JPS61117680A - Generating system of dictionary for recognition of hand-written character - Google Patents

Abstract

PURPOSE:To obtain a dictionary for recognition of hand-written characters which is suitable for a user, by synthesizing deformed patterns of characters and selecting the character pattern resembling hand-written characters of the user from them the registering its feature vectors in the dictionary for recognition of hand-written characters. CONSTITUTION:In the learning mode operation, a dictionary retrieving part 11 retrieves the dictionary similarly to the normal mode. If the recognition result is rejected, the operation is performed as follows; a control part 7 displays an input character pattern and a message to an operator on a display device 6, and the operator discriminates a character in accordance with the displayed input pattern and inputs its character code from a key input device 3. This character code is sent to a recognizing part 4 and is given to a dictionary register part 13 through the dictionary retrieving part 11. If there is one candidate character for the input character pattern and the recognition result is rejected because the distance between its feature vectors exceeds the allowable error range of the candidate character, the dictionary register part 13 combines feature vectors of the input character pattern with the key-inputted character code, etc. and registers additionally the result in a dictionary memory 14.

Description

[Detailed Description of the Invention] [Technical Field] The present invention provides a literary magazine 51 that can recognize draft characters! Regarding the device 1. Specifically, the present invention relates to a method for creating a manual character recognition dictionary optimal for each user in such a character recognition device.

[Prior art]

In character recognition devices for recognizing handwritten characters,
It is necessary to load the most suitable dictionary for the user into the dictionary memory. Conventionally, the following methods were used: (1) the manufacturer created a standard that was considered optimal for a specific user from a standard dictionary (1) and then the needs loaded the dictionary into dictionary memory and used it, (2) printed a variety of modified characters. The user selects a sheet that seems similar from among them.
(3) A method in which the user writes handwritten characters on the learning sheet and (1) selects a dictionary corresponding to the selected sheet from the dictionary and loads it into the dictionary memory; Recognize it literally, 1
! A method is adopted in which the device is made to read (1) and learn, and the minimum necessary poems suitable for the user are loaded into the reading memory.

However, even with the ``Kaji'', ``A'', and ``tL'' methods, it is necessary to create a large-scale marking Iτ dictionary 1;・ that can accommodate a variety of modified characters. And to create a standard dictionary.

Collect handwritten letters from many people and enter them into a calculator.
This requires a huge amount of time and manpower. For this reason, the cost of creating a dictionary increased by L, and it was not possible to respond quickly to user needs. Furthermore, it is not always possible to obtain a dictionary that is most suitable for the user.

〔the purpose〕

The present invention has been made in view of the above circumstances, and its purpose is to provide a method for quickly creating a dictionary for handwritten character recognition suitable for each user without having to prepare a large-scale standard dictionary. It is about providing.

〔composition〕

In order to achieve the above object, the present invention is capable of recognizing f, IF characters, and extracting feature vectors of input character patterns to create handwritten characters! ! A character recognition device equipped with a device for registering characters in a common dictionary uses a two-character synthesizer to synthesize transformed patterns of characters.
e! By selecting and inputting a character pattern similar to a character, the feature vector of that character pattern can be applied to the handwritten character!
! This feature is characterized in that a handwritten character recognition dictionary optimal for each user is created by registering the handwritten character in a common dictionary.

An embodiment of the present invention will be described below with reference to two drawings.

FIG. 1 is a schematic block diagram of an f, v character recognition device according to an embodiment of the present invention. The handwritten character recognition device shown here uses a character pattern input means to optically scan a document, read the character pattern, binarize it, cut out one character, and normalize it.

In addition to a character reading section 1 that performs preprocessing such as noise removal, it is provided with a character synthesis section 2 that synthesizes and inputs character patterns when creating a dictionary for handwritten character recognition. Additionally, key human power devices used for operator input! It also has 3. 4 is a recognition part.

