TW200937308A - Method and apparatus for constructing HMM models for recognizing online eastern asian characters - Google Patents

Abstract

The present invention provides a scheme for designing a HMM model for recognizing online East-Asian characters and a method for constructing such a HMM model. With full consideration of the features of East-Asian characters such as the large number of strokes, the diversity of writing order, the complexity of structure, the diversity of writing styles, and the uncertainty between strokes, the present invention addresses the above issues by introducing turning states, providing multi-paths and parallel state in HMM topology. The present invention also reduces data quantity and operation complexity by clustering and merging.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a handwriting recognition technique, and more particularly to a recognition technique for writing East Asian characters.

V ^ [Prior Art] Handwriting input recognition has always been an important research direction in computer application technology. Unlike static image or text recognition, online handwriting input recognition is a time-random process, not a static object. Therefore, in the field of online handwriting recognition, Hidden Markov (HMM) models are often used. First, a brief introduction to the HMM model, the HMM model is a recessive Markov model, which is a kind of Markov model. The Markov model can be used to predict the probability that a certain event will occur in a future state, and this probability will be based only on the current state of the event. The Markov model behaves as a finite state automaton, states can be converted to each other, and each time ¥ is switched from one state to the next (may be other states, it may be the state itself). - The Markov model can be divided into two types, one called the dominant Markov model. In the dominant Markov model, the order of transition between states is known. The other is called the hidden Markov model (HMM), in which the order of transition between states is unknown, and what is known is only the transition between states. Therefore, the HMM model can be defined as having the following characteristics: 1) is a finite state automaton, states can be converted to each other, and each time from one state to the next state (may be other states, also It may be the state itself); 2) The transition between states is determined by a set of transition probabilities, and the probability of occurrence of a set of observed events (observation sequences) is determined by the state-dependent transition probabilities.

For example, give an example to illustrate the HMM model. There is a beverage vending machine offering two drinks, cola and tea. This beverage vending opportunity has two states, the status of "preferable cola" and the state of "preference tea". When investing in a grazing coin to purchase a beverage, the vending opportunity determines the beverage to be sold according to its state, that is, Cola will be sold in the "preferable cola" state, while tea will be sold in the "preference tea" state. The vending machine converts between the two states after each sale, and the probability of conversion is as follows: When in the "Preference Coke" state, the probability of transitioning to the "Preference Tea" state is 30%, remaining in "Preference Coke" The probability of the state is 70%; out of the state of "preference tea", the probability of switching to the "preferable cola" state is 50%, and the probability of staying in the "preference tea" state is also 50%. For the above-described beverage vending machine, if it is desired to determine the probability of a particular one of the beverage sequences being sold, an HMM model can be established for the beverage vending machine described above. The model can be referred to Figure 1 b, which includes two states: "Preference cola" status and "Preference Tea" status. According to the state transition probability listed above, the probability of each transition path is determined. The lb diagram shows. When it is necessary to determine a particular sale sequence, it is possible to obtain the probability of occurrence of 14 a particular sale sequence by calculating the probability of all the paths for which the above results will occur in 200937308 and summing them. The HMM topology reflects the transfer order connection between states in the hmm model. The application of the HMM model and the HMM topology involves three main types of problems: 1) When an HMM model is given, the probability of occurrence of a particular “viewing event (corresponding to a state transition order) is determined. 2) Given a hmm model and a specific "observation event", select a state transition sequence that best describes the particular "observation event". 3) When a specific "observation event" is given and a set of possible HMM model spaces, the best hmm model is determined to describe the "observation event". The above characteristics of the HMM model make it particularly useful when solving the following problems: the probability of occurrence of a potential event affects the observed event. A typical application is in the field of recognition, especially speech recognition, including speech recognition technology and handwriting recognition technology. The HMM model is a trainable model. Through a large amount of data training, it is possible to obtain a group that is most suitable for representing "observation matters." The HMM model parameters of the V piece, after which the HMM model obtained from the data training can be used to determine the degree of matching between the other “observation event” and the model. According to the matching degree, an unknown “observation event” can be determined. "Which model is most likely to belong to this, it achieves the purpose of identification. 8 200937308 At present, although a variety of identification techniques using HMM models have been developed, they are mostly used in Western languages, but for Eastern languages, especially those used in East Asia, such as Chinese characters, they cannot be effectively identified. This is mainly due to Western languages such as English and East Asian characters, such as the difference in the glyph structure of Chinese characters.

For Western languages, each word is composed of letters, and the individual letters are simple in structure, usually one can be completed, there is no problem of variable stroke order, and the similarity between letters and letters is relative. Lower, except for individual letters, most letters have their own distinct characteristics. These features bring a lot of convenience to handwriting recognition. Therefore, the HMM models developed with Western language as the main identification object generally have left-to-right HMM topologies. The characteristics of general western letters can be described by defining the starting and ending points of the HMM topology. However, East Asian characters have distinct characteristics. Take Chinese characters as an example: 1) There are many strokes, and there are different stroke sequences in the handwriting process; 2) The structure is complicated, which makes the fonts change more, which is reflected in the handwriting. Handwriting style; 3) There is a problem of connection between strokes. Due to the large number of Chinese strokes and complicated structure, coupled with personal handwriting habits, there will be a lot of uncertainty in the connection between strokes; 4) Large amount of data, Chinese characters Each word is an independent individual, rather than a limited number of units that can be further split into letters like English. Therefore, for Chinese characters, the amount of data in the handwriting recognition model is very large. 9 200937308 Through the above analysis, we can see that there are many obvious differences in the characteristics of East Asian characters, which leads to the current HMM model for Western language development is not applicable to East Asian characters, especially the chess type can not Effectively solve the stroke sequence, the writing style of the question provides a solution between the East Asian character strokes, and can not be well provided 2: the program's, thus leading to the current handwritten East Asian character recognition, there is no one Very good identification technology. ❹ ❹ [Summary] The object of the present invention is to provide a method for constructing a handwritten East Asian word and a method for identifying the East Asian character for the handwritten East Asian character. In one aspect of the invention, a method for designing an HMM topology suitable for handwriting East Asian characters is established in accordance with a method of writing a handwritten East Asian character stroke, a pen, and a connection with a complicated relationship. The HMM topology provides a description of the handwritten East Asian character continuous strokes and describes the corner state of the handwritten East Asian character strokes. The corner state of the corner between the strokes is entered to make the entire topology report well. The characteristics of East Asian characters. In an implementation of the moon, the HMM topology is designed as a left topology. From a starting state to a state of termination, the state of the topology can be transferred to the next state or transferred. @角角状态 can only be transferred to the lower state 'cannot be transferred; and the continuous state and the corner state alternately exist. 10 200937308 According to this issue The other _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ Corresponding to one of the Sj sequences in the handwriting; each of the poles of the structure that goes to the structure of the tenth, the path of the HMM extension and the mother of the four mothers corresponds to one of the handwritten East Asian characters. Sequence or hand integration into a ΗΜΜ topology, the writing style of the path character stroke order is diverse and handwritten to solve the problem of handwritten East Asian and handwritten styles. In the present invention - a - path is left to right path _ Every end-to-end state in the structure ends....The structure of the flap, starting from a starting state, is used to describe the handwritten East Asia ΗΜΜ ΗΜΜ topology so that the same type of East Asian characters continue to be written. State Μ and description of the corner state of the handwritten East include the holding of the strong state of the performance state can be Han # $ (4) I state and corner state, can only be transferred to the next ιΓ state or self-transfer, corner Taste, and can not self-transition; Thorens # makeup holding state is read sequentially and alternately present can corners STATUS, to the same - "shape: all paths starting from the two ends with the Taguchi ~ state. At the same time, the path in the topology of the present invention is subjected to a merge process to control the number of paths according to another state of the present invention #, for the uncertainty of the hand-to-connection transition, g, sub-70 strokes The method of 疋 即 即 即 即 即 即 即 即 即 即 即 即 即 即 即 即 即 本 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供 提供Partially, or,: 2:: can be applied to the application of multi-space probability distribution (MSD), corresponding to Α 结果 结果 亚 亚 亚 亚

200937308 can be a virtual stroke, or it may be part of a real stroke. According to another aspect of the present invention, in order to reduce the complexity of redundant low cost operations, the present invention also clusters HMM topologies such that at least one set of states in the HMM topology is $ and for a set of shared parameters State, a set of parameters in the HMM topology. The present invention also provides an implementation method of a pounce structure suitable for handwritten East Asian character recognition. The multi-path HMM topology is automatically constructed from training materials of characters; according to handwritten East Asian character order or writing style, application-machine self-learning Automatic classification and training of materials to gather m 'Each-category data corresponds to one or one writing style. —_ ... ..., 〇 〇 疋, π 训练 training data self-turning + up, _ Beiding inch visual generation without the need for artificial dry materials including handwritten East Asian characters harassing sample t70 handwriting samples, handwriting sample package Preface or different handwriting style

The handwriting sample, in the process of constructing H., may include & T ^ in the following stages: for each - according to the arc length, it is divided into spare parts, knife attack several segments, each segment will be The segmentation feature sequence is ..,, and the negative sequence is arranged into the entire handwriting sample. The integration of each cluster corresponds to a plug-in stroke order or the data of each cluster = the cluster determines the topology and initial parameters of the path. ^Style of the HMM topology; the month of the month - a real view 'also establishes the corresponding; the amount of material, the state of the drop t with parameters, only save the HMM extension handwriting East Asia stroke stroke method for the training writing sequence handwriting East Asia training笔 笔 顺 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 拓扑 ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹ ❹

