke and describe corner between the hand-written East Asia character stroke equally and comprise persistent state and corner state; Wherein persistent state can be transferred to NextState or shift certainly; The corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state also are alternately to exist successively; And wherein all paths originate in same entry state, to same discharge state end.Simultaneously, the present invention also merges processing to the path of multichannel in the HMM topological structure, with the quantity in control path.

According to a further aspect in the invention, to the uncertainty that connects transition between the hand-written East Asia character stroke, promptly existing promptly possibly be real pen; It also possibly be the part of empty pen; The present invention provides the method that solves equally, such as, in the HMM topological structure, parastate is provided; In the corresponding hand-written East Asia character both possibly be empty stroke, also possibly be the part of real stroke.Perhaps, the HMM topological structure being used many spatial probability distribution (MSD), both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

According to a further aspect in the invention; In order to reduce redundant data volume; Reduce the complexity of computing, the present invention also carries out cluster to the state in the HMM topological structure, makes at least one group of state common parameter in the HMM topological structure; And, in the HMM topological structure, only preserve set of parameter for the state of one group of common parameter.

The present invention provides the implementation method of the HMM topological structure that is suitable for hand-written East Asia character identification equally, and this multipath HMM topological structure makes up from the training data of hand-written East Asia character automatically; According to the order of strokes or the writing style of hand-written East Asia character, the automatic classification method of using a machine self study carries out cluster to training data, and the data of each classification are corresponding to a sequential write or a writing style.

According to a kind of implementation; This HMM topological structure is according to automatic generation of training data of hand-written East Asia character and the not artificial intervention of needs; This training data comprises the hand-written East Asia character writing sample, and writing sample comprises the writing sample of different order of strokes or different handwriting styles; In the process that makes up the HMM topological structure, can comprise following several stages: for each writing sample, be divided into several segmentations by arc length, a characteristic is extracted in each segmentation respectively; A thereby characteristic the segmentation feature series arrangement being formed whole writing sample; Above-mentioned characteristic is carried out cluster, and each cluster is corresponding to a kind of order of strokes or a kind of handwriting style; Data based on each cluster are confirmed topology and initial parameter corresponding to the path in the HMM topological structure of each order of strokes or each handwriting style.

In a realization of the present invention, also set up subsequence direction histogram vector, to help to set up this HMM topological structure corresponding to segmentation and characteristic.

In a realization of the present invention, also the path in the multipath HMM topological structure is merged, with the quantity in control path.

Because the data volume of hand-written East Asia character is huge, in order to reduce data volume, reduce the complexity that realizes, in a realization of the present invention, the state measuring similarity between per two states determines whether to carry out the operation of cluster in the HMM topological structure through calculating.

The present invention provides the system that can realize above-mentioned various aspects equally, utilizes the HMM model to come hand-written East Asia character is discerned.

The HMM topological structure that the characteristics that the present invention is directed to hand-written East Asia character provide a kind of foundation to be applicable to hand-written East Asia character is discerned; Taken into full account that the hand-written East Asia character stroke is many, order of strokes is various, complex structure, writing style are various, connect uncertain characteristics between stroke; Through in the HMM topological structure, introducing the corner state, multipath is provided, provides the means of parastate to solve above-mentioned problem; And; Operation through cluster and merging reduces data volume, reduces the complexity of computing.

Description of drawings

Above-mentioned and other characteristic of the present invention, character and advantage will be through becoming more obvious below in conjunction with accompanying drawing to the description of embodiment, and in the accompanying drawings, identical Reference numeral is represented identical characteristic all the time, wherein,

Fig. 1 a is an example that can realize suitable computingasystem environment of the present invention;

Fig. 1 b shows the example of a HMM model;

Fig. 2 a shows " corner " in the hand-written East Asia character;

Fig. 2 b shows the HMM topological structure that can embody corner characteristic in the hand-written East Asia character;

Fig. 2 c shows the distinguishing rule that continues stroke and corner in the hand-written East Asia character;

Fig. 3 a shows the diversity of the order of strokes of hand-written East Asia character;

Fig. 3 b shows the diversity of the handwriting style of hand-written East Asia character;

Fig. 4 a shows the structural drawing according to the multipath HMM topological structure of one embodiment of the invention;

Fig. 4 b shows the example of the multipath HMM topological structure of the corresponding a kind of order of strokes in each path;

Fig. 4 c shows the example of the multipath HMM topological structure of the corresponding a kind of handwriting style in each path;

Fig. 5 a shows real stroke in the hand-written East Asia character writing process/probabilistic situation of empty stroke;

Fig. 5 b shows the synoptic diagram according to the HMM topological structure with parastate of one embodiment of the invention;

Fig. 5 c shows the synoptic diagram of the HMM topology mechanism with parastate that makes up according to the handwriting sample shown in Fig. 5 a;

Fig. 6 a shows an instance of the sample application subsequence direction histogram of identifying the handwriting;

Fig. 6 b shows an instance that training data is carried out cluster operation;

Fig. 6 c shows an instance that adopts the double gauss mixture model to realize parastate;

Fig. 7 shows the structural drawing that utilizes the system that the HMM model discerns hand-written East Asia character according to one embodiment of the invention;

Fig. 8 shows the structural drawing that utilizes the system that the HMM model discerns hand-written East Asia character according to another embodiment of the present invention.

Embodiment

The present invention is directed to the characteristics of hand-written East Asia character: stroke is many, order of strokes is various, complex structure, writing style are various, it is uncertain to connect between stroke.Providing a kind of utilizes improved HMM model to come the scheme that hand-written East Asia character is discerned; The characteristics of HMM topological structure have been improved; Through in the HMM topological structure, adding the corner state, multipath is provided, provides the means of parastate to come to reflect well the These characteristics of hand-written East Asia character; And, reduce data volume through the operation of cluster and merging, reduce the complexity of computing.HMM topological structure of the present invention is from training data, to generate voluntarily, does not need manual intervention.

What need explanation a bit is, below among the embodiment that will describe, be that example describes with the Chinese character; But scope of the present invention is not limited to Chinese character; But have all East Asia alphabetic characters with the Chinese character similar features, comprise kanji, assumed name of Japan or the like.

Suitable realization environment

Fig. 1 a has explained an example of suitable computingasystem environment 100, wherein can realize the present invention.Computingasystem environment 100 only is an example of suitable computing environment and is not intention restriction usable range of the present invention or function.Computing environment 100 should not be interpreted as to be had with the arbitrary of the assembly described in the exemplary operation environment 100 or makes up relevant dependency or requirement.

Those skilled in the art will appreciate that computing machine or other client computer or server apparatus can be used as the part computer network and adopt, and perhaps are used for DCE.In this, the invention belongs to any computer system with any amount internal memory or storage unit, and the application program and the process that occur in any amount on any amount storage unit or the capacity, they can use with the present invention.The present invention can be applied in network environment or DCE, adopt the environment of server computer and client computers.The present invention can also be used for independent computing equipment, has the programming language function and produces, receives and launch information interpreting and executive capability with long-range or local service.

The present invention can use multiple other general or special-purpose computing system environment or configuration to operate.The example of known computing system, environment and/or the configuration that can be fit to use with the present invention including, but not limited to: personal computer, server computer, portable or portable set, multicomputer system, based on microprocessor system, STB, programmable user electronic equipment, network PC, small-size computer, mainframe computer, comprise DCE of arbitrary said system or the like.

The present invention can describe with the general context of computer executable instructions, the program module of for example being carried out by computing machine.Generally speaking, program module comprises routine, program, object, assembly, data structure etc., and they are carried out particular task or realize specific abstract data type.The present invention can also be actually used in the DCE, is wherein executed the task by the teleprocessing equipment that connects through communication network or other data transmission medium.In DCE, program module and other data can be arranged in local and remote storage medium, comprise memory storage device.Distributed Calculation is convenient to share computer resource and service through the direct exchange between computing equipment and system.These resources comprise information, high-speed cache with service, reach the exchange of file disk storage.Distributed Calculation is utilized the network connectivity, allows their collective's effect of subscriber computer performance to help whole company.In this, plurality of devices can have application program, object or resource, and they can utilize technology of the present invention.