Extract feature vectors from human character patterns and compare them with feature vectors in handwritten character recognition dictionaries to perform character recognition, register additional characters in the dictionary (as described later, create new handwritten character dictionaries, and (including updates). 5 is an external storage device for storing various dictionaries and single character recognition results. Reference numeral 6 denotes a display device for displaying character recognition results, single character patterns, messages to the operator, and the like. Reference numeral 7 denotes a control section that performs sequence control and data transfer control of each section, and is connected to other sections via a system bus 8.

The configuration of the recognition unit 11 will be explained with reference to FIG. As shown in the figure, the recognition unit 4 includes a feature extraction unit 10, a dictionary search unit 11, and a dictionary 1!
It consists of a registration section 13 and a dictionary memory 14. The handwritten character recognition device of this embodiment can operate in one of the normal mode, learning mode, and dictionary creation mode, which can be designated by the operator through the key manual device 3.

First, the operation of the recognition unit 4 in the normal mode will be explained. The feature extraction unit 10 extracts a feature vector vp of the input character pattern. A handwritten character recognition dictionary is loaded into the dictionary memory 14 from the external storage device 5 under the control of the control unit 7 . The dictionary search unit 11 sequentially compares the feature vector Vp of the input character pattern with the feature vector Vd of the dictionary 7F, searches for a candidate character with the minimum distance between the feature vectors, uses the character code as a recognition result, and uses the character code in the system. Output to bus 8. However, if the minimum distance between feature vectors exceeds the tolerance range (added to the feature vector) for that candidate character, or if there are two or more candidate characters whose distances are within the tolerance range, The input character pattern is rejected, and the dictionary search section 11 outputs a reject code.

The recognition result is stored in the external storage device P5 under the control of the control section 7.

In the learning mode operation, the dictionary search unit 11 performs a dictionary search in the same way as in the normal mode, searches for candidate characters, and outputs the results to the system bus 8.

If the recognition result is rejected, the operation is as follows, unlike the 1 normal mode. First, the control unit 7 displays the input character pattern and a message to the operator on the display device i!
6. The operator is.

The character is determined from the displayed input cover pattern, and the character code is inputted from the key input device 3. This character code is sent to the recognition unit 4 under the control of the control unit 7, and is sent to the dictionary search unit 1.
1 to the letter V registration unit 13. There is one candidate character for the input character pattern, and the distance between its feature vectors exceeds the tolerance range (added to the feature vector of the candidate character) for that candidate character, so it is rejected. is reported from the dictionary search unit 11 to the dictionary IF registration unit 13), and the dictionary IF registration unit 13 converts the feature vector of the input character pattern into the character code manually entered by the key,
and the standard tolerance range Δ0, and are additionally registered in the dictionary (in the dictionary memory 14).

A message to that effect is sent to the system bus 8 via the dictionary search section 11. This registration notification is displayed on the display device 6 under the control of the control unit 7. On the other hand, if the candidate character is rejected because there are two or more candidate characters whose distances from the input character pattern are within the respective allowable error ranges, the dictionary registration unit 13 selects the candidate character whose distance from the input character pattern is the shortest among the candidate characters. Select a character and compare its distance Δ' with the minimum allowable value ΔL. and.

If Δ′>ΔL, the distance is too short, so the feature vector of that input character pattern is not added to the dictionary.

This is notified to the control unit 7 and displayed. on the other hand.

If Δ'≧ΔL, the feature vector of the input character pattern is additionally registered in the dictionary, and a message to that effect is displayed.

At this time, the permissible error range Δ is calculated using a function T(Δ') as shown in FIG.

As shown in 2, determining whether additional registration is possible and determining the allowable error range between the feature vectors of additional IJII characters is necessary to avoid the increase in similar characters due to additional registration in the dictionary and the resulting misrecognition. This is to avoid an increase.

In the dictionary creation mode, a single character pattern is input from the character synthesis section 2, but the operation of the recognition section 4 is the same as in the learning mode.

Next, the character synthesis section 2 will be explained. This character synthesis part 2
is a device that synthesizes characters by stroke composition,
The configuration is shown in FIG.

