WO2015161823A1 - 手写识别方法和设备 - Google Patents

手写识别方法和设备 Download PDF

Info

Publication number
WO2015161823A1
WO2015161823A1 PCT/CN2015/077367 CN2015077367W WO2015161823A1 WO 2015161823 A1 WO2015161823 A1 WO 2015161823A1 CN 2015077367 W CN2015077367 W CN 2015077367W WO 2015161823 A1 WO2015161823 A1 WO 2015161823A1
Authority
WO
WIPO (PCT)
Prior art keywords
stroke
segmentation
character
strokes
input
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2015/077367
Other languages
English (en)
French (fr)
Inventor
江淑红
吴波
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sharp Corp
Original Assignee
Sharp Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sharp Corp filed Critical Sharp Corp
Publication of WO2015161823A1 publication Critical patent/WO2015161823A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing

Definitions

  • the present application generally relates to the field of human-computer interaction technology, and in particular to handwriting recognition.
  • an overlapping handwriting input recognition method is needed to identify a plurality of characters that are continuously input by the user in an overlapping manner in the same input area.
  • Chinese Patent No. CN102141892 B entitled “Overlay Handwriting Input Display Method and System” discloses a scheme in which the affiliation relationship of a stroke is determined according to the handwritten feature of the stroke and the positional relationship between adjacent strokes. And, based on the pause time between adjacent strokes, it is judged whether the input strokes constitute the same character.
  • segmentation hyphenation based on the pause time between adjacent strokes is not precise enough. For example, a user may pause or think a little while in the process of entering a complex character. Splitting the word break based on the pause will result in an erroneous recognition result. Although it is possible to distinguish between inter-character pauses and intra-character pauses by forcing the user to wait for a relatively long time before entering the next character, this does not conform to the handwriting habits of people's daily continuous input of characters, and is bound to reduce the speed and efficiency of handwriting input. .
  • the object of the present invention is to propose a new overlapping handwriting recognition scheme, which can not only recognize a plurality of characters continuously input by a user in an overlapping manner in the same input area, but also ensure comparison. High segmentation hyphenation accuracy and handwriting input efficiency.
  • a handwriting recognition method comprising: receiving a sequence of handwritten strokes continuously input by a user in the same input area; and segmenting the received sequence of handwritten strokes based on word confidence.
  • the segmentation hyphenation of the received handwritten stroke sequence based on the word confidence may include forward segmentation and/or reverse segmentation.
  • the forward segmentation determines the segmentation point of the received handwritten stroke sequence in the same order as the stroke input.
  • the reverse segmentation determines the segmentation point of the received handwritten stroke sequence in the reverse order of the stroke input.
  • the forward segmentation may include: reading a stroke after the upper point of the received handwritten stroke sequence into the forward segmentation set; and calculating, for each stroke in the forward segmentation set, the stroke and the stroke
  • the former stroke forms the credibility of the word; the gap between the stroke with the greatest degree of credibility and the subsequent stroke is determined as the cut point; and the above three steps are repeated.
  • the upper all points are located before the first input stroke.
  • the reverse segmentation may include: reading a stroke of the received handwritten stroke sequence before the upper point of the division into a reverse segmentation set; for each stroke in the reverse segmentation set, calculating the stroke and the stroke
  • the posterior stroke forms the credibility of the word; the gap between the stroke with the greatest degree of credibility and the previous stroke is determined as the cut point; and the above three steps are repeated.
  • the upper all points are located after the last input stroke.
  • the fine segmentation may include enumerating all the segmentation possibilities of the stroke, wherein each segmentation may correspond to a segmentation point configuration related to the number and position of the segmentation points; Possibly, calculating the credibility of the stroke between the cut points to form a single word, and determining the total credibility of the cut according to the calculated reliability of the single word; and the possible maximum score of the total credibility
  • the segmentation point configuration is determined as the segmentation result.
  • the method may further include determining whether there is an overlap region between the single words formed by the strokes between the cut points and a size of the overlap region; and determining the word based on the determining Whether it is a composite word.
  • the method may further include displaying or blanking the recognized complete character in a light color when the user inputs the stroke.
  • Displaying or blanking the recognized full character in a light color when the user inputs the stroke may include: after the user newly inputs a stroke, handwriting recognition of the stroke sequence input by the user, thereby identifying the character string; if newly input A stroke is the first stroke of the last character in the string and the second-to-last character in the string is the same as the last character of the string recognized by the user after the last stroke, or if a new one is entered
  • the stroke is not the first stroke of the last character in the string and the second-to-last character in the string is the same as the second-to-last character of the character string recognized by the user after inputting the last stroke, then the second to last Whether the number of strokes of the characters is greater than 2; and if the number of strokes of the second to last character is greater than 2, the second to last character and its previous characters are lightly displayed or blanked.
  • the segmentation of the received handwritten stroke sequence may also be based on the degree of matching of some or all of the strokes in the received handwritten stroke sequence with the overlapping character template.
  • Each overlapping character template can be composed of two overlapping characters.
  • text recognition is aided by language and/or writing rules.
  • a handwriting recognition apparatus comprising: receiving means for receiving a sequence of handwritten strokes continuously input by a user in the same input area; and cutting means for correcting based on word confidence The received sequence of handwritten strokes is segmented and broken.
  • the slitting device may comprise a forward slitting device and/or a reverse slitting device.
  • the forward severing means is operative to determine a puncturing point of the received handwritten stroke sequence in the same order as the stroke input.
  • the reverse slicing device is operative to determine a cut point of the received handwritten stroke sequence in an order opposite to the stroke input.
  • the forward segmentation device may include: a forward segmentation set forming unit, configured to read a stroke of the received handwritten stroke sequence after the upper point of the entry into the forward segmentation set; the word confidence calculation unit For calculating the credibility of the stroke and the strokes of the preceding stroke for each stroke in the forward segmentation set; the segmentation point determination unit is configured to use the stroke with the greatest degree of credibility and the subsequent stroke The gap between the gaps is determined as a split point; and a control unit for controlling the above three units to repeatedly perform respective functions.
  • Forward segmentation set forming unit When the function is first executed, the upper point is placed before the first input stroke.
  • the reverse segmentation device may include: a reverse segmentation set forming unit, configured to read a stroke of the received handwritten stroke sequence before the upper point of the entry into the reverse segmentation set; the word confidence calculation unit For calculating the credibility of the stroke and the subsequent stroke to form a single word for each stroke in the reverse segmentation set; the segmentation point determination unit is configured to use the stroke with the greatest degree of credibility and the previous stroke The gap between the gaps is determined as a split point; and a control unit for controlling the above three units to repeatedly perform respective functions.