Subsequence Direction Histogram Vector' to help establish the HMM In an implementation of the present invention, the paths of the multipath faces are also combined to control the number of streams. Since the amount of data written by East Asian characters is large, in order to reduce the complexity of implementation', in one implementation of the present invention, the state between every two states in the abdominal BM topology is similarly determined as the operation of clustering. The use of Erming also provides the ability to implement the above-mentioned various criteria using the H_model to identify handwritten East Asian characters. The invention provides the characteristics of handwritten East Asian characters. - East: Characters for knowledge _ topological knots, handwritten East Asia to children signing eve styles, strokes, door series: sequence diversity, phase-to-phase stroke connection uncertainty In the structure, the corner state 1 is introduced for multipath to solve the above problem, the amount of data, the complexity of the low operation, and the like. [Embodiment] The present invention is directed to the handwriting of East Asian characters, the complexity of the structure, and the writing style. 7 A name, m strokes between strokes, "Γ, Γ..._ model to handwriting (4) The text has entered the characteristics of the topology, adding a corner state to the heartbeat structure, providing multiple channels, providing a flat section is very good It reflects the above characteristics of handwritten East Asian characters: , ; flapping structure. The small amount of data in the topological structure, the system determined by the calculation degree, the establishment of the appropriate kind, the full consideration of the complex, writing and topological state means to reduce The order of the strokes is uncertain. The characters are identified by the hand in the state of the extension line, and the amount of data is reduced by the operation of the class 13 200937308 and the merge operation. The complexity of the operation. The HMM topology of the present invention is self-generated from the training data and does not require manual intervention. It should be noted that in the following embodiments, Chinese characters are taken as an example, but The scope of the present invention is not limited to Chinese characters, but all East Asian characters having similar features to Chinese characters, including kanji Chinese characters, katakana's pseudonyms, and the like.

Appropriate Implementation Environment Figure 1a illustrates an example of a suitable computing system environment 100 in which the present invention may be implemented. The computing system environment 100 is only one example of a suitable computing environment and is not intended to limit the scope of use or functionality of the present invention. The computing environment 100 should not be construed as having a dependency or requirement relating to any or combination of the elements described in the exemplary computing environment 100. Those skilled in the art will appreciate that a computer or other client or server device can be employed as part of a computer network or in a distributed computing environment. In this regard, the present invention pertains to any computer system having any number of memory or storage units, and any number of applications and processes occurring in any number of storage units or capacities that can be used with the present invention. The present invention can be applied to an environment in which a server computer and a client computer are employed in a network environment or a distributed computing environment. The present invention can also be used in stand-alone computing devices, with programming language functions, and interpretation and execution capabilities for generating, receiving, and transmitting information with remote or local services. The invention can operate in a variety of other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems that may be suitable for use with the present invention, 14 200937308 environments and/or configurations include, but are not limited to, personal computers, servos, portable or portable devices, multi-processor systems, micro-based systems, Set-top box, programmable user electronics, network PC, small

Brain, large computer, decentralized computing environment including any of the above systems I* The present invention can be described in the general context of computer executable instructions, such as a program module executed by a computer. In general, program modules include packages, programs, objects, components, data structures, etc. that perform specific Α or implement specific abstract data types. The present invention can also be practiced in a computing environment where tasks are performed by remote processing devices that are connected through a communications network or other data transfer. In a decentralized computing environment, program modules and other data can be stored in local and remote storage media, including memory storage devices. Decentralized computing is directly exchanged between computing devices and systems to facilitate sharing of computer resources and services. These resources are exchanged for information, cache memory, and file disk storage. Decentralized computing uses network connectivity to allow users to leverage their collective benefits for the entire company. In this regard, a variety of devices may have application objects or resources that may utilize the techniques of the present invention. Referring to Figure 1a, an exemplary system for implementing the present invention includes a general purpose computing device for computer 110. The components of computer 110 may include, but are not limited to, processing unit 120, system memory 130, and various system components, including w memory, coupled to stream 1 1 1 of processing unit 120. The system bus 1 1 1 can be any of a variety of types of bus bars, including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus structures. Transceiver processor type: etc. In the "distributed media", the package-based package, the form, the system, and the system are all provided by the company. This structure includes the industry standard structure (ISA). Busbar, Microchannel Structure (MCA) Bus, Reluctant IS A (EIS A) Bus, Video Electronics Standards Alliance (VESA) Local Bus, and Peripheral Component Mutual (PCI) Bus (also known as Mezzanine) Bus bar).

Computer 110 typically includes a variety of computer readable media. The computer readable medium can be any available media that can be accessed by computer 110 and includes both volatile and non-volatile media, extractable and non-extractable media. By way of example and not limitation, computer readable media may include computer storage media and communication media. Computer storage media includes volatile and non-volatile, extractable and non-extractable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, or other materials. Computer storage media includes, but is not limited to, RAM, R〇M, EEpR〇M' flash memory or other storage technology, CDR (10), home digital compact disc (DVD) or other disc memory, (4) box, tape, disk memory Body or other magnetic storage device, or any other medium used to store desired information and accessible by the computer. Communication media typically includes a computer readable finger +, data structure: a program chess group or other material' and includes any information delivery media, such as a carrier or other transmission mechanism. A data has been tuned to mean a signal that has one or more characteristics set or changed in such a way as to encode information within the signal. By way of example and not limitation, all media, including wired media such as a wired network or direct-line connection, and any combination of wireless media such as voice, RF, infrared, and other wireless media, should be Media 16 200937308 System Memory 130 includes computer storage media in the form of volatile and 'or non-volatile memory such as read only memory (R〇M) 13 1 and random access memory (RAM) 132. The basic input/output system i33 (BI〇s) is normally stored in the ROM 131. It contains, for example, a basic routine for facilitating the transfer of information between components within the computer ι1〇 during startup. RAM 132 generally contains data > and/or program modules' which are immediately accessible and/or currently being operated by processing unit 120. By way of example and not limitation, Section 1 & 圊 illustrates Q operating system 134, application 135, other programming modules 136, and program information 1 37. Computer 110 may also include other extractable/non-extractable, volatile/non-volatile computer storage media. By way of example only, FIG. 1 illustrates a hard disk drive for reading and writing non-extractable, non-volatile magnetic media, a disk drive 151 for reading and writing non-volatile magnetic sheets 152, and a removable extractor. The disc drive 155 for reading and writing the non-volatile optical disc 156, such as a CD ROM or other optical medium. Other extractable/non-extractable, volatile/non-volatile, non-extractable, non-volatile, and non-volatile memory available in the exemplary operating environment. Media includes, but is not limited to, tape cartridges, flash memory cards, devices, s Dual-certified universal disk, digital video tape, solid-state RAM, solid-state ROM, etc. The hard disk drive 141 is typically connected to the system bus bar 121 by a non-extractable memory interface such as interface 140, and the disk drive 151 and the optical disk drive 155 are typically interfaced with a removable memory interface such as interface 150. Connected to the system 121^尔玩汇排排•121. The drives and their associated computer storage media discussed above and illustrated in FIG. 1 provide computer 110 with computer readable instructions, data source 17 200937308 program modules and other data storage. In the figure la, for example, the hard disk drive 141 stores the operating system 144, the application program 145, the other program modules 146, and the program data 148. It is noted that these components may be the same as or different from the operating system 134, the application 135, the other program modules 136 and the program data 137. Here, the operating system 1 44, the application 145, the other program modules 146, and the program data 147 are given different numbers to indicate that they are at least different copies. The user can enter commands and information into the computer 110 through input devices such as a keyboard 162 and pointing device 161, commonly referred to as a mouse, trackball or touch pad. Other input devices (not shown) may include a microphone, joystick, game board, satellite dish, scanner, and the like. These and other input devices are often coupled to the processing unit 120 via a user input interface 160 coupled to the system bus 121, but may be connected by other interfaces and bus structures, such as parallel ports, gaming ports, or universal serial busses (USB). ). A monitor 191 or other type of display device is also coupled to the system bus 121 via an interface such as video interface 190. In addition to the monitor 19 1 , the computer may include other external devices such as a speaker 197 and a printer 196, which may be connected by an output peripheral interface 195. The computer 110 can operate in a networked environment that uses a logical connection with one or more remote computers, such as the remote computer 180. The remote computer 1 80 can be a personal computer, server, router, network PC, peer device, or other public network node, and generally includes many or all of the components described above in connection with the computer 110, although only illustrated in FIG. A memory storage device 181. The logical connections described in Figure la include the area 18 200937308 Network (LAN) 1 7 1 and Wide Area Network (WAN) 1 73, but may also include other networks. This networking environment is commonplace in offices, enterprise-wide computer networks, corporate intranets, and the Internet.