With reference to figure 1a, be used to realize that example system of the present invention comprises that form is the universal computing device of computing machine 110.The assembly of computing machine 110 can be including, but not limited to: processing unit 120, Installed System Memory 130, and comprising that the various system components of Installed System Memory are coupled to the system bus 121 of processing unit 120.System bus 121 can be polytype bus-structured any, comprise rambus or Memory Controller Hub, peripheral bus, and use arbitrary multiple bus-structured local bus.But unrestricted, this structure comprises industrial standard architectures (ISA) bus, MCA (MCA) bus, enhancement mode ISA (EISA) bus, video electronics standard alliance (VESA) local bus, reaches periphery component interconnection (PCI) bus (being also referred to as the Mezzanine bus) through example.

Computing machine 110 generally comprises various computer-readable mediums.Computer-readable medium can be and to comprise volatibility and non-volatile medium, removable and not removable medium by any available media of computing machine 110 visit.But unrestricted, computer-readable medium can comprise computer storage media and communication media through example.Computer storage media comprises volatibility and non-volatile, removable and not removable medium, and they are realized with any means or the technology that is used to store such as the such information of computer-readable instruction, data structure, program module or other data.Computer storage media is stored expectation information and can be by any other medium of computing machine 110 visits including, but not limited to: RAM, ROM, EEPROM, flash memory or other memory technology, CDROM, digital versatile disc (DVD) or other optical disc memory, tape cassete, tape, magnetic disk memory or other magnetic storage apparatus or be used to.Communication media generally comprises computer-readable instruction, data structure, program module or other data in the modulated data signal such such as carrier wave or other transmission mechanism, and comprises any information delivery medium.Term " modulated data signal " means the signal that its one or more characteristics are set up or change with the mode that the signal internal information is encoded.Through example but unrestricted, communication media comprise such as cable network or directly line connect so wired medium and such as sound, RF, infrared such wireless media and other wireless media.Above-mentioned combination in any should be included in the scope of computer-readable medium.

Installed System Memory 130 comprises computer storage media, and its form is volatibility and/or Nonvolatile memory, such as read-only memory (ROM) 131 and random access memory (RAM) 132.Basic input/output 133 (BIOS) generally is stored in the ROM131, and it comprises the basic routine of the inter-module transmission information that for example helps between the starting period in computing machine 110.RAM132 generally comprises data and/or program module, and they can zero accesses and/or are currently operated above that by processing unit 120.But unrestricted, Fig. 1 a has explained operating system 134, application program 135, other program module 136 and routine data 137 through example.

Computing machine 110 can also comprise other removable/not removable, volatile/nonvolatile computer storage media.Only pass through example; Fig. 1 explained hard disk drive 141 that not removable, non-volatile magnetic medium is read and write, to disc driver 151 removable, that non-volatile magnetic disk 152 is read and write and CD drive 155 that removable, non-volatile CD 156 is read and write, such as CD ROM or other optical medium.In the exemplary operation environment available other removable/not removable, volatile, nonvolatile calculates storage media including, but not limited to tape cassete, flash card, digital universal disc, digital video tape, solid-state RAM, solid-state ROM or the like.The hard disk drive 141 general not removable memory interfaces that pass through as interface 140 link to each other with system bus 121, and disc driver 151 generally links to each other with system bus 121 with the removable memory interface as interface 150 with CD drive 155.

Driver of discussing above and in Fig. 1 a, explaining and their correlation computer storage media are the storage that computing machine 110 provides computer-readable instruction, data structure, program module and other data.In Fig. 1, for example, said hard disk drive 141 storage operating systems 144, application program 145, other program module 146 and routine data 147.Notice these assemblies or can be identical with routine data 137, perhaps different with them with operating system 134, application program 135, other program module 136.Here provide different digital to explain that they are different copies at least for operating system 144, application program 145, other program module 146 with routine data 147.The user can be through being input to order and information in the computing machine 110 such as keyboard 162 and indicating equipment 161 such input equipments, and input equipment is commonly referred to mouse, trace ball or touch panel.Other input equipment (not shown) can comprise microphone, joystick, cribbage-board, satellite dish, scanner or the like.These often link to each other with processing unit 120 through the user's input interface 160 with system bus 121 couplings with other input equipment, but also can use other interface to be connected with bus structure, such as parallel port, game port or USB (USB).The display device of monitor 191 or other type is also through linking to each other with system bus 121 such as video interface 190 such interfaces.Except monitor 191, computing machine can also comprise other external unit, and like loudspeaker 197 and printer 196, they can connect through output Peripheral Interface 190.

Computing machine 110 can be operated in the networked environment, and this environment uses and is connected such as the logic between the such one or more remote computers of remote computer 180.Remote computer 180 can be personal computer, server, router, network PC, peer device or other common network node; And generally comprise the relevant many or whole elements of above-mentioned and computing machine 110, although memory storage device 181 only has been described among Fig. 1.The described logic of Fig. 1 a connects and comprises Local Area Network 171 and wide area network (WAN) 173, but can also comprise other network.This networked environment is common in office, enterprise-wide. computer networks, corporate intranet and internet.

When being used for the LAN networked environment, computing machine 110 links to each other with LAN171 through network interface or adapter 170.When being used for the WAN networked environment, computing machine 110 generally comprises modulator-demodular unit 172 or other device that is used on the WAN173 such such as the internet, setting up communication.Modulator-demodular unit 172 can be inner or outside, and it can link to each other with system bus 121 through user's input interface 160 or other suitable mechanism.In networked environment, can be stored in the remote memory storage devices about computing machine 110 described program modules or its part.But unrestricted, Fig. 1 a has explained the remote application 185 that resides on the memory device 181 through example.It is exemplary that network shown in being appreciated that connects, and also can use other device that establishes a communications link at intercomputer.

Term definition

In order more succinctly, clearly to describe the present invention, in this article, following term refers in particular to following implication,

Stroke: the movement locus when writing.For the identification sampling, when writing, when nib contacts with contact plate, will stay " person's handwriting ", when not contacting, just can not stay " person's handwriting " with contact plate.But, if nib then is a continuous track in the motion in space when considering to write.In the present invention, the track of this motion will be made a general reference in term " stroke ", and no matter whether it stays " person's handwriting ".

Real stroke: real stroke refers to stay the stroke of person's handwriting, promptly when writing, and the movement locus when nib contacts with contact plate.

Empty stroke: empty stroke refers to not stay the stroke of person's handwriting, promptly when writing, and the movement locus of nib when contact plate does not contact; Empty stroke is used for embodying the connection between the real stroke more; Need to prove that among the present invention, empty stroke mainly embodies the movement tendency and the direction of nib; For empty stroke, might not require to reflect fully nib this stage the actual path of process.

Continue stroke: in a period of time, the stroke that direction is constant basically.Lasting stroke can be real stroke, also can be empty stroke, and what continue the stroke reflection is the motion of the nib on basic identical direction continuously.

Corner: the stroke of direction significant change.Corner is the transfer process that continues between the stroke, and the corner reflection has the nib motion that remarkable direction changes.In the stroke conversion is the notable feature that corner is a hand-written East Asia character to occur, and the present invention provides " corner " to come to reflect better this characteristic for this reason specially.

Person's handwriting: person's handwriting is meant the concrete vestige that hand-written East Asia character is left.

Radical: the stroke of a part or combination.

General introduction

Although the HMM model has been widely used in the identification field of the hand-written character of line, for the identification of online hand-written East Asia character, still have the challenge of two main aspects: the order of strokes of hand-written East Asia character is various various with writing style.The HMM topological structure that typically is used for hand-written script identification is right (left-to-right) HMM topological structure of left-hand.In the present invention, relate to improvement, mainly comprise the HMM topological structure:

In the HMM topological structure, add " corner state ", with " corner " between the hand-written process stroke of reflection hand-written East Asia character.Certainly the transfer of " corner state " is restricted, and " corner state " is alternately to occur with " persistent state " that reflect lasting stroke in the hand-written process of hand-written East Asia character.