  • the reverse-segment integration forming unit performs its function for the first time, the upper all points are located after the last input stroke.
  • the apparatus may further include: a fine segmentation device, configured to: before the non-coincidence segmentation point, the two segmentation points, in the case that the segmentation points determined by the forward segmentation and the reverse segmentation do not coincide
  • the strokes between the executions are finely divided.
  • the fine slicing device may comprise: a split possible enumeration unit for enumerating all the segmentation possibilities of the stroke, wherein each segmentation may correspond to a slice related to the number and position of the segmentation points.
  • the point-to-point configuration; the credibility calculation unit is configured to calculate the credibility of the strokes formed by the strokes between the segmentation points for each segmentation possibility, and determine the possible segmentation according to the calculated word confidence.
  • the total credibility; and the segmentation result determining unit is configured to determine the segmentation point configuration corresponding to the segmentation with the largest total credibility as the segmentation result.
  • the device may also include a post-processing device.
  • the post-processing device includes: an overlap region evaluation unit configured to determine whether an overlap region and a size of the overlap region exist between the words formed by the strokes between the segment points; and a synthesized word determining unit for determining, based on the determination, It is judged whether the word is a composite word.
  • the post-processing device may be further configured to display or blank out the recognized complete characters in a light color when the user inputs the stroke.
  • the post-processing device may further include: a character string identifying unit, configured to perform handwriting recognition on the stroke sequence input by the user after the user newly inputs a stroke, thereby identifying the character string; and the determining unit is configured to be newly input
  • a stroke is the first stroke of the last character in the string and the second-to-last character in the string is the same as the last character of the string recognized by the user after the last stroke, or in the new
  • the input stroke is not the first stroke of the last character in the string and the second-to-last character in the string is the same as the second-to-last character of the string recognized by the user after the last stroke.
  • the segmentation device may further perform segmentation and word segmentation on the received handwritten stroke sequence based on the degree of matching of some or all of the strokes in the received handwritten stroke sequence with the overlapping character template.
  • Each overlapping character template can be composed of two overlapping characters.
  • the post-processing device can be configured to utilize language and/or writing rules to aid in text recognition.
  • FIG. 1 is a flow chart showing a handwriting recognition method according to the present invention
  • FIG. 2 is a flow chart showing a forward severing operation in accordance with the present invention.
  • Figure 3 is a flow chart showing a reverse dicing operation in accordance with the present invention.
  • Figure 4 is a flow chart showing a fine segmentation operation in accordance with the present invention.
  • Figure 5 shows an exploded stroke of a single character "cut” and "minute”
  • Figure 6 shows the effect of the characters "cut” and "minute” overlap input
  • Figure 7 shows an example of a segmentation of a segmentation operation according to the present invention applied to an input stroke sequence of a string "segmentation";
  • Figure 8 is a flow chart showing the display or blanking of the identified complete characters in a light color when the user performs stroke input in accordance with the present invention
  • FIG. 9 shows the actual effect of applying the pre-word fade processing according to the present invention to the Japanese input " ⁇ " of the overlap input
  • Fig. 10 is a block diagram showing an example structure of a handwriting recognition apparatus according to the present invention.
  • the handwriting recognition method 100 starts at step s110, and receives a sequence of handwritten strokes that the user continuously inputs in the same input area.
  • the received handwritten stroke sequence is segmented and word-break based on the word confidence.
  • the template matching method may be used for single word recognition, and the matching distance is determined as the single word reliability in step s120.
  • the feature template of the template matching method can be generated using a sample training method based on a learning strategy such as generalized learning vector quantization (GLVQ).
  • GLVQ generalized learning vector quantization
  • Features used in single character recognition may include, for example, stroke direction distribution features, grid stroke features, perimeter orientation features, and the like.
  • Pre-processing before feature extraction may include, for example, equidistant smoothing, centroid-based linear normalization, nonlinear normalization, etc., to normalize all features.
  • a multi-stage cascade matching method can be employed.
  • the above-mentioned contents regarding the template matching method can be found in the Chinese patent CN 101354749 B entitled "Dictionary Making Method, Handwriting Input Method and Apparatus", and will not be described again here.
  • segmentation of the received handwritten stroke sequence based on the word confidence can significantly improve the segmentation hyphenation accuracy and the handwriting input efficiency.
  • step s120 may include forward segmentation and/or reverse segmentation (collectively referred to as coarse segmentation).
  • the forward segmentation determines the segmentation point of the received handwritten stroke sequence in the same order as the stroke input.
  • the reverse segmentation determines the segmentation point of the received handwritten stroke sequence in the reverse order of the stroke input.
  • the forward severing starts at step s201.
  • the forward slice set is set to an empty set.
  • the counter i is initialized to zero.
  • step s203 the counter i is incremented by one.
  • step s204 the stroke s i in the handwritten stroke sequence is added to the forward segmentation set S.
  • step s206 it is judged whether or not the counter i is equal to the total number L of strokes in the received handwritten stroke sequence.
  • step s206 the process proceeds to step s207, the search for the maximum value P k max ⁇ P k ⁇ .
  • step s208 the stroke index K corresponding to max ⁇ P k ⁇ is recorded, and the gap between the stroke and the subsequent stroke is determined to be recorded as a forward segmentation point.
  • step s209 the forward slice set S is emptied.
  • step s210 the counter i is set to K, and the process returns to step s203.
  • the reverse segmentation starts at step s301.
  • the reverse slice set is set to an empty set.
  • the counter i is initialized to L+1.
  • step s303 the counter i is decremented by one.
  • step s304 the strokes s i in the sequence of handwritten strokes are added to the inverse segmentation set S.
  • step s306 it is judged whether or not the counter i is equal to 1.
  • step s306 the process proceeds to step s307, the search for the maximum value P k max ⁇ P k ⁇ .
  • step s308 the stroke index K corresponding to max ⁇ P k ⁇ is recorded, and the gap between the stroke and the previous stroke is determined to be recorded as a reverse segmentation point.
  • step s309 the forward slice set S is emptied.
  • step s310 the counter i is set to K, and the process returns to step s303.
  • the segmentation points can be temporarily fixed. However, there may be cases where the forward severing and the reverse severing points do not completely coincide. In this case, preferably, the fine cut is performed for the stroke between the two cut points before and after the non-coinciding cut point.
  • step s401 the fine cut starts at step s401.
  • all the segmentation possibilities of the stroke are enumerated, wherein each of the segments may correspond to a segmentation point configuration related to the number and position of the segmentation points.