When used in a LAN networking environment, the computer 110 is coupled to the LAN 171 via a network interface or adapter 170. When used in a WAN networking environment, computer 110 typically includes a data machine 172 or other device for establishing communications over a WAN 173, such as the Internet. The data engine 172 can be internal or external and can be coupled to the system bus 1 1 1 via a user input interface 160 or other suitable mechanism. In a networked environment, the program modules described in relation to computer 101 may be stored in a remote memory storage device. By way of example and not limitation, FIG. 1a illustrates a remote application 185 residing on a memory device 181. It will be appreciated that the network connections shown are exemplary and other means of establishing a communication connection between computers may be used. Definition of Terms In order to describe the present invention in a more concise and clear manner, the following terms are specifically referred to herein as strokes: the trajectory of motion when writing. For the identification sampling, when the pen tip is in contact with the contact plate during writing, "handwriting" is left, and when it is not in contact with the contact plate, "writing" is not left. However, if you consider the movement of the nib in space while writing, it is a continuous trajectory. In the present invention, the term "stroke" will generally refer to the trajectory of such motion, whether or not it leaves a "handwriting". Real stroke: The real stroke refers to the stroke that leaves the handwriting, that is, when writing, the movement track of the tip of the pen 19 200937308 when it contacts the contact plate. Virtual strokes: virtual strokes refer to strokes that do not leave handwriting, that is, the movement trajectory when the nib is in contact with the non-contact plate during writing, and the virtual strokes are mostly used for the connection between the realistic strokes. It should be noted that, in the present invention, The virtual stroke mainly reflects the movement trend and direction of the nib. For the virtual stroke, it is not necessarily required to fully reflect the actual trajectory of the nib at this stage. Continuous stroke: A stroke that has a constant orientation over a period of time. The continuous stroke can be a real stroke or a virtual stroke, and the continuous stroke reflects the movement of the nib in the same direction in the same direction. Corner: A stroke that changes direction significantly. The corner is the transition between continuous strokes, and the corner reflects the movement of the tip with a significant change in direction. In the case of stroke conversion, it is a remarkable feature that the corner is a handwritten East Asian character. The present invention specifically provides a "turning angle" to better reflect this feature. Handwriting: Handwriting refers to the specific traces left by handwritten East Asian characters. Root: A part of a stroke or combination.概述 Overview Although the HMM model has been widely used in the 'identification field of online handwriting, the recognition of online handwritten East Asian characters is still there. There are two main challenges: the stroke order of handwritten East Asian characters. Diverse and writing styles. A typical HMM topology for handwritten font recognition is a left-to-right HMM topology. In the present invention, it relates to the improvement of the HMM topology, which mainly includes: 20 200937308 Add a "corner state" in the HMM topology to reflect the "corner" between the handwritten process strokes of the handwritten East Asian character. The self-transition of the "corner state" is restricted, and the "corner state" alternates with the "continuous state" that reflects the continuous stroke of the handwritten East Asian character handwriting. Multiple paths are used in one HMM to reflect the different stroke order and writing style of the same handwritten East Asian character. In the process of constructing such an HMM topology, the present invention considers the following questions: First, how many paths are appropriate in an HMM model. For this problem, the present invention proposes a new solution that uses a Subsequence Direction Histogram Vector to cluster the paths derived from the character data to determine the appropriate number of paths. Second, how many states are appropriate in a path. For this problem, the present invention utilizes the "turning angle" feature in the handwritten East Asian character, and uses the connection relationship of the corners in the curvature scale space to determine the number of "corner states" in a path, and then uses the "corner state" and "continuation". The state is an alternating occurrence of this feature, determining the appropriate number of states in the path. Third, for the continuous uncertainty between the strokes unique to handwritten East Asian characters, the present invention employs a method of setting a parallel state (e.g., a double Gaussian mixture model) in the HMM topology, or applying an MSD method. The above problems will be described in detail below. twenty one

200937308 HMM Topology Suitable for Recognizing Handwritten East Asian Characters The present invention is mainly directed to the recognition technique of handwritten East Asian characters. The characteristics of East Asian characters can be summarized as the following main aspects: 1) The stroke sequence is diverse and the writing style is diverse. Just as the S handwritten East Asian character has many strokes and the font structure is complicated, the word will appear in multiple stroke sequences, and the final written word will have multiple fonts or fonts, which is called handwriting style. 2) Due to the diversity of writing styles, the difference between the same word is as follows: 2a) For different simplified writing of local continuous strokes, the writing style of the guide may have different stroke numbers. 2b) In the process of connecting between real strokes, it can be connected in real or virtual strokes, and for real East Asian users, regardless of which strokes are connected, recognized characters, this is It is difficult to identify the machine. The first problem to be solved in handwritten East Asian character recognition is that the stroke sequence is diverse and the handwriting style is diverse. The same Chinese character, the writing habits of different people determine the order in which the strokes will be drawn. Although the written pens are fixed according to the standard of Chinese characters, the actual statistics show that most Chinese characters have several writing sequences in the application. And, every type of writing population has a considerable proportion. As a handwriting recognition, it is necessary to consider this situation. According to an example, there are two different stroke sequences in the "nine" character in Figure 3a. Among them, the handwriting and analysis of the same presentation method will result in different strokes and characters. 1) Now that the various sequences of the actual sequence are shown, the connection between the two strokes is indicated by the dashed line 22 200937308. Similarly, the handwriting of handwritten East Asian characters is more diverse. For example, referring to Figure 3b, the same Chinese character “complex” may have many different writing styles.

One implementation method is to construct an HMM topology for each stroke order and each handwriting style. The data volume and operation complexity are relatively large, and another problem is that such a scheme will bring certain implementation and training. Difficulty. To this end, the present invention provides a multipath HMM topology in which only one HMM topology (HMM model) is provided for the same character, using the paths therein to represent different stroke sequences or handwritten styles. Referring to Fig. 4a, Fig. 4a shows a block diagram of a proposed multipath HMM topology suitable for handwritten East Asian character recognition, in accordance with an embodiment of the present invention. Referring to FIG. 4a, the multipath HMM topology corresponds to a handwritten East Asian character; each path in the multipath HMM topology corresponds to one of a plurality of stroke sequences of handwritten East Asian characters; or a multipath HMM topology Each of the paths corresponds to one of a plurality of handwritten styles of handwritten East Asian characters. For example, referring to FIG. 4b, the multi-path HMM topology constructed according to the two handwritten examples of the Chinese character "nine" in FIG. 3a has two paths, each of which corresponds to a stroke of the "nine" word. order. Referring to FIG. 4c, according to the multiple handwriting examples of the Chinese character "complex" shown in FIG. 3b, after constructing the multipath HMM topology for the Chinese character "complex", two handwritten styles with typical features are clustered. (The clustering 23 200937308 process will be described in detail below), such that the multipath HMM topology has two paths, each of which corresponds to a handwritten style "complex" word. As mentioned above, one of the problems caused by the diversity of writing styles is 2 a) different simplified writing of local continuous strokes, resulting in different writing styles may have different stroke numbers. The number of states on each path is determined by the number of strokes in the writing style corresponding to this path. East Asian characters have obvious directional transition characteristics between adjacent strokes, that is, "turning angles", while continuous strokes and corners alternate. According to the present invention, it is first necessary to detect the number of strokes of the writing style by detecting the corners of the training samples corresponding to a certain path, thereby determining the number of states of the path, and the specific steps will be detailed below. description. It is precisely because of the fact that continuous strokes and corners alternate, the coarse-to-fine algorithm in the curvature scale space (C S S ) can be used to determine the number of states in a path. For each particular path, the present invention also introduces a "corner state" in the path to describe the short stroke transition feature at the corner. On a path corresponding to the same writing style, the continuous stroke corresponds to the "continuous state", the corner corresponds to the "corner state", and the continuous state and the corner state alternate. Due to the short-term nature at the corner of the stroke, the present invention sets the "corner state" to be non-rotating. This non-rotation characteristic combined with the difference between the continuous stroke feature and the corner feature can significantly improve the relationship between the feature sequence and the path state of the training sample. The correct correspondence. 24 200937308 After studying the characteristics of handwritten East Asian characters, we can find that the handwritten East Asian characters will have a significant turning process before completing a stroke and starting the next stroke, which is the "turning angle" mentioned above. Referring to Fig. 2a, Fig. 2a is an example of a handwritten example of a very simple Chinese character "上,". In the example of Fig. 2a, the real stroke of the word "up" is 3 strokes, which is the 2a. In the handwriting shown in the figure, the virtual stroke has 2 strokes, which is indicated by the dotted line with arrows in Figure 2a. Thus, between the total of 5 strokes (3 real strokes and 2 virtual strokes), 4 appear. The corner 2〇2a_2〇2d is represented by a circle with a dashed box in the figure. Thus, the present invention provides an improved HMM model with which the angle of the East Asian character can be well modeled. The present invention provides a method for recognizing handwritten East Asian characters using the HM Μ model to provide a ΗΜΜ topology corresponding to handwritten East Asian characters; wherein a continuation state describing handwritten East Asian character continuation strokes is provided in the ΗΜΜ topology; The corner state describing the corner between the handwritten East S-character strokes is provided in the topological structure. The above-mentioned ΗΜΜ topology is constructed as a left-to-right ΗΜΜ topology starting from the initial state to a termination state The continuous state in the topology can be transferred to the next state or self-transition, and the corner state can only be transferred to the next state, which cannot be self-transferred. The continuous state and the corner state alternately exist one after the other. The structure conforms to the characteristics of handwritten East Asian characters and can well reflect the characteristics of handwritten East Asian characters.