In a HMM, use mulitpath, to reflect the different order of strokes and the writing style of same hand-written East Asia character.

In the process that makes up this HMM topological structure, the present invention has considered following problem:

At first, in a HMM model, the path of how much quantity is suitable.Problem hereto; The present invention proposes a kind of new solution; Adopt subsequence direction histogram vector (Subsequence Direction Histogram Vector) that cluster is carried out in the path that from character data, obtains, confirmed suitable number of paths.

The second, in a path, the state of how much quantity is suitable.Problem hereto; The present invention utilizes " corner " this characteristic in the hand-written East Asia character; Utilize the annexation of corner in the curvature scale space to confirm the quantity of " corner state " in the path; Utilizing " corner state " and " persistent state " again is this characteristic alternately to occur, confirms proper state number in the path.

The 3rd, for continuous uncertainty between distinctive stroke in the hand-written East Asia character, the present invention has adopted the method that parastate (such as the double gauss mixture model) is set in the HMM topological structure, perhaps use the method for MSD and solve.

To at length describe with regard to above-mentioned problem respectively below.

Be suitable for discerning the HMM topological structure of hand-written East Asia character

The present invention is primarily aimed at the recognition technology of hand-written East Asia character, analyzes the characteristics of hand-written East Asia character, can be summed up as following main aspect:

1) order of strokes is various, and writing style is various.Analyze as top, the stroke of hand-written East Asia character is many, and font structure is complicated, caused same word multiple order of strokes can occur, and the word of finally having write can demonstrate multiple font or font, just so-called handwriting style.

2) owing to the diversity of writing style, following two kinds of typical differences can appear in the different literary styles of same word:

2a), cause different writing styles possibly have different stroke numbers to the different simplification literary style of local continuous stroke.

2b) in the connection procedure between real stroke, can link to each other, also can link to each other with empty stroke with real stroke; And for the user of real East Asia character; No matter with which kind of stroke link to each other, all be recognized as correct character, this just individual machine recognition has been brought difficulty.

Hand-written East Asia character identification first problem to be solved, promptly problem 1) order of strokes is various, handwriting style is various.

Same Chinese character; Different people's writing style has determined to occur the order of multiple stroke, though according to the standard of standard Chinese character, the order of strokes of writing fixes; But actual statistics shows; There are several sequential writes in most Chinese character in the application of reality, and, adopt the crowd of each sequential write all to occupy suitable ratio.Like this towards masses' technology, must consider this situation as handwriting recognition.According to an example, shown in figure 3a, " nine " word wherein just has two kinds of different order of strokes, wherein, dots two connections between the stroke.

Equally, hand-written East Asia character to write lattice varied especially, for example, shown in figure 3b, same Chinese character " is answered ", just various writing style possibly occur.

A kind of implementation is that data volume and the computational complexity done like this are bigger for each order of strokes, each handwriting style all make up a HMM topological structure, and another problem is that such scheme will be to realizing and training and all bring certain degree of difficulty.

For this reason, the present invention provides a kind of HMM topological structure of multipath, for identical character, a HMM topological structure (HMM model) only is provided, and uses path wherein to show different order of strokes or handwriting style.

With reference to figure 4a, Fig. 4 a shows according to one embodiment of the invention, a kind of structural drawing that is applicable to the multipath HMM topological structure that carries out hand-written East Asia character identification that is proposed.Shown in figure 4a, this multipath HMM topological structure is corresponding to a hand-written East Asia character; In the multiple order of strokes of the corresponding hand-written East Asia character of each paths in the multipath HMM topological structure one; Perhaps in the multiple handwriting style of the corresponding hand-written East Asia character of each paths in the multipath HMM topological structure.

Such as, with reference to figure 4b, the multipath HMM topological structure constructed according to two handwriting samples of the Chinese character " nine " of being used among Fig. 3 a has 2 paths, a kind of order of strokes of corresponding " nine " word of each bar wherein.

With reference to figure 4c; A plurality of handwriting samples of " answering " according to the shown Chinese character of Fig. 3 b; After structure was used for multipath HMM topological structure that Chinese character " answers ", cluster went out two kinds of handwriting styles (process of cluster will be discussed in more detail below) with characteristic feature, like this; This multipath HMM topological structure has two paths, " answering " word of the corresponding a kind of handwriting style of each bar wherein.

As top said, because the problem that the diversity of writing style is brought, one of them promptly is 2a) to the different simplification literary style of local continuous stroke, cause different writing styles possibly have different stroke numbers.Status number on each paths is to be confirmed by the stroke number of the pairing writing style of this paths.East Asia character has tangible direction transformation characteristic between adjacent strokes, i.e. and " corner ", and lasting stroke and corner are alternately to occur.

According to the present invention, at first need count the typical stroke number of this writing style, thereby confirm the status number of this paths through the pairing training sample in a certain path is carried out the detection of corner, concrete step can be carried out detailed description below.Just because of lasting stroke and corner is this characteristics alternately to occur, can utilize the algorithm that becomes more meticulous (coarse-to-fine algorithm) in the curvature scale space (CSS) to realize confirming of number of states in the path.

For the concrete path of each bar, the present invention has also introduced the of short duration stroke transformation characteristic that " corner state " is used for describing the corner in the path.On a pairing paths with a kind of writing style, continue stroke corresponding to " persistent state ", corner is corresponding to " corner state ", and persistent state and corner state are alternately to occur.Because the short-time characteristic of stroke corner; The present invention is made as not rotation with " corner state "; The characteristic of this not rotation combines to continue the difference of stroke feature and corner characteristic, can obviously improve the characteristic sequence of training sample and the correct corresponding relation between the path status.

Study the characteristic of hand-written East Asia character again, can find, hand-written East Asia character begins before the next stroke accomplishing a stroke, a process that turns to significantly can occur, just above said " corner ".Shown in figure 2a, Fig. 2 a with a foolproof Chinese character " on " a handwriting sample be example.In the example of Fig. 2 a, " on " the real stroke of word is 3 strokes, is exactly the person's handwriting that is shown among Fig. 2 a, empty stroke has 2 strokes, in Fig. 2 a, representes with the dotted line of band arrow.Like this,, 4 corner 202a-202d occurred, used the circle of frame of broken lines to represent in the drawings altogether between 5 strokes (3 real strokes, 2 empty strokes).

So, the invention provides a kind of improved HMM model, utilize this HMM model, can be well the corner of East Asia character be carried out modeling.

According to the present invention, a kind of method of the HMM of utilization Model Identification hand-written East Asia character is provided, the HMM topological structure corresponding to hand-written East Asia character is provided; Wherein

In the HMM topological structure, provide a description the persistent state that hand-written East Asia character continues stroke; And, in the HMM topological structure, provide a description the corner state of corner between the hand-written East Asia character stroke.

Above-mentioned HMM topological structure is built into the right HMM topological structure of left-hand, since an initial state, finishes to a final state.Persistent state in the HMM topological structure can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting.Persistent state and corner state alternately exist successively.

The improved HMM topological structure of above-mentioned this meets the characteristics of hand-written East Asia character, can reflect the characteristic of hand-written East Asia character well.

With reference to figure 2b, Fig. 2 b show above-mentioned principle according to the present invention from Fig. 2 a " on " make up the word person's handwriting corresponding to " on " the HMM topological structure of word.

Shown in Fig. 2 a " on " 5 strokes in the word handwriting sample, comprising 3 real strokes and 2 empty strokes 5 persistent state 204a-204e among the corresponding diagram 2b respectively, each of persistent state 204a-204e can be from shifting or to next state transitions.

Continue with reference to figure 2b, 4 corner state 206a-206d among 4 corner 202a-202d difference corresponding diagram 2b shown in Fig. 2 a, corner state 206a-206d can not can only transfer to NextState from shifting.And corner state 206a-206d is inserted between the persistent state 204a-204e.

In another embodiment, this HMM topological structure shown in Fig. 2 b can also originate in an initial state, and this state is corresponding to the moment of the first stroke of a Chinese character, and if there is initial state, this initial state can not be from shifting.