  • step s402 for each of the segmentation possibilities, the credibility of the stroke forming word between the segment points is calculated, and the total credibility of the segmentation may be determined according to the calculated word confidence.
  • step s403 the segmentation point configuration corresponding to the segmentation with the largest total reliability is determined as the fine segmentation result.
  • the forward segmentation points a 1 and a 2 can be obtained by performing the forward segmentation operation as shown in Fig. 7(a).
  • the reverse segmentation points b 1 , b 2 and b 3 can be obtained by performing a reverse dicing operation.
  • the forward segmentation points a 1 and a 2 coincide with the reverse segmentation points b 2 and b 1 , respectively, so that these segmentation points can be temporarily fixed.
  • the inverse segmentation point b 3 has no corresponding forward segmentation point, and thus two stroke subsequences that are not broken by the segmentation point before and after (ie, "- "with” ⁇ ") as a whole (ie, "- ⁇ ") Perform fine cuts.
  • first two potential cut points of s 1 and s 2 are added as shown in Fig. 7(c). All strokes are combined arbitrarily to form a combination of strokes such as C 1 , C 2 , ..., C 9 . Then, list all possible segmentation paths, such as (1) C 1 ; (2) C 2 C 9 ; (3) C 4 C 5 ; (4) C 4 C 8 C 9 ; For each possible segmentation path, first, each combination constituting the path is single-word recognized and its word confidence is calculated. Then, the total confidence of the segmentation path is calculated. Next, the segmentation path with the largest total reliability is selected, and the corresponding segmentation point is determined as the segmentation result. Among them, the method of calculating the optimal path can be performed by dynamic programming, N-best algorithm, and the like.
  • the N-best method is to select the best N paths, so that the sum of the values of the cost functions of all the paths passed is the smallest, the second smallest... the Nth is small.
  • the N-Best method can be implemented in a variety of ways, for example, combining a dynamic programming (DP) method with a stack algorithm to generate multiple candidates, and so on.
  • the N-Best method includes two steps: the forward search process adopts an improved Viterbi algorithm (the Viterbi algorithm is a dynamic plan for finding the most likely implicit state sequence).
  • the backward search process uses a stack algorithm based on the A* algorithm, for each A node m whose heuristic function is the sum of the following two functions: one is the "path cost function", which represents the sum of the cost function values of the shortest path from the starting point to the mth node, and the second is "inspiration".
  • the estimation function represents an estimate of the path cost from the mth node to the target node.
  • the path score in the stack is the calculated full path score, and the optimal path is always at the top of the stack. Therefore, the algorithm is a global optimal algorithm.
  • segmentation path C 1 has a greater overall reliability. Therefore, the corresponding cut point is selected as the fine cut result. Further, the segmentation point b 3 obtained in the reverse dicing operation is eliminated.
  • language recognition can be assisted by language and/or writing rules.
  • ⁇ and ⁇ promotion
  • ⁇ and ⁇ ⁇
  • ⁇ ⁇
  • ⁇ ⁇
  • the handwriting recognition method of the present invention supports displaying or blanking the recognized complete character in a light color when the user inputs the stroke. An example flow for implementing this function is described below with reference to FIG.
  • step s801 the counter n is initialized to zero.
  • step s802 the user is waited for a new stroke to be input, and after the user inputs a new stroke, the character string C 1 C 2 . . . C k is recognized by handwriting recognition of the stroke sequence input by the user.
  • step s803 it is judged that the newly input one stroke is the first stroke of the last character C k in the character string. If yes, go to step s804, otherwise go to step s805.
  • step s804 it is determined whether the second-to-last character C k-1 in the character string is the same as the last character C' k of the character string recognized by the user after the last stroke. If the same, step s806 is performed, otherwise step s809 is performed.
  • step s805 it is determined whether the second-to-last character C k-1 in the character string is the same as the second-to-last character C' k-1 of the character string recognized by the user after the last stroke. If the same, step s806 is performed, otherwise step s809 is performed.
  • step s806 the counter n is set to 1.
  • step s807 is performed to determine whether the number of strokes of the penultimate character C k-1 is greater than two. If yes, go to step s808, otherwise go back to step s802.
  • step s808 the penultimate character C k-1 and its previous characters are lightly displayed or blanked.
  • step s809 n is reset to zero. Then, it returns to step s802.
  • Table 1 gives a breakdown of the pre-word desalination process for the Japanese input " ⁇ " of the overlap input in tabular form.
  • the serial number column in Table 1 indicates the number of strokes input by the user (i.e., the number of rounds in which step s802 is performed).
  • step s807 it is determined that the fade processing is performed when the number of strokes of the second-to-last character is greater than 2: based on the following considerations: during the segmentation process, when the input strokes are small, the stroke sequence is often mis-cut into Single-stroke or two-stroke words (for example, Chinese characters "one", "two”, etc.). At this time, if the previous word is faded, it will result in an incorrect display. For example, in line No. 2 of Table 1, it will result in " "It is faded out.
  • Fig. 9 shows the actual effect of the above fade display processing. As can be seen from the figure, the user can clearly distinguish between the stroke of the entered character and the stroke of the character currently being written.
  • FIG. 10 shows a schematic block diagram of such a handwriting recognition device 1000.
  • the handwriting recognition apparatus 1000 includes a handwriting input device 1100, a handwritten information storage device 1200, a handwritten character string recognition device 1300, an identification candidate selection device 1400, and a display control device 1500.
  • the handwriting input device 1100 is configured to receive a sequence of strokes input by the user and digitize them to obtain handwritten handwriting for use by other devices.
  • the handwritten information storage device 1200 is used to store handwritten handwriting and other information generated during the handwriting process.
  • the handwritten character string recognition device 1300 may include a handwriting segmentation unit 1310, a single character/overlapping character recognition unit 1320, and a post processing unit 1330.
  • the handwriting segmentation unit 1310 can invoke the single character/overlapping character recognition unit 1320 to receive the received word based on the word confidence and also based on the degree of matching of some or all of the strokes in the received handwritten stroke sequence with the overlapping character template.
  • the sequence of handwritten strokes is used to segment and break words.
  • the post-processing unit 1330 may determine whether the recognized word constitutes a synthesized word; correct the recognition result based on language and/or writing rules; and/or display or blank the identified completeness in a light color when the user inputs the stroke character.
  • the recognition candidate selection means 1400 provides the user with an identification candidate for the user to select the correct recognition result.
  • the display control device 1500 controls display of display contents whose contents are constantly changing as the handwriting, the recognition candidate, and the final recognition result.
  • the handwriting recognition method and apparatus according to the present invention can be applied to various electronic devices that support handwriting input, such as an electronic whiteboard, a tablet computer, a desktop computer, a laptop computer, a personal digital assistant, a mobile phone, and the like.
  • electronic whiteboard a tablet computer
  • desktop computer a laptop computer
  • personal digital assistant a mobile phone
  • the principle applies to Chinese characters and Japanese, and also applies. In a variety of other languages (such as Korean).