200937308 Referring to Figure 2b, Figure 2b shows the above-described "upper" HMM topology constructed from the "upper" handwriting of Figure 2a in accordance with the present invention. The five strokes in the "upper" handwriting example shown in Fig. 2a '3 real strokes and 2 virtual strokes respectively correspond to the five holding states 204a-204e in the second drawing, and each of the holding states 204a-204e One can be transferred from the next state or the next state. Continue to refer to the four corners 202a shown in Fig. 2b's Fig. 2a - corresponding to the four corner states 206a-206d in Fig. 2b, respectively. The corners 206a-206d cannot be transferred. 'Can only be transferred to the next state. And the angular states 206a-: 206d are inserted between the persistent states 204a-204e. In another embodiment, the HMM topology shown in FIG. 2b may also start in a starting state 'this state corresponds to the styling, and if there is an initial state, the starting state is not transferred. . It is also possible to end in the termination state. The termination state represents the time of the receipt. If the termination state is used, the termination state cannot be transferred. Thus, the handwriting example on the "upper" shown in Fig. 2a is constructed as the HMM topological graph shown in Fig. 2b, which is called the feature of the topological structure of the ΗMM: the left-to-right ΗΜΜ 揸 揸 ,, 扑 结构 structure From the initial state to the end state, the continuous state in the mouth, the mouth state can go to the next state or self-transition, and the m ^ φ ^ ^ corner state can only be transferred to i, 'can't Self-transition, continuous state and corner state are alternately stored> The principle word includes the continuation transition 202d state, and the time of the structure can be self-executed, such as the beginning of the merger, the transfer of the HMM topology shown in Figure 2b Another 26 of the structure 200937308 It should be noted that in the handwritten example of the "upper" word shown in Fig. 2a, the number of real strokes defined according to the present invention is equal to the number of strokes in the canonical Chinese character writing, however, For some Chinese characters or Chinese character strokes, the number of strokes (including real strokes and virtual strokes) defined in accordance with the present invention may not be equal to the number of strokes in standard Chinese characters. Reference. The example shown in Fig. 2c, the word "mouth", according to the canonical Chinese character writing, has a total of three strokes, but according to the definition of the present invention, the real stroke has a total of 4 strokes, and the virtual stroke has a total of 2 strokes. The main difference is the "π" pen, which is a stroke in the canonical Chinese character ,, and according to the definition of the present invention, "Π" contains 2 real strokes and 1 corner. Thus, in accordance with the definition of the present invention, handwritten examples of the "mouth" word shown in Figure 2c include real strokes 212a-212d, virtual strokes 214a and 214b, and corners 2 1 6a-2 1 6e. The order of arrangement is: real stroke 2 1 2a, corner 216a, virtual stroke 214a, corner 216b, real stroke 212b, corner 216c, real stroke 212c, corner 216d, virtual stroke 214b, corner 216e, real stroke 2 1 2 d. © In the present invention, the stroke is a stroke defined in the definition of the above terms. Returning to the multipath HMΜ topology mentioned above, since it is mainly suitable for the recognition of handwritten East Asian characters, 'the above-mentioned 'corner' state will still be provided in each of the ΗΜΜ paths. Returning to Figure 4a, in the multipath ΗΜΜ topology, since each path corresponds to handwritten East Asian characters, each of these paths has the following characteristics: 27 200937308 Each path is left to right The HMM topology starts from an initial state and ends at a termination state; the ΗMM topology includes a description of the continuation state of the handwritten East Asian character continuation stroke, and a description of the corner state of the handwritten East Asian character stroke between corners, wherein the continuation state can be transferred to In the next state or self-transition, the corner state can only be transferred to the next state, and cannot be self-transferred; the continuous state and the corner state alternately exist in sequence; for the entire multi-path topology, in which all paths start in the same entry state, End to the same exit state. In Fig. 4a, each circle represents a state, and the line segment with an arrow indicates a transition, wherein the arrow pointing to itself indicates self-transition, the state with state self-transition is a continuous state, and the state not capable of self-transition is a corner state. All paths begin in a public entry state and end in a public exit state. It should be noted that although in the example shown above, each path in the multipath HMM topology is a path that conforms to the previously defined features, and the path in the multipath HMM topology is shown as being used Consistently corresponds to multiple stroke sequences or handwritten styles. However, those skilled in the art will appreciate that the scope of the present invention is by no means limited thereto. The broad definition of the multipath HMM topology in the present invention should be: At least one path is a left-to-right HMM topology, together The start state begins and ends with a termination state; for the entire multipath HMM topology, where all paths start at the same entry state and end to the same exit state. 28

200937308 Also, it is described that a variety of different stroke sequences and different handwritten style paths can be placed in the same multi-path HMM topology, that is, in a multi-path HMM topology, partial strokes can correspond to different stroke sequences. And the other part of the path corresponds to a different writing style. In one case, due to the "turning" characteristic of the handwritten East Asian character, at least one path includes a description of the continuous stroke state of the handwritten East Asian character, and a corner state describing the angle between the handwritten East Asian character strokes, wherein the continuous state can be transferred. To the next state or self-transition, the transition state can only be transferred to the next state, and cannot be self-transferred; the persistent state and the transition state alternately exist in turn. Since the HMM topology of the present invention is self-contained from the training data, the samples in the training data may have great redundancy, or some extremely low probability contingency data may be recorded, so that the obtained In the multi-path HMM topology, there are redundant paths or paths with only a small amount of handwritten data, which will increase the model data and increase the complexity of the entire HMM topology. To this end, the invention also needs to merge the paths to eliminate redundant paths and paths that are not used by the base. The specific implementation of path merging will be described in detail below. It should also be noted that, for those skilled in the art, multipath can be implemented in an HMM topology by existing means. Therefore, the present invention mainly provides a type of HMM topology and Realizing this new type of HMM topology multi-path HMM model, therefore, those skilled in the art can fully understand the angle of the road hand, and can read the present invention at 29 200937308. The multipath model described herein is implemented after the specification without any additional creative effort. Now, the method of the present invention has solved the problem of inter-stroke connection of handwritten East Asian characters by providing a "corner" state, and provides a multi-channel HMM topology to solve various writing sequences and various writing styles. The last problem that needs to be solved, that is, problem 2 b), can be connected by real strokes or connected by strokes during the connection between strokes, and for users of true East Asian characters. Regardless of the type of strokes, it is recognized as the correct character. This kind of writing variable machine recognition brings difficulty. For example, referring to Figure 5a, for the word "up", two connections can be made between the completion of "|" and the next "one", and the strokes are connected, for example, in Figure 5a. Real strokes 5 02, or virtual strokes, such as strokes 504 shown in Figure 5 a, of course, there may be other connections, such as real strokes and some virtual strokes, here, for the sake of concise Say that the two situations are taken as an example for the time being.情况 The situation shown in Figure 5a can be called “virtual stroke/real stroke certainty”. This is also one of the problems to be solved by the present invention. For this case, the present invention provides two solutions. ♦ The first way is to provide a parallel state in the HMM topology. The handwritten East Asian characters may be either virtual strokes or real strokes. Referring to Figure 5b, there is shown a block diagram of an HMM topology having parallel states in accordance with an implementation of the present invention. Where the HMM is very good, the question of the pen is given by the character, and the imaginary part of the figure can be shown, in the case of the right part, at 30

200937308 State 5 0 6 and 5 0 8 State 5 1 0 has a flat 5 1 Ob, state 5 0 6 can be transferred to state 5 1 0 or 5 1 Ob one, state 5 1 0 or 5 1 0 b Any one of them can also be 50 8 . It should be noted that the HMM with the continuous state and the corner state proposed in the parallel state proposed by the present invention is used in combination, for example, any one of the continuous corner states in the HMM topology may have a parallel state. The concept of the parallel state described above can also be applied to a multi-path flap structure, that is, in any of the multi-path HMM topologies, there may be one or several states having parallel states. FIG. 5c illustrates A schematic diagram of a HMM topology mechanism in a parallel state constructed by the handwriting example of Figure 5a, in which only the relevant portions of the state are shown. Among them, the continuous state 5 1 2 represents the real stroke. The continuous state 5 1 8 represents the real stroke "one". For the part of Figure 5 that can be a real stroke or a virtual stroke, in the 5c | HMM topology, the persistent state 520 represents the 5th stroke 502, and the persistent state 520b represents the virtual capture in the 5b diagram. In the embodiment shown in Fig. 5c, the corner states 5 1 4 and 5 1 6 have a parallel corner state 5 1 4b and 5 16b. The corner state is a state in which the parallel state and its corresponding corner are closer to the position of the corner of the corner in the handwriting, which may be a real stroke or a virtual point. Then the corner state corresponding to the corner will also have a state, correspondingly, If the position of the corner is far from any state in the state that may be a real stroke, it may be linked to the state of the structure or the HMM. Built in the figure shown in the parallel diagram: t>] 5 04 ° also has a correlation, such as the stroke of the parallel part may also be 31 200937308 virtual stroke part, then The transition state corresponding to this corner has no parallel state. In general, in the Η Μ topology of the case shown, for example, in Fig. 5c, the corner states 5 1 4 and 5 16 each have a parallel state 5 1 4b and 5 1 6b. In other embodiments, it may be that 5 16 has a parallel state and 5 14 does not, or 5 14 has a parallel state and 5 16 does not. Of course, the present invention does not exclude the case where both corner states are not parallel.