In this figure, reference numeral 20 is an input port that takes in character codes and the like from the system bus 8 to the bus 21. 22 is a main processing section that controls the entire device and performs arithmetic processing. 23 is a character dictionary memory, 24 is a stroke dictionary memory, and 25 is a decoration dictionary memory.

Under the control of the control section 7 and the main processing section 22, respectively,
A character dictionary, a stroke dictionary, and a decoration dictionary are loaded from the external storage device 5.

The character dictionary contains the basic data for synthesizing character and symbol patterns (drifting patterns and deformed patterns) including kanji. 2 should be combined! ! Character code indicating the character (
(Identification number information for the standard pattern and each modified pattern is added) Number of strokes that make up one character, Stroke code that specifies stroke dictionary data, Decoration code that specifies data within the decoration, Stroke starting point degree It consists of proportions that indicate the length of the stroke.

For example, there are 80 types of strokes that make up the core of a character! a
It has been standardized to a certain degree, and data regarding each stroke is registered in a stroke dictionary. As conceptually shown in Figure 6, the data in this stroke dictionary includes a stroke coat that specifies a stroke, coordinates of the starting point of the stroke, proportions that indicate the length of the stroke, thickening information for fleshing out the stroke, and stroke code that specifies the stroke. Consists of a chain coat for development into a bit pattern. The second chain code expands the stroke a bit.

The direction of transition from one dot to the next is expressed as a code from 0 to 7 as shown in Figure 13.
8 is a stop code. As shown in FIG. 8, the information consists of a feature coat, a nodal seat ring, and thickness data (L, R) for one stroke coat. The feature code is a code that specifies the fleshed out part of the loaf.
/E indicates the stroke of the part to be drawn etc.)
The 1-node coordinate indicates the X-direction coordinate at which the thickness of the stroke changes. The thickness data indicates the width of the center line of the stroke, that is, the bare bone 00, to be filled left and right in the Y direction.

In the decoration dictionary, there are 8 data showing the protrusions at the beginning and end of the horizontal bar and vertical bar of Mincho typeface kanji, the protrusions on the corner of a key, and the decorations such as ゛.

Second decoration: IF data is as shown in Figure 7.

It consists of a decorative coat, a one-sided drift of the starting point S of the decoration, and an I-soto pattern of the decoration. Part 4 is shown in Fig. 10. 5. At the right end of the stroke of the horizontal bar shown in Fig. 11, the decoration shown in Fig. 1O is added by aligning the reference point St and the starting point S. The stroke pattern shown in Fig. 12 is can get.

If ¥bi? 5. Refer to PJ/1 diagram. Stroke generator 2
6 is.

The dot pattern of the stroke is developed on the output buffer 36 by inputting stroke data. The decoration generating section 28 is a device that receives the decoration data as input and synthesizes the decoration pattern with the stroke pattern on the cover sofa 36. Input data to the stroke generating section 26 and the decoration generating section 28 are read out from each dictionary memory 23°, 24, and 25 by the main processing section 22. The contents of the output buffer 3G are outputted under the control of the output control section 30, smoothed by the smoothing section 38 to remove diagonal edges, etc., and then sent to the system bus 8. Please refer to the flowchart.

Character synthesis processing by the character synthesis section 2 will be explained. First, the operator specifies the review creation mode using the key manual device 3. Then, the control section 7 puts the character composition section 2 into operation. Next, the operator inputs the code of the character to be generated from the key-powered device 3 (5). Input (block lot>.

The main processing unit 22 uses the input character food to create a character word! F(23) is searched and data corresponding to one pattern of the corresponding character (standard pattern for i first, each modified pattern thereafter) is read out (Prosogame 102). The main processing unit 22 generates a stroke dictionary v(21
1), reads the stroke data related to the corresponding stroke, and inputs it to the stroke generation section 2G (block l'04).The main processing section 22 also checks whether the stroke has a decoration code (block 107).
If there is m, use that decoration code to create decoration 1F (2
5) and reads out the corresponding decoration data and inputs it to the decoration generation section 28 (block 108).