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Character Discrimination (AREA)

Abstract

本申请公开了一种手写识别方法以及相应的手写识别设备。所述方法包括：接收用户在同一输入区域连续输入的手写笔画序列；以及基于单字可信度，对所接收的手写笔画序列进行切分断字。所公开的手写识别方法以及相应的手写识别设备不但能够识别用户在同一输入区域内以重叠覆盖的方式连续输入的多个字符，还能保障较高的切分断字精度和手写输入效率。

Description

手写识别方法和设备 技术领域
本申请总体涉及人机交互技术领域，具体涉及手写识别。
背景技术
随着移动通信技术的日益发展，具有触摸屏的智能终端变得越来越普及。为了借助触摸屏以手写方式输入信息的需要，手写识别技术在这些终端上得到了广泛的应用。
传统上，屏幕尺寸有限的智能终端采用基于单字符输入的手写识别技术。即，用户在预定的书写区内(如预置的书写框或整个屏幕上)逐字输入，在单字结束后提笔等待系统反馈。在得到系统反馈的单字识别结果后，清空书写屏幕继续下个字符的输入。然而，这样的输入方式不符合人们日常连续输入字符的书写习惯，而且提笔轮候和等待识别影响了输入效率。
为了改善用户手写体验和提高手写输入效率，需要一种重叠手写输入识别方法，用以识别用户在同一输入区域内以重叠覆盖的方式连续输入的多个字符。
为此，题为“叠加手写输入显示方法及系统”的中国专利CN102141892 B公开了一种方案，其中，根据笔画的手写特征和相邻笔画之间的位置关系，来确定笔画的隶属关系。并且，根据相邻笔画之间的停顿时间，来判断所输入的笔画是否组成同一字符。
然而，基于相邻笔画之间的停顿时间来进行切分断字不够精确。例如，用户在输入一个复杂字符的过程中可能会中途停下稍作思考或休息。基于该停顿进行切分断字将导致错误的识别结果。虽然能够通过强制用户在输入下一字符前等待相对较长的时间来区分字符间停顿和字符内停顿，但是这样做不符合人们日常连续输入字符的手写习惯，且势必降低手写输入的速度和效率。
发明内容
鉴于现有技术的上述问题和缺陷，本发明的目的在于提出一种新的重叠手写识别方案，不但能够识别用户在同一输入区域内以重叠覆盖的方式连续输入的多个字符，还能保障较高的切分断字精度和手写输入效率。
根据本发明的第一方面，提供了一种手写识别方法，包括：接收用户在同一输入区域连续输入的手写笔画序列；以及基于单字可信度，对所接收的手写笔画序列进行切分断字。
所述基于单字可信度对所接收的手写笔画序列进行切分断字可以包括前向切分和/或反向切分。所述前向切分按与笔画输入相同的顺序，确定所接收的手写笔画序列的切分点。所述反向切分按与笔画输入相反的顺序，确定所接收的手写笔画序列的切分点。
所述前向切分可以包括：将所接收的手写笔画序列中位于上一切分点之后的笔画读入前向切分集合；针对前向切分集合中的每个笔画，计算该笔画及其前笔画形成单字的可信度；将单字可信度最大的笔画与其后一笔画之间的间隙确定为切分点；以及重复执行上述三个步骤。当首次执行所述三个步骤时，所述上一切分点位于最先输入的笔画之前。
所述反向切分可以包括：将所接收的手写笔画序列中位于上一切分点之前的笔画读入反向切分集合；针对反向切分集合中的每个笔画，计算该笔画及其后笔画形成单字的可信度；将单字可信度最大的笔画与其前一笔画之间的间隙确定为切分点；以及重复执行上述三个步骤。当首次执行所述三个步骤时，所述上一切分点位于最后输入的笔画之后。
如果所述前向切分和反向切分确定的切分点不重合，可以针对非重合切分点前后两个切分点之间的笔画执行细切分。所述细切分可以包括：列举所述笔画的所有切分可能，其中，每一种切分可能对应于一种与切分点数目和位置有关的切分点配置；针对每一种切分可能，计算切分点之间的笔画形成单字的可信度，并根据所计算的单字可信度确定该切分可能的总可信度；以及将总可信度最大的切分可能所对应的切分点配置确定为切分结果。
所述方法还可以包括：确定切分点之间的笔画形成的单字之间是否存在重叠区域以及重叠区域的大小；以及基于所述确定，判断所述单字 是否是构成一个合成字。
所述方法还可以包括：在用户进行笔画输入时，以淡色显示或消隐已识别出的完整字符。
在用户进行笔画输入时以淡色显示或消隐已识别出的完整字符可以包括：在用户新输入了一个笔画后，对用户输入的笔画序列进行手写识别，从而识别出字符串；如果新输入的一个笔画是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的最后一个字符相同，或者如果新输入的一个笔画不是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的倒数第二个字符相同，则判断倒数第二个字符的笔画数是否大于2；以及如果所述倒数第二个字符的笔画数大于2，则对所述倒数第二个字符及其之前的字符进行淡色显示或消隐。
对所接收的手写笔画序列进行切分断字还可以基于所接收的手写笔画序列中的部分或全部笔画与重叠字符模板的匹配程度。每个重叠字符模板可以由两个重叠的字符构成。
优选地，利用语言和/或书写规则，来辅助文字识别。
根据本发明的第二方面，提供了一种手写识别设备，包括：接收装置，用于接收用户在同一输入区域连续输入的手写笔画序列；以及切分装置，用于基于单字可信度，对所接收的手写笔画序列进行切分断字。
所述切分装置可以包括前向切分装置和/或反向切分装置。所述前向切分装置用于按与笔画输入相同的顺序，确定所接收的手写笔画序列的切分点。所述反向切分装置用于按与笔画输入相反的顺序，确定所接收的手写笔画序列的切分点。
所述前向切分装置可以包括：前向切分集合形成单元，用于将所接收的手写笔画序列中位于上一切分点之后的笔画读入前向切分集合；单字可信度计算单元，用于针对前向切分集合中的每个笔画，计算该笔画及其前笔画形成单字的可信度；切分点确定单元，用于将单字可信度最大的笔画与其后一笔画之间的间隙确定为切分点；以及控制单元，用于控制上述三个单元重复执行各自的功能。当所述前向切分集合形成单元 首次执行其功能时，所述上一切分点位于最先输入的笔画之前。
所述反向切分装置可以包括：反向切分集合形成单元，用于将所接收的手写笔画序列中位于上一切分点之前的笔画读入反向切分集合；单字可信度计算单元，用于针对反向切分集合中的每个笔画，计算该笔画及其后笔画形成单字的可信度；切分点确定单元，用于将单字可信度最大的笔画与其前一笔画之间的间隙确定为切分点；以及控制单元，用于控制上述三个单元重复执行各自的功能。当所述反向切分集成形成单元首次执行其功能时，所述上一切分点位于最后输入的笔画之后。
所述设备还可以包括：细切分装置，用于在所述前向切分和反向切分确定的切分点不重合的情况下，针对非重合切分点前后两个切分点之间的笔画执行细切分。所述细切分装置可以包括：切分可能枚举单元，用于列举所述笔画的所有切分可能，其中，每一种切分可能对应于一种与切分点数目和位置有关的切分点配置；可信度计算单元，用于针对每一种切分可能，计算切分点之间的笔画形成单字的可信度，并根据所计算的单字可信度确定该切分可能的总可信度；以及切分结果确定单元，用于将总可信度最大的切分可能所对应的切分点配置确定为切分结果。
所述设备还可以包括后处理装置。所述后处理装置包括：重叠区域评估单元，用于确定切分点之间的笔画形成的单字之间是否存在重叠区域以及重叠区域的大小；以及合成字判定单元，用于基于所述确定，判断所述单字是否是构成一个合成字。
所述后处理装置还可以被配置为：在用户进行笔画输入时，以淡色显示或消隐已识别出的完整字符。
