JP2002063548A - Handwritten character recognizing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a handwritten character recognizing method which can accurately segment even obliquely handwritten characters and handwritten characters having narrow character intervals and recognize the handwritten characters in an arbitrary line together at a time according to the segmentation result. SOLUTION: Strokes which are smaller in distance, obtained by evaluating the distance between strokes constituting stroke groups according to a predetermined relational expression, than a threshold for temporary combination are combined and the stroke groups are divided into character elements. Then, the circumscribed rectangle of each character element is found and the maximum height or minimum width of the circumscribed rectangle is estimated as the standard character size of a handwritten character; and character elements having parameters, representing the relation between adjacent character elements in a space of the estimated standard character size, smaller than the threshold for temporary combination are combined, character elements are divided into character element sets, and the character recognition of the character element set is carried out by using a dictionary.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a handwritten character string recognition method for recognizing handwritten characters input from a handwritten character input device such as a tablet or an electronic blackboard.

[0002]

2. Description of the Related Art Heretofore, the following has been proposed as this kind of handwritten character recognition method or recognition device. (1) JP-A-61-29982 (name: online handwritten character string recognition method) (2) JP-A-5-174185 (name: Japanese character recognition device) (3) JP-A-6-162269 (name: Handwritten character recognition device) (4) JP-A-8-50632 (name: handwritten character cutout method and device)

The on-line handwritten character string recognition method disclosed in Japanese Patent Application Laid-Open No. 61-29982 solves a restriction in recognizing a character string written in a free form on a data tablet, and performs character segmentation. In order to perform correctly, the stroke sequence input from the data tablet is divided into a plurality of basic segment sequences, then the basic segments are combined to generate candidate characters, and then the generated candidate characters are standardized. The process of successively recognizing characters by collation with characters and accumulating the character names and the degree of difference of the recognition result is repeatedly executed for all candidate characters, and the character name of the character sequence that minimizes the sum of the degrees of difference for the input stroke sequence A sequence is assigned using a minimum route search algorithm.

The Japanese character recognition device disclosed in Japanese Patent Application Laid-Open No. 5-174185 aims at minimizing erroneous segmentation and erroneous recognition of a Japanese character string input online or offline from a scanner or the like. Detects a range of character strings where there may be separation characters or half-width characters, finds all cutout candidates in that range, performs recognition, and determines the cutoff priority and the recognition similarity the most. Extracts circumscribed figures of connected parts of character parts in order to output unique recognition character codes, and integrates when adjacent circumscribed figures overlap in the vertical direction for horizontal writing documents and in the horizontal direction for vertical writing documents To create a basic rectangle, determine whether the basic rectangle can be determined as one character by itself, if not, detect the range of the basic rectangle, In the range of, a combination of integration of adjacent basic rectangles is obtained as a cutout candidate, priorities are assigned to each of them, all cutout candidates are recognized, and a recognized character code that is more certain than the cutout priority and the recognition similarity is output. It is like that.

A handwritten character recognition device disclosed in Japanese Patent Laid-Open No. 6-162269 has an object to enable a handwritten character to be smoothly input at an arbitrary position at an arbitrary speed. , The position of the starting point is detected, the coordinate data is identified in character units, and the character represented by the stroke is recognized by the coordinate data in character units.

A method and apparatus for extracting a handwritten character disclosed in Japanese Patent Application Laid-Open No. Hei 8-50632 is intended to enable the extraction of a character without providing an input frame. Is determined based on the character string height H, a range of the width L in the horizontal direction from the base point O is set as a preliminary search range, and the number of strokes S, the maximum height h, and the like within the preliminary search range. The shape feature x (maximum blank length) is obtained, a search range is determined according to the variables S, h, x, and a section where the histogram takes the minimum value within the search range is searched. One character is cut out assuming that the longest one is a break between subsequent characters.

[0007]

However, in the handwritten character recognition methods described in the above publications, the writing direction is specified in advance as horizontal writing or vertical writing or fixed. Since it is assumed that the line feed position is also specified, a handwritten character document in which the writing direction and line feed position are not specified, for example, a multi-line handwritten document written on an electronic blackboard is imported online, and There is a problem that it cannot be recognized collectively.

In the online handwritten character string recognition system disclosed in Japanese Patent Application Laid-Open No. 61-29982,
As a method of dividing the input stroke sequence into basic segment sequences, for the horizontal handwritten input character pattern, the ratio of the overlapping degree of projection of each stroke to the horizontal axis and the height of the circumscribed figure of the handwritten input character pattern Since the stroke is divided by comparing it with the threshold value and each set of divided strokes is used as a basic segment, when a handwritten character is written in a diagonal direction, the height of the circumscribed figure is abnormally larger than the character height. As a result, as a result, the set is divided into a set of basic segment strings including the segment strings constituting the adjacent characters. As a result, there is a problem that a handwritten input character written in a slanted direction cannot be correctly recognized.