The stroke generator 26 develops a stroke dot pattern on the output cover sofa 36 based on the input stroke data (starting point coordinates, proportion, chain code). That is, the bare bones of the stroke are drawn in the direction specified by the chain code by the specified proportion from the specified starting point (block 105). For example, the 11th
In the example of the horizontal bar shown in the figure, a core wire 41 is drawn. Next, the stroke generation unit 26 adds thick line information to the bare bones of the created stroke, and accordingly adds flesh to the bare bones (block 106). In the example in Figure 11, the fleshing is partially 42.4
3 is added to the bare bone 41. The decoration generating part 28 is
Based on ornament data.

Output basso 7736. ): adds a decorative pattern to the stroke pattern developed in (block 109).

For example, in the example of Fig. 11, the decorative pattern of Fig. 1O is
The starting point S is added 5 so as to match the reference position ISl of the bare bone 4I, thereby synthesizing the horizontal bar pattern shown in FIG. 12.

When the composition of one stroke pattern is completed.

The main processing unit 22 checks whether there are any remaining strokes to be synthesized (block 103), and if there are any remaining strokes, the same process is executed for the next stroke. After completing the synthesis of all strokes, the main processing section 22 outputs the output control section 3.
0 and causes the output cover sofa 36 to output the synthesized character pattern to the system bus 8 (block 110
). The control unit 7 displays the character pattern data carried on the system bus 8! ! Enter 68 and display 9.

Next, the main processing unit 22 checks whether a combination pattern of the specified character remains (block ilN).

If the remaining data remains, the data is read from the character dictionary 23 (block 102), and the processing described in the same article is performed to synthesize the character pattern.1 The synthesized character pattern is similarly displayed on the display device 6. Ru. When the synthesis of all patterns is completed, the main processing section 22 sends a completion report to the control section 7 through the output control section 30.

2 and 2 with a standard pattern of one character.

The one and two modified patterns are sequentially synthesized by the character synthesis section 2 and displayed on the display device 6. Note that the control unit 7
Add an identification number to each pattern of the same character and display it. Further, each pattern is stored in a memory inside the display device 6. Upon receiving the completion report, the control section 7 causes the display device 16 to display a text pattern selection water increase for the operator.

Next, the creation of a dictionary for 9 handwritten character recognition will be explained with reference to the flowchart of FIG. In this case, the operator specifies the dictionary creation mode using the key manual device 3.

Next, the operator inputs the code of the character to be registered in the dictionary using the keypad w3 (block 201).

When the second character code is input to the character synthesis section 2, the standard pattern and one or more modified patterns of that character are sequentially synthesized by the character synthesis section 2 and displayed on the first display device 6 (
block 202). When the control section 7 receives the report of completion of synthesis of all patterns from the character synthesis section 2, it causes the display device 6 to display a pattern selection message. The operator selects the number of one or more patterns from among the displayed patterns that are similar to his f, ilF character (as mentioned above, each pattern is accompanied by an identification number). )
is input from the key operator [3], and the control unit 7 displays the pattern selected by the operator on the display! ! 6 from the internal memory and input to the recognition unit 4 (block 203)
.

The recognition unit 4 operates in the learning mode (block 204
), that is, the feature vector of the input character pattern is extracted by the feature extraction unit 1O, and the dictionary search unit 11 searches for the feature vector having the minimum distance from the feature vector. If the input character pattern can be recognized, there is no need to additionally register the feature vector of the deformed pattern of the character input at that time in the handwritten character recognition dictionary (memory 14), and if it is registered, the recognition rate may decrease. . Therefore, in this case, only a message to that effect is displayed on the display device 6 under the control of the control section 7, and the processing for the input character pattern ends. If rejected, 11
Additional registration is performed by the F registration unit 12. First, if a similar character is rejected because it is not registered in the dictionary, the feature vector of the input character pattern is the permissible error range △0 of the signboard and the character code (input from the control unit 7). are combined and registered in the dictionary IF (1/l).

If the character is rejected because there are two or more similar candidate characters, the distance between the feature vectors between the most similar candidate character and the input character is checked, and if it exceeds the minimum allowable value, it is added to the dictionary. No additional registration was made.