所述后处理装置还可以包括：字符串识别单元，用于在用户新输入了一个笔画后，对用户输入的笔画序列进行手写识别，从而识别出字符串；判断单元，用于在新输入的一个笔画是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的最后一个字符相同的情况下，或者在新输入的一个笔画不是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的倒数第二个字符相同的情况下，判断倒数第二个字符的笔画数是否大于2；以及淡色显示或消隐单元， 用于在所述倒数第二个字符的笔画数大于2的情况下，对所述倒数第二个字符及其之前的字符进行淡色显示或消隐。
所述切分装置还可以基于所接收的手写笔画序列中的部分或全部笔画与重叠字符模板的匹配程度，对所接收的手写笔画序列进行切分断字。每个重叠字符模板可以由两个重叠的字符构成。
后处理装置可以被配置为利用语言和/或书写规则，来辅助文字识别。
附图说明
通过下面结合附图说明本发明的优选实施例，将使本发明的上述及其它目的、特征和优点更加清楚，其中：
图1是示出了根据本发明的手写识别方法的流程图；
图2是示出了根据本发明的前向切分操作的流程图；
图3是示出了根据本发明的反向切分操作的流程图；
图4是示出了根据本发明的细切分操作的流程图；
图5示出了单个字符“切”和“分”的分解笔画；
图6示出了字符“切”和“分”重叠输入的效果；
图7示出了对字符串“切分”的输入笔画序列应用根据本发明的切分操作的切分示例；
图8示出了根据本发明的在用户进行笔画输入时以淡色显示或消隐已识别出的完整字符的流程图；
图9示出了对重叠输入的日语“にほん”应用根据本发明的前字淡化处理的实际效果；以及
图10示出了根据本发明的手写识别设备的示例结构的框图。
具体实施方式
下面参照附图对本发明的优选实施例进行详细说明，在描述过程中省略了对于本发明来说是不必要的细节和功能，以防止对本发明的理解造成混淆。
首先，参照图1，对根据本发明的手写识别方法100的过程进行描 述。如图1所示，手写识别方法100起始于步骤s110，接收用户在同一输入区域连续输入的手写笔画序列。接着，在步骤s120中，基于单字可信度，对所接收的手写笔画序列进行切分断字。为了实现步骤s120，可以采用模板匹配法进行单字识别，并将匹配距离确定为步骤s120中的单字可信度。
模板匹配法的特征模板可以基于学习策略(如广义学习矢量量化GLVQ)使用样本训练法来产生。在单字符识别中使用的特征可以包括例如笔画方向分布特征、网格笔画特征、周边方向特征等。特征提取前的预处理可以包括例如等距离平滑处理、基于质心的线性归一化、非线性归一化等，用以对全部特征进行归一化。为了提高识别速度，可以采用多级级联匹配法。关于模板匹配法的上述内容可在题为“字典制作方法、手写输入方法和设备”的中国专利CN 101354749 B中找到，此处不再赘述。
与基于相邻笔画之间的停顿时间来进行切分断字的现有技术相比，基于单字可信度对所接收的手写笔画序列进行切分断字能够显著提高切分断字精度和手写输入效率。
在具体实现中，步骤s120可以包括前向切分和/或反向切分(统称粗切分)。所述前向切分按与笔画输入相同的顺序，确定所接收的手写笔画序列的切分点。所述反向切分按与笔画输入相反的顺序，确定所接收的手写笔画序列的切分点。
以下参照图2和图3，描述前向切分和反向切分操作的示例实现。如图2所示，前向切分起始于步骤s201。在该步骤中，将前向切分集合设置为空集。在步骤s202中，将计数器i初始化为0。
接着，在步骤s203中，令计数器i加1。在步骤s204中，将手写笔画序列中的笔画si添加至前向切分集合S中。在步骤s205中，针对前向切分集合中的每个笔画sk(k＝1，……，i-1，i)，对该笔画及其前笔画形成单字进行单字识别，并计算其单字可信度Pk。在步骤s206中，判断计数器i是否等于所接收的手写笔画序列中的笔画总数L。
如果步骤s206中的判断结果为是，则前进至步骤s207，在Pk中搜索最大值max{Pk}。接着，在步骤s208中，记录max{Pk}所对应的笔画索 引K，将该笔画与其后一笔画之间的间隙确定为记录为前向切分点。在步骤s209中，将前向切分集合S清空。在步骤s210中，将计数器i设置为K，返回至步骤s203，此时，在步骤s205中，k不是从1开始了，而是从切分点后的笔画K+1开始，即，k＝K+1，……，i-1，i。如果步骤s206中的判断结果为否，则返回至步骤s203。
如图3所示，反向切分起始于步骤s301。在该步骤中，将反向切分集合设置为空集。在步骤s302中，将计数器i初始化为L+1。
接着，在步骤s303中，令计数器i减1。在步骤s304中，将手写笔画序列中的笔画si添加至反向切分集合S中。在步骤s305中，针对反向切分集合中的每个笔画sk(k＝i，i+1，……，L)，对该笔画及其后笔画形成单字进行单字识别，并计算其单字可信度Pk。在步骤s306中，判断计数器i是否等于1。
如果步骤s306中的判断结果为是，则前进至步骤s307，在Pk中搜索最大值max{Pk}。接着，在步骤s308中，记录max{Pk}所对应的笔画索引K，将该笔画与其前一笔画之间的间隙确定为记录为反向切分点。在步骤s309中，将前向切分集合S清空。在步骤s310中，将计数器i设置为K，返回至步骤s303，此时，在步骤s305中，k不是到L结束了，而是到反向切分点前的笔画K-1结束，即，k＝i，i+1，……，K-1。如果步骤s306中的判断结果为否，则返回至步骤s303。
在既执行前向切分又执行反向切分的情况下，如果前向切分点和反向切分点完全重合，则可以暂时固定住这些切分点。然而，可能存在前向切分和反向切分点不完全重合的情形。在该情形下，优选地，针对非重合切分点前后两个切分点之间的笔画执行细切分。
以下参考图4，描述根据本发明的细切分操作的流程图。如图所示，细切分起始于步骤s401。在该步骤中，列举所述笔画的所有切分可能，其中，每一种切分可能对应于一种与切分点数目和位置有关的切分点配置。接着，在步骤s402中，针对每一种切分可能，计算切分点之间的笔画形成单字的可信度，并根据所计算的单字可信度确定该切分可能的总可信度。最后，在步骤s403中，将总可信度最大的切分可能所对应的切分点配置确定为细切分结果。
下面，以对重叠输入的手写字符串“切分”的手写笔画序列进行切分断字为例，阐述根据本发明的切分断字操作。作为示意，图5示出了单个字符“切”和“分”的分解笔画，图6示出了重叠输入的效果。
假设通过执行前向切分操作，可以得到前向切分点a1和a2，如图7(a)所示。通过执行反向切分操作，可以得到反向切分点b1、b2和b3。前向切分点a1和a2分别与反向切分点b2和b1重合，因而可以暂时固定住这些切分点。反向切分点b3无对应的前向切分点，因而将其前后未被切分点断开的两段笔画子序列(即，“-
Figure PCTCN2015077367-appb-000001
”和“
Figure PCTCN2015077367-appb-000002
丿”)作为一个整体(即，“-
Figure PCTCN2015077367-appb-000003
Figure PCTCN2015077367-appb-000004
丿”)执行细切分。
为此，首先如图7(c)所示补充s1和s2两个潜在切分点。对所有笔画进行任意组合，形成C1，C2，……，C9等笔画组合。然后，列出所有可能的切分路径，如(1)C1；(2)C2C9；(3)C4C5；(4)C4C8C9；……等。针对每一种可能的切分路径，首先，对构成该路径的每个组合进行单字识别并计算其单字可信度。然后，计算该切分路径的总可信度。接着，选择总可信度最大的切分路径，并将其所对应的切分点确定为切分结果。其中，计算切分最佳路径的方法可以用动态规划，也可以用N-best算法，等等。