Further, in the Japanese character recognition device disclosed in Japanese Patent Laid-Open No. 5-174185, strokes overlapping in the vertical direction for horizontal writing and in the horizontal direction for vertical writing are combined with one another. Since the basic segment can be composed of two characters, that is, it is divided into basic segments by a deterministic method of determining whether there is an overlap or not, the circumscribed figure between adjacent characters overlaps with a narrow character interval. In such a case, there is a risk that the strokes of a plurality of characters are integrated into a basic segment of one character, and there is a risk that handwritten input characters with a narrow character interval cannot be correctly recognized.

In the handwritten character recognition device disclosed in Japanese Patent Application Laid-Open No. 6-162269, when cutting out a plurality of handwritten characters one by one, attention is paid to the starting point of the first stroke, and The start point position of the last stroke is below a predetermined threshold and the first point of the character is
If it is detected that the starting point position of the stroke is above the threshold value, this part is set as a candidate for the boundary of one character, and then the distance and direction between the starting point of the first stroke of the immediately preceding character and the first stroke of the target character If the distance between the starting points is larger than the threshold value and is the same as the character input direction,
A cutout candidate for one character is determined, a circumscribing box of the cutout candidate is created, and an overlapping relationship with the circumscribing box created immediately before is checked. If it does not, the stroke group of the previous cut candidate is cut out as one character, so the vertical stroke format where the start point of the first stroke is always lower than the start point of the last stroke of the previous character There is a problem that it cannot be applied to handwritten character recognition. Even in the horizontal writing format, for the same reason, in the case where the characters on the entire line are handwritten characters in the diagonal writing format in which the characters are inclined to the lower right, the starting point of the first stroke of the target character is set to the position of the immediately preceding character. If the first stroke is lower than the starting point of the last stroke, the first stroke is classified as a stroke constituting the immediately preceding character, and there is a problem that the clipping is not performed correctly.

In the handwritten character extracting method disclosed in Japanese Patent Application Laid-Open No. 8-50632, a height H of an input handwritten character string is obtained, and a width L is determined based on the character string height H. Is determined, and the range of the width L in the horizontal direction from the base point O is set as the preliminary search range, and within the preliminary search range, the number S of strokes, the maximum height h, and the shape feature x (the maximum blank length) are obtained. A search range is determined according to the variables S, h, and x, and a section where the histogram takes the minimum value is searched within the search range, and the longest one of the sections is a break between the subsequent character. Since one character is cut out, for example, when a three-digit number "111" is written vertically, these numbers are cut out as a stroke sequence that constitutes one character, and are called a kanji character "kawa". Misrecognized by letters Mau there is a risk. In addition, for handwritten characters of a plurality of lines, line division is performed at line break positions, but how to detect line break positions is not taken into consideration. Therefore, there is a problem that handwritten characters written over a plurality of lines cannot be collectively recognized for each line.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and it is an object of the present invention to accurately cut out each character even if it is a diagonal writing or a handwritten character with a narrow character spacing. It is another object of the present invention to provide a handwritten character recognition method capable of collectively recognizing handwritten characters on an arbitrary line according to the cutout result.

Another object of the present invention is to take in a handwritten character written on an electronic blackboard or the like without specifying a writing direction online, accurately determine the writing direction of the handwritten character, and according to the determination result. An object of the present invention is to provide a handwritten character recognition method capable of recognizing handwritten characters collectively.

Another object of the present invention is to take in a handwritten character written on an electronic blackboard or the like without specifying a line feed position online, determine the line feed position of the handwritten character accurately, and according to the determination result. An object of the present invention is to provide a handwritten character recognition method capable of collectively recognizing handwritten characters over a plurality of lines.

Another object of the present invention is to fetch online handwritten characters on an electronic blackboard or the like and collectively write the handwritten characters irrespective of the type of vertical and horizontal writing, the number of lines, and the presence or absence of a writing frame. It is an object of the present invention to provide a handwritten character recognizing method capable of recognizing by handwriting. The above and other objects and novel features of the present invention will become apparent from the description of the present specification and the accompanying drawings.

[0016]

SUMMARY OF THE INVENTION Among the inventions disclosed in the present application, the outline of a representative one will be briefly described.
It is as follows. That is, the present invention is a handwritten character recognition method for recognizing a plurality of handwritten character strings composed of a plurality of stroke groups input in the order of strokes from a handwritten character input device, wherein a distance between the strokes constituting the plurality of stroke groups is determined. A plurality of stroke groups are divided into a plurality of character elements by repeating a process of combining strokes whose evaluated distance is smaller than a threshold for provisional combination until there are no more connectable strokes, according to a predetermined relational expression. After that, the circumscribed rectangle of each character element is obtained, and the maximum value or average value of the height of the circumscribed rectangle and the maximum value or average value of the width of the circumscribed rectangle are estimated as the standard character size of the handwritten character. A parameter representing the relationship between adjacent character elements in the standard character size space is calculated according to a predetermined relational expression. By repeating the process of combining character elements whose parameters are smaller than the provisional combination threshold until there are no more character elements that can be combined, multiple character elements are divided into multiple character element sets, and the character element set It is characterized in that a dictionary is searched, and a character having a maximum evaluation value for a handwritten character pattern registered in the dictionary is output as a recognition result.

In a preferred embodiment of the present invention, an attribute flag is set for each character element whose parameter representing the relationship between the character elements is larger than a threshold for provisional division, indicating that the character is a delimiter. The method is characterized in that a plurality of character elements are classified into a state where character delimitation is clear by the attribute flag, and a dictionary is searched with reference to the classification.