If it is less than or equal to the minimum allowable value, the feature vector of the input character pattern is registered in the dictionary along with the allowable error range and character code calculated as described above using it. A message to the effect of registration is displayed on the display device 6.

When the above processing is completed for all selected patterns, a confirmation message to continue is displayed (block 205).
). Operators must be registered separately.

If there is a character you want, press the continue instruction key, and if you want to finish, press the end instruction key. The control unit 7 determines whether the input instruction is an end instruction (block 206), and if it is an end instruction, shifts from the dictionary creation mode to the normal mode7. If the instruction is to continue, it will wait for the character code input. In addition, the dictionary creation mode is used to update (additionally register) an already created user dictionary, as well as to create a completely new dictionary for the 5th Kneeze. Also used for

As described above, it is possible to efficiently create a dictionary that is suitable for recognizing the user's handwritten characters more quickly than the method of writing the user's handwritten characters and having the character reading unit read the learning seeds. Of course, you can easily create a dictionary and store it in the external storage space 5.

Although the embodiments have been described above, the present invention is not limited thereto.

For example, in the dictionary creation process, an operator corrects the pattern synthesized by the single character synthesis section 2 while looking at the display screen, and inputs the corrected pattern into the recognition section.
The pattern modification described in (2) may be performed directly on the display screen, or, for example, by displaying data read out from a character dictionary and Input the correction of the starting point ring and proportion of the stroke from the key human power device 5, send the corrected data to the g processing section of the character synthesis section, and synthesize the 1 & character pattern 11#jσ with the modified IF data, etc. By the way? One is fine.

〔effect〕

As explained above, according to the present invention, it is possible to create (update) a dictionary for handwritten character recognition that is optimal for each user more efficiently than before, without preparing a large-scale +1 dictionary. .

[Brief explanation of the drawing]

FIG. 1 is a schematic block diagram of a handwritten character recognition device according to an embodiment of the present invention, FIG. 2 is a block diagram of a recognition unit, and FIG. 3 is a graph of a function for determining an allowable error range of a feature vector. Fig. 4 is a block diagram of the character synthesis section, Fig. 5 is a conceptual diagram showing the structure of character dictionary data, Fig. 6 is a conceptual diagram showing the structure of stroke dictionary data, and Fig. 7 is a conceptual diagram showing the structure of decoration dictionary data. Conceptual diagram, Figure 8 is a conceptual diagram showing the contents of the thick line information, Country 9 is a diagram showing an example of a stroke pattern to explain the meaning of the thick line information, and Figure 10 is a diagram showing the meaning of the thick line information. - Examples of explanatory diagrams, Figures 11 and 12, are bare bones and fleshed-out strokes. Figure 1:1 is an explanatory diagram of the chain needle, Figure 14 is a flowchart showing pattern synthesis processing by the character composition unit, Figure 15 is the creation of a dictionary for recognition of handwritten characters. It is a flowchart showing (update) processing. 1... Character reading section, 2... Character synthesis section. 3... Key human power device, 4... Recognition section, 5...
External storage device, 6...Display device, 7...Control unit. IO: Feature extraction unit; 11: Dictionary search unit. 13...Dictionary IF registration section, 14...Dictionary memory. 20... Input port, 22... Main processing unit, 23
... Character dictionary memory, 24... Stroke dictionary memory, 25... Decoration dictionary memory. 26... Stroke generation part, 28... Decoration generation part. , 30... Output control unit, 36... Output cover sofa. 38 - Smoothing processing section. Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15

Claims (1)

[Claims] (1) In a character recognition device that can recognize handwritten characters and is equipped with a device that extracts feature vectors of input character patterns and registers them in a handwritten character recognition dictionary, a character synthesis device is used to synthesize transformed patterns of characters. A method for creating a dictionary for handwritten character recognition, characterized in that by selecting and inputting a character pattern similar to a user's handwritten character from among them, the feature vector of that character pattern is registered in the dictionary for handwritten character recognition. .