JPH10198761A - Character recognizing method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recognize characters with high segmentation accuracy by checking the mutual connecting relation between the codes of candidate character patterns having the adjacent relation in character string patterns. SOLUTION: This device includes an image input part 10, a candidate character pattern extraction part 12, a character recognition part 14, a candidate path production part 16 and a best result selection part 18. The part 16 reads a 1st character code out of a character code memory 22 for an extracted 1st candidate character pattern. In the same way, a 2nd character code is read out for a 2nd candidate character pattern that is adjacent to the 1st one. If the connecting relation between the 1st and 2nd character codes satisfies the connection rule that is previously set, the 1st character code is recorded in a path memory. Thus, it's possible to judge whether or not the characters form a candidate path by checking the connecting relation between the character codes of the candidate character patterns which are adjacent to each other in a character string pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for recognizing handwritten characters.

[0002]

2. Description of the Related Art Handwritten characters have a large change in character spacing and character shape as compared with standard characters. Therefore, conventional methods such as "characters are cut out at regular intervals" cannot provide sufficient character cutout accuracy. However, erroneous recognition occurs. As a solution to this problem, for example, Reference 1 “Technical Report of the Institute of Electronics and Communication Engineers PRL84-13 pp67-76” and Reference 2 “Transactions of the Institute of Electronics and Communication Engineers '86 / 9 Vol.J69-D No.9 pp1292-1301 And Japanese Unexamined Patent Application Publication No. 4-160582.

According to the character recognition method disclosed in Document 1, characters of extraction candidates are extracted from a character string pattern, and the best extraction position is selected so as to be the most character-like combination by using the recognition result. . However, in this method, since the candidate character pattern is normalized as preprocessing, the graphic characteristics of the pattern are lost. Even though there are actually a plurality of characters, one character has a high standard pattern. In some cases, recognition is performed based on the similarity. Thus, an effective character cut-out result cannot be obtained simply by using the recognition result.

On the other hand, the character recognition method disclosed in Document 2 is a method of selecting a recognition result satisfying linguistic and grammatical (that is, linguistic) connectivity as the best cutout position. . In other words, it is assumed that not only the character recognition result of the candidate character pattern but also the relevance with the recognition result of the adjacent candidate character pattern is better in a linguistic manner.

[0005] The character recognition method disclosed in Reference 3 is a method that introduces the continuity of graphic features. When a graphic connection is introduced, a method of cutting out such that a boundary line for cutting out a character does not cross the character itself as much as possible is adopted as a good cutting method. That is, a cutout position that cuts characters too much is not selected as the best cutout position.

[0006]

However, the character recognition method disclosed in the above-mentioned document 2 is difficult when the character string pattern to be recognized is as short as about one word and there are a plurality of possible words. However, there is a problem that a cut-out separation error is inevitable. For example, when recognizing a very short character string pattern such as "action", both "action" and "gravity work" are hit as a result of word matching, and it is not possible to determine which is the correct answer. In addition, as for the linguistic concatenation, as the dictionary of concatenation information increases, the concatenation information between basic words constituting a compound word becomes weak constraint information, and it is difficult to cut out the apartment name and the number of issues appearing in the address etc. . In other words, apartment names are often compound words or coined words,
Any word may be used, and for these word combinations, the effect of the linguistic concatenation information is negligible.

Further, in the character recognition method disclosed in the above-mentioned reference 3, in the case of a word composed of characters separated into a plurality of parts, such as "Ogawa", the expected extraction candidate position is
In order to come to a position passing between the vertical bars, that is, a position where the stroke is not divided, it is difficult to cut out the character even if a degree of graphic coupling is introduced.

To summarize the above: (1) When recognizing a word (such as “Ogawa”) separated into a plurality of parts or a short character string pattern, linguistic information (Reference 2) and graphic connectivity ( If only the reference 3) is used, a separation error occurs for the above-described reason, so that the cutout accuracy is reduced and erroneous recognition occurs.

(2) Words such as an apartment name, an address, and a number are not linguistically limited, and therefore, can be said to be a word string having very many linguistic combinations and very weak linguistic connection.
In such a case, the knowledge processing used in Literature 2 does not function, and the extraction accuracy is reduced.

(3) In the address character string pattern,
Characters that are keys for address recognition such as “town” and “village” are characters that are easily separated side by side, and that are formed as characters even if separated, so that graphic connectivity (Reference 3) It is difficult to judge separation / non-separation only by considering ()). For this reason, the cutting accuracy is reduced.

Therefore, there has been a demand for a character recognizing method capable of recognizing with higher extraction accuracy than the conventional method and a character recognizing apparatus for performing a character recognizing process according to the character recognizing method.

[0012]

According to the character recognition method of the present invention, (a) a plurality of consecutive candidate character patterns are extracted from a character string pattern to be recognized, and the position information is recorded in a coordinate memory. (B) recognizing the extracted candidate character pattern to determine an appropriate number of character codes, and recording these character codes in a character code memory; and (c) based on the recorded position information. A character recognition method comprising: arranging the recorded character codes to create a candidate path; and (d) selecting a best path from the created candidate paths based on the recognition result. The step (c) comprises:
(C1) designating the extracted candidate character pattern as a first candidate character pattern; and (c2) using the character code determined for the first candidate character pattern as a first character code from the character code memory. (C3) specifying a candidate character pattern adjacent to the first candidate character pattern as a second candidate character pattern based on the recorded position information; and (c4) specifying a second candidate character pattern as the second candidate character pattern. Reading from the character code memory the character code obtained as a second character code, and (c5) reading the first character code and the second character code.
Recording the first character code in a path memory when the connection relation of the character codes satisfies a predetermined connection rule. The step (c1) to the step (c5) are repeated to execute the path memory. The sequence of character codes recorded in the candidate path is set as the candidate path.

As described above, the character codes of the candidate character patterns having the adjacent relation in the character string pattern are read out from the character code memory, and the connection relation between them is examined. You can determine if there is. Therefore, if the connection rule is set, the character string pattern can be recognized more accurately than before. Here, the connection relation described above includes, for example, a relation between character types and a relation between character shapes. And
When the rule that “character types are the same” is set as the connection rule, the connection rule is satisfied when the read character codes have the same character type.

In the character recognition method according to the present invention, preferably, the connection rule is satisfied when the first character code and the second character code have the same connection type. Is good.

If the connection rule is set in this way, a character string pattern composed of character codes of the same character type is output as a recognition result. This concatenation rule is effective when recognizing a character string pattern that is often represented by the same character type, such as first and last names.

In the character recognition method according to the present invention,
Preferably, the first character code is a character code belonging to the classification of homomorphic characters.

Thus, in this example, the first character code and the second character code are character codes of the same character type,
Further, there is a case where the first character code is a character code belonging to the classification of homomorphic characters. Here, the homomorphic character refers to a character having the same shape, for example, a horizontal bar character such as a hyphen "-", a long sound "-", a Chinese numeral "1", and the like, and a numeral "0", It is a round character such as the letter "O".
In addition, vertical bar characters such as the numeral “1” and the alphabetical character “l” also apply to this. For example, normally, when a horizontal bar character is read, any of the character codes of the above-mentioned hyphen “−”, long sound “−”, and Chinese numeral “1” forms a character string with the same probability. In many cases, the data is output as a result, resulting in erroneous recognition. On the other hand, according to the method of the present invention, a character having the same character type as the adjacent character code is selected, so that erroneous recognition is reduced.

In the character recognition method according to the present invention, preferably, when the first character code and the second character code have a specific combination, the second character code is deleted from the path memory. Preferably, a character code corresponding to the specific combination is recorded in the path memory instead of the first character code.

As described above, in this example, when two character codes having a continuous positional relationship in the character string pattern are specific character codes, they are set in advance instead of these character codes. The recorded character code is recorded in the path memory as a character constituting the candidate path. Therefore,
If the character is a character that is often included in the character string pattern to be recognized and that character is easily separated into two parts, and the separated part itself is a character, as described above, this character is By setting in advance a character code corresponding to a specific combination, a separation error can be avoided, and erroneous recognition hardly occurs.