One implementation that provides a parallel state is to use a double Gaussian mixture model (GMM), which uses the distribution characteristics of the Gaussian function such that the real and virtual strokes correspond to the peaks of different Gaussian functions, respectively. The second method to solve the real stroke/virtual stroke uncertainty is to apply a multi-space probability distribution (MSD) to the HMM topology, which may be either a virtual stroke or a real stroke in the handwritten East Asian character. MSD is a commonly used algorithm, so the detailed principle is not specifically described in the present invention. Applying the MSD to the HMM topology can be implemented according to the following method: For each HMM topology, two spaces are defined. , corresponding to real strokes and virtual strokes. The MSD can calculate the multi-space probability distribution in two spaces for the uncertain part, and obtain the first metric value corresponding to the real stroke and the second metric value corresponding to the virtual stroke, which is solved by processing the metric value. The problem of real stroke/virtual stroke uncertainty. Thus, the present invention effectively provides an identification technique for the main features of handwritten East Asian characters, which is well adapted to the characteristics of handwritten East Asian characters. However, one problem that is still faced is the problem of the amount of data in the model. As described above, for each handwritten East Asian character, a multipath ΗΜΜ topology is provided, and each path provides 32 200937308 states for each stroke, and between strokes. It is necessary to provide a "corner" state, and for the part where there is uncertainty in the real stroke/virtual stroke, it is also necessary to provide a parallel state or apply MSD technology. These will result in a very large amount of data for the HMM model. In order to effectively save the data space and reduce the implementation cost of the present invention, it is also necessary to consider the problem of data compression. Although the similarity between East Asian characters is not obvious, if the East Asian characters are divided into several parts, these partial structures still have many similarities. With this feature, clustering of states in the HMM topology can be achieved, thereby reducing the amount of data and reducing the complexity of the HMM topology. According to the present invention, after constructing the HMM topology, the states in the HMM topology topology are also clustered such that at least one set of states in the HMM topology topology share parameters, and for a set of shared parameter states, in the HMM topology Only one set of parameters is saved in the structure. The clustering of states may be between states on the same path, or between states on different paths, and is considered to be a clusterable state. Only one set of all these states will be reserved in the HMM topology. parameter. This can effectively reduce the amount of data and reduce the complexity of the HMM topology. Implementation Method for HMM Topology Suitable for Recognizing Handwritten East Asian Characters For the HMM topology described above, it can be implemented as described below: 33 200937308 1 ) Generate root training data before the HMM training process is required Marking data, for East Asian children, 'data is usually marked with characters instead of roots. For the identification method of the present invention, it is more necessary to use the data marked by the root in the established HMm topology. An additional effect of Viterbi decoding provides a correspondence between handwriting samples and hmm states. Therefore, with the Viterbi decoding, the handwriting data corresponding to the root can be automatically obtained from the character data, for example, the handwriting sample is decomposed by referring to the boundary point of the corresponding relationship in the HMM model of the corresponding root. In order to do the above-mentioned segmentation to obtain the root data, we must first obtain a preliminary HMM model. Although this model may not be able to identify the handwriting very accurately, it can obtain the demarcation point of the radical relatively accurately. The preliminary HMM model is built from a single path and gradually divides more and more paths until the number of paths is sufficient, and whether the number of paths is sufficient to measure by the convergence measure of the path. An example of this method is as follows:

a) Initialize an HMM with a single path (n=l); b) Calculate their convergence metrics C(Pl), c(P2)_..C for the n paths P1, P2 Pn already existing in the HMM Pn), and select the path Pj in which the convergence metric C(Pj) is the largest; c) If C(Pj) < T, T is the predetermined convergence threshold, this HMM topology is described for the current training data. The number of paths in the path is already enough, so you don't have to continue to split the path; d) If C(Pj) T ', copy the path Pj, add noise, and then build a new road, Jiang, & μ At this point, the ΗΜΜ model has n+1 paths;) 1 way to train the HMM model, until it can no longer be raised to identify the elite 蛊 _ $ ^ J inertia, so far, has been obtained The preliminary path HMM model, gland _ Λ Λ 丨 τ τ τ 资料 资料 资料 资料 τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ τ The steps are used for training to obtain a topology-optimized HMM model for identifying the year-old g-scores & 2) Determination of the optimal number of paths According to the present method, a statistical feature is also provided to solve the problem of determining the optimal number of paths, which is called "subsequence direction histogram vector". For each handwriting sample, it is divided into several segments by arc length, and each segment is extracted with a feature. Here, the arc length referred to herein refers to the sum of the lengths of all the strokes 'that is the sum of the lengths of all the real strokes plus all the virtual strokes'. The process of segmentation is based on the arc length after this summation. Among them are the shape features of the segment. Therefore, after the operation of segmenting and extracting features, it is possible to obtain a -^^^-shaped feature of a handwriting sample on each segment. The features on each segment are connected in order to obtain a feature of the handwriting sample. The parent shape of the handwriting sample can be implemented as a subsequence direction histogram vector. The segment of the handwriting sample is further divided into a plurality of sub-segments, each of which determines a quantized directional feature, and each of the quantized directional features corresponds to a direction of a predetermined angular range. The shape feature on each segment is the histogram sequence direction histogram, and the feature of each handwriting sample is a subsequence direction histogram vector ° 35 200937308

Referring to Figure 6a, the process of establishing a subsequence direction histogram is illustrated. First, the handwriting sample "King" 6 0 2 is divided into several segments according to the arc length, where the arc length is the total arc length 6 04 after the real stroke and the virtual stroke are connected. Among these segments, some are only Including real strokes, such as segment 604a, some include only virtual strokes, such as segment 604c, and some have both real strokes and virtual strokes, such as segment 604b. Taking segment 604b as an example, it can be further divided into a number of word segments for which a subsequence direction histogram is obtained based on their direction (reference direction indicator 606) and stroke density. For example, the subsequence direction histogram corresponding to one segment 605a of the virtual stroke portion of 604b is 605b, and the subsequence direction histogram corresponding to one segment 607a of the real stroke portion of 604b is 607b. Combining the subsequence direction histograms of all sub-segments, the sub-sequence direction histogram corresponding to each segment can be obtained, and the sub-sequence direction histogram of each segment can be further combined to obtain the corresponding handwriting. The subsequence direction histogram vector of the sample, such as 608 in Figure 6a. After obtaining the sub-sequence direction histogram vector, the above features can be easily clustered, each cluster corresponding to a stroke order or a handwritten style. The operation of clustering can be achieved by, for example, a Guassian Mixture Model. Finally, the results of m clusters will be obtained, and m is also the number of optimized paths corresponding to a handwritten East Asian character. Then, based on the data of each cluster, parameters of a path in the HMM topology corresponding to each stroke order or each handwritten style are determined. Referring to Figure 6b, Figure 6b is an automatic classification of training data, more 36 200937308 is exactly an example of clustering operations. In Figure 6b, a total of nine sample data were provided from the training materials. By automatically classifying the data, they are clustered into two typical classes, representing the two complex words of the written style after clustering. '3) Determination of the optimal number of states • After determining the number of paths, it is also necessary to determine the number of states in the path. According to the present invention, the number of states of the machine self-learning from the data is still used to determine the number of states. When determining the number of states in a path, the data used is the handwriting sample from the same class in step 2) above. The state in the HMM reflects the shape and variation of the handwriting sample. Therefore, the state includes two categories: the part that is not obvious in direction change (the curvature is not large), called the "continuous stroke", corresponding to the "continuous state" in the HMM. ": For the part where the direction changes significantly (large curvature), it is the "corner", which corresponds to the "corner state". The number of states in a path is the sum of the number of "continuous strokes" and "corners", that is, the sum of the number of "continuous states" and the number of "corner states". The gradually refined texture-to-fine Curvature Scale Space can be used to detect the corners to determine the number of corners. At the same time, because the corners and continuous strokes alternate, The number of continuous strokes is derived from the number of corners so that the number of states can be determined. 4) State Connection Design As described above, the problem of "virtual stroke/real stroke uncertainty", the present invention provides two solutions. The first way is to provide a parallel state in the HMM topology. For the 2009 2009308, you should write a part of the East Asian character that may be either a virtual stroke or a real stroke. One implementation form that provides a parallel state is to use a double Gaussian mixture model (GMM), that is, using the distribution characteristics of the Gaussian function, so that the real stroke and the virtual stroke correspond to the peaks of different Gaussian functions, respectively, and by using such a double Gaussian mixture model, Solve the problem of solving the uncertainty of real strokes/virtual strokes. Referring to Figure 6c, for the word "wood", there are two ways of writing a pen and a pen, and the writing of the pen corresponds to the case of "real strokes", and the case of not writing a pen corresponds to the case of "virtual strokes". In the 6 c-picture, the real stroke and the virtual stroke have their own Gaussian function distributions, and their peaks are different, so they can be combined to obtain a double Gaussian mixture model with two different peaks to achieve a parallel state. In the second method, a multi-space probability distribution (MSD) is applied to the HMM topology, which may be a virtual stroke or a part of a real stroke in the handwritten East Asian character. 5) Path Merging After the above steps are completed, a preliminary multipath HMM topology is obtained, which includes paths corresponding to all cases in the training data, including paths reflecting stroke order and handwritten style, but due to training data Redundant information and very low probability information exist in the process, and the obtained paths need to be combined. The merging of paths can include two aspects: removing paths with very low probability of occurrence; and merging similar paths. A path with a very low probability will appear, that is, the amount of data in the training data corresponding to the path is small, and the proportion of the total amount of training data is removed.