Equally, also can end at final state among another embodiment of the HMM topological structure shown in Fig. 2 b, in the moment that pen is received in the final state representative, if use final state, final state can not shift certainly.

Like this, shown in Fig. 2 a " on " handwriting sample just be built into the HMM topological structure shown in Fig. 2 b, this HMM topological structure meets following characteristic: be the right HMM topological structure of left-hand; Since an initial state; To final state end, the persistent state in the HMM topological structure can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState; Can not be from shifting, persistent state and corner state alternately exist successively.

Need to prove; Shown in Fig. 2 a " on " in the handwriting sample of word; The quantity of the real stroke that defines according to the present invention equates with the stroke quantity in the standard Chinese character literary style; But for some Chinese characters or Chinese character stroke are arranged, the stroke quantity in defined stroke quantity according to the present invention (comprising real stroke and empty stroke) and the standard Chinese character literary style possibly be unequal.With reference to the shown example of figure 2c, " mouth " word according to the standard Chinese character literary style, has 3 strokes, but according to definition of the present invention, real stroke has 4 strokes, and empty stroke has 2 strokes.Main difference is " ┓ " this pen, and in standard Chinese character, this is a stroke, and according to definition of the present invention, " ┓ " comprised 2 real strokes and 1 corner.

Therefore, according to definition of the present invention, the handwriting sample of " mouth " word shown in Fig. 2 c comprises real stroke 212a-212d, empty stroke 214a and 214b, and corner 216a-216e.The order of arranging is: real stroke 212a, corner 216a, empty stroke 214a, corner 216b, real stroke 212b, corner 216c, real stroke 212c, corner 216d, empty stroke 214b, corner 216e, real stroke 212d.

In the present invention, said stroke is according to defined stroke in the top term definition.

Get back to the HMM topological structure of multipath noted earlier,, therefore in each bar HMM path, above-mentioned " corner " state can be provided still owing to mainly be applicable to the identification of hand-written East Asia character.

Get back to Fig. 4 a, in this multipath HMM topological structure, because each path all is corresponding to hand-written East Asia literal, so each bar in these paths all has following characteristics:

Each paths is the right HMM topological structure of left-hand, since an initial state, finishes to a final state;

The HMM topological structure comprises describes the persistent state that hand-written East Asia character continues stroke; And the corner state of describing corner between the hand-written East Asia character stroke; Wherein persistent state can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively;

For whole multipath HMM topological structure, wherein all paths originate in same entry state, to same discharge state end.

In Fig. 4 a, each circle is represented a state, and the line segment of band arrow is represented to shift, and wherein, point to the arrow of oneself and represent that from shifting, having state is persistent state from the state that shifts, can not be the corner state from the state that shifts.All paths all start from public entry state, and end at public discharge state.

Need to prove; Though in the example in the above; Each path in the multipath HMM topological structure all is the path that meets the defined characteristics in front, and the path in the multipath HMM topological structure is illustrated as and is used for as one man corresponding multiple order of strokes or handwriting style.But, those skilled in the art will appreciate that scope of the present invention never limits therewith, the wide in range qualification of the multipath HMM topological structure among the present invention should be:

Having a paths at least is the right HMM topological structure of left-hand, since an initial state, finishes to a final state;

And; The path of describing multiple different order of strokes and different handwriting style can be placed in the same multipath HMM topological structure; That is to say; In a multipath HMM topological structure, can be by the corresponding different order of strokes in the path of part, and the corresponding different handwriting style in the path of another part.

In a kind of situation; Because the effect of " corner " characteristic in the hand-written East Asia character; Have at least a paths to comprise and describe the persistent state that hand-written East Asia character continues stroke, and the corner state of describing corner between the hand-written East Asia character stroke, wherein persistent state can be transferred to NextState or shift certainly; The corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively.

Because HMM topological structure of the present invention is from training data, to generate voluntarily, and possibly there is very big redundancy in the sample in the training data, perhaps wherein possibly write down the contingency data of some extremely low probability; Like this; Just possibly make in the resulting multipath HMM topological structure, have redundant path, the path of perhaps having only a small amount of hand-written data to cover; This will make the model data amount roll up, and increase the complexity of whole HMM topological structure.For this reason, the present invention also need merge processing to the path, eliminates redundant path and the path that can not use basically.Concrete implementation method about the path merges can be described in detail below.What also need explain a bit is; For a person skilled in the art; Can in a HMM topological structure, realize multipath through existing means fully; Therefore; The multipath HMM model that the present invention mainly provides a kind of novel HMM topological structure herein and realizes this novel HMM topological structure, therefore, those skilled in the art can realize multipath HMM model described here fully and not need to make any creative work again after having read instructions of the present invention.

Now, method of the present invention is through provide connectivity problem between the stroke that " corner " solved hand-written East Asia character in good shapely, provides the HMM topological structure of multipath to solve the problem of multiple sequential write and multiple writing style.The problem that last need solve, just problem 2b), in the connection procedure between real stroke; Can link to each other with real stroke; Also can link to each other, and, no matter link to each other with which kind of stroke for the user of real East Asia character with empty stroke; All be recognized as correct character, this changeability of writing has been brought difficulty to machine recognition.For example, as shown in reference to Figure 5a, for the "on" character, in the finished "Shu" to the next record, "a" between the two connections may occur, real stroke connection, for example in Figure 5a 502 strokes shown in real or virtual stroke connections, such as shown in Figure 5a imaginary stroke 504 and, of course, there may be other connection methods, such as the real part of the imaginary part of the stroke plus stroke, here, for simplicity be described, being an example will be described in two.

Situation shown in Fig. 5 a can be referred to as " uncertainty of empty stroke/real stroke ".This also is one of problem to be solved by this invention.

For this situation, the present invention provides the scheme of two kinds of solutions.

First kind of mode is in the HMM topological structure, parastate to be provided, and both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

Shown in figure 5b, it shows a kind of structural drawing with HMM topological structure of parastate according to one embodiment of the invention.Wherein, the state 510 between

state

506 and 508 has a parallel state 510b, and state 506 can be transferred to any one among state 510 or the 510b, and any one among state 510 or the 510b also can be transferred to state 508.

Need to prove; Can being used in combination of parastate proposed by the invention with the HMM topological structure that is proposed before with persistent state and corner state; Such as, any one persistent state or corner state in the HMM topological structure can have parallel state.

The notion of above-mentioned parastate also can be applied in the multipath HMM topological structure, that is, in any paths of multipath HMM topological structure, the state that all has or several has parastate.

Fig. 5 c shows the synoptic diagram according to the constructed HMM topology mechanism with parastate of the handwriting sample of Fig. 5 a, wherein only shows the relevant portion of parastate.Among them, the last representative of the real state of 512 strokes, "Shu", continuous state 518 represents real stroke "a."For shown in Fig. 5 a maybe be for real stroke also possibly be the part of empty stroke, in the HMM topological structure shown in Fig. 5 c, adopt the real stroke 502 among the persistent state 520 representative graph 5a, with the empty stroke 504 among the persistent state 520b representative graph 5b.In the embodiment shown in Fig. 5 c, corner state 514 and 516 has parallel a corner state 514b and a 516b equally.It is relevant with its pairing corner residing position in writing whether the corner state has a parastate; If the position of corner is more near possibly also being the part of empty stroke for real stroke; The pairing corner state of this corner also can have a parastate so; Accordingly, if the position of corner is away from maybe be for real stroke also possibly be the part of empty stroke, the pairing transfering state of this corner does not just have parastate so.Usually, in the HMM topological structure of the for example situation shown in Fig. 5 c, corner state 514 and 516 has a parastate 514b and 516b separately.In other embodiment, possibly be 516 to have parallel state and 514 do not have, perhaps 514 have parastate and 516 do not have.Certainly, the present invention does not get rid of the situation that two corner states all do not have parastate yet.

A kind of way of realization that parastate is provided is to adopt double gauss mixture model (GMM), promptly utilizes the distribution character of Gaussian function, makes real stroke and empty stroke distinguish the peak value of corresponding different Gaussian functions.