[0018]

Embodiments of the present invention will be described below in detail with reference to the drawings. In all the drawings for describing the embodiments, components having the same functions are denoted by the same reference numerals, and repeated description thereof will be omitted. FIG. 1 is a block diagram showing an embodiment of a handwritten character recognition apparatus to which the present invention is applied, which is constituted by a tablet or an electronic blackboard, etc. A handwritten character input device 2 for outputting stroke order, a display device 3 for displaying a recognition result of handwritten characters,
A central processing unit (CPU) 4 for combining / dividing a plurality of stroke groups of handwritten characters input from the handwritten character input device 2 for each character element candidate and recognizing them by collating with a dictionary;
It comprises a keyboard 5 for inputting various parameters and commands necessary for the recognition process, and a storage device 6 for storing a handwritten character recognition program 61, a dictionary 62 and the like.

Here, the handwritten character input device 2 is not limited to an electronic blackboard or a tablet, and any device can be used as long as it has a configuration in which the brush point coordinates of handwritten characters are output in the order of strokes. Further, an input device having a structure in which a display screen is mounted on the lower surface of the transparent tablet can be used.

In the handwritten character recognition device of the present embodiment, as shown in FIG. 2, no input frame for handwritten characters is provided on the input surface 21 of the handwritten character input device 2. After inputting, for example, an arbitrary handwritten character “recognition of a frameless handwritten character” as shown in FIG. 2 over a plurality of lines by the pen 1 at an arbitrary position, the “recognition” command button 22 is selected. When operated, input surface 2
The handwritten characters written on 1 are collectively recognized, and the recognition result is displayed on the display screen of the display device 3 as characters. In this case, if there is an error in the recognition result, a series of processes starting from the process of recognizing the writing direction are executed again by selecting and operating the "re-recognize" command button 23, and the re-recognition result is displayed. You. When an erroneous handwritten character is written, it can be canceled in units of one character by selecting and operating the "cancel" command button 24.

Here, definitions of terms used in this specification will be described. (1) Stroke The stroke means one handwritten line written from when the pen 1 comes into contact with the input surface 21 of the input device 2 until the pen 1 leaves, and corresponds to "one stroke" in Japanese. . One handwritten character is composed of a plurality of strokes except for punctuation marks. (2) Brush Point A brush point means a point of a minimum unit constituting each stroke, and a coordinate value of a pen 1 pressed on the input surface 21;
Alternatively, it is expressed by a logical coordinate value derived from the pressed coordinate value, and has attributes such as a start point and an end point of the stroke.

(3) Character element A character element refers to a set of strokes clearly included in one character, and is a combination of strokes having intersections with an arbitrary set of strokes, and a stroke having a short distance. It is obtained through a process such as bonding between them. FIG. 3 illustrates the distinction between a stroke and a character element. (4) Handwritten Pattern The handwritten pattern refers to the entire stroke group that constitutes the recognition target handwritten character written on the input surface of the input device 2 as illustrated in FIG. The user may explicitly indicate that a break is to be performed by using a button or menu, or the pen 1
There is a method in which a point in time when the contact operation has not been performed for a certain period of time apart from 1 is set as a break.

(5) Back stroke The back stroke means a vector from the end point of a certain stroke to the start point of the next stroke. In the present invention, the back stroke is subdivided into a back stroke inside a character, a back stroke between characters, and a back stroke between line feeds. . (6) Inner-character back stroke The inner-character back stroke means a back stroke generated between two consecutive strokes included in one character. (7) Inter-character back stroke The inter-character back stroke means a back stroke generated between the end point of the last stroke of a certain character and the start point of the next character. (8) New Line Back Stroke The new line back stroke means a back stroke generated between the end point of the last stroke of the last character of a certain line and the start point of the first stroke of the first character of the next line.

FIG. 4 is a functional block diagram of the handwritten character input device of the present embodiment. When a handwritten character is written on the input surface 21 of the input device 2, a plurality of brushes constituting each stroke of the handwritten character are displayed. A coordinate data sequence of points is output from the input device 2 in the order of strokes. The brush point coordinate data sequence of each stroke is sequentially stored in the storage device 6. When the input of an arbitrary handwritten character is completed and the user selects and operates the “recognition” command button 22, the handwritten character recognition program 61 is started, and the pen point coordinate data sequence of the handwritten character stored in the storage device 6 is read. , Writing direction determination processing, line feed position determination processing, character size determination processing, stroke group division / combination processing, character element division / combination processing, dictionary 6
2 is performed. The handwritten character recognition program 61 includes a writing direction acquisition unit 611, a line feed position acquisition unit 61.
2, a standard character size acquisition unit 613 and a frameless handwritten character string recognition unit 614. Among them, the frameless handwritten character string recognition unit 614 includes a temporary combination processing unit 615, a temporary division processing unit 616, and an evaluation / search processing unit 6 as shown in FIG.
17 is comprised.