200937308 In addition to these paths can be achieved by setting a predetermined threshold. The similar path merging operation can be implemented by measuring the path meter. The similarity metric is used to represent the path degree. When the two paths are sufficiently similar, the path merging similarity measure can use the Kullback-Leibler difference. The difference between the Kullback-Leibler and the Kullbaek-Leibler difference indicates that the difference is less than a predetermined value, indicating that the two are similar, then they can be combined, and the merge operation will process the two paths to obtain A path that best reflects the main features of the original two. 6) After state clustering, it is also necessary to aggregate the states in the HMM topology to divide the East Asian characters into several parts, and these local structures have few similarities. With this feature, clustering of states in the HMM can be achieved, thereby reducing the amount of data and reducing the complexity of the HMM. The clustering of states may be between states on the same path or between states on different paths. It is considered to be the state. In the HMM topology, only one set of parameters will be reserved for all belonging to the same state. According to the present invention, the clustering operation includes: calculating a state similarity measure between each two states of the HMM extension, and when the states are similar, indicating that the two states are sufficiently similar, the two states share the parameter when two or more states are mutually When the state similarity measures are between, the parameters are shared by the states. A phase between the similarities in the present invention. For example, - indicates, changes. If the path is sufficient for the class in the flat path. If there is an inter-topological structure with a non-topological structure, the clustered cluster-like structure can be clustered; wherein, in a sufficiently similar implementation, 39 200937308 state similarity measure is represented by Kullback-Leibler difference, Kullback-Leibler difference A value below a predetermined value indicates that the two states are sufficiently similar. For example, the process of a specific clustering operation is shown as follows: Initialization phase: Set the total number of all states to one, and classify the mth state into the mth class.

For any two states m and η, calculate the Kullback-Leibler difference between them and denote it as D( m,n). Clustering phase: Finding a set of states (m', n') with the smallest Kullback-Leibler difference can be achieved by the following function: (m', n') = argmin( m, n). The m'th and η' states are merged by adding rows and columns corresponding to m' and η' in the Kullback-Leibler difference matrix. Decrement the total number of states 1 by 1. Loop operation: If the number of defects is greater than a predetermined value, repeat the above operation, otherwise the merge operation is completed. Through the above steps 1)· 5), a multi-path ΗΜΜ topology corresponding to handwritten East Asian characters can be automatically constructed from the training materials of handwritten East Asian characters, each of which corresponds to multiple stroke sequences of handwritten East Asian characters. One of them, or each of the paths corresponds to one of a plurality of handwritten styles of handwritten East Asian characters; providing a description of the persistence state of the handwritten East Asian character continuous stroke 40 200937308 in the topological structure, and describing the handwritten East Asian character strokes The corner state of the corners; wherein each path in the multipath HMM topology is a left-to-right HMM topology, starting from an initial state to a termination state; wherein the persistent state can be transferred to the next state or self-transfer The corner state can only be transferred to the next state and cannot be self-transferred; the persistence in the HMM topology • the state and the corner state alternately exist in turn; and all of the paths start at the same entry state and end at the same exit state. 〇 Possible hardware implementations The present invention can be implemented in the form of software, such as by running a software that implements the method of the present invention by a general purpose computing system. The present invention can also be implemented in the form of instructions or programs which can be stored on a storage medium which can be implemented by a computing device after it has been executed and executed from the storage medium. In addition, the present invention can also be implemented in the form of a hardware. It should be noted that it will be apparent to those skilled in the art that in the field of the present invention, the conversion between soft and hard bodies has various forms. That is, having different forms of hardware can achieve the same function, and therefore, the possible hardware implementations listed below of the present invention are functions that define the hardware, and are not limited to the specific implementation thereof, and are known to those skilled in the art. In fact, it is obvious to implement various forms of functions based on these functions. Referring to FIG. 7, a system 700 for recognizing handwritten East Asian characters using an HMM model according to an example of the present invention includes: 41 ❹ 200937308 HMM topology construction device 7〇2, constructing an HMM topology corresponding to handwritten elements; The persistent state setting means 7〇4, in the HMM topology, mentions the continuous state of the handwritten East Asian character continuous stroke; the corner state setting means 7〇6 provides the corner state of the handwritten East Asian character stroke between the corners in the HMM topology. The Ημμ structure constructed by the HMM topology device 702 is a left-to-right HMM structure, starting from an initial state, ending the state; and, the continuation state in the MM topology can be moved to the next- The state or self-transition 'corner state can only be transferred to the lower state, and cannot be self-transferred; the continuous state and the corner state are alternately present in another embodiment. The H-blue topology construction device 702 has a multi-path HMM topology corresponding to handwriting. One path of the East Asian character corresponds to one of the various stroke order orders of the handwritten East Asian car, the singer, and the endogenous squad, each of which corresponds to one of the various handwriting of the handwritten East Asian character. In the middle of the , , , H H H H H H 牡 牡 牡 牡 牡 牡 牡 牡 构建 构建 构建 构建 构建 构建 构建 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 702 , to the end of the strong &&& Dan, the end state; and, where the continuous can be transferred to the next - state or self-transition, the corner state can only be transferred - the state 'can not be transferred; ΗΜΜ topology to make the state And the transition state alternates in turn, and all the paths in it start at the same state, and the same exit state ends. In one embodiment, the surviving μ system 7〇〇 also includes a parallel state-sub-word "description describing the topology to a per-turn-like configuration: or a style ημμ from a state to a lower-angle mouth-like supply 42 200937308 708, providing a parallel state in the HMM topology, corresponding to the handwritten East Asian character may be either a virtual stroke or a real stroke part. The parallel state providing device 708 can apply a double Gaussian mixture model to achieve parallel Alternatively, a multi-space probability distribution (MSD) processing device 710 is applied to apply a multi-space probability distribution to the HMM topology, which may be a virtual stroke or a part of a real stroke in the handwritten East Asian character.

It should be noted that the parallel state providing means 708 and the multi-space probability distribution processing means 710 are selected one by one. In one embodiment, the system 700 further includes a state clustering device 712 that clusters states in the HMM topology such that at least one set of states in the HMM topology share parameters and for the set of shared parameters State, only one set of parameters is saved in the HMM topology. It should be noted that the various devices of the system 700 for identifying handwritten East Asian characters using the HMM model described herein can be used to implement the method described above in connection with FIG. 2.5, wherein each detail feature corresponds, so Description will be repeated again. 8 is a block diagram showing a system 800 for identifying handwritten East Asian characters using an HMM model, the system 800 comprising: an HMM topology construction device 802, which is constructed to correspond to handwritten East Asian characters, in accordance with another embodiment of the present invention. A multi-path HMM topology in which each path corresponds to one of a plurality of stroke sequences of handwritten East Asian characters, or each of which corresponds to one of a plurality of handwritten styles of handwritten East Asian characters; 43 200937308 Continuous state setting device 804, setting a continuous state describing the handwritten East Asian character continuous stroke in the HMM topology; the corner state setting device 806, setting and describing a corner state of the handwritten East Asian character stroke between the corners in the HMM topology; wherein, the multipath HM Μ Each path in the topology is a left-to-right ΗΜΜ topology, starting from an initial state to a termination state; the continuation state can be transferred to the next state or self-transition, and the corner state can only be transferred to the next state, Self-transition; continuous state and corner state alternately exist; and all paths start In the same entry state, to the end of the same exit state; and one of the following two devices: Parallel state providing device 808, providing parallel states in the topological structure, corresponding to handwritten East Asian characters may be either virtual strokes, or It is a part of the real stroke; the parallel state providing device 800 can adopt a double Gaussian mixture model (GMM), that is, using the distribution characteristic of the Gaussian function, so that the real stroke and the virtual stroke respectively correspond to the peaks of different Gaussian functions, using such a double The Gaussian mixture model can also solve the problem of solving the uncertainty of real stroke/virtual stroke. The multi-space probability distribution processing device 8 10 0 applies a multi-space probability distribution to the topological structure, which may be a virtual stroke or a part of a real stroke in the handwritten East Asian character. According to the system 800 shown in FIG. 8, the ΗΜΜ topology construction apparatus 802 automatically constructs a multi-path ΗΜΜ topology from the training material of the handwritten East Asian character; and the ΗΜΜ topology construction apparatus 802 writes the East Asian character according to 44 200937308 The automatic classification of the habits of the '$ person's writing style, should be, using the - machine self-study method to classify the training materials. Continue to refer to Figure 8 to write the East Asian character handwriting 搂 The training data used by the system 8〇0 includes the handwriting style handwriting style handwriting sample including different stroke sequences or not included, the track sample; the MM topology construction device 8〇2