Second kind solves real stroke/probabilistic method of empty stroke is that the HMM topological structure is used many spatial probability distribution (MSD), both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.MSD is a kind of algorithm commonly used; Therefore do not specifically describe its detailed principle in the present invention, MSD is applied on the HMM topological structure, can realize according to following mode: for each HMM topological structure; Define two spaces, respectively corresponding real stroke and empty stroke.Can in two spaces, carry out the calculating of many spatial probability distribution to uncertain part respectively through MSD; Obtain first metric of corresponding real stroke and second metric of corresponding empty stroke, solve real stroke/probabilistic problem of empty stroke through processing for metric.

Like this, the present invention provides a kind of technology of identification effectively to the principal feature of hand-written East Asia character, can adapt to the characteristic of hand-written East Asia character well.But, a problem that problem is the model data amount that also faces.Just as described above, for each hand-written East Asia character, the HMM topological structure of a multipath can be provided; And; In each paths, all can provide several states to come corresponding each stroke, " corner " state also need be provided between stroke; For there being real stroke/probabilistic part of empty stroke, parastate also need being provided or using the MSD technology.These all will cause the data volume of HMM model very huge.In order to save data space effectively, reduce realization cost of the present invention, also need consider the problem of data compression.

Though East Asia character similarity each other is not clearly, if East Asia character is divided into several parts, these local structures still have many similarities.Utilize this characteristics, can realize the cluster of state in the HMM topological structure, thereby reduce data volume, reduce the complexity of HMM topological structure.

According to the present invention; After making up the HMM topological structure, also can carry out cluster to the state in the HMM topology topological structure; Make at least one group of state common parameter in the HMM topology topological structure, and, in the HMM topological structure, only preserve set of parameter for the state of one group of common parameter.

The cluster of state possibly carried out between the state on the same path, also possibly between the state on the different paths, carry out, be considered to can cluster state, will only be all these states reservation set of parameters in the HMM topological structure.

So just can reduce data volume effectively, reduce the complexity of HMM topological structure.

Be suitable for discerning the implementation method of the HMM topological structure of hand-written East Asia character

For the top HMM topological structure of introducing, can realize according to following described method:

1) produces the radical training data

Before carrying out the HMM training process, need carry out mark to data, for East Asia character, data normally adopt character rather than radical to carry out mark.And for recognition methods of the present invention, what more need utilize in the HMM topological structure of foundation is the data of being carried out mark by radical.An additional effect of Viterbi (Viterbi) decoding just can provide the corresponding relation between writing sample and the HMM state.Therefore,, just can from character data, obtain handwriting data automatically corresponding to radical by Veterbi decoding, such as, the separation with reference to corresponding relation in the HMM model of corresponding radical decomposes writing sample.

Obtain etymon data in order to do above-mentioned cutting, at first will obtain a preliminary HMM model.The very accurate recognition hand-written character though this model possibly differ surely, the separation that obtains radical that can be relatively accurate.This preliminary HMM model begins to make up from single-pathway, and little by little is partitioned into increasing path till number of path is enough, and whether number of path enough adopts the convergence in path to measure is weighed.An example of this method is following:

A) initialization one has the HMM (n=1) of single-pathway;

B), calculate their convergence tolerance C (P1), C (P2) ... C (Pn), and select the maximum path P j of wherein convergence tolerance C (Pj) for the n paths P1, the P2...Pn that have existed among the HMM;

C) if C (Pi) T, T explains then for present training data that for predetermined convergence threshold value the number of path in this HMM topological structure is abundant, needn't continue split path;

D) if C (Pj)>T, then path P j is duplicated, make up a new path behind the increase noise, at this moment, had the n+1 paths in the HMM model;

E) on the basis of n+1 paths, carry out the training of HMM model, till can't improving accuracy of identification again; So far, obtain preliminary multipath HMM model, character data has been done radical alignment and cutting, obtained relatively correct etymon data.The etymon data that obtains will be used to train the HMM model of the topological structure optimization that obtains discerning hand-written character by the step of back.

2) the optimal path number is confirmed

Bright according to this law, also provide a kind of statistical nature to solve the problem that the optimal path number is confirmed, this characteristic is referred to as " subsequence direction histogram vector ".

For each writing sample, be divided into several segmentations by arc length, a characteristic is extracted in each segmentation respectively.Wherein, said here arc length is meant all stroke length sums, is exactly the length sum that all real strokes add all empty strokes, and the process of segmentation is based on the arc length after this summation.The shape facility of this segmentation of characteristic wherein.Therefore, after the operation through segmentation and extraction characteristic, can obtain a kind of shape facility of a writing sample in each segmentation.Characteristic in each segmentation is linked in sequence, thereby obtains a characteristic of this writing sample.

Shape facility in each segmentation of writing sample can be implemented as subsequence direction histogram vector.The segmentation of writing sample further is divided into several sub-section, and each sub-section is confirmed the direction character of a quantification, the direction of the corresponding predetermined angular range of the direction character of each quantification.Shape facility in each segmentation is a subsequence direction histogram, each writing sample be characterized as subsequence direction histogram vector.

Shown in figure 6a, the process of setting up of subsequence direction histogram has been described.At first, writing sample " king " 602 has been divided into several segmentations according to arc length, and wherein this arc length is the total arc length 604 after real stroke links to each other with empty stroke; In these segmentations, there are some to include only real stroke, such as segmentation 604a; There are some to include only empty stroke,, also have some existing real strokes such as segmentation 604c; Empty stroke is also arranged, such as segmentation 604b.With segmentation 604b is example, and it can be divided into several word segmentations again, for this a little segmentation, has obtained the subsequence direction histogram according to their direction (reference direction pointer 606) and stroke density.Such as, the corresponding subsequence direction histogram of a segmentation 605a of the empty stroke part of 604b is 605b, the corresponding subsequence direction histogram of a segmentation 607a of the real stroke part of 604b is 607b.The subsequence direction histogram of all sub-segmentations is combined; Just can obtain subsequence direction histogram corresponding to each segmentation; Further the subsequence direction histogram with each segmentation makes up; Just can obtain the subsequence direction histogram vector of corresponding writing sample, such as 608 among Fig. 6 a.

After having obtained subsequence direction histogram vector, just can easily carry out cluster to above-mentioned characteristic, each cluster is corresponding to a kind of order of strokes or a kind of handwriting style.The operation of cluster can be through realizing such as gauss hybrid models (Guassian Mixture Model).

Finally, with the result after m cluster of acquisition, m also is the number of paths corresponding to the optimization of a hand-written East Asia character.Then, confirm parameter based on the data of each cluster corresponding to the path in the HMM topological structure of each order of strokes or each handwriting style.

With reference to figure 6b, Fig. 6 b carries out training data to classify automatically, passes through the instance of cluster operation more precisely.Among Fig. 6 b, 9 sample datas are provided altogether from training data.Through data automatic classification, they are clustered into 2 typical type, represent two " answering " words through the writing style after the clustering processing.

3) optimum state quantity confirms

After the quantity of having confirmed the path, also need confirm amount of state in the path, according to the present invention, still adopt from the method for the machine self study of data and confirm amount of state.When confirming a amount of state in the path, the data of employing are from above-mentioned step 2) in writing sample in same type.What the state among the HMM reflected is the form and the variation of writing sample, and therefore, state comprises two types: change the part of not obvious (flexibility is little) for direction, be referred to as " continuing stroke ", " persistent state " among the corresponding HMM; Changing the obviously part of (flexibility greatly) for direction, is exactly " corner ", corresponding to " corner state ".Amount of state is exactly the quantity sum of " continuing stroke " and " corner ", just the quantity sum of the quantity of " persistent state " and " corner state " in the path.The curvature scale space algorithm that becomes more meticulous gradually (coarse-to-fineCurvature Scale Space) can be used to carry out the detection of corner; Thereby confirm the quantity of corner; Simultaneously, because corner and lasting stroke are alternately to occur, can extrapolate the quantity that continues stroke from the quantity of corner again; Like this, amount of state just can be determined.

4) state connects design

Like the problem of foregoing " uncertainty of empty stroke/real stroke ", the present invention provides the scheme of two kinds of solutions.