Hereinafter, the handwritten character recognition program 61
Will be described in detail with respect to the configuration and the processing content of each unit constituting. (1) Configuration of Pen Point Coordinate Data Sequence Stored in Storage Device 6 As shown in FIG. 6, each stroke pen point coordinate data sequence of handwritten characters output from the input device 2 basically has a stroke number 631. And the x and y coordinate values 632 of each brush point. The stroke relation attribute 63 such as the character number of each stroke in the recognition process.
3. A line feed position flag 634 indicating that the stroke corresponds to the line feed position is added.

(2) Writing Direction Acquisition Unit 611 The writing direction acquisition unit 611 determines whether the handwritten pattern is vertical writing or horizontal writing according to the procedure shown in FIGS. FIG. 7 is an explanatory diagram of the back stroke and the vertical / horizontal writing determination vector. As described above, the back stroke is a vector from the end point of a certain stroke to the start point of the next stroke. Intuitively, the movement of the pen 1 away from the tablet during the input of the handwritten pattern is a back stroke, and the back stroke can be further classified into a back stroke inside a character and a back stroke between characters. A back stroke within a character is a back stroke generated between strokes included in one character, and a back stroke between characters is a back stroke from the end point of the last stroke of a certain character to the start point of the first stroke of the next character. is there. In the handwritten pattern of FIG. 7, BS1, BS2, BS4, and BS6 are the back strokes inside the character, and BS3 and BS5 are the back strokes between the characters.

The writing direction acquisition unit 611 adds only rightward components R3, R4, R5 and downward component D6 included in each back stroke to all stroke groups of the handwritten pattern to be recognized, Find the vertical writing horizontal writing discrimination vector. In FIG. 7, Vtotal is a vertical writing horizontal writing determination vector. In the case of Japanese, the backstroke between characters includes many rightward components in a horizontally written character string, and the backstroke between characters includes many downward components in a vertically written character string. Utilizing this property, the writing direction acquisition unit 611 performs
Vertical / horizontal writing is determined by the following procedure.

First, the vertical / horizontal writing discrimination vector shown in FIG. 7 is obtained (step 801). Next, a value A (the ratio of the rightward component to the downward component) obtained by dividing the rightward component of the vertical writing horizontal writing determination vector by the downward component, a threshold Th for horizontal writing determination, and a threshold Tv for vertical writing determination Are determined as horizontal writing if the value A is equal to or greater than Th, and vertical writing if the value A is equal to or less than Tv (step 802). If the determination cannot be made in the above-described processing, it is determined that the number of written characters is small, and it is determined whether the aspect ratio (the ratio of the width to the height) of the circumscribed rectangle of the entire written handwritten pattern is “1” or more. It is determined that horizontal writing is performed when the value is "1" or more, and vertical writing is performed when the value is less than "1" (step 803). Therefore, as shown in FIG. 7, if the ratio of the rightward component to the downward component of the discrimination vector exceeds the threshold Th for horizontal writing determination, it is accurately determined as “horizontal writing”. By judging horizontal writing or vertical writing in this way,
The user does not need to specify the writing direction in advance, and is free from the trouble of writing handwritten characters.

(3) Line Feed Position Acquisition Unit 612 The line feed position acquisition unit 612 obtains a histogram in the writing direction for a plurality of strokes of the handwritten character input from the input device 2, and determines the writing point by the histogram. A small portion is selected as a line feed position candidate, and at the stroke input time in the stroke group, a vector from the end point to the start point of an adjacent stroke and the average of the lengths of the vectors are obtained. The length and the average of the vector lengths are compared, and the position of the vector exceeding the threshold for line feed determination is determined as a line feed position. That is, as shown in the flowchart of FIG. 11, the line feed position obtaining unit 612 obtains a histogram of the stroke group in the writing direction based on the determination result of the writing direction determined by the writing direction obtaining unit 611 (step 1101). In the case of horizontal writing, as shown in FIG. 9, a position corresponding to a “valley” in the histogram 901 is estimated to be a line feed position.
Therefore, a back stroke that straddles a portion (valley portion) where the brush point distribution frequency is small in the histogram 901 is selected as a line feed position candidate (step 1102).

The line feed back stroke is a kind of inter-character back stroke, and as shown in FIG. 10, the first stroke of the first character of the next character from the end point of the last stroke of the last character of a certain line. This means a back stroke to the starting point. In the case of Japanese, the back stroke of the line feed in the horizontal writing is the lower left direction, and the back stroke in the vertical writing is the upper left direction. Therefore, in the case of vertical writing, the histogram 9
At 01, a back stroke in the upper left direction straddling a portion (valley portion) where the brush point distribution frequency is small, and in the case of horizontal writing, a back stroke in the lower left direction is selected as a line feed back stroke candidate.
Next, in the case of horizontal writing, if the left horizontal component Wcr of the back stroke selected in the above process exceeds the threshold for line feed determination, it is determined as a line feed back stroke. In the case of vertical writing, the back stroke selected in the above process is determined. Upward vertical component Hcr of stroke
Is determined to be a line feed back stroke (step 1103).