Handwriting sample segmentation I ❹ Divide it into several segments 1 , for a parent handwriting sample, according to the arc length sum, that is, all the actual two said arc length refers to the segment length of all strokes based on this : The sum of the lengths of some virtual strokes, the length of the clothes and the length of the arc. The feature extraction means 822 extracts the features of the shape of each segment for each of the eight features; wherein, after the operation, the K-shaped/over-segmented and extracted feature features are extracted. A feature of the handwriting sample on the ::: segments on each segment.筏 筏 连接 , , , , , , 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 824 ^ 824 ^ ^ ^ The base corresponds to each stroke order, and the data of each cluster is a parameter of a path in the structure.悝Write style HMM extension

The topology construction apparatus 8〇2 may further include an embodiment, the HMM vertical apparatus 826, which implements the above-described handwriting===column direction histogram vector as a subsequence direction histogram vector^ on the parent segment Shape feature For each handwriting sample, the real stroke, the handwriting sample segmentation device is connected to the virtual stroke and the virtual stroke is 45

200937308 The arc length after the stroke and the virtual stroke are connected into a plurality of segments; and the feature is the direction of quantization, the feature extraction means 822 causes each sign to correspond to a direction of a predetermined angle range; the sub-sequence direction histogram vector establishing means 826 establishes a correspondence corresponding to a subsequence direction histogram vector of segments and features, the subsequence direction direct direction amount establishing means 286 further dividing the segment of the handwriting sample into a plurality of segments, each sub segment determining a quantized directional feature, each quantized The feature corresponds to a direction of a predetermined range of angles. The shape feature on each segment is a subsequence direction histogram, and each hand sample feature is a subsequence direction histogram vector. With continued reference to Fig. 8, in one embodiment, the system 800 also includes a path merging device 814 that combines the paths in the multipath HMM topology to control the number of paths. The path merging device 8 1 4 determines the data amount of the training data corresponding to the path, deletes the path of the corresponding data amount smaller than the predetermined value; calculates the path similarity measure between each two paths in the multi-path HMM topology, when The path similarity measure indicates that the two paths are merged when the two states are sufficiently similar, wherein the paths are merged when the path similarity measures of the two or more paths are sufficiently similar. In one implementation, the path similarity metric is represented by a Kullback-Leibler difference. When the Kullback-Leibler difference is below a predetermined value, the paths are sufficiently similar and the paths can be merged. The system 800 can also include a state clustering device 8 1 2 that clusters states in the HMM topology to make at least one configuration common parameter in the HMM topology, and for a set of shared parameter states, in HMM TNT The sub-picture sub-path is broken in the road path. The path is only 46. 37308 Only one set of parameters is saved in the structure. The state clustering means 812 calculates a state similarity measure between each two states in the HMM topology, and when the state similarity measure indicates that the two states are sufficiently similar, the two states share parameters; wherein 'when more than two When the state similarity metrics are sufficiently similar to each other, the states are all shared parameters "and the state clustering means 812 state represents the similarity metric by the Kullback-Leibler difference, when the Kullback-Leibler difference is lower than one When the value is predetermined, it means that these states are sufficiently similar that they can be merged. It should be noted that the various devices of the system 800 for recognizing handwritten East Asian characters using the HMN1 model herein can be used to implement the method described above in connection with FIG. 6 including step 1)_5) described above to construct a multipath uplift. The structure in which each detail feature corresponds is therefore not described repeatedly here. The invention provides a scheme for recognizing handwritten East Asian characters by using the HMM model of improved Q for the characteristics of handwritten East Asian characters, and fully considers that the handwritten East Asian character has many strokes, various stroke sequences, complicated structure, and diverse writing styles. The characteristics of the connection between the strokes are uncertain, and the characteristics of the HM • • • structure are improved. By adding the corner state in the HMM topology, providing multiple paths and providing parallel states to solve the above problems, and 'through clustering and The combined operation reduces the amount of data and reduces the complexity of the operation. The present invention has been described in detail above with reference to an embodiment of the present invention, but it should be noted that this is not to be construed as limiting the scope of the invention to any of the limitations of the above-mentioned embodiments. It is intended that the present invention be construed as being limited to the scope of the invention as defined by the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS The above and other features, aspects and advantages of the present invention will become more apparent from the description of the appended claims Features, wherein Figure 1a is an example of a suitable computing system environment in which the present invention may be implemented; Figure lb shows an example of an HMM model; Figure 2a shows "turning corners" in handwritten East Asian characters Figure 2b shows the HMM topology that reflects the cornering characteristics of handwritten East Asian characters; Figure 2c shows the distinction between continuous strokes and corners in handwritten East Asian characters; Figure 3a shows handwritten East Asian characters Diversity of stroke sequences; Figure 3b shows the diversity of handwritten styles of handwritten East Asian characters; Figure 4a shows the structure of a multipath HMM topology according to an embodiment of the invention; An example of a multi-path HMM topology in which each path corresponds to one stroke order is shown; 48 200937308 Figure 4c shows an example of a multi-path HMM topology corresponding to one hand-written style for each path; Figure 5a shows a case where the real stroke/virtual stroke uncertainty is written in the East Asian character writing process; Figure 5b shows a schematic diagram of the HMM topology having a parallel state according to an embodiment of the present invention; Figure 5c shows a schematic diagram of a HMM topology with parallel states constructed according to the handwritten example shown in Figure 5a; Figure 6a shows an example of applying a subsequence direction histogram to handwriting samples; Figure 6b shows An example of a clustering operation on training data; Figure 6c shows an example of implementing a parallel state using a double Gaussian mixture model; Figure 7 shows a nest using a HMM model according to an embodiment of the present invention. A structural diagram of a system for identifying East Asian characters; FIG. 8 is a structural diagram of a system for recognizing handwritten East Asian characters using the HMM model according to another embodiment of the present invention. [Main component symbol description] 1 3 0 system memory 1 3 4 operating system 1 35 application 1 3 6 other program module 49 200937308 137 program data 144 operating system 1 4 5 application 1 3 6 other program module 147 program data 120 processing unit 140 is not extractable, non-volatile memory interface

1 5 0 extractable, non-volatile memory interface 1 90 video interface 1 60 user input interface 1 9 5 output peripheral interface 1 7 0 network interface 172 data machine 162 keyboard 1 6 1 indicating device 163 microphone 1 9 1 monitoring 196 printer 197 speaker 1 7 1 area network 1 7 3 wide area network 180 remote computer 50 200937308 185 remote application 7 02 HMM topology construction device 7 04 continuous state setting device 706 corner state setting device 708 parallel state Providing device 710 multi-space probability distribution processing device 7 1 2 state clustering device 8 02 HMM topology construction device 8 2 handwriting sample segmentation device 8 22 feature extraction device 824 clustering device 826 sub-sequence direction histogram vector establishing device 804 Continuous state setting means 86 6 corner state setting means 8 0 8 parallel state providing means 810 multi-space probability distribution processing means 8 1 4 path combining means 8 1 2 state clustering means 51

Claims (1)