在采用N-Best方法的情况下，计算最可能的N种切分路径。定义每个笔画的开始点为一个基元节点，基元或基元组合构成的路径即为对应的笔画组合，每个部分路径的代价函数为：C(Y)＝1-f(Y)，也就是说，切分可信度越高，部分路径的代价函数值越小。N-best方法就是要选取最佳的N种路径，使所经过的所有路径的代价函数的数值之和最小、第二小……第N小。
N-Best方法可以用多种方式实现，例如，把动态规划(DP)方法与堆栈(Stack)算法相结合来产生多个候选项，等等。本发明实施例中，N-Best方法包括两个步骤：前向搜索过程采用一种改进的维特比(Viterbi)算法(维特比算法就是一种用于查找最可能的隐含状态序列的动态规划方法)，用来记录转移到每个基元节点的最优N个部分路径的状态(即为所经过路径的代价函数值之和)；第m个基元节点的状态只和第m-1个基元结点的状态有关；后向搜索过程采用一种基于A*算法的堆栈算法，对每 一个节点m，它的启发函数(heuristic function)为下列两个函数的和：一是“路径代价函数”，表示从起始点到第m节点的最短路径的代价函数值之和，二是“启发估计函数”，表示从第m节点到目标节点的路径代价的估计。在后向搜索过程中，堆栈中的路径得分是计算的全路径得分，且最优的路径总是位于栈顶，所以，该算法是一种全局最优算法。
以图7(c)所示的情形为例，与其他切分路径相比，切分路径C1具有更大的总可信度。因此，选择与其对应的切分点作为细切分结果。进而，消除了在反向切分操作中获得的切分点b3。
在确定了切分点之后，可以读取在执行切分操作的过程中识别的单字，作为手写识别的结果。仍以图7为例，在经过粗切分之后，确定了切分点a1＝b2以及a2＝b1。经过细切分之后，未增加新的切分点。进而，手写识别结果可以被读取为在执行粗切分操作的过程中识别出的单字“切”、“八”、“刀”。可以对手写识别结果进行后处理，以优化识别准确度。
在具体实现中，可以确定切分点之间的笔画形成的单字之间是否存在重叠区域以及重叠区域的大小。基于确定结果，判断所述单字是否是构成一个合成字。通常，重叠区域越小，构成合成字的可能性就越大；重叠区域越大，构成合成字的可能性就越小。例如，根据“八”和“刀”无重叠区域或重叠区域极小，可以判断出两者构成合成字“分”。
此外，可以利用语言和/或书写规则，来辅助文字识别。例如，当对重叠输入的平假名序列进行识别时，可以采用以下方式来区分大写和小写假名：つ和っ(促音)；以及やゆよ和ゃゅょ(拗音)。具体地，对于ゃゅょ(拗音)，如果其之前输入的字符是“きぎしじちぢにひぴびみり”之一并且其尺寸明显小于之前输入的字符，则将其确定为小写字符。否则，将其确定为大小字符。对于っ(促音)，可以首先将其尺寸与其上下文的多个字符进行比较，然后利用一些规则(如字典匹配规则)来确定其是小写字符还是大写字符。
为了进一步提高识别准确度，可以考虑训练重叠字符模板，并基于所接收的手写笔画序列中的部分或全部笔画与重叠字符模板的匹配程度，对手写笔画序列进行切分断字。以两字符平假名重叠为例，可以将 84个平假名中的每一个与84个平假名依此进行组合，形成84*84个重叠字符模板“ああ”，“あい”，“あう”，“あえ”，“あお”，…，“あん”，..，“いう”，“いえ”，“いお”，…“いん”，...，等。
为了便于用户区分已录入的字符的笔画和当前正在书写的字符的笔画，本发明的手写识别方法支持在用户进行笔画输入时，以淡色显示或消隐已识别出的完整字符。下面参照图8描述实现该功能的示例流程。
首先，在步骤s801中，将计数器n初始化为0。在步骤s802中，等待用户输入新笔画，并在用户输入了新笔画后，通过对用户输入的笔画序列进行手写识别，识别出字符串C1C2……Ck。
接着，在步骤s803中，判断新输入的一个笔画是所述字符串中最后一个字符Ck的第一笔。如果是则执行步骤s804，否则执行步骤s805。在步骤s804中，判断所述字符串中倒数第二个字符Ck-1是否与用户输入上一笔画后识别出的字符串的最后一个字符C’k相同。如果相同则执行步骤s806，否则执行步骤s809。在步骤s805中，判断所述字符串中倒数第二个字符Ck-1是否与用户输入上一笔画后识别出的字符串的倒数第二个字符C’k-1相同。如果相同则执行步骤s806，否则执行步骤s809。
在步骤s806中，将计数器n设置为1。接着，执行步骤s807，判断倒数第二个字符Ck-1的笔画数是否大于2。如果是则执行步骤s808，否则返回步骤s802。
在步骤s808中，将倒数第二个字符Ck-1及其之前的字符进行淡色显示或消隐。在步骤s809中，将n重置为0。然后，返回步骤s802。
表1以表格形式给出了对重叠输入的日语“にはん”进行前字淡化处理的分解过程。表1中的序号栏指示用户输入的笔画数(即执行步骤s802的回合数)。
表1
Figure PCTCN2015077367-appb-000005
Figure PCTCN2015077367-appb-000006
需要说明的是，在步骤s807中判断倒数第二个字符的笔画数大于2时才进行淡化处理是基于以下考虑：在切分过程中，当输入笔画较少时，笔画序列常会被误切分成单笔画或两笔画的字(如，汉字“一”、“二”等)。此时，如果把前字淡化显示，将导致不正确显示效果。例如，在表1的No.2一行，将导致“
Figure PCTCN2015077367-appb-000007
”被淡化显示。
图9给出了上述淡化显示处理的实际效果。由图可见，用户能够清楚地区分已录入的字符的笔画和当前正在书写的字符的笔画。
与上述手写识别方法相对应地，本发明还提出了相关的手写识别设备。图10示出了这样的手写识别设备1000的示意结构方框图。
如图所示，根据本发明的手写识别设备1000包括手写笔迹输入装置1100、手写信息存储装置1200、手写字符串识别装置1300、识别候选选择装置1400以及显示控制装置1500。
手写笔迹输入装置1100用于接收用户输入的笔画序列，并对其进行数字化，以得到手写笔迹，供其他装置使用。手写信息存储装置1200用于存储手写笔迹以及在手写过程中产生的其他信息。
手写字符串识别装置1300可以包括手写切分单元1310、单字符/重叠字符识别单元1320以及后处理单元1330。手写切分单元1310可以调用单字符/重叠字符识别单元1320，以基于单字可信度以及还可以基于所接收的手写笔画序列中的部分或全部笔画与重叠字符模板的匹配程度，对所接收的手写笔画序列进行切分断字。后处理单元1330可以判断所识别的单字是否构成合成字；基于语言和/或书写规则，对识别结果进行校正；和/或在用户进行笔画输入时，以淡色显示或消隐已识别出的完整字符。
识别候选选择装置1400向用户提供识别候选，供用户从中选择正确的识别结果。显示控制装置1500控制显示随手写笔迹、识别候选和最终识别结果等内容不断变化的显示内容。
根据本发明的手写识别方法和设备可以应用于支持手写输入的多种电子设备，如，电子白板、平板计算机、台式计算机、膝上型计算机、个人数字助理、移动电话等。此外，其原理适用于汉字和日文，还适用 于其他多种文字(例如韩文)。
应当注意的是，在以上的描述中，仅以示例的方式，示出了本发明的技术方案，但并不意味着本发明局限于上述步骤和单元结构。在可能的情形下，可以根据需要对步骤和单元结构进行调整和取舍。因此，某些步骤和单元并非实施本发明的总体发明思想所必需的元素。因此，本发明所必需的技术特征仅受限于能够实现本发明的总体发明思想的最低要求，而不受以上具体实例的限制。
至此已经结合优选实施例对本发明进行了描述。应该理解，本领域技术人员在不脱离本发明的精神和范围的情况下，可以进行各种其它的改变、替换和添加。因此，本发明的范围不局限于上述特定实施例，而应由所附权利要求所限定。