In this case, the horizontal component W of the line feed back stroke is
The size of cr and the vertical component Hcr differs depending on the number of characters in one line. Therefore, when the standard size of one handwritten character is known or can be estimated as shown in FIG. 10, when the value obtained by dividing the horizontal component Wcr by the width Ws of the standard character size exceeds the threshold, the horizontal writing is performed. By selecting the stroke as the line feed backstroke in the case, and selecting the one in which the value obtained by dividing the vertical component Hcr by the height Hs of the standard character size exceeds the threshold value as the linefeed back stroke in the case of vertical writing,
The judgment accuracy is further improved.

When a handwritten character is written in an oblique direction, a horizontal component Wcr and a vertical component Hcr are written.
May not be calculated, but in the case of diagonal writing,
The horizontal component Wcr and the vertical component H are obtained by performing a correction process for converting a handwritten character pattern into an orthonormal coordinate system.
It is possible to calculate cr normally. In this case, whether or not the writing is oblique can be determined, for example, by obtaining a line connecting the centers of the circumscribed rectangles of the respective handwritten characters, and determining the inclination of the line. By determining the line feed position in this way, the user does not need to specify the line feed position during writing, and is free from the trouble of writing handwritten characters.

(4) Standard Character Size Acquisition Unit 613 The standard character size acquisition unit 613 evaluates the distance between each stroke constituting a plurality of stroke groups of the handwritten character input from the input device 2 according to a predetermined relational expression. Then, by repeating the temporary joining process of joining strokes whose evaluated distances are smaller than the threshold for temporary joining until there are no more joinable strokes, after dividing the plurality of stroke groups into a plurality of character elements, A circumscribed rectangle of the character element is obtained, and the maximum value or average value of the height and the maximum value or average value of the width of the circumscribed rectangle are estimated as the standard character size of the handwritten character. The distance between the strokes in the temporary connection processing is evaluated by a value obtained by multiplying each parameter by a coefficient as shown in FIGS.

Here, L is a standard size (length of one side) of one stroke as shown in FIG.
The standard area of one stroke. For the standard size L and standard area S of one stroke, a circumscribed rectangle of each stroke as shown by a broken line in FIG. 12B is obtained, and only the longer one of the height and width of the circumscribed rectangle is selected. Then, the largest one of the heights and widths of all the strokes is selected, and the standard size L and standard area S of one stroke are estimated from this. In the combining process between the character elements described later, L is the standard size of one character element, and S is the standard area of one character element.

(A) Evaluation parameter = d / L As shown in FIG. 12B, the ratio of the displacement d in the writing direction of the circumscribed figure (shown by a broken line) of adjacent strokes to the standard size L of one character, (B) Evaluation parameter = c / S As shown in FIG. 12C, the ratio of the area c of the overlapping part of the circumscribed figure (shown by a broken line) of the adjacent stroke to the standard area S of one character, (c) Evaluation parameter = d / L As shown in FIG. 12 (d), the ratio of the Euclidean distance d of the barycentric coordinates of adjacent strokes to the standard size L of one character, (d) evaluation parameter = d / L ), The ratio of the displacement d in the writing direction of the barycentric coordinates of adjacent strokes to the standard size L of one character. (E) Evaluation parameter = d / L As shown in FIG. End of stroke Euclidean distance d of the beginning of the writing point of the stroke and after the writing point
(F) Evaluation parameter = d / L As shown in FIG. 13C, the writing direction of the last writing point of the previous stroke and the writing point of the first writing point of the subsequent stroke Is the ratio of the displacement d to the standard size L of one character.

At least two of these evaluation parameters are selected in advance, and when the evaluation values based on the selected plurality of evaluation parameters are obtained, each of the evaluation values is multiplied by a predetermined coefficient and added. The sum is compared with a threshold value for provisional combination. As a result of this comparison processing, a small addition value is determined to be included in one character, and the pair of strokes is combined into the same set and selected as one character element candidate. This temporary combining process is recursively repeated until all strokes equal to or less than the threshold value are combined with any of the character elements.

For example, when a kana character “soft” is input as shown in FIG. 14A, strokes ST ₁ to ST ₅ constituting the kana character are connected with each other by the adjacent strokes shown in FIG. 13) to 13 (c), comprehensive evaluation is performed using all the evaluation parameters, and it is determined which strokes are combined into one character element. FIG. 14B shows an example of the value of each evaluation parameter. Here, FIG.
The evaluation parameters (a) to (c) in (b) are shown in FIG.
The evaluation parameters 2 (a) to (c) and the evaluation parameters (d) to (f) correspond to the evaluation parameters in FIGS. 13 (a) to 13 (c). The smaller the calculated evaluation parameters, the stronger the degree of coupling.

With respect to the evaluation parameters shown in FIG.
“Temporary combination threshold = −4.0”, “Temporary division threshold = −5.
If "0" is set, the overall evaluation is the stroke ST₁, S
T₂The interval is "-3.2", stroke ST₂, ST₃Between
"-5.45", stroke ST₃, ST₄The interval is "-
7.4 ”, stroke ST₄, ST₅The interval is "-1.4
1 ”, the stroke ST₁, ST₂In the interval,
Go ", stroke ST₂, ST ₃Between the "split", str
Ark ST₃, ST₄"Division" between strokes, ST₄,
ST₅The space is a "bond".