First kind of mode is in the HMM topological structure, parastate to be provided, and both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.A kind of way of realization that parastate is provided is to adopt double gauss mixture model (GMM); Promptly utilize the distribution character of Gaussian function; Make real stroke and empty stroke distinguish the peak value of corresponding different Gaussian functions; Utilize such double gauss mixture model, just can solve real stroke/probabilistic problem of empty stroke.With reference to figure 6c,, there is Lian Bihe not connect two kinds of literary styles of pen for " wood " word; Connect the situation of the literary style corresponding " real stroke " of pen, and do not connect the situation of the literary style corresponding " empty stroke " of pen, with reference to figure 6c; The Gaussian function that real stroke and empty stroke have respectively separately distributes, and their peak value is different, therefore their combinations can be obtained the double gauss mixture model; Have two different peak value, thereby realize parastate.

Second method is used many spatial probability distribution (MSD) to the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

5) path merges

After above-mentioned step is accomplished; Obtained the HMM topological structure of preliminary multipath; Comprising corresponding to the path of all situations in the training data, comprise the path of reflection order of strokes and handwriting style, still; Because the information of redundant information that exists in the training data and extremely low probability need merge processing to the path that is obtained.The merging in path is handled can comprise two aspects: remove in the path that probability of occurrence is extremely low; And, similar path is merged.

The path that probability of occurrence is extremely low, just the data volume in the pairing training data in this path is less, and the ratio that accounts for the training data total amount is low to be removed, and removes these paths and can realize through the method that predetermined threshold is set.

And similarly the operation of path merging can be represented the similarity degree between the path through measuring similarity through for the mode of path computing measuring similarity realizes, when two paths are enough similar, just merges the path.Such as, measuring similarity can represent that the Kullback-Leibler difference is represented degree similar between the path with the Kullback-Leibler difference.If the Kullback-Leibler difference is lower than a predetermined value; Just represent that two paths are enough similar; Then can merge them, union operation will carry out the processing of equilibrating to two paths, to obtain the path that can reflect the principal feature in original two paths well.

6) state clustering

Afterwards, also need carry out cluster to state in the HMM topological structure.If East Asia character is divided into several parts, these local structures still have many similarities.Utilize this characteristics, can realize the cluster of state in the HMM topological structure, thereby reduce data volume, reduce the complexity of HMM topological structure.The cluster of state possibly carried out between the state on the same path, also possibly between the state on the different paths, carry out, be considered to can cluster state, will only belong to states reservation set of parameter of same cluster in the HMM topological structure for all.

According to the present invention, cluster operation comprises: calculate the state measuring similarity between per two states in this HMM topological structure, when the state measuring similarity representes that two states are enough similar, make this two state common parameters; Wherein, when plural state state measuring similarity each other is all enough similar, make all common parameters of these these states.In a realization of the present invention, the state measuring similarity representes that through the Kullback-Leibler difference Kullback-Leibler difference is lower than a predetermined value, representes that then these two states are enough similar.Process such as a concrete cluster operation is shown as follows:

Initial phase:

That sets all states adds up to M, m state is referred in m the class (1≤m≤M);

For two state m and n arbitrarily, calculate the Kullback-Leibler difference between them, and be expressed as D (m, n).

The cluster stage:

Searching has one group of state (m ', n ') of minimum Kullback-Leibler difference, can realize through following function: (m ', n ')=argmin (m, n).

Through in Kullback-Le ibler matrix of differences, merging m ' and n ' state corresponding to row and the row addition of m ' and n '.

The sum of state M is subtracted 1.

Cycling:

If the quantity of M greater than a predetermined value, then repeats top operation, otherwise accomplish union operation.

Through top step 1)-5); Just can from the training data of hand-written East Asia character, make up multipath HMM topological structure automatically corresponding to hand-written East Asia character; In the multiple order of strokes of the corresponding hand-written East Asia character of each paths wherein one; Perhaps, in wherein the multiple handwriting style of the corresponding hand-written East Asia character of each paths;

In the HMM topological structure, provide a description the persistent state that hand-written East Asia character continues stroke, and describe the corner state of corner between the hand-written East Asia character stroke;

Wherein, each paths in the multipath HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; Wherein persistent state can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively in the HMM topological structure; And wherein all paths originate in same entry state, to same discharge state end.

Possible hardware way of realization

The present invention can realize through the form of software, such as the software of realizing method of the present invention through general-purpose computing system operation, just can realize the present invention.The present invention also can be implemented with the form of instruction or program, and these instructions or program can be kept on the storage medium, when a computing equipment obtains these instructions or program and carries out from storage medium after, just can realize the present invention.

In addition, the present invention also can use the form of hardware to realize, need to prove; For a person skilled in the art, in the residing field of the present invention, the conversion between the software and hardware has various ways obviously; Promptly have multi-form hardware and can realize identical functions; Therefore, cited possible hardware way of realization is the function that limits hardware below the present invention, and does not limit its concrete way of realization; For a person skilled in the art, realize that according to these functions various forms of functions are to show and suggestion.

With reference to shown in Figure 7, the system 700 that utilizes HMM Model Identification hand-written East Asia character of an example of the present invention comprises:

HMM topological structure construction device 702 makes up the HMM topological structure corresponding to hand-written East Asia character;

Persistent state setting device 704 provides a description the persistent state that hand-written East Asia character continues stroke in the HMM topological structure;

Corner state setting device 706 provides a description the corner state of corner between the hand-written East Asia character stroke in the HMM topological structure.

Wherein, the HMM topological structure that HMM topological structure construction device 702 makes up is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; And the persistent state in the HMM topological structure can be transferred to NextState or from shifting, the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively.

In another embodiment, this HMM topological structure construction device 702 makes up multipath HMM topological structure, corresponding to hand-written East Asia character; In the multiple order of strokes of the corresponding hand-written East Asia character of each paths wherein one; Perhaps, in wherein the multiple handwriting style of the corresponding hand-written East Asia character of each paths.

Wherein, each paths in the multipath HMM topological structure that HMM topological structure construction device 702 makes up is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; And wherein persistent state can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively in the HMM topological structure; And wherein all paths originate in same entry state, to same discharge state end.

In one embodiment, this system 700 also comprises parastate generator 708, in the HMM topological structure, parastate is provided, and both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.This parastate generator 708 can be used the double gauss mixture model and realize parastate.

Perhaps, comprising many spatial probability distribution (MSD) treating apparatus 710, the HMM topological structure is used many spatial probability distribution, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

Need to prove that parastate generator 708 and many spatial probability distribution treating apparatus 710 are two to select one of which.

In one embodiment; This system 700 also comprises state clustering device 712, and the state in the HMM topological structure is carried out cluster, makes at least one group of state common parameter in the HMM topological structure; And, in the HMM topological structure, only preserve set of parameter for the state of said one group of common parameter.

Need to prove; Each device of the system 700 of the HMM of utilization Model Identification hand-written East Asia character described herein combines the described method of accompanying drawing 2-5 above can be used for realizing; Wherein each minutia is all corresponding, therefore just no longer repeatedly describes here.

Fig. 8 shows the structural drawing of the system 800 that utilizes HMM Model Identification hand-written East Asia character according to another embodiment of the present invention, and this system 800 comprises:

HMM topological structure construction device 802; Structure is corresponding to the multipath HMM topological structure of hand-written East Asia character; In the multiple order of strokes of the corresponding hand-written East Asia character of each paths wherein one; Perhaps, in wherein the multiple handwriting style of the corresponding hand-written East Asia character of each paths;

Persistent state setting device 804 is provided with in the HMM topological structure and describes the persistent state that hand-written East Asia character continues stroke;

Corner state setting device 806, the corner state of corner between setting and description hand-written East Asia character stroke in the HMM topological structure;

Wherein, each paths in the multipath HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; Persistent state can be transferred to NextState or from shifting, the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively; And all paths originate in same entry state, to same discharge state end; And

One of them of following two devices:

Parastate generator 808 provides parastate in the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke; Parastate generator 808 can adopt double gauss mixture model (GMM); Promptly utilize the distribution character of Gaussian function; Make real stroke and empty stroke distinguish the peak value of corresponding different Gaussian functions; Utilize such double gauss mixture model, also can solve real stroke/probabilistic problem of empty stroke.