Claims (22)

  1. 一种手写识别方法，包括：
    接收用户在同一输入区域连续输入的手写笔画序列；以及
    基于单字可信度，对所接收的手写笔画序列进行切分断字。
  2. 根据权利要求1所述的方法，其中，所述基于单字可信度对所接收的手写笔画序列进行切分断字包括前向切分和/或反向切分，
    所述前向切分按与笔画输入相同的顺序，确定所接收的手写笔画序列的切分点，
    所述反向切分按与笔画输入相反的顺序，确定所接收的手写笔画序列的切分点。
  3. 根据权利要求2所述的方法，其中，所述前向切分包括：
    将所接收的手写笔画序列中位于上一切分点之后的笔画读入前向切分集合；
    针对前向切分集合中的每个笔画，计算该笔画及其前笔画形成单字的可信度；
    将单字可信度最大的笔画与其后一笔画之间的间隙确定为切分点；以及
    重复执行上述三个步骤，
    其中，当首次执行所述三个步骤时，所述上一切分点位于最先输入的笔画之前。
  4. 根据权利要求2所述的方法，其中，所述反向切分包括：
    将所接收的手写笔画序列中位于上一切分点之前的笔画读入反向切分集合；
    针对反向切分集合中的每个笔画，计算该笔画及其后笔画形成单字的可信度；
    将单字可信度最大的笔画与其前一笔画之间的间隙确定为切分点；以及
    重复执行上述三个步骤，
    其中，当首次执行所述三个步骤时，所述上一切分点位于最后输入的笔画之后。
  5. 根据权利要求2所述的方法，其中，如果所述前向切分和反向切分确定的切分点不重合，则针对非重合切分点前后两个切分点之间的笔画执行细切分，其中，所述细切分包括：
    列举所述笔画的所有切分可能，其中，每一种切分可能对应于一种与切分点数目和位置有关的切分点配置；
    针对每一种切分可能，计算切分点之间的笔画形成单字的可信度，并根据所计算的单字可信度确定该切分可能的总可信度；以及
    将总可信度最大的切分可能所对应的切分点配置确定为切分结果。
  6. 根据权利要求1所述的方法，还包括：
    确定切分点之间的笔画形成的单字之间是否存在重叠区域以及重叠区域的大小；以及
    基于所述确定，判断所述单字是否是构成一个合成字。
  7. 根据权利要求1所述的方法，还包括：在用户进行笔画输入时，以淡色显示或消隐已识别出的完整字符。
  8. 根据权利要求7所述的方法，其中，在用户进行笔画输入时以淡色显示或消隐已识别出的完整字符包括：
    在用户新输入了一个笔画后，对用户输入的笔画序列进行手写识别，从而识别出字符串；
    如果新输入的一个笔画是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的最后一个字符相同，或者如果新输入的一个笔画不是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的倒数第二个字符相同，则判断倒数第二个字符的笔画数是否大于2；以及
    如果所述倒数第二个字符的笔画数大于2，则对所述倒数第二个字符及其之前的字符进行淡色显示或消隐。
  9. 根据权利要求1所述的方法，其中，对所接收的手写笔画序列进行切分断字还基于所接收的手写笔画序列中的部分或全部笔画与重叠字符模板的匹配程度。
  10. 根据权利要求9所述的方法，其中，每个重叠字符模板由两个重叠的字符构成。
  11. 根据权利要求1所述的方法，其中，利用语言和/或书写规则，来辅助文字识别。
  12. 一种手写识别设备，包括：
    接收装置，用于接收用户在同一输入区域连续输入的手写笔画序列；以及
    切分装置，用于基于单字可信度，对所接收的手写笔画序列进行切分断字。
  13. 根据权利要求12所述的设备，其中，所述切分装置包括前向切分装置和/或反向切分装置，
    所述前向切分装置用于按与笔画输入相同的顺序，确定所接收的手写笔画序列的切分点，
    所述反向切分装置用于按与笔画输入相反的顺序，确定所接收的手写笔画序列的切分点。
  14. 根据权利要求13所述的设备，其中，所述前向切分装置包括：
    前向切分集合形成单元，用于将所接收的手写笔画序列中位于上一切分点之后的笔画读入前向切分集合；
    单字可信度计算单元，用于针对前向切分集合中的每个笔画，计算该笔画及其前笔画形成单字的可信度；
    切分点确定单元，用于将单字可信度最大的笔画与其后一笔画之间的间隙确定为切分点；以及
    控制单元，用于控制上述三个单元重复执行各自的功能，
    其中，当所述前向切分集合形成单元首次执行其功能时，所述上一切分点位于最先输入的笔画之前。
  15. 根据权利要求13所述的设备，其中，所述反向切分装置包括：
    反向切分集合形成单元，用于将所接收的手写笔画序列中位于上一切分点之前的笔画读入反向切分集合；
    单字可信度计算单元，用于针对反向切分集合中的每个笔画，计算该笔画及其后笔画形成单字的可信度；
    切分点确定单元，用于将单字可信度最大的笔画与其前一笔画之间的间隙确定为切分点；以及
    控制单元，用于控制上述三个单元重复执行各自的功能，
    其中，当所述反向切分集成形成单元首次执行其功能时，所述上一切分点位于最后输入的笔画之后。
  16. 根据权利要求13所述的设备，还包括：细切分装置，用于在所述前向切分和反向切分确定的切分点不重合的情况下，针对非重合切分点前后两个切分点之间的笔画执行细切分，其中，
    所述细切分装置包括：
    切分可能枚举单元，用于列举所述笔画的所有切分可能，其中，每一种切分可能对应于一种与切分点数目和位置有关的切分点配置；
    可信度计算单元，用于针对每一种切分可能，计算切分点之间的笔画形成单字的可信度，并根据所计算的单字可信度确定该切分可能的总可信度；以及
    切分结果确定单元，用于将总可信度最大的切分可能所对应的切分点配置确定为切分结果。
  17. 根据权利要求12所述的设备，还包括后处理装置，所述后处理装置包括：
    重叠区域评估单元，用于确定切分点之间的笔画形成的单字之间是否存在重叠区域以及重叠区域的大小；以及
    合成字判定单元，用于基于所述确定，判断所述单字是否是构成一个合成字。
  18. 根据权利要求12所述的设备，其中，所述后处理装置还被配置为：在用户进行笔画输入时，以淡色显示或消隐已识别出的完整字符。
  19. 根据权利要求18所述的设备，其中，所述后处理装置还包括：
    字符串识别单元，用于在用户新输入了一个笔画后，对用户输入的笔画序列进行手写识别，从而识别出字符串；
    判断单元，用于在新输入的一个笔画是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的最后一个字符相同的情况下，或者在新输入的一个笔画不是所述字符串中最后一个字符的第一笔且所述字符串中倒数第二个字符与用户输入上一笔画后识别出的字符串的倒数第二个字符相同的情况下，判断倒数第二个字符的笔画数是否大于2；以及
    淡色显示或消隐单元，用于在所述倒数第二个字符的笔画数大于2的情况下，对所述倒数第二个字符及其之前的字符进行淡色显示或消隐。
  20. 根据权利要求12所述的设备，其中，所述切分装置还基于所接收的手写笔画序列中的部分或全部笔画与重叠字符模板的匹配程度，对所接收的手写笔画序列进行切分断字。
  21. 根据权利要求20所述的设备，其中，每个重叠字符模板由两个重叠的字符构成。
  22. 根据权利要求12所述的设备，还包括：后处理装置，被配置为利用语言和/或书写规则，来辅助文字识别。
PCT/CN2015/077367 2014-04-25 2015-04-24 手写识别方法和设备 Ceased WO2015161823A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201410171650.2A CN105095924A (zh) 2014-04-25 2014-04-25 手写识别方法和设备
CN201410171650.2 2014-04-25