Here, a simple X-axis direction (horizontal writing direction)
Depending on the degree of overlap, whether it is "combined" or "divided"
If you decide to use a deterministic method,
If the stroke ST₂, ST₃Less than the distance d2 between
Is set to the threshold for temporary connection, the stroke
ST₄, ST₅Since the distance d3 between them is d2> d3,
Stroke ST₄, ST₅Becomes "join". But,
Stroke ST₁, ST₂The distance d1 between them is d1> d2
Therefore, these strokes ST₁, ST₂The interval is "minutes
%) And the stroke ST₂, ST ₃Between the "join"
The strokes are joined and split correctly
Disappears.

On the other hand, as in the present invention, the connection and division of strokes are determined by comprehensive evaluation of a plurality of evaluation parameters, so that the connection and division of strokes can be performed accurately. The standard character size acquisition unit 613 combines and divides strokes as described above and determines candidates to be character elements. As a result, a plurality of stroke groups of handwritten characters input from the input device 2 are indicated by broken lines in FIG. It is divided into a plurality of character elements, as shown by surrounding with.

Then, next, as shown by a broken line in FIG.
Finds the circumscribed rectangle of each character element and determines whether the size of the circumscribed rectangle is
Estimate the size of one character. The size of the character is height and
Calculate the width separately and calculate the height of each circumscribed rectangle.
Only the longer value of the height and width is used. FIG.
When a handwritten pattern like is given, calculate the height
H₁, H ₃, H₄, H₅, H₆To calculate the width
W₂, W₇Use Selected data for calculation
Then, find the average value and standard deviation of each data,
If the difference between the value and the standard deviation is greater than or equal to the threshold,
And delete it from the data. Most
The maximum or average value of the remaining data is
Estimate height or width.

In this case, the data is finally insufficient and the height H
Alternatively, if one of the widths W cannot be calculated, the value of the calculated width is also used for the value of the uncalculated one. For example, when only the height H can be calculated and the width W is not obtained, the width W is equal to the height H. In the example of FIG. 14, the calculation is performed with the height of the character = H ₆ and the width = W ₇ . By doing so, even if the writing direction and the number of lines are not specified,
The size of the character can be estimated. If the information on the writing direction and the number of lines is found in the writing direction discrimination processing and the line feed position discrimination processing, the accuracy of the temporary combination processing is further improved, and as a result, the accuracy of estimating the standard size of handwritten characters is improved. There is.

In particular, it is possible to accurately estimate the standard character size for cutting out each character element, even for oblique writing or handwritten characters with a narrow character interval. For example, when a diagonally-written handwritten character is input as shown in FIG. 16A, the circumscription of the character element is performed in units of character elements which are combined or divided as shown in FIG. Since a rectangle is obtained and the size of one character is estimated from the size of the circumscribed rectangle, the standard character size can be accurately estimated even in the case of oblique writing.

(5) Frameless Handwritten Character String Recognition Unit 614 The frameless handwritten character string recognition unit 614 includes a temporary combination processing unit 615, a temporary division processing unit 616, an evaluation / search processing unit, as shown in detail in FIG. 617. Temporary join processing unit 6
The processing in 15 is exactly the same as the temporary combination processing in the standard character size acquisition unit 613. However, the temporary combination processing in the standard character size acquisition unit 613 is to combine individual strokes to create a “character element in a state clearly included in one character”. The provisional combination processing in 615 is to further combine each character element with reference to the estimated value of the standard character size.
In this case, the evaluation parameters and the procedure used when combining the character elements can be exactly the same as the temporary combination processing in the standard character size acquisition unit 613. The difference is that the standard size L uses the larger value of the length of the circumscribed rectangle of one character element, and the standard area S uses the square area of the standard size L. Note that an evaluation parameter set specifically for combining character elements may be used.

By the recursive provisional combination processing of the character elements, for example, as shown in FIG. 17, for the kanji character “Q”, the character element “mouth” in the “monarchy” is replaced with the last written character element. Despite this, it is combined in the "gate stance" to form a single kanji character element set of "question". When the character elements are further combined and a new character element set is created, the provisional division processing unit 616 performs provisional division processing. Temporary division processing is to evaluate the distance between character elements, and between character elements with a distance larger than the threshold for temporary division,
This is a process of setting an attribute flag indicating that there is a character delimiter. In this case, the method of evaluating the distance between the character elements is the same as in the above-described provisional combination processing.

As a result of this processing, between the last stroke of the previously written character element and the first stroke of the later written character element of the two character elements for which the character delimiter attribute flag is set, a “character” It is a state in which it is clear that it is a delimiter. In FIG. 6, the attribute flag is illustrated by a character sequence number. As a method of expressing the attribute flag, any other method may be used. The provisional combining process and the provisional division process in the frameless character string recognition unit 614 are processes for reducing the search space in the subsequent evaluation / search processing unit 617, so that the processing time does not matter (high-speed processing). If no time is required) can be omitted.