Many spatial probability distribution treating apparatus 810 is used many spatial probability distribution to the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

According to system shown in Figure 8 800, HMM topological structure construction device 802 wherein makes up multipath HMM topological structure automatically from the training data of hand-written East Asia character; And this HMM topological structure construction device 802 is used the automatic classification method of a machine self study training data is classified according to the order of strokes or the writing style of hand-written East Asia character.

Continuation is with reference to figure 8, and the training data that this system 800 adopts comprises the hand-written East Asia character writing sample, and writing sample comprises the writing sample of different order of strokes or different handwriting styles; This HMM topological structure construction device 802 comprises,

Writing sample sectioning 820 for each writing sample, is divided into several segmentations by arc length; Here said arc length is meant all stroke length sums, is exactly the length sum that all real strokes add all empty strokes, and the process of segmentation is based on the arc length after this summation.

Feature deriving means 822 extracts a characteristic respectively to each segmentation; The shape facility of this segmentation of characteristic wherein.Therefore, after the operation through segmentation and extraction characteristic, can obtain a kind of shape facility of a writing sample in each segmentation.Characteristic in each segmentation is linked in sequence, thereby obtains a characteristic of this writing sample.

Clustering apparatus 824 carries out cluster to above-mentioned characteristic, and each cluster is corresponding to a kind of order of strokes or a kind of handwriting style;

HMM topological structure construction device 802 is confirmed the parameter corresponding to the path in the HMM topological structure of each order of strokes or each handwriting style based on the data of each cluster.Further, according to an embodiment, HMM topological structure construction device 802 also can comprise subsequence direction histogram vector apparatus for establishing 826, and the shape facility in each segmentation of above-mentioned writing sample is embodied as subsequence direction histogram vector.Wherein, the writing sample sectioning links to each other real stroke for each writing sample with empty stroke, and real stroke is divided into several sections with arc length after empty stroke links to each other;

And being characterized as the direction of quantification, feature deriving means 822 makes the direction of the corresponding predetermined angular range of each characteristic;

This subsequence direction histogram vector apparatus for establishing 826 is set up the subsequence direction histogram vector corresponding to segmentation and characteristic; This subsequence direction histogram vector apparatus for establishing 826 further is divided into several sub-section with the segmentation of writing sample; Each sub-section is confirmed the direction character of a quantification, the direction of the corresponding predetermined angular range of the direction character of each quantification.Shape facility in each segmentation is a subsequence direction histogram, each writing sample be characterized as subsequence direction histogram vector.

Continuation is with reference to figure 8, and in one embodiment, this system 800 also comprises

The path merges device 814, the path in the multipath HMM topological structure is merged, with the quantity in control path.This path merges the data volume of the pairing training data in device 814 judgement paths, and the data volume of deletion correspondence is less than the path of a predetermined value; Calculate the path measuring similarity between per two paths in the multipath HMM topological structure; When the path measuring similarity representes that two states are enough similar, merge this two paths, wherein; When the path in plural path measuring similarity is all enough similar, merge these paths.In a realization, this path measuring similarity representes that with the Kullback-Leibler difference when the Kullback-Leibler difference was lower than a predetermined value, the expression path was enough similar, can merge these paths.

This system 800 also can comprise state clustering device 812, and state in the HMM topological structure is carried out cluster, makes at least one group of state common parameter in the HMM topological structure, and for the state of one group of common parameter, in the HMM topological structure, only preserves set of parameter.State clustering device 812 calculates the state measuring similarity between per two states in this HMM topological structure, when the state measuring similarity representes that two states are enough similar, makes this two state common parameters; Wherein, when plural state state measuring similarity each other is all enough similar, make all common parameters of these these states.And state clustering device 812 states are represented measuring similarity through the Kullback-Leibler difference, when the Kullback-Leibler difference is lower than a predetermined value, represent that these states are enough similar, can merge these states.

Need to prove; Utilize each device of the system 800 of HMM Model Identification hand-written East Asia character to can be used for combining accompanying drawing 6 described methods above the realization here; Comprise top described step 1)-5) make up multipath HMM topological structure; Wherein each minutia is all corresponding, therefore just no longer repeatedly describes here.

The characteristics that the present invention is directed to hand-written East Asia character provide a kind of and have utilized improved HMM model to come the scheme that hand-written East Asia character is discerned; Taken into full account that the hand-written East Asia character stroke is many, order of strokes is various, complex structure, writing style are various, connect uncertain characteristics between stroke; The characteristics of HMM topological structure have been improved; Through in the HMM topological structure, adding the corner state, multipath is provided, provides the means of parastate to solve above-mentioned problem; And, reduce data volume through the operation of cluster and merging, reduce the complexity of computing.

Combine one embodiment of the invention that the present invention has been carried out detailed description above; But need to prove that this is not to make any restriction for scope of the present invention, for the various variations of having done that do not need creative work of top said embodiment; Revise; All should be considered to be within scope of the present invention, for the present invention, should expand to the most wide in range scope that meets the inventive principle that claim limits.

Claims

1. a foundation is applicable to the method for the hidden markov models HMM of identification hand-written East Asia character, it is characterized in that,

Multipath HMM is provided topological structure; Corresponding to hand-written East Asia character; Each paths in the wherein said multipath HMM topological structure is built as in the multiple order of strokes of hand-written East Asia character, and each paths in the perhaps said multipath HMM topological structure is built as in the multiple handwriting style of hand-written East Asia character;

In the HMM topological structure, persistent state is provided, this persistent state is used to describe the lasting stroke of hand-written East Asia character;

The corner state is provided in the HMM topological structure, and this corner state is used to describe corner between the stroke of hand-written East Asia character.

2. the method for claim 1 is characterized in that,

Constructed HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state;

Persistent state in the said HMM topological structure is set to be transferred to NextState or from shifting, the corner state is set to be transferred to NextState, can not be from shifting;

Persistent state and corner state are set to alternately exist successively in the said HMM topological structure.

3. the method for claim 1 is characterized in that,

Each paths in the said multipath HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state;

Said HMM topological structure comprises persistent state and corner state, and wherein persistent state can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting;

Persistent state and corner state alternately exist successively in the said HMM topological structure; And

Wherein all paths originate in same entry state, to same discharge state end.

4. the method for claim 1 is characterized in that, also comprises:

In the HMM topological structure, parastate being provided, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

5. the method for claim 1 is characterized in that, also comprises:

The HMM topological structure is used many spatial probability distribution MSD, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

6. the method for claim 1 is characterized in that, also comprises:

State in the HMM topological structure is carried out cluster, make at least one group of state common parameter in the said HMM topological structure, and, in the HMM topological structure, only preserve set of parameter for the state of said one group of common parameter.

7. a method of utilizing hidden markov models HMM identification hand-written East Asia character is characterized in that,

Multipath HMM topological structure corresponding to hand-written East Asia character is provided, in the multiple order of strokes of the corresponding hand-written East Asia character of each paths wherein one, perhaps, in the multiple handwriting style of the corresponding hand-written East Asia character of each paths wherein one;

Wherein, each paths in the said multipath HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; Wherein persistent state can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively in the said HMM topological structure; And wherein all paths originate in same entry state, to same discharge state end;

In the HMM topological structure, parastate being provided, perhaps the HMM topological structure being used many spatial probability distribution MSD, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

8. method as claimed in claim 7 is characterized in that,

Said multipath HMM topological structure makes up from the training data of hand-written East Asia character automatically;

According to the order of strokes or the writing style of hand-written East Asia character, use the automatic classification method of a machine self study said training data is classified.

9. method as claimed in claim 8 is characterized in that,

Said training data comprises the hand-written East Asia character writing sample, and said writing sample comprises the writing sample of different order of strokes or different handwriting styles;

For each writing sample, be divided into several segmentations by arc length, a characteristic is extracted in each segmentation respectively, the characteristic in each segmentation is linked in sequence, thereby obtains a characteristic of this writing sample;

Above-mentioned characteristic is carried out cluster, and each cluster is corresponding to a kind of order of strokes or a kind of handwriting style;

Data based on each cluster are confirmed the topological sum initial parameter corresponding to the path in the HMM topological structure of each order of strokes or each handwriting style.