Publications (1)

Publication Number Publication Date
WO2015161823A1 true WO2015161823A1 (zh) 2015-10-29

Family

ID=54331772

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/077367 Ceased WO2015161823A1 (zh) 2014-04-25 2015-04-24 手写识别方法和设备

Country Status (2)

Country Link
CN (1) CN105095924A (zh)
WO (1) WO2015161823A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112613512A (zh) * 2020-12-29 2021-04-06 西北民族大学 基于结构属性的乌金体藏文古籍字符切分方法及系统
CN113468972A (zh) * 2021-06-07 2021-10-01 中金金融认证中心有限公司 用于复杂场景手写识别的手写轨迹切分方法及计算机产品
CN113641253A (zh) * 2021-07-09 2021-11-12 北京搜狗科技发展有限公司 筛选候选项的方法、装置和介质
CN115100748A (zh) * 2022-07-05 2022-09-23 重庆傲雄在线信息技术有限公司 一种基于签名笔画序列的单字提取方法及系统
CN117519515A (zh) * 2024-01-05 2024-02-06 深圳市方成教学设备有限公司 一种用于记忆黑板的文字识别方法、装置和记忆黑板

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105654087B (zh) * 2015-12-30 2019-03-12 李宇 一种基于彩色模板的脱机手写字符提取方法
CN107368248B (zh) * 2017-06-19 2020-03-17 广东小天才科技有限公司 一种书写笔迹的回放方法及装置
CN111931710B (zh) * 2020-09-17 2021-03-30 开立生物医疗科技(武汉)有限公司 一种联机手写文字识别方法、装置、电子设备及存储介质
CN118887688B (zh) * 2024-06-29 2025-07-18 华为技术有限公司 一种手写笔画的识别方法以及电子设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2007188512A (ja) * 2000-09-29 2007-07-26 Japan Science & Technology Agency 文字認識方法、文字認識プログラム及び文字認識プログラムを記録したコンピュータ読み取り可能な記録媒体
CN101484907A (zh) * 2006-07-06 2009-07-15 辛纳普蒂克斯公司 用于手写体符号的识别的方法和设备
CN102156577A (zh) * 2011-03-28 2011-08-17 安徽科大讯飞信息科技股份有限公司 实现连续手写识别输入的方法及系统
CN102855082A (zh) * 2011-06-13 2013-01-02 谷歌公司 用于重叠文本用户输入的字符识别
CN103080878A (zh) * 2010-08-24 2013-05-01 诺基亚公司 用于将重叠手写的笔画分割成一组或多组的方法和装置

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102394061B (zh) * 2011-11-08 2013-01-02 中国农业大学 基于语义检索的文语转换方法及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2007188512A (ja) * 2000-09-29 2007-07-26 Japan Science & Technology Agency 文字認識方法、文字認識プログラム及び文字認識プログラムを記録したコンピュータ読み取り可能な記録媒体
CN101484907A (zh) * 2006-07-06 2009-07-15 辛纳普蒂克斯公司 用于手写体符号的识别的方法和设备
CN103080878A (zh) * 2010-08-24 2013-05-01 诺基亚公司 用于将重叠手写的笔画分割成一组或多组的方法和装置
CN102156577A (zh) * 2011-03-28 2011-08-17 安徽科大讯飞信息科技股份有限公司 实现连续手写识别输入的方法及系统
CN102855082A (zh) * 2011-06-13 2013-01-02 谷歌公司 用于重叠文本用户输入的字符识别

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112613512A (zh) * 2020-12-29 2021-04-06 西北民族大学 基于结构属性的乌金体藏文古籍字符切分方法及系统
CN112613512B (zh) * 2020-12-29 2022-08-12 西北民族大学 基于结构属性的乌金体藏文古籍字符切分方法及系统
CN113468972A (zh) * 2021-06-07 2021-10-01 中金金融认证中心有限公司 用于复杂场景手写识别的手写轨迹切分方法及计算机产品
CN113468972B (zh) * 2021-06-07 2024-02-27 中金金融认证中心有限公司 用于复杂场景手写识别的手写轨迹切分方法及计算机产品
CN113641253A (zh) * 2021-07-09 2021-11-12 北京搜狗科技发展有限公司 筛选候选项的方法、装置和介质
CN115100748A (zh) * 2022-07-05 2022-09-23 重庆傲雄在线信息技术有限公司 一种基于签名笔画序列的单字提取方法及系统
CN115100748B (zh) * 2022-07-05 2025-07-29 重庆亲笔签数字科技有限公司 一种基于签名笔画序列的单字提取方法及系统
CN117519515A (zh) * 2024-01-05 2024-02-06 深圳市方成教学设备有限公司 一种用于记忆黑板的文字识别方法、装置和记忆黑板
CN117519515B (zh) * 2024-01-05 2024-05-28 深圳市方成教学设备有限公司 一种用于记忆黑板的文字识别方法、装置和记忆黑板

Also Published As

Publication number Publication date
CN105095924A (zh) 2015-11-25

Similar Documents

Publication Publication Date Title
WO2015161823A1 (zh) 手写识别方法和设备
US20210406578A1 (en) Handwriting-based predictive population of partial virtual keyboards
US11640503B2 (en) Input method, input device and apparatus for input
CN114365075B (zh) 用于选择图形对象的方法和对应装置
US7778464B2 (en) Apparatus and method for searching for digital ink query
US10325018B2 (en) Techniques for scheduling language models and character recognition models for handwriting inputs
JP2002203208A (ja) オンライン文字認識装置及び方法並びにコンピュータ読み取り可能な記憶媒体及びオンライン文字認識プログラム
KR20120011010A (ko) 수기 인식 방법 및 장치
CN107346183B (zh) 一种词汇推荐方法和电子设备
US20140184610A1 (en) Shaping device and shaping method
US20140297276A1 (en) Editing apparatus, editing method, and computer program product
US20120109633A1 (en) Method and system for diacritizing arabic language text
JP2013206141A (ja) 文字入力装置、文字入力方法、及び文字入力プログラム
JP2012173959A (ja) 文字認識装置及びプログラム
US9384304B2 (en) Document search apparatus, document search method, and program product
CN115004262B (zh) 处理手写中列表的方法和计算装置
JP6735775B2 (ja) 重畳手書き入力認識技術のためのシステム及び方法
JP2003196593A (ja) 文字認識装置および文字認識方法および文字認識プログラム
CN104635949A (zh) 一种汉字输入装置及方法
JP2012108893A (ja) 手描き入力方法
JP6772629B2 (ja) 情報処理装置、文字入力プログラムおよび文字入力方法
CN106293368B (zh) 一种数据处理方法及电子设备
CN105683873A (zh) 容错输入法编辑器
Nguyen et al. Semi-incremental recognition of on-line handwritten Japanese text
CN105094544B (zh) 一种颜文字的获取方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15783294

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15783294

Country of ref document: EP

Kind code of ref document: A1