Next, in the evaluation / search processing unit 617,
The dictionary 62 is searched by each character element set, a character having the maximum evaluation value for the handwritten character pattern registered in the dictionary 62 is determined, and the code of the character is output to the display device 3 as a recognition result. To display the character corresponding to the character code. The above-mentioned provisional division processing unit 616
Is completed, the state between all adjacent strokes included in the handwritten pattern input from the input device 2 is “a state clearly included in one character”,
Either "a state where it is clear that a character is delimited" or "an ambiguous state". One “cutout pattern” can be defined depending on whether the “ambiguous state” existing at this stage is included in one character or a character delimiter. The number of “cutout patterns” in the search space is 2 to the power of n, where n is the number of “ambiguous states”.

The evaluation / search process is a process of searching for a “cut pattern” that maximizes an evaluation value described below from “all cut patterns” included in the search space. As a search method in this case, an existing search method such as a dynamic programming method, a full search, or a beam search can be used. In the present embodiment, as shown in FIG.
It is represented by a binary tree, and a beam search for the binary tree is performed. As the evaluation value of the cutout pattern, a value obtained by multiplying the following evaluation parameter by a coefficient and adding the result is used.

(A) an evaluation parameter obtained from the distance between each cut-out handwritten pattern and the handwritten pattern registered in the dictionary; (b) an evaluation parameter obtained from the transition probability between each recognition result character; An evaluation parameter obtained from a ratio of the size of each cut-out handwritten pattern to a standard character size; (d) an evaluation value of a distance between character elements between adjacent strokes determined to be included in one character; (E) An evaluation value obtained from the evaluation value of the distance between character elements between adjacent strokes determined to be character delimiters, and an evaluation parameter obtained from the threshold of the temporary division processing.

In FIG. 18, a dashed line indicates a portion where it is ambiguous whether or not it becomes a break, a broken line arrow indicates that each character element is divided or combined by a combining process, and a solid arrow indicates that each character element is combined or divided by a combining process. For example, the handwritten character "sunny"
After combining at the ambiguous part of and then dividing at the ambiguous part, the character is recognized as "sunny". However, if the ambiguous part is also combined, it indicates that it cannot be recognized. The evaluation / search processing unit 617 proceeds in order from left to right in the part where the connection relation between the character elements is ambiguous, and proceeds to the left when it is determined to be a character delimiter, and to the right when it is determined that it is included in one character. While evaluating the certainty of each node of the binary tree in FIG. 18 as a Japanese character string by the method described below, the most probable leaf is searched from the leaves of the binary tree, and the character corresponding to the leaf is searched. Let the sequence be the recognition result.
This corresponds to the evaluation method (c). The probability that a certain handwritten pattern X is a character string C can be expressed by the following “Equation 1” according to Bayes' theorem.

[0051]

(Equation 1)

Where P (X) is the probability of occurrence of event X,
P (X | Y) is the conditional probability that event X occurs under event Y. That is, P (X | C); the probability that the character string C is written like the pattern X, P (C); the probability that the character string C is written, P (X); the probability that the pattern X is written; Are independent and are considered as constants). Here, P (C) can be approximately represented by “Equation 2”.

[0053]

(Equation 2)

[0054] _{However, P (C i + 1 │C} i) , by the probability that the i-th character and (i + 1) th character is written continuously determined from the table there are prepared taking advance statistics.
N is the number of characters. P (X│C) is approximately “Equation 3”
Can be represented by

[0055]

(Equation 3)

Here, P (X _i | C _i ) is the _i in the character string C.
Is the probability that the i-th character Ci is written like the i-th hand-written pattern Xi in which the hand-written pattern X is divided for each character, and recognizes a dictionary pattern corresponding to the character Ci and the hand-written pattern Xi with an online frame. It is determined by comparing with the device. P (delimiter or join | dk)
Is the probability that, when the distance between the k-th character element and the (k + 1) -th character element is dk, the two character elements form a character delimiter or are included in one character. . Which probability is determined depends on how to divide the handwritten pattern X.

In the division method of the handwritten pattern under evaluation, if the k-th character element and the (k + 1) -th character element are not included in one character, the probability of becoming a character is determined as follows. Find the probability of being included in one character. P (SIZE
_i | standard size) is the certainty of the size SIZEi of the i-th character when the standard size of one character is the standard size. Next, in consideration of calculation by a computer, “Equation 3” requires many multiplications and (2i + k) multiplications. Therefore, "Equation 3" is replaced by a calculation formula having a logarithmic term as shown in "Equation 4".
Is used as the statistical evaluation value.

[0058]

(Equation 4)

As described above, the likelihood of Japanese is evaluated, and the character having the largest evaluation value is output as a recognition result, so that handwritten characters with irregular character spacing and handwritten characters written diagonally. Even if exists,
A recognition result suitable for the context of a character string over a plurality of lines can be obtained, and a high-precision recognition result that cannot be obtained by character-by-character recognition can be obtained collectively. For example, the handwritten characters in FIG. 16A are correctly recognized by combining the character elements as shown in FIG.

It should be noted that the present invention is not limited to the above embodiment, and the processing in the writing direction acquisition unit 611, line feed position acquisition unit 612, standard character size acquisition unit 612, and frameless handwritten character recognition unit 614 is newly performed. As an elemental technology of, it can be configured by being incorporated into existing character recognition processing. In addition, the handwritten character recognition program
It is stored in a recording medium such as a ROM and provided to the user.
Alternatively, it is provided for a fee through a communication medium such as the Internet.