10. method as claimed in claim 9 is characterized in that, also comprises:

Foundation is vectorial corresponding to the subsequence direction histogram of segmentation and characteristic,

Wherein,, real stroke is linked to each other with empty stroke, and real stroke is divided into several sections with arc length after empty stroke links to each other for each writing sample,

Each sub-section is confirmed the direction character of a quantification, the direction of the corresponding predetermined angular range of the direction character of each quantification;

Shape facility in each segmentation is a subsequence direction histogram, each writing sample be characterized as subsequence direction histogram vector.

11. method as claimed in claim 8 is characterized in that, also comprises:

Path in the multipath HMM topological structure merges, with the quantity in control path.

12. method as claimed in claim 11 is characterized in that, said path merges and comprises:

Judge the data volume of the pairing training data in path, the data volume of deletion correspondence is less than the path of a predetermined value;

Calculate the path measuring similarity between per two paths in the multipath HMM topological structure; When the path measuring similarity representes that two paths are enough similar; Merge this two paths; Wherein, when two path measuring similarities with upper pathway represent that these paths are all enough similar, merge these paths.

13. method as claimed in claim 12 is characterized in that, this path measuring similarity uses the Kullback-Leibler difference to represent, when

When the Kullback-Leibler difference is lower than a predetermined value, represent that said path is enough similar.

14. method as claimed in claim 11 is characterized in that, also comprises:

15. method as claimed in claim 14 is characterized in that, the state in a plurality of HMM topological structures of cluster comprises:

Calculate the state measuring similarity between per two states in this HMM topological structure, when the state measuring similarity representes that two states are enough similar, make this two state common parameters;

Wherein, when plural state state measuring similarity each other is all enough similar, make all common parameters of these these states.

16. method as claimed in claim 15 is characterized in that,

The state measuring similarity is represented through the Kullback-Leibler difference, when the Kullback-Leibler difference is lower than a predetermined value, representes that said state is enough similar.

17. a foundation is applicable to the system of the hidden markov models HMM of identification hand-written East Asia character, it is characterized in that, comprising:

HMM topological structure construction device makes up the HMM topological structure corresponding to hand-written East Asia character, and wherein said HMM topological structure construction device makes up multipath HMM topological structure, corresponding to hand-written East Asia character; In the multiple order of strokes of the corresponding hand-written East Asia character of each paths wherein one; Perhaps in wherein the multiple handwriting style of the corresponding hand-written East Asia character of each paths;

The persistent state setting device provides a description the persistent state that hand-written East Asia character continues stroke in the HMM topological structure;

Corner state setting device provides a description the corner state of corner between the hand-written East Asia character stroke in the HMM topological structure.

18. system as claimed in claim 17 is characterized in that,

The HMM topological structure that HMM topological structure construction device makes up is the right HMM topological structure of left-hand, since an initial state, finishes to a final state;

And the persistent state in the HMM topological structure is set to be transferred to NextState or from shifting, the corner state is set to be transferred to NextState, can not be from shifting;

19. system as claimed in claim 17 is characterized in that,

Each paths in the multipath HMM topological structure that HMM topological structure construction device makes up is the right HMM topological structure of left-hand, since an initial state, finishes to a final state;

And wherein persistent state can be transferred to NextState or shift certainly, and the corner state can only be transferred to NextState, can not be from shifting;

Wherein all paths originate in same entry state, to same discharge state end.

20. system as claimed in claim 17 is characterized in that, also comprises:

The parastate generator provides parastate in the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

21. system as claimed in claim 17 is characterized in that, also comprises:

Many spatial probability distribution MSD treating apparatus is used many spatial probability distribution to the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

22. system as claimed in claim 17 is characterized in that, also comprises:

The state clustering device carries out cluster to the state in the HMM topological structure, makes at least one group of state common parameter in the said HMM topological structure, and for the state of said one group of common parameter, in the HMM topological structure, only preserves set of parameter.

23. a system that utilizes hidden markov models HMM identification hand-written East Asia character is characterized in that, comprising:

HMM topological structure construction device; Structure is corresponding to the multipath HMM topological structure of hand-written East Asia character; In the multiple order of strokes of the corresponding hand-written East Asia character of each paths wherein one, perhaps, in the multiple handwriting style of the corresponding hand-written East Asia character of each paths wherein one;

The persistent state setting device is provided with in the HMM topological structure and describes the persistent state that hand-written East Asia character continues stroke;

Corner state setting device, the corner state of corner between setting and description hand-written East Asia character stroke in the HMM topological structure;

Wherein, each paths in the said multipath HMM topological structure is the right HMM topological structure of left-hand, since an initial state, finishes to a final state; Persistent state can be transferred to NextState or from shifting, the corner state can only be transferred to NextState, can not be from shifting; Persistent state and corner state alternately exist successively; And all paths originate in same entry state, to same discharge state end; And

One of them of following two devices:

The parastate generator provides parastate in the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke;

Many spatial probability distribution treating apparatus is used many spatial probability distribution to the HMM topological structure, both possibly be empty stroke in the corresponding hand-written East Asia character, also possibly be the part of real stroke.

24. system as claimed in claim 23 is characterized in that,

HMM topological structure construction device makes up multipath HMM topological structure automatically from the training data of hand-written East Asia character; And

This HMM topological structure construction device is used the automatic classification method of a machine self study said training data is classified according to the order of strokes or the writing style of hand-written East Asia character.

25. system as claimed in claim 24 is characterized in that,

Said training data comprises the hand-written East Asia character writing sample, and said writing sample comprises the writing sample of different order of strokes or different handwriting styles; This HMM topological structure construction device comprises,

The writing sample sectioning for each writing sample, is divided into several segmentations by arc length, the characteristic in each segmentation is linked in sequence, thereby obtains a characteristic of this writing sample;

Feature deriving means extracts a characteristic respectively to each segmentation;

Clustering apparatus carries out cluster to above-mentioned segmentation and characteristic, and each cluster is corresponding to a kind of order of strokes or a kind of handwriting style;

HMM topological structure construction device is confirmed the topological sum initial parameter corresponding to the path in the HMM topological structure of each order of strokes or each handwriting style based on the data of each cluster.

26. system as claimed in claim 25 is characterized in that, also comprises:

Subsequence direction histogram vector apparatus for establishing,

Wherein, the writing sample sectioning links to each other real stroke for each writing sample with empty stroke, and real stroke is divided into several sections with arc length after empty stroke links to each other;

The shape facility that this subsequence direction histogram vector apparatus for establishing is set up in each segmentation is a subsequence direction histogram, each writing sample be characterized as subsequence direction histogram vector.

27. system as claimed in claim 23 is characterized in that, also comprises:

The path merges device, the path in the multipath HMM topological structure is merged, with the quantity in control path.

28. system as claimed in claim 27 is characterized in that, this path merges device

29. system as claimed in claim 28 is characterized in that, this path merges the device path and representes measuring similarity with the Kullback-Leibler difference, when the Kullback-Leibler difference is lower than a predetermined value, representes that said path is enough similar.

30. system as claimed in claim 27 is characterized in that, also comprises:

The state clustering device carries out cluster to state in the HMM topological structure, makes at least one group of state common parameter in the said HMM topological structure, and for the state of said one group of common parameter, in the HMM topological structure, only preserves set of parameter.

31. system as claimed in claim 30 is characterized in that,

The state clustering device calculates the state measuring similarity between per two states in this HMM topological structure, when the state measuring similarity representes that two states are enough similar, makes this two state common parameters; Wherein, when plural state state measuring similarity each other is all enough similar, make all common parameters of these these states.

32. system as claimed in claim 31 is characterized in that,

The state clustering device is represented the state measuring similarity through the Kullback-Leibler difference, when the Kullback-Leibler difference is lower than a predetermined value, representes that said state is enough similar.