200937308 X. Patent application scope: 1. A method for establishing a hidden Markov (HMM) model suitable for recognizing handwritten East Asian characters, characterized in that: an HMM topology is provided, which is used to identify handwritten East Asia Character; * Provides a persistent state in the HMM topology that describes the continuous stroke of the handwritten East Asian character; © Provides a corner state in the HMM topology that describes the handwritten East Asian character The corner between the strokes. 2. The method of claim 1, wherein the HMM topology is a left-to-right HMM topology, starting from an initial state to a termination state; a persistent state in the HMM topology It is set to be able to shift to the next state or self-transition, and the corner state is set to be transferred only to the next φ state, and cannot be self-transferred; the persistent state and the corner state in the HMM topology are set to alternately exist in sequence. 3. The method of claim 1, further comprising: providing a multipath HMM topology corresponding to handwritten East Asian characters; each path in the multipath HMM topology is constructed as Hand 52 200937308 Write one of a plurality of stroke sequences for East Asian characters; or each of the multipath HMM topologies is constructed as one of a plurality of handwritten styles of handwritten East Asian characters. 4. The method of claim 3, wherein each path in the multipath HMM topology is a left-to-right HMM topology, starting from an initial state to a termination state; The HMM topology includes a persistent state and a corner state, wherein the persistent state can be transferred to a next state or a self-transition, the corner state can only be transferred to the next state, and cannot be self-transferred; the persistent state in the HMM topology and The corner states alternate in sequence; and all of the paths start at the same entry state and end at the same exit state. The method of claim 3, further comprising: providing a parallel state in the HMM topology, which may be a virtual stroke or a real one in the handwritten East Asian character. The part of the stroke. 6. The method of claim 3, further comprising: applying a multi-space probability distribution (MSD) to the HMM topology, the pair of 2009 2009308 handwritten East Asian characters may be virtual strokes, It may also be part of a real stroke. 7. The method of claim 3, further comprising: • clustering states in the HMM topology such that at least one set of states in the HMM topology share parameters, and The state of the set of shared parameters, only one set of parameters is saved in the HMM topology. 8 - A method for recognizing handwritten East Asian characters using an HMM model, characterized in that: a multi-path HMM topology corresponding to handwritten East Asian characters is provided, wherein each path corresponds to a plurality of stroke sequences of handwritten East Asian characters One, or each of the paths corresponds to one of a plurality of handwritten styles of handwritten East Asian characters; q providing a continuous state describing the handwritten East Asian character continuous stroke in the HMM topology, and describing the handwritten East Asian character strokes a corner state of the corner; • wherein each path in the multipath HMM topology is a leftward/right HMM topology, starting from an initial state to an end state; wherein the persistent state can be transferred to the next state State or self-transition, the corner state can only be transferred to the next state, and cannot be self-transferred; the persistent state and the corner state in the HMM topology alternately exist in turn; and wherein 54 200937308 all paths start at the same entry state, to the same Ending the exit state; providing a parallel state in the HMM topology, or The spatial structure of a multi-topology application HMΜ probability distribution (M S D), corresponding to the handwritten characters in both East may be a dummy stroke, it may be part of the real stroke. • The method of claim 8, wherein the multi-path topology is automatically constructed from handwritten East Asian character training material ©; according to the stroke order or writing style of the handwritten East Asian character, The training data is classified by an automatic classification method of machine self-learning. The method of claim 9, wherein the training material comprises a handwritten East Asian character handwriting sample, the handwriting sample comprising handwriting samples of different stroke sequences or different handwriting styles; for each handwriting sample Dividing it into several segments according to the arc length, extracting a feature for each segment of φ, and connecting the features on each segment in order, thereby obtaining a feature of the handwriting sample; clustering the feature Each cluster corresponds to a stroke order or a handwritten style; 'Based on each clustered data, the topology and initial parameters of a path in the topological structure corresponding to each stroke order or each handwritten style are determined. The method of claim 10, further comprising: establishing a subsequence direction histogram vector corresponding to the segment and the feature, wherein for each handwriting sample, the real stroke is The virtual strokes are connected, * and the arc length after the real stroke and the virtual stroke are connected is divided into several segments, * each sub-segment determines a quantized directional feature, and each quantized directional feature corresponds to a predetermined angular range direction; The shape feature on the segment is a subsequence direction histogram, and each hand sample feature is a subsequence direction histogram vector. 12. The method of claim 9, further comprising: merging the paths in the multipath ΗΜΜ topology to control the number of paths. The method of claim 12, wherein the merging of the path comprises: determining a data amount of the training material corresponding to the path, and deleting a path corresponding to the data amount less than a predetermined value; A path similarity measure between each two paths in the multipath topology, when the path similarity measure indicates that the two paths are sufficiently similar, the two paths are merged, wherein when two or more paths have path similarity measures Merge these paths when they are sufficiently similar. The method of claim 13 is characterized in that the path similarity measure is represented by a Kullback-Leibler difference value, and when the Kullback-Leibler difference is lower than a predetermined value, the path is sufficient similar. The method of claim 12, further comprising: clustering states in the HMM topology to make at least one set of states in the HMM topology share parameters, and The state of a set of shared parameters in which the HMM topology allows only one set of parameters to be saved. 16. The method of claim 15, wherein the clustering the states in the plurality of HMM topologies comprises: calculating a state similarity 0 degree metric between each of the two states in the HMM topology, when The state similarity measure indicates that when the two states are sufficiently similar, the two states share parameters; wherein when the state similarity measures ' between two or more states are sufficiently similar, the states are shared . 17. The method of claim 16, wherein the state similarity measure is represented by a Kullback-Leibler difference value, and when the Kullback-Leibler difference is lower than a predetermined value, the state is sufficiently phased. 57 200937308 Like. ...-HMM topology construction device, which constructs an HMM topology corresponding to the handwritten East Asian sub-element; Ο a continuous state sag device, which provides a continuous state describing the handwritten East Asian character continuous stroke in the HMM topology; The corner state setting device provides a corner state of the handwritten East Asian character stroke between the HMM topologies. The system of claim 18, wherein: the hmm topology constructed by the HMM topology construction device is a left-to-right HMM topology starting from an initial state and ending at a termination state; The persistent state in the HMM topology is set to be able to transition to the next state or the self-transition 'corner state is set to only transfer to the next ~ state' cannot be self-transferred; the persistent state and the corner state in the HMM topology are Set to alternate in order. The system of claim U is characterized in that: the HMM topology construction device constructs a multipath topology node 58 200937308, corresponding to handwritten East Asian characters; each of the paths corresponds to a plurality of handwritten East Asian characters One of the stroke sequences; or each of the paths corresponds to one of a plurality of handwritten styles of handwritten East Asian characters. 2 1 · The system of claim 20, characterized in that: * each path in the multi-path HMM topology constructed by the HMM topology construction device is a left-to-right HMM topology, starting from The 〇 state begins and ends at a termination state; and, wherein the continuation state can be transferred to the next state or self-transition, the corner state can only be transferred to the next state, and cannot be self-transferred; the persistent state and the corner state in the HMM topology are sequentially Alternate; and where all paths start at the same entry state, ending at the same exit state. φ 22. The system of claim 20, further comprising: a parallel state providing device that provides a parallel state in the HMM topology, which may be a virtual stroke in the corresponding handwritten East Asian character It may also be a part of real and stroked. The system according to claim 20, characterized in that it further comprises: 59 200937308 A multi-space probability distribution (MSD) processing device, which applies a multi-space probability distribution to the HMM topology, corresponding to the handwritten East Asian character It may be a virtual stroke or a part of a real stroke. 24. The system of claim 20, wherein the method further comprises: a state clustering device that aggregates states in the HMM topology to at least one of the HMM topologies The group state shares parameters, and for the state of the set of shared parameters, only one set of parameters is saved in the HMM topology. A system for recognizing handwritten East Asian characters using an HMM model, comprising: an HMM topology construction device, which constructs a multipath HMM topology corresponding to handwritten East Asian characters, each of which paths Corresponding to one of the various stroke sequences of the handwritten East Asian character, or each of the paths corresponds to one of a plurality of handwritten styles of handwritten East Asian characters; a continuous state setting device, setting a description in the HMM topology • Handwriting a continuous state of the East Asian character continuation stroke; - a corner state setting means for setting and describing a corner state of the handwritten East Asian character stroke between the HMM topologies; wherein each of the multipath HMM topologies The path is a left-to-right HMM topology, starting from an initial state to a termination state junction 60 200937308 bundle; the persistent state can be transferred to the next state or self-transition, and the state can only be transferred to the next state, and cannot be self-transferred; The state of the persistent state alternates in turn; and all paths start at the same entrance to the same The exit state ends; and one of the following two devices: a parallel state providing device that raises the state in the HMM topology, corresponding to the portion of the handwritten East Asian character that may be both a virtual stroke and a real stroke; The spatial probability distribution processing device, the multi-space probability distribution of the HMM topology, corresponds to a part of the handwritten East Asian character that may be either a virtual or a real stroke. 26. The system of claim 25, wherein the HMM topology construction device automatically constructs a multi-path HMM topology from handwritten East Asian characters; and the HMM topology construction device is based on handwritten East Asian characters Meta-order or writing style, using a machine self-learning automatic classification of the training materials for classification. The system according to claim 26, wherein the training material comprises a handwritten East Asian character handwriting sample, the pen comprising a handwriting sample of different stroke order or different handwriting style; the metatop topology construction device comprises: The corner and corner state, for a flat may be the application stroke, in: Training stroke method for: Trace sample HMM 61 200937308 One trace sample segmentation device, for each handwriting sample, divide it into several points according to the arc length a segment, the features on each segment are sequentially connected to obtain a feature of the handwriting sample; a feature extraction device that extracts a feature for each segment; a clustering device that performs the segment and feature Clustering, each cluster corresponds to a stroke order or a handwritten style; ❹ the HMΜ topology construction device determines one of the topologies corresponding to each stroke order or each handwritten style based on the data of each cluster The topology and initial parameters of the path. The system of claim 27, further comprising: a subsequence direction histogram vector establishing device, wherein the handwriting sample segmentation device performs a real stroke for each handwriting sample Connected to the virtual stroke, and divide the arc length after the real stroke and the virtual stroke are divided into several segments; each sub-segment determines a quantized directional feature, and each quantized directional feature corresponds to a predetermined angular range direction; the sub-sequence direction The histogram vector establishing means establishes a shape feature on each segment as a subsequence direction histogram, and each of the handwriting samples has a subsequence direction histogram vector. The system of claim 25, wherein the system further comprises: a path merging device that merges the paths in the multipath Η MM topology to control the number of paths. The system according to claim 29, wherein the path merging device: determining the amount of the training material corresponding to the path, deleting the path of the corresponding resource amount less than a predetermined value; A path similarity measure between every two paths in a multipath HMM topology. When the path similarity measure indicates that the two paths are sufficiently similar, the two paths are merged, wherein when two or more paths have path similarity metrics When these paths are sufficiently similar, merge the paths. The system according to claim 30, wherein the path merging device path uses a Kullback-Leibler difference to represent a similarity metric, and when the Kullback-Leibler difference is lower than a predetermined value, The path is similar enough. The system of claim 29, further comprising: a state clustering device that clusters states in the HMM topology to make at least one of the HMM topologies The state shares parameters, and for the state of the set of shared parameters, only 63 200937308 sets of parameters are saved in the HMM topology. 3. The system of claim 3, wherein the state clustering device calculates a state similarity measure between each of the two states in the HMM topology, and the state similarity measure represents two When the states* are sufficiently similar, the two states share the parameters; wherein when the state similarity metrics of the two or more states are sufficiently similar to each other, the states are shared. 3: The system according to claim 3, wherein the state clustering device indicates the state similarity measure by the Kullback-Leibler difference, and when the Kullback-Leibler difference is lower than a predetermined value, This state is sufficiently similar. 64