As described above, according to the present embodiment, the writing direction of a handwritten character written on an electronic blackboard or the like without specifying the writing direction is accurately determined, and the handwritten character is recognized according to the determination result. can do. In addition, it is possible to accurately determine a line feed position of a handwritten character written without specifying a line feed position on an electronic blackboard or the like, and recognize a handwritten character over a plurality of lines according to the determination result. Furthermore, even for diagonal writing or handwritten characters with a narrow character interval, it is possible to accurately cut out each character element and recognize handwritten characters on an arbitrary line according to the cutout result. Also,
Regardless of the type of vertical and horizontal writing, the number of lines, and the presence or absence of a writing frame,
Handwritten characters written on an electronic blackboard or the like can be recognized with high accuracy. As described above, the invention made by the inventor has been specifically described based on the embodiment. However, the present invention is not limited to the embodiment, and can be variously modified without departing from the gist of the invention. Of course, it is.

[0062]

The effects obtained by typical ones of the inventions disclosed in the present application will be briefly described as follows. ADVANTAGE OF THE INVENTION According to this invention, even if it is a diagonal writing or a handwritten character with a narrow character spacing, it becomes possible to cut out each character element accurately, and to recognize the handwritten character of an arbitrary line according to the cutout result.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a handwritten character recognition apparatus to which the present invention is applied.

FIG. 2 is an explanatory diagram illustrating an example of a handwritten character written on an input surface of a handwritten character input device.

FIG. 3 is an explanatory diagram showing a unit of data in a handwritten character.

FIG. 4 is a functional configuration diagram of the handwritten character recognition device of FIG. 1;

FIG. 5 is a detailed configuration diagram of a frameless character string recognition unit.

FIG. 6 is a diagram illustrating an example of a data configuration of a handwritten character stored in a storage device.

FIG. 7 is an explanatory diagram of a vertical / horizontal writing discrimination vector.

FIG. 8 is a flowchart illustrating a process of determining vertical writing and horizontal writing.

FIG. 9 is an explanatory diagram showing an example of a histogram used for determining a line feed position.

FIG. 10 is an explanatory diagram of a line feed back stroke.

FIG. 11 is a flowchart illustrating a line feed position determination process.

FIG. 12 is an explanatory diagram of evaluation parameters used for a temporary connection process between strokes.

FIG. 13 is an explanatory diagram of evaluation parameters used for a temporary connection process between strokes.

FIG. 14 is an explanatory diagram showing an example of an input stroke to be subjected to a temporary connection process of strokes and an example of calculation of an evaluation parameter.

FIG. 15 is an explanatory diagram of a process of estimating a standard character size from a circumscribed rectangle of a character element.

FIG. 16 is a diagram illustrating an example of a process of temporarily combining obliquely written handwritten characters into character elements.

FIG. 17 is an explanatory diagram showing an example of handwritten characters that can be combined by recursive processing of character elements.

FIG. 18 illustrates a method for searching for a handwritten character in a dictionary.
It is explanatory drawing which shows an example of a branch tree.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Pen, 2 ... Handwritten character input device, 3 ... Display device, 4
... CPU, 6 ... Storage device, 21 ... Input surface for handwritten characters,
61: handwritten character recognition program, 62: dictionary, 611
... writing direction acquisition unit, 612 ... line feed position acquisition unit, 613 ...
Standard character size acquisition unit, 614: frameless handwritten character string recognition unit, 615: temporary combination processing unit, 616: temporary division processing unit
617 ... Evaluation / search processing unit.

 ──────────────────────────────────────────────────続 き Continued on the front page F-term (reference) 5B029 AA01 CC21 EE08 5B064 AB04 BA05 DD06 EA34

Claims (2)

[Claims] 1. A handwritten character recognition method for recognizing a plurality of handwritten character strings composed of a plurality of stroke groups input in the order of strokes from a handwritten character input device, wherein a distance between strokes constituting the plurality of stroke groups is determined in advance. A plurality of stroke groups were divided into a plurality of character elements by repeating the process of combining strokes whose evaluated distances were smaller than the threshold for provisional combination until there were no more connectable strokes, according to the determined relational expression. Then, the circumscribed rectangle of each character element is obtained, and the maximum or average value of the height of the circumscribed rectangle and the maximum or average value of the width of the circumscribed rectangle are estimated as the standard character size of the handwritten character. Calculate a parameter representing the relationship between adjacent character elements in a standard character size space according to a predetermined relational expression, and calculate the By repeating the process of combining character elements whose parameters are smaller than the threshold for provisional combination until there are no more character elements that can be combined, multiple character elements are divided into multiple character element sets, and A handwritten character recognition method comprising: searching a dictionary; and outputting, as a recognition result, a character having a maximum evaluation value for a handwritten character pattern registered in the dictionary. 2. An attribute flag indicating a character delimiter is set for each character element whose parameter indicating the relationship between the character elements is larger than a threshold for provisional division, and the plurality of character elements are assigned the attribute 2. The handwritten character recognition method according to claim 1, wherein the flag is divided into a state in which character delimitation is clear, and a dictionary is searched with reference to this division.