For example, the specific combination is a combination in which the second character code is “ta” and the first character code is “cho”, and the character code corresponding to this combination is “machi”. Is the case. As described above, the character “machi” is easily separated into the characters “ta” and “cho”, and therefore, a character string pattern including this character is easily erroneously recognized. However, it is a character that frequently appears in a character string pattern such as an address. Therefore, as described above, when the characters “ta” and “cho” are sequentially read out,
Since "town" is recorded instead of "and" in the path memory, accurate character string patterns can be recognized.

Further, without being limited to the above-described example, for example, the specific combination is a combination when the second character code is “tree” and the first character code is “dimension”. The character code corresponding to this combination may be “village”. The word "village"
This is a character that often appears in a character string pattern indicating an address. Therefore, if such characters are registered in advance, the address can be accurately recognized.

Next, in carrying out the character recognition method of the present invention, preferably, the extraction of the candidate character pattern is performed by (a
1) extracting a black block area of the character string pattern as a segment and recording the position coordinates of this segment in the coordinate memory; and (a2) integrating the extracted segments based on their respective position coordinates. Generating a new segment and appending the position coordinates of the new segment to the coordinate memory.
It is preferable that the position coordinates of the segment and the new segment recorded in the coordinate memory by executing the steps 1) and (a2) are set as the position coordinates of the candidate character pattern.

Here, the above-mentioned black block area is a rectangular area circumscribing a character pattern constituting a character string pattern read as a recognition target. However, the "character pattern" in this case is not guaranteed to be a meaningful one-character pattern, and may be only a partial part or only a side part. As described above, a new segment can be generated by extracting segments and integrating the extracted segments, and a candidate character pattern can be extracted.

In the character recognition method of the present invention,
Preferably, the generation of the new segment includes the steps of: reading out position coordinates of the recorded segment #Sn (n is an integer) from the coordinate memory;
determining position coordinates of another segment #Sk the n (k is an integer), a step of reading from the coordinate memory, the distance value D _nk between segments #Sn and segment #Sk from position coordinates of each read the And a step of comparing a constant e (e is a positive real number) times the line height L of the character string pattern with the obtained distance value D _{nk. When} the result of the comparison is D _nk ≦ e · L And segment #Sn
And the step of integrating the segment #Sk with segments between these segments to generate the new segment.

As described above, by comparing the distance value with the row height based on the distance value between the adjacent segments, it is possible to determine whether or not to integrate the set of segments. Here, the distance value is an amount proportional to the distance between the segments on the image or the information medium, and may be, for example, a length of a straight line connecting the start end positions of the segments and being parallel to the character string direction. As the line height, the maximum value of the length of the black block area in the direction perpendicular to the character string direction is used. In this way, when the distance between the segments is smaller than the height of the segments, the segments are set to be integrated.

Next, according to the character recognition device of the present invention,
It reads an original image including a character string pattern to be recognized, an image input unit having an image memory for storing the original image, and extracts a plurality of continuous candidate character patterns from the character string pattern, A candidate character pattern extraction unit that records the character code in a coordinate memory, a character recognition unit that recognizes the extracted candidate character pattern to determine an appropriate number of character codes, and records these character codes in a character code memory. Arranging the recorded character codes on the basis of the obtained position information to create a candidate path, and selecting a best path from the created candidate paths based on the recognition result. In the character recognition device including a best result selection unit, the candidate path creation unit is configured to specify the extracted candidate character pattern as a first candidate character pattern. A first character reading unit for reading the character code determined for the first candidate character pattern from the character code memory as the first character code,
A second specifying unit that specifies a candidate character pattern adjacent to the first candidate character pattern as a second candidate character pattern based on the recorded position information; and A second character reading unit that reads a character code from the character code memory as a second character code, and the first character code is connected to the first character code and the second character code when the connection relationship satisfies a predetermined connection rule.
A connection rule checking unit that records a character code in a path memory.

As described above, the character recognition device having this configuration has the following features.
Reading character codes of candidate character patterns having an adjacent relationship in a character string pattern from a character code memory, and examining a connection relationship therebetween to determine whether or not the character is a character constituting a candidate path. Can be. Therefore, if a connection rule is set, character strings can be recognized more accurately than in the past.

In the character recognition device according to the present invention, it is preferable that the connection rule is satisfied when the first character code and the second character code have the same connection type.

By setting the connection rule in this way, a character string composed of character codes of the same character type is output as a recognition result. As mentioned earlier, this association rule
This is effective when recognizing a character string often represented by the same character type, such as a first and last name.

Further, in the character recognition device of the present invention,
Preferably, the first character code is a character code belonging to the classification of homomorphic characters.

As described above, in this configuration example, the first character code and the second character code are character codes of the same character type, and the first character code is a character code belonging to the classification of homomorphic characters. It is. As described above, the homomorphic characters include horizontal bar characters, round characters, vertical bar characters, and the like. Then, for example, normally, when a horizontal bar character is read, a hyphen "-", a long sound "-", and a Chinese numeral "one"
Are output as constituting a character string with the same probability, and therefore, there are many cases where erroneous recognition occurs. On the other hand, according to the apparatus of the present invention, a character having the same character type as the adjacent character code is selected, so that erroneous recognition is reduced.

In the character recognition device of the present invention,
Preferably, the connection rule checking unit deletes the second character code from the path memory when the first character code and the second character code have a specific combination, and corresponds to the specific combination. It is preferable that the character code is recorded in the path memory instead of the first character code.

As described above, the connection rule checking unit of this example
When two character codes having a continuous positional relationship in the character string pattern are specific character codes, a character code set in advance is replaced with a character code that constitutes a candidate path instead of these character codes. Means for recording in the path memory. Therefore, if the character is a character that is often included in the character string pattern to be recognized, and the character is easily separated into two parts, and the separated part itself is a character, as described above, By preliminarily setting characters as character codes corresponding to specific combinations, separation errors can be avoided and erroneous recognition can be prevented.

For example, the specific combination is a combination when the second character code is “ta” and the first character code is “cho”, and the character code corresponding to this combination is “machi”. Is the case. With this setting, as described above, a character string pattern indicating an address can be accurately recognized.

Further, without being limited to the above-described example, for example, the specific combination is a combination when the second character code is “tree” and the first character code is “dimension”. The character code corresponding to this combination may be “village”. With this setting, a character string pattern indicating an address can be accurately recognized as described above.

Next, in the character recognition device of the present invention,
Preferably, the candidate character pattern extraction unit extracts a black block region of the character string pattern as a segment, and records a position coordinate of the segment in the coordinate memory, and separates the extracted segment from each other. And a segment integrating unit that adds the position coordinates of the new segment to the coordinate memory, based on the position coordinates of the candidate character pattern. It is better to extract as position coordinates.

As described above, a new segment is generated by extracting a segment and integrating the extracted segments, and a candidate character pattern can be extracted.

In the above-described character recognition device, preferably, the segment integrating section includes a first coordinate reading section that reads out the position coordinates of the recorded segment #Sn (n is an integer) from the coordinate memory; The segment #S
a second coordinate reading unit for reading the position coordinates of a segment #Sk (k is an integer) different from n from the coordinate memory; and a distance value D _nk between the segment #Sn and the segment #Sk.
Is compared with the distance value detection unit that obtains from the read position coordinates, a constant e (e is a positive real number) times the line height L of the character string pattern, and the calculated distance value _Dnk. And a generating unit that integrates the segment #Sn, the segment #Sk, and the segment between these segments to generate the new segment when the result of the comparison is D _nk ≦ e · L, Preferably, a writing unit for adding and recording the generated position information of the new segment to the coordinate memory is provided.

As described above, by comparing the distance value with the row height based on the distance value between the adjacent segments, it is possible to determine whether or not to integrate the set of segments. In this way, when the distance between the segments is smaller than the height of the segments, the segments are set to be integrated.

[0040]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. The figures are schematically shown to the extent that the configuration, arrangement, and operation of the present invention can be understood, and the numerical conditions and the like described below are merely preferred examples. Therefore, the present invention is not limited to this embodiment.

[First Embodiment] FIG. 1 is a block diagram showing the configuration of a character recognition device according to the first embodiment.
As shown in FIG. 1, the character recognition device of this embodiment
An image input unit 10, a candidate character pattern extraction unit 12, a character recognition unit 14, a candidate path creation unit 16, and a best result selection unit 18 are provided. Further, the character recognition device includes a coordinate memory 20 for recording position information and a character code memory 22 for storing character codes as storage means. The character recognition device includes the image input unit 10, the candidate character pattern extraction unit 12, the character recognition unit 14, the candidate path creation unit 16, and the best result selection unit 1 described above.
8. The control unit 24 manages the operation timing and data input / output of the coordinate memory 20 and the character code memory 22. The character recognition device mainly has the above-described configurations. With these configurations, a character string pattern is read from an information medium, and recognition is performed by converting the character string pattern into a character code.

The character code (or cut-out position) which is the output result of this character recognition device is output from the above-described best result selection unit 18, and this output is input to, for example, a so-called computer device where the character recognition information is output. It is utilized as. Further, the above-described candidate character pattern extraction unit 1
2. Character recognition unit 14, candidate path creation unit 16, best result selection unit 18, coordinate memory 20, character code memory 22,
The control unit 24 may configure the hardware of each element as a computer device including a central processing unit (CPU), an input / output device, and a storage unit, or each unit described above as one computer device. It is good also as a collective structure.

FIG. 2 is a flowchart showing the operation flow of the character recognition device of this embodiment. As shown in FIG. 2, the character recognition process performed by the character recognition device includes character type connection rule input (S1 in FIG. 2) and image input (S1 in FIG. 2).
2), segment extraction (S3 in FIG. 2), segment integration (S4 in FIG. 2), character recognition (S5 in FIG. 2), candidate path creation (S6 in FIG. 2), evaluation value calculation (S7 in FIG. 2), The sorting (S8 in FIG. 2) and the output (8 in FIG. 2) are sequentially performed. Hereinafter, the procedure of the character recognition process will be described along with the configuration of each unit and the operation thereof according to the flow of FIG.

<Input of Concatenation Rule> The character recognition device of this embodiment has a special feature in that processing using the concatenation rule is performed. This processing is performed by a candidate path creation unit 1 described later.
6, the connection rule is first input to the candidate path creation unit 16 (S1 in FIG. 2). The connection rule is determined by the connection rule inspection unit 2 provided in the candidate path creation unit 16.
6 (FIG. 12) is set by input means such as a keyboard. Then, the connection rule inspection unit 26 stores the input connection rule in a freely readable manner.

The concatenation rule is a rule determined to make the concatenation relationship between adjacent characters specific. In this embodiment, rules are set such that adjacent character codes have the same character type in a recognized character string. In this embodiment, as character types, numerals and symbols,
Four types of characters are assumed: katakana, hiragana, and kanji. And, for example, in the sequence of character codes constituting a character string, if the first character code is a kanji,
For the character code adjacent to the character code, the kanji of the same character type is selected next. That is, the character codes constituting the character string are recognized so that they all have the same character type.

FIG. 3 is a diagram conceptually showing a state of the connection rule set in connection rule inspection section 26 of this embodiment. The character type connection rule 28 is composed of three rules. Rule 1 (Kanji → Kanji) is a rule that “if the first character of two adjacent characters is a kanji, the next character must also be a kanji”. Rule 2
(Flat → Flat) means “Of the two adjacent characters,
If the first character is flat, the next character must be flat. " Similarly, rule 3
(Katakana → Katakana) is a rule that “if the first character of two adjacent characters is katakana, the next character must also be katakana”. Applying these three rules is effective, for example, in recognizing a character string of a name. This is based on the fact that character strings representing names usually consist of the same character type in most cases. Also, since there are almost no names using symbols and numbers, no rules are set for this.

It should be noted that the above-described connection rule can be set for each partial area of the document area to be subjected to character recognition.
For example, when a form on which a table or the like is described is to be recognized, a target area for character recognition is divided for each item. Then, when recognizing the column of “name” of the item, it is configured to apply the above-described connection rule,
Also, when recognizing another item, it may be configured so that the corresponding connection rule is applied. For this reason, the connection rule inspection unit 26 can set a connection rule for each partial area, and can store the connection rule in association with the position coordinates in the document area.

The character recognition operation for recognizing the character string pattern shown in FIG. 4 will be described below. FIG. 4 shows a handwritten character pattern (character image) of “Kozo” as a character string pattern (input image) 30. As shown in the figure, in this embodiment, it is assumed that the character string is written horizontally, but the present invention is not limited to this and may be written vertically. Hereinafter, the steps described above will be described for the case where the column in which the name “Kozo” is described is to be recognized. For the above-described reason, it is assumed that the above three rules (FIG. 3) are set in the connection rule inspection unit 26 as the connection rules for the character string pattern 30 shown in FIG.

<Input of Image> Next, the step of inputting an image (S2 in FIG. 2) will be described. In this step, the image input unit 10 performs a process of optically reading the target document area and recording the read image data in the storage unit. This process is performed after inputting the connection rules.
The process is started when an operator gives an instruction to the image input unit 10 via the control unit 24.

The above-described image input unit 10 has an image memory 32 for reading an original image including a character string to be recognized and storing the original image. Further, similarly to the conventionally known configuration, although not shown, a scanning mechanism (scanner) as a means for moving an information medium such as a form or a document in the main scanning direction and the sub-scanning direction, and an original image from the information medium. It comprises a photoelectric conversion unit for optically converting it into an electrical signal to be read, and a cutout unit for cutting out a predetermined area from the read original image. With the above configuration, the image input unit 10 reads the original image from the information medium and stores the original image in the image memory 3.
2 and a predetermined area from the original image (character string pattern)
And cut out. The reading of the original image from the information medium may be performed as a binary image or a multi-valued image. As the image input unit 10, for example, a so-called tablet may be used.

<Extraction of Candidate Character Pattern> Next, the steps of extracting candidate character patterns (S3 and S4 in FIG. 2) will be described. The process of this step is performed by the candidate character pattern extraction unit 12. The candidate character extracting unit 12 extracts a plurality of consecutive candidate character patterns from the character string pattern to be recognized, and stores the position information thereof in the coordinate memory 20.
This is a means for recording. And in this embodiment,
The candidate character pattern extraction unit 12 includes a segment extraction unit 34
And a segment integration unit 36. First, the segment extraction unit 34 extracts a segment (S in FIG. 2).
3) Subsequently, the segment integrating unit 36 integrates the segments (generates a new segment) (S4 in FIG. 2).

When the image input unit 10 notifies the control unit 24 that the image input has been completed, the control unit 24 starts the operation of the segment extraction unit 34. Alternatively, the operator instructs the control unit 24 to output the candidate character pattern extraction unit 12
A configuration may be employed in which an instruction is directly given so as to start the operation of (1).

The above-described segment extracting section 34 is means for extracting a black block area of the character string pattern obtained by the image input section 10 as a segment, and recording the position coordinates of this segment in the coordinate memory 20. Hereinafter, the case where the segment extraction of the character string pattern 30 shown in FIG. 4 is performed will be described. In FIG. 4, the character string direction is the x coordinate direction (main scanning direction) in the figure. In FIG. 4, the above-mentioned black block area is a rectangular area (referred to as a segment) circumscribing each character pattern.

The character string pattern 30 is a character pattern "bow".
Segment # S0 circumscribing ("Hiro"), segment # S1 circumscribing the character pattern "mu" (next to "Hiro"),
And the segment # S2 circumscribing the character pattern "3"
Is a character string pattern consisting of These segments are sequentially arranged in the character string direction, such as # S0, # S1, and # S2.

In order to extract a segment, first, scanning is performed in the character string direction (x direction) to obtain a projection distribution, that is, a histogram of black spots, in which the character string pattern 30 is projected on the X axis. Then, the minimum point of this histogram is x
This is the cutting position on the axis. Next, x of the character string pattern 30
By scanning in the y-axis direction perpendicular to the axis, a histogram is similarly obtained. Then, the minimum point of the histogram is set as a cutout position on the y-axis. In this way,
A rectangular area surrounded by cutout positions on the x-axis and the y-axis is obtained. In the present embodiment, obtaining this rectangular area is referred to as segment extraction, and can be performed by the above-described known method.

As described above, the extraction of the segment is performed by detecting the coordinates (pixel positions) of four points (corresponding to each vertex of the segment) which are the intersections between the obtained cutout positions. The coordinates of the detected segment are stored in the coordinate memory 20 as a segment coordinate table.

FIG. 5 shows a storage state (internal state) of the coordinate memory 20, that is, a segment coordinate table. On the left side of the figure, the character pattern "mu", that is, the segment # S1, is shown, and on the right side of the figure, the segments # S0, # S1, and #
Respective coordinates of S2 (the x-coordinate component X _s and X _e in FIG. 5, represented by the y coordinate component Y _s and Y _e.) Shows a segment coordinate table describes. For example, (shown surrounded by a broken line a.) Coordinate components of the segment # S0 is, X _s is 1, X _e is 36, Y _s is 1 and Y _e
Is 84 and so on. As described above, the segment extracting unit 34 coordinates the four vertexes of the segment in a predetermined storage location corresponding to the segment number (for example, the suffix 0 of the symbol # S0 representing the segment # S0) in the coordinate memory 20. It is recorded as a component value in a readable manner. Then, when the segment extraction is completed, the segment extraction unit 34 issues an instruction to the segment integration unit 36 via the control unit 24 to start the operation.

The above-mentioned segment integrating section 36 integrates the segments extracted by the segment extracting section 34 based on the respective position coordinates to generate a new segment, and stores the position coordinates of the new segment in the coordinate memory 20. This is a means for appending. FIG. 6 is a block diagram showing a detailed configuration of the segment integration unit 36. The segment integrating unit 36 of this embodiment includes a first coordinate reading unit 38, a second coordinate reading unit 40, a distance value detecting unit 42, a comparing unit 44, and a generating unit 46.
And a writing unit 48.

The operation of the segment integrating unit 36 when the character string pattern 30 shown in FIG. 4 is to be processed will be described below. As described later, in the example of the character string pattern 30 of FIG. 4, the segment # S0 and the segment # S1 are integrated to generate a new segment # S3. As described above, generation of a new segment is performed by the segment extraction unit 34.
Is performed by integrating the segments extracted by the above based on the respective pieces of position information. In describing this integration processing, a flowchart shown in FIG. 7 will be referred to.

First, the first coordinate reading unit 38 reads the position coordinates of the segment #Sn recorded by the segment extracting unit 34 from the coordinate memory 20 (S10 in FIG. 7).
Next, the second coordinate reading unit 40 sets a segment # different from the segment #Sn called by the first coordinate reading unit 38.
The position coordinates of Sk are read from the coordinate memory 20 (FIG. 7).
S11). In FIG. 7, a hexagonal flow chart symbol S10
The step marked with is the hexagonal flow diagram symbol S1
This means that steps between the step indicated by No. 6 and the step indicated by S6 are repeatedly performed until the condition described in S10: Loop 1 is satisfied. Similarly, the hexagonal flow chart symbol S1
The step indicated by 1 means that the steps between the flowchart symbol S15 and the flow chart symbol S15 are repeatedly performed until the condition described in S11: the loop 2 is satisfied. In this flow, the loop 1 is continued until the position information of all the segments recorded in the coordinate memory 20 is read by the first coordinate reading unit 38. Loop 2 is
The process is continued until the position information of all the segments recorded in the coordinate memory 20 is read by the second coordinate reading unit 40.

The segment position information read by the first and second coordinate reading units 38 and 40 is input to a distance value detecting unit 42 described below and used there. Since the loop 1 includes the loop 2 as a nested loop, all combinations of segments are input to the distance value detection unit 42. In order to input the position information of the coordinate memory 20 in this manner, the control unit 24
Has two counters for executing the integration processing, designates a segment number by these count numbers n and k, and specifies the first and second coordinate reading units 38 and 40.
Give instructions to However, the control unit 24 always counts the number of counts such that n <k, by the first and second coordinate reading units 3.
8 and 40.

Next, the steps executed for the above-mentioned loop 1 and loop 2 will be described with reference to the following (I), (II) and (III).

(I) The distance value detecting section 42 outputs the segment #
A distance value D _nk between Sn and #Sk is obtained from each position information read by the first and second coordinate reading units 38 and 40 (S12 in FIG. 7). Here, the distance value D _nk between the segments #Sn and #Sk is, as shown in FIG.
starting end position of k is defined as the linear distance character string direction connecting to each other (the position of the above-mentioned X _s). For example, according to the segment coordinate table shown in FIG. 5, D ₀₁ = 36.
As described above, the distance value detection unit 42 is configured to detect the distance value by detecting the difference between the x-coordinate components at the start end position of each segment. For example, the distance value detection unit 4
As 2, a normal difference calculation circuit can be used.

(II) Then, the distance _Dnk detected by the distance value detecting section 42 is transmitted to the comparing section 44. The comparison unit 44 compares the line height L of the character string pattern with a constant e times (e is a positive real number) and the distance value D _nk obtained by the distance value detection unit 42 (S13 in FIG. 7). ). Here, the line height L refers to the length of each segment constituting the character string pattern 30 in the direction (y direction) perpendicular to the character string direction. However, in this embodiment, the maximum value of the above-described length is defined as the line height L of the character string pattern. Therefore, the maximum value may be detected in advance from the position coordinates of each segment stored in the coordinate memory 20 and stored in the comparing unit 44. For example, in the segment coordinate table shown in FIG. 5, L = 87.

In this embodiment, 1.2 is set as the constant e. The value “1.2” is a value determined empirically by performing repeated tests. The value of the constant e is set in advance in a readable memory means provided in the comparing section 44 by an input means such as a keyboard. The comparison unit 44 is configured to read the value of the constant e set from the above-described memory unit at the input timing of the distance value _Dnk from the distance value detection unit 42.

With such a configuration, the comparison unit 44
Can determine the magnitude relationship between the input distance value D _nk and the value 1.2L. Then, the comparison unit 44 outputs a signal corresponding to the obtained magnitude relationship to the generation unit 46.

(III) Subsequently, the generation unit 46 sets the comparison unit 44
Is a signal corresponding to D _nk ≦ e · L, a new segment is generated by integrating the segment #Sn, the segment #Sk, and the segment between the segments #Sn and #Sk (FIG. 7 S14). In this example, since D ₀₁ = 36 and L = 87, the comparison result of the comparison unit 44 is D ₀₁ ≦ e · L, and the generation unit 46 sends the segment from the first and second coordinate reading units 38 and 40 to the segment. # S0
And the position information of the segment # S1 is input, and the position information of these segments is integrated. As shown in FIG. 4, a new segment # S3 is generated by integrating the segments # S0 and # S1 (the segment # S3 is a rectangular area circumscribing the character pattern "Hiro"). In addition, not only the case where two segments are integrated, but also the case where three or more consecutive segments are integrated. That is, when the distance value between the segment #Sn read by the first coordinate reading unit 38 and the segment #Sk read by the second coordinate reading unit 40 is smaller than eL, the segment #Sn and the segment #S
k and all the segments existing between the segments #Sn and #Sk are integrated as one new segment.

The new segment generated in this manner is equivalent to newly setting an area on a character string pattern including both the segment #Sn and the segment #Sk. As shown in the segment coordinate table after integration in FIG. 9, one of the coordinate component values corresponding to the segment #Sn and the segment #Sk is selected as the coordinate component value of the generated new segment. Is set. Then, the selection is made such that the new segment to be generated occupies as large an area in the character string pattern as possible. Taking the segment # S0 and the segment # S1 as an example, for X _s component, segment # S0
But 1, segment # S1 is 37, 1 is set as the coordinate component X _s new segment # S3. Also, X
The _e component, the segment # S0 is 36, the segment # S1 is 106, 106 is set as the coordinate component X _e new segment # S3.

The above treatments (I), (II) and (III)
Is repeatedly performed until the conditions of the loops 1 and 2 described above are satisfied, so that the integration process is performed on all sets of segments. Then, the writing unit 48
The position information of the new segment generated by the generation unit 46 is additionally recorded in the coordinate memory 20 (S1 in FIG. 7).
7). As shown in FIG. 9, a new segment # S3 is additionally written.

As described above, the segment # S0, # S1, and # S2 and the new segment # S3 are obtained by the candidate character pattern extraction unit 12. Hereinafter, these segments # S0, # S1 and # S2 and the new segment # S3 are collectively referred to as candidate character patterns.

<Recognition of Candidate Character Pattern> Next, a step of performing character recognition of the extracted candidate character pattern will be described (S5 in FIG. 2). The processing performed in this step is
This is performed by the character recognition unit 14. The character recognizing unit 14 of this embodiment is means for recognizing the extracted candidate character pattern, obtaining an appropriate number of character codes, and recording the character codes in the character code memory 22. Here, a character code corresponding to the candidate character pattern, which is a certain number of upper character codes in ascending order of the degree of difference, is obtained and recorded in the character code memory 22 for each candidate character pattern. As a result of this character recognition, the candidate character pattern that is the image data is encoded into a character code.

First, the control unit 24 controls the segment integrating unit 3
In response to the operation end signal from step 6, the position information of the candidate character pattern stored in the coordinate memory 20 is read.
The control unit 24 cuts out an area (referred to as a character pattern) corresponding to the candidate character pattern from the original image (character string) stored in the image memory 32 based on the read position information. The extraction of the character pattern may be performed by the extraction unit included in the image input unit 10 described above, or the control unit 24 may include an extraction unit for this purpose. The cut-out character pattern is transmitted to the character recognition unit 14. Then, the character recognition unit 14 performs a normal character recognition process on the captured character pattern. Therefore,
The character recognizing unit 14 includes extraction means for extracting features from the extracted character patterns, and a dictionary in which the characteristics of the standard characters are stored in advance. And matching means for comparing the character with the standard character. The degree of difference between features is calculated by this matching means. Then, the top 10 character codes are obtained for one candidate character pattern in ascending order of the degree of difference.

Here, the degree of difference is an amount indicating the degree of difference between the feature of a candidate character pattern to be recognized and the feature of a character code converted from the candidate character pattern by recognition. In this embodiment, a distance (different from the distance _Dnk between segments described above) is detected as the degree of difference. This is a quantity representing the closeness between feature vectors. In this embodiment, there is no particular limitation on the method for obtaining the distance between features. Also, the scale used for comparing the features may be arbitrarily set.

FIG. 10 shows a part of the recognition result. Segment # S0 is shown on the left side of the figure, and the top 10 character codes obtained for segment # S0 are shown in the table on the right side of the figure in association with their order and distance. FIG.
As shown in the correspondence relation of 0, the character recognizing unit 14 converts the acquired character code (broken line b) and the distance to the character code (broken line c) into the character codes corresponding to the candidate character numbers in ascending order of the distance. In the storage location (address) of the memory for use 22. FIG. 11 shows the character code and the distance thus obtained for each candidate character pattern. FIG. 11A shows the recognition result of # S0, and FIG.
1 and the recognition result of # S2 in FIG.
FIG. 11D shows the recognition result of # S3 in a table.

<Creation of Candidate Path> Next, the operation of the candidate path creation section 16 will be described (S6 in FIG. 2). The configuration of the candidate path creation unit 16 is shown in the block diagram of FIG. The candidate path creation unit 16 is means for arranging the character codes recorded in the character code memory 22 based on the position information recorded in the coordinate memory 20 and creating a candidate path. Therefore, the candidate path creating unit 16 includes the first specifying unit 50, the first character reading unit 52, the second specifying unit 54, and the second character reading unit 5
6 and a connection rule checking unit 26. Also, the first
And the second specification sections 50 and 54 are means for specifying the extracted candidate character patterns, respectively. In particular, the first specification section 50 specifies this specification as a cutout candidate point #Ci (i is an integer). It is done by doing. For this reason, the candidate path creation unit 16 includes a table creation unit 58 for creating the correspondence between the cutout candidate point #Ci and the candidate character pattern. In addition, the candidate path creation unit 16 includes a path memory 60 for recording the created candidate paths. still,
The candidate path is an array of character codes. In this case, the candidate path means an array having a possibility of converting a character string pattern into a character code.

First, the table creation unit 58 will be described. The table creation unit 58 reads out the position information of the candidate character pattern recorded in the coordinate memory 20 and obtains the start point position and the end point position of each candidate character pattern as the cutout candidate point #Ci based on the position information. This is a means for recording the correspondence between the pattern and the extraction candidate point.
A character pattern can be cut out from a character string pattern by designating two points on the x-axis as a start point position and an end point position, respectively. The table creation unit 58 designates all positions that can be specified for extracting a character pattern from the character string pattern 30 as extraction candidate points. FIG. 4 shows candidate extraction points of the character string pattern 30. String pattern 3
In the case of 0, four extraction candidate points # C0, # C1, #
C2 and # C3 can be specified.

On the other hand, if the extraction candidate points can be designated in this way, by selecting two of these extraction candidate points, an arbitrary candidate character pattern can be extracted. In this embodiment, the cutout candidate points are numbered in order in the main scanning direction. For example, by selecting the extraction candidate point # C0 and the extraction candidate point # C1,
Candidate character # S0 having # C0 as the start point position and # C1 as the end point position can be designated. Also, by selecting a candidate clipping point # C0 and a candidate clipping point # C2, a candidate character # S3 having # C0 as a start point position and # C2 as an end point position can be designated.

The extraction candidate points are extracted based on both end positions of the character string pattern and boundary positions between adjacent candidate character patterns. In the example of the character string pattern 30, the table creation unit 58 determines that the extraction candidate points # C0, # C1, and # C2
A candidate character # S0, and extracts the readout of the starting end position of # S1 and # S2 (X _s component). The extraction candidate point # C3 is set to the last segment #S of the character string pattern.
It is extracted by reading the end position ( _Xe component) of No. 2.

Next, the table creating section 58 obtains the correspondence between the set extraction candidate point and the candidate character pattern, and records this in the memory means. FIG. 13 shows, as a segment table, the internal state of the memory means in which the correspondence is recorded. This segment table is represented by an adjacency matrix commonly used in the field of graph theory. In the figure, the cutout candidate point #Ci (i = 0, 1, 2, 3) as the start point position is described in the item box of the row column, and the cutout candidate point as the end point position is described in the item box of the column column. Point #Ci (i = 0, 1, 2, 3) is described. Then, in the data frame located at the intersection of the row and the column, candidate character numbers #Sn corresponding to the start point position and the end point position are described. If there is no candidate character pattern corresponding to the extraction candidate point, a blank character (NULL) is included in the data frame.
Is set. In this way, by specifying the extraction candidate point as the start point position and the extraction candidate point as the end point position, the extraction target candidate character pattern and candidate character number can be specified from this table.

The first specifying section 50 is means for specifying the extracted candidate character pattern as a first candidate character pattern. As described above, the first specifying unit 50 specifies the candidate character pattern by specifying the extraction candidate point with reference to the segment table created by the table creating unit 58. The first character reading unit 52 is a unit that reads the character code obtained for the first candidate character pattern by the character recognizing unit 14 from the character code memory 22 as a first character code.

Further, the second specifying section 54 is provided in the coordinate memory 20.
Is a means for designating a candidate character pattern adjacent to the first candidate character pattern designated by the first designation unit 50 as a second candidate character pattern based on the position information recorded in the first candidate character pattern.
The second character reading unit 56 is means for reading the character code obtained for the second candidate character pattern from the character code memory 22 as a second character code.

Further, since the character recognition device of this embodiment includes not only the above-described configuration but also the connection rule checking unit 26, only the candidate paths satisfying the predetermined condition can be stored. ing. Connection rule inspection unit 26
Is a means for recording the first character code in the path memory 60 when the connection relation between the first character code and the second character code satisfies a preset connection rule. In this embodiment, the connection rule is satisfied when the first character code and the second character code have the same connection type and the connection type is the same. As the connection rules, the rules regarding the three character types described above are set.

The first specifying section 50 and the first character reading section 5
2, the second designation unit 54, the second character reading unit 56, and the connection rule checking unit 26 operate together to create a candidate path satisfying the connection rule. The operation is based on a graph search algorithm. Hereinafter, this operation will be described with reference to the candidate path creation function shown in the flowchart of FIG.

Candidate path creation function Funct (#Ci, #S
k, Mk) are the extraction candidate point #Ci and the candidate character #Sk
And the character code M that is the recognition result of the candidate character pattern
This is a function with k as an argument. First, the path memory 60
Is cleared, and # Ci = # C0, #S as initial values
k = # S0 and Mk = NULL are set. When the processing of this function is started, first, the first argument #C
The first specifying unit 50 determines whether or not the extraction candidate point specified by i is the last position in the character string pattern.
(S18 in FIG. 14). If #Ci is the last position, that is, the rightmost end, the processing of the candidate path creation function using #Ci as an argument ends. If the specified extraction candidate point is not the rightmost end, the processing of loop 1 and its nested loop, loop 2, is executed.

Processing of loop 1 (S19 and S19 in FIG. 14)
26 are continued until all of the candidate extraction points #Cj (j is an integer satisfying j> i) located on the right side of the candidate extraction point #Ci are designated by the first designation unit 50 ( For each # Cj∈ (Adj (# Ci # Cj)! = NULL)). Further, processing of loop 2 (processing between S21 and S25 in FIG. 14)
Indicates that the character code (Mk + 1∈ (candidate character of # Sk + 1)) obtained for candidate character # Sk + 1 is sequentially entered one by one in each loop, and all are in the first character reading unit 52. Until it is read. Here, candidate character # Sk +
1 represents the above-described first candidate character pattern specified by the cutout candidate points #Ci and #Cj. The selection of the first candidate character pattern is performed by the first specifying unit 50 with reference to the segment table (S20 in FIG. 14: # Sk + 1 ← Adj (# Ci # Cj)
). The character code Mk + 1 is the first character code described above.

In response to the designation of the first candidate character pattern by the first designation unit 50 and the reading of the first character code by the first character reading unit 52, the second designation unit 54 is adjacent to the left of the first candidate character # Sk + 1. Is designated as a second candidate character pattern. This designation can be made by referring to the coordinate memory 20. Then, the second specifying unit 54 instructs the second character reading unit 56 to read the character code Mk belonging to the second candidate character #Sk from the character code memory 22 as the second character code.

The above-described reading of the first and second character codes is performed such that one pair is formed for each loop, different pairs are created for each loop, and all combinations are finally realized. Done in Then, each pair is input to the connection rule checking unit 26. Connection rule inspection unit 2
6 determines whether or not the input first character code Mk + 1 and second character code Mk satisfy the connection rule (S22 in FIG. 14). In this way, the connection relation of all combinations is checked.

The flow proceeds to the following two steps or according to the inspection result of the connection inspection unit 26. First, if the connection rule is not satisfied, the process goes out of the loop 2. Next, when the connection rule is satisfied, the first character code Mk + 1 is added as a candidate path, and the path memory 60 is added.
(S23 in FIG. 14). Then, the processing of the candidate path creation function in which #Cj, # Sk + 1, and Mk + 1 are substituted as arguments is started (S24 in FIG. 14: #Cj,
# Sk + 1, Mk + 1 → Funct (# Cj, # Sk + 1, Mk + 1)). in this way,
By performing a process of recursively calling the candidate path creation function, a candidate path satisfying the connection rule is recorded in the path memory 60.

FIG. 15 shows the internal state of the path memory 60 after execution of the above-described processing. In the upper part of the figure, the character string pattern 30
And a part of the candidate path formed in the form of a two-terminal directed graph is shown on the lower side in the figure. In addition, character string pattern 3
The extraction candidate point of 0 is associated with the terminal of the graph. The array of candidate character patterns in this example is # S0,
An array of # S1 and # S2 and an array of # S3 and # S2 are conceivable. As an example of the former array, for example,
A candidate path of “many, six, three” (broken line d in FIG. 15) is recorded in the path memory 60. Also, the candidate path of “many, six, two” (broken line e in FIG. 15) is recorded. And, as a constituent of the latter arrangement, for example, "Hiro,
A candidate path “3” (broken line f in FIG. 15) is recorded.

As shown in FIG. 15, when the concatenation rule of this embodiment is used, only an arrangement of the same character type is obtained (an arrangement of kanji and not shown in FIG. 15,
The sequence of katakana is obtained. ). Further, since a character string of a name is considered as a recognition target, there are no rules for symbols and numbers. Therefore, only candidate paths that are possible as names remain, and the number of candidate paths can be reduced, so that separation errors can be reduced.

<Selection of Best Result> The process of this step is performed by the best result selection unit 18 (S7 in FIG. 2,
S8 and S9). As shown in FIG. 1, the best result selection unit 18 includes an evaluation value calculation unit 62 and a cutout position output unit 64, and selects the best path from the created candidate paths based on the recognition result. .

The above-described evaluation value calculation unit 62 evaluates each candidate path (S7 in FIG. 2). The evaluation of the candidate path is performed by calculating the sum of the distances (degrees of difference) of the character codes constituting the candidate path. The evaluation value calculation unit 62
The distance (degree of difference) between the candidate character patterns stored in the character code memory 22 is added in accordance with the candidate paths stored in the path memory 60. Then, the sum of the distances is sorted in ascending order (S8 in FIG. 2). Then, the cutout position output unit 64 detects a candidate path having the minimum distance sum from the sorting result, and outputs the detected path as the best path (S9 in FIG. 2). The cutout position output unit 64 outputs the cutout position and the character code to the outside.

In the candidate paths shown in FIG.
The candidate path “3” is the path with the shortest distance.
Therefore, this candidate path is selected as the best path and output to an external computer device or the like. Further, since the best path is determined in this manner, the cutout positions of the character string pattern 30 are the above-described cutout candidate positions # C0, # C2, and #C.
3 is determined.

[Second Embodiment] In a character recognition device according to a second embodiment, the contents of the connection rules set in the connection rule checking unit 26 and the candidate path creation unit 16
Is different from the configuration of the first embodiment. Therefore, different configurations will be mainly described, and redundant description will be omitted. The same components are denoted by the same reference numerals. Further, the processing of the character recognition device of this embodiment is also performed according to the flow shown in FIG. 2, so that FIG. 2 is referred to as necessary.

<Input of Concatenation Rule> FIG. 16 shows a character string pattern to be recognized in this embodiment. A character string pattern 30 shown in FIG. 16 is a character string pattern indicating an address, and “Park Avenue 1-2-32” is written by hand. When recognizing such an address character string pattern, it is preferable to set a connection rule as shown in the table of FIG. FIG. 17A shows a character type concatenation rule (rule 1) similar to that of the first embodiment, and FIG. 17B shows a concatenation rule (rule 2) relating to a character shape.

In rule 1, symbols, alphanumeric characters, katakana,
Five types of characters, Hiragana and Kanji, are assumed. Rule 1 is a rule that “adjacent candidate characters must be of the same character type”. This rule is based on the fact that many words consist of characters of the same character type. Rule 2 is a rule related to a character shape, and particularly a rule related to a horizontal bar character. That is, a hyphen "-" follows symbols and alphanumeric characters,
The rule is that the long sound "-" must follow the katakana and hiragana characters, and the kanji "1" must follow the kanji. A character string indicating an address, especially words below the municipalities are composed of apartment names and addresses, and are compound words combining the same character type or word, such as "~ heights" or "~ sho" Often. Rule 2 is a rule based on this.

The connection rule inspection unit 26 operates based on the connection rules described above. As in the case of the first embodiment, the connection rule is satisfied when the first character code and the second character code have the same character type. In this embodiment, it is assumed that the first character code is a character code belonging to the classification of homomorphic characters. The homomorphic characters targeted in this embodiment are a hyphen "-" and a long sound "-".
And a horizontal bar character such as the Chinese numeral “I”.

<Extraction of Candidate Character Pattern> Extraction of candidate character patterns can be performed in the same manner as in the first embodiment. Although not shown in the character string pattern 30 shown in FIG. 16, the segment # S0 and the segment # S1 are integrated to generate a new segment # S16. Also, segment #S
0, # S1 and # S2 are integrated to generate a new segment # S17. In this embodiment, 1.0 is used as the value N used in the segment integration processing.

<Recognition of Candidate Character Pattern> A part of the recognition result of the candidate character pattern is shown in FIG. The segment # S0 is shown on the left side of the figure, and the segment #S
The character codes obtained for 0 are shown in association with their ranks and distances. Although only two character codes corresponded to this segment # S0,
As in the first embodiment, the upper ten character codes are determined in ascending order of the degree of difference.

<Creation of Candidate Path> Next, the operation of the candidate path creation unit 16 will be described (S6 in FIG. 2). The configuration of the candidate path creation unit 16 is shown in the block diagram of FIG. The candidate path creation unit 16 of this embodiment refers to the path memory 60, and counts the length (number of character codes) of the path (array of character codes) recorded therein. A part 66 is provided. Each time one character code is added to the path memory 60, the path length counting unit 66 counts the length of the recorded path and selects the longest path. Hereinafter, the operation of the candidate path creation unit 16 will be described with reference to FIG.
This will be described with reference to the flowchart of the candidate path creation process shown in FIG. The candidate path creation function Fun that appears in the flow
ct (#Ci, #Sk, Mk) corresponds to steps S19 to S26 in the flow shown in FIG.
26) is a function performed in this order.

Before starting the candidate path creation processing shown in FIG. 20, the table creation unit 58 first creates a segment table. In the case of the character string pattern 30 shown in FIG. 16, a table shown in FIG. 21 is obtained. It should be noted that the symbol of the segment symbol and the symbol of the extraction candidate point symbol are omitted in the table.
It is assumed that NULL is set in the blank data frame. Then, a candidate path creation process is performed using this segment table.

First, the path memory 60 is cleared (S27 in FIG. 20), and # Ci = #
C0, # Sk = # S0, and Mk = NULL are set (S28 in FIG. 20). And these #Ci, #Sk,
Candidate path creation function Funct (#Ci,
#Sk, Mk) starts (S29 in FIG. 20).
As described above, this creation function represents steps 19 to 26 shown in FIG. 14 (S19 to S26 in FIG. 14).

Then, when the processing of the creation function is completed, next, the first specifying unit 50 determines whether or not the extraction candidate point #Ci is the rightmost end (S30 in FIG. 20). As a result of this determination,
When #Ci is the rightmost end, the candidate path creation processing ends. If #Ci is not the rightmost end, then the path length counting unit 66 counts the length of the path recorded in the path memory 60 (S31 in FIG. 20). In the counting result, the longest path is selected, the position of the end of the path is set to #Ci, and the candidate character pattern of the end of the path is set to #S.
k, and the character code belonging to the candidate character pattern is Mk
(S32 in FIG. 20), the creation function Funct (#C
i, #Sk, Mk) is started (S2 in FIG. 20).
9).

As a result of the above processing, candidate paths shown in FIG. 22 are created. Note that FIG. 22 also shows paths whose growth has been interrupted halfway (ends are indicated by symbols g, h, and i).
The path interrupted on the way is a path that has not been selected as the longest path by the path length counting unit 66. In this example, finally, eight paths formed between the start point j and the end point k are obtained as candidate paths. Also, in FIG.
The candidate paths are shown in a two-terminal directed graph in correspondence with the extraction candidate points of the character string pattern.

As in the first embodiment, the best path can be determined by selecting the path that minimizes the sum of the differences. In this way, the path length counting unit 66 preferentially grows the longest path at that time. Therefore, a path that is interrupted in the middle, that is, a path in which the connection rule is not satisfied in the middle is excluded in the creation process, so that the number of candidate paths can be reduced. Therefore, even a character string in which a separation error is likely to occur, such as an apartment name, can lead to a correct result without erroneous recognition.

In this embodiment, the horizontal bar character is taken as an isomorphic character. However, the present invention is not limited to this.
It may be a round character such as the numeral “0” or the alphabetic character “O”. Further, the isomorphic character may be a vertical bar character such as a numeral “1” or an alphabetical character “l”.

[Third Embodiment] The character recognition device of the third embodiment differs from the first embodiment in the content of the connection rule set in the connection rule inspection unit 26. As shown in the block diagram of FIG. 24, the best result selecting unit 18 includes a word matching unit 68. Other configurations are the same as those of the first embodiment. FIG. 25
The flowchart of [1] shows the operation flow of the character recognition device of this embodiment.

First, the connection rule according to the third embodiment will be described. In this embodiment, an address as shown in FIG. 26, in particular, a town name, a village name and the like are to be recognized. FIG.
3 shows a handwritten character “Higashiasakawa-cho” as a character string pattern 30 of the input image. In such an address character string pattern, characters such as "town" and "village" always appear. But,
“Town” is easily separated into “field” and “cho”, and “village” is easily separated into “tree” and “dimension”, and separation errors are likely to occur.

For this reason, the connection rule checking unit 26 of this embodiment deletes the second character code from the path memory 60 when the first character code and the second character code are in a specific combination, and The character code corresponding to the combination is recorded in the path memory 60 instead of the first character code. The above-mentioned specific combination is a combination in which the second character code is “ta” and the first character code is “cho”, and the character code corresponding to this combination is “machi”. FIG. 27 shows a character type connection rule according to this embodiment. Since "town" and "village" are key characters in a character string such as an address, they are also called key characters,
This concatenation rule relating to key characters is also called a key character concatenation rule. Then, when performing character recognition, first, the key character rules are input to the connection rule checking unit 26 (S33 in FIG. 25). Hereinafter, a process in the case of using this key character rule will be described.

In the same manner as in the first embodiment, image input is performed (S34 in FIG. 25), extraction of candidate character patterns, that is, extraction of segments (S35 in FIG. 25), and integration of segments (S36 in FIG. 25). I do. In this embodiment, the constant N used for the segment integration process is 1.2
Is set.

Subsequently, the candidate character pattern is recognized (S37 in FIG. 25). FIG. 28 shows a part of the recognition result of the character string pattern 30. In this example, only the character code having the smallest difference is obtained and recorded in the character code memory 22.

Next, a candidate path is created (FIG. 25).
S38). First, the table creation unit 58 creates a segment table (FIG. 29). Then, a candidate path is created using this segment table.
The candidate path creation function Funct (#Ci, #C
(Sk, Mk) is shown in FIG. Hereinafter, this function will be described.

First, the path memory 60 is cleared, and # Ci = # C0, # Sk = # S0, M
k = NULL is set in advance. Note that a part of the processing of the creation function according to the present embodiment performs processing overlapping with the creation function according to the first embodiment (FIG. 14). Specifically, FIG.
(The process between S41 and S49 in FIG. 30) overlaps. However, since processing different from the processing of the first embodiment is inserted in the processing of this loop 1, mainly the point will be described.

Processing of loop 1 (S41 and S41 in FIG. 30)
49) are continued until all of the extraction candidate points #Cj (j is an integer satisfying j> i) located on the right side of the extraction candidate point #Ci are specified by the first specifying unit 50 ( For each # Cj∈ (Adj (# Ci # Cj)! = NULL)). The processing of loop 2 (processing between S42 and S48 in FIG. 30)
Indicates that the character code (Mk + 1∈ (candidate character of # Sk + 1)) obtained for candidate character # Sk + 1 is sequentially entered one by one in each loop, and all are in the first character reading unit 52. Until it is read. Here, candidate character # Sk +
1 is a first candidate character pattern specified by the cutout candidate points #Ci and #Cj. The selection of the first candidate character pattern is performed by the first specifying unit 50 with reference to the segment table (S42 in FIG. 30: # Sk + 1 ← Adj (# Ci # Cj)). The character code Mk + 1 is the read first character code.

In response to the designation of the first candidate character pattern by the first designation unit 50 and the reading of the first character code by the first character reading unit 52, the second designation unit 54 is adjacent to the left of the first candidate character # Sk + 1. Is designated as a second candidate character pattern. This designation can be made by referring to the coordinate memory 20. Then, the second specifying unit 54 instructs the second character reading unit 56 to read the character code Mk belonging to the second candidate character #Sk from the character code memory 22 as the second character code.

The above-described reading of the first and second character codes is performed such that one pair is formed for each loop, different pairs are created for each loop, and all combinations are finally realized. Done in Then, each pair is input to the connection rule checking unit 26. Connection rule inspection unit 2
6 determines whether or not the inputted first character code Mk + 1 and second character code Mk satisfy the connection rule (S44 in FIG. 30). In this way, the connection relation of all combinations is checked.

The process proceeds to the following two steps or according to the inspection result of the connection inspection unit 26. First, if the connection rule is not satisfied, the first character code Mk + 1
Is added to the candidate path (S50 in FIG. 14), and then processing of a candidate path creation function in which #Cj, # Sk + 1, and Mk + 1 are substituted as arguments is started (S47 in FIG. 30).
# Cj, # Sk + 1, Mk + 1 → Funct (# Cj, # Sk + 1, Mk + 1)).

Next, when the connection rule is satisfied,
The second character code Mk is deleted from the candidate path (S in FIG. 30).
45), the first character code Mk + 1 is added to the candidate path (S46 in FIG. 30). That is, the second character code Mk is deleted from the path memory 60, and the first character code Mk + 1 is deleted.
Is recorded in the path memory 60. And as arguments
The processing of the candidate path creation function to which #Cj, # Sk + 1, and Mk + 1 are substituted starts (S47 in FIG. 30: #Cj, # Sk +
1, Mk + 1 → Funct (# Cj, # Sk + 1, Mk + 1)). As described above, by performing the process of recursively calling the candidate path creation function, the path memory 60 stores the connection rule (key character rule).
Are recorded.

FIG. 31 shows the result of creating a candidate path. In FIG. 31, the character string pattern 30 is shown in the upper part of the figure, and the lower part of the figure is made to correspond to the extraction candidate points of the character string pattern 30.
The candidate paths are shown in the form of a terminal directed graph. However, paths that do not satisfy the key character rules are also shown.

Next, the best path is selected from these candidate paths. The best result selection unit 18 includes a word matching unit 68 and an evaluation value calculation unit 62 for this purpose. In this example, the word matching unit 68 performs matching between the candidate path and the word dictionary (S39 in FIG. 25), and the evaluation value calculation unit 62 sets the evaluation value when the same word as the candidate path exists in the dictionary. Is given to the candidate path. If the same word as the candidate path does not exist in the dictionary, “0” is given to the candidate path as an evaluation value. And
The cutout position output unit 64 sets the candidate path to which “1” has been assigned as the best path, and outputs the cutout position and the character code to the outside (S40 in FIG. 25). FIG. 31 shows a candidate path creation result together with a character string pattern 30. In this example, since the candidate path 1 shown in FIG. 31 is the same as a word in the address dictionary, this is selected as the best path.

As described above, in this embodiment, by setting a key character and setting a connection rule relating to the key character, it is possible to reliably recognize a key character that is easily separated. Therefore, the character string pattern can be correctly recognized. Further, since these key characters are important when recognizing an address, it is significant to correctly recognize them. In addition, if the key characters can be correctly recognized, there is an effect that the search range of the dictionary can be limited.

It is to be noted that not only "town" is set as a key character but "village" may be set. In this case, in the specific combination described above, the second character code is “tree” and the first character code is “tree”.
This is a combination when the character code is “dimension”, and the character code corresponding to this combination may be “mura”.

[0123]

According to the character recognition method and the character recognition device of the present invention, the character codes corresponding to the candidate character patterns having the adjacent relation in the character string pattern are read out from the character code memory, respectively. By examining the connection relation, it can be determined whether or not the character is a character constituting a candidate path. Therefore, if the connection rules are set appropriately, the character string pattern can be recognized more accurately than in the past.

Further, according to the preferred embodiment of the character recognition method and apparatus of the present invention, a character string composed of character codes of the same character type is output as a recognition result. The concatenation rule set at this time is effective when recognizing a character string pattern that is often represented by the same character type, such as first and last names.

Further, according to the preferred embodiment of the character recognition method and apparatus of the present invention, when two character codes having a continuous positional relationship in a character string pattern are both specific character codes, Instead of the character code, a previously set character code is recorded in the path memory as a character constituting the candidate path. Therefore, it is a character that is frequently included in the character string pattern to be recognized, and the character is 2
In the case where it is easy to separate into two parts and the separated part itself is a character, as described above, this character is set in advance as a character code corresponding to a specific combination, so that the separation error Can be avoided, and erroneous recognition is less likely to occur.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration of a character recognition device according to a first embodiment.

FIG. 2 is a diagram illustrating a character recognition flow according to the first embodiment.

FIG. 3 is a diagram illustrating a character type connection rule according to the first embodiment;

FIG. 4 is a diagram illustrating an input image according to the first embodiment.

FIG. 5 is a diagram illustrating a segment coordinate table before integration according to the first embodiment;

FIG. 6 is a diagram illustrating a configuration of a segment integration unit according to the embodiment;

FIG. 7 is a diagram illustrating a segment integration process according to the first embodiment;

FIG. 8 is a diagram for explaining the definition of the distance between segments.

FIG. 9 is a diagram illustrating a segment coordinate table after integration according to the first embodiment;

FIG. 10 is a diagram illustrating a character recognition result according to the first embodiment.

FIG. 11 is a diagram illustrating a recognition result of each candidate character according to the first embodiment.

FIG. 12 is a diagram illustrating a configuration of a candidate path creation unit according to the first embodiment.

FIG. 13 illustrates a segment table according to the first embodiment.

FIG. 14 is a diagram illustrating processing of a candidate path creation function according to the first embodiment;

FIG. 15 is a diagram illustrating a candidate path creation result according to the first embodiment.

FIG. 16 is a diagram illustrating an input image according to the second embodiment.

FIG. 17 is a diagram illustrating a character type connection rule according to the second embodiment;

FIG. 18 is a diagram illustrating a character recognition result according to the second embodiment.

FIG. 19 is a diagram illustrating a configuration of a candidate path creation unit according to the second embodiment.

FIG. 20 is a diagram illustrating a candidate path creation process according to the second embodiment.

FIG. 21 illustrates a segment table according to the second embodiment.

FIG. 22 is a diagram provided for describing rules application and candidate path creation according to the second embodiment.

FIG. 23 is a diagram illustrating a candidate path creation result according to the second embodiment.

FIG. 24 is a diagram illustrating a configuration of a character recognition device according to a third embodiment.

FIG. 25 is a diagram illustrating a character recognition flow according to the third embodiment.

FIG. 26 is a diagram illustrating an input image according to the third embodiment.

FIG. 27 is a diagram illustrating a character type connection rule according to the third embodiment;

FIG. 28 is a diagram illustrating a character recognition result according to the third embodiment.

FIG. 29 is a diagram illustrating a segment table according to the third embodiment.

FIG. 30 is a diagram illustrating processing of a candidate path creation function according to the third embodiment.

FIG. 31 is a diagram illustrating a candidate path creation result according to the third embodiment;

[Explanation of symbols]

 10: Image input unit 12: Candidate character pattern extraction unit 14: Character recognition unit 16: Candidate path creation unit 18: Best result selection unit 20: Coordinate memory 22: Character code memory 24: Control unit 26: Connection rule inspection unit 28 : Character type concatenation rule 30: Character string pattern 32: Image memory 34: Segment extraction unit 36: Segment integration unit 38: First coordinate reading unit 40: Second coordinate reading unit 42: Distance value detection unit 44: Comparison unit 46: Generation Unit 48: writing unit 50: first designation unit 52: first character reading unit 54: second designation unit 56: second character reading unit 58: table creation unit 60: path memory 62: evaluation value calculation unit 64: cutout Position output unit 66: path length counting unit 68: word matching unit

Claims (22)

[Claims] 1. A method comprising: (a) extracting a plurality of consecutive candidate character patterns from a character string pattern to be recognized; and recording position information thereof in a coordinate memory; and (b) extracting the extracted candidate character patterns. Recognizing an appropriate number of character codes and recording the character codes in a character code memory; and (c) arranging the recorded character codes based on the recorded position information to create a candidate path. And (d) selecting a best path from the created candidate paths based on the result of the recognition. In the character recognition method, the step (c) includes: (c1) the extracted candidate path; Specifying a character pattern as a first candidate character pattern; and (c2) using the character code determined for the first candidate character pattern as a first character code, Reading from a character code memory; (c3) specifying a candidate character pattern adjacent to the first candidate character pattern as a second candidate character pattern based on the recorded position information; Reading the character code obtained for the two candidate character patterns from the character code memory as a second character code; and (c5) a connection in which the connection relationship between the first character code and the second character code is set in advance. Recording the first character code in a path memory when the rule is satisfied, and repeating the steps (c1) to (c5) to change the character code array recorded in the path memory to the A character recognition method characterized by using a candidate path. 2. The character recognition method according to claim 1, wherein the connection rule is satisfied when the first character code and the second character code have the same connection type. Character recognition method. 3. The character recognition method according to claim 2, wherein the first character code is a character code belonging to a class of homomorphic characters. 4. The character recognition method according to claim 3, wherein the isomorphic characters are horizontal bar characters such as a hyphen "-", a long sound "-", and a Chinese numeral "1". Method. 5. The character recognition method according to claim 3, wherein the isomorphic characters are round characters such as a numeral “0” and an alphabetic character “O”. 6. The character recognition method according to claim 3, wherein the isomorphic characters are vertical bar characters such as a numeral “1” and an alphabetical character “l”. 7. The character recognition method according to claim 1, wherein when the first character code and the second character code have a specific combination, the second character code is deleted from the path memory. A character recognition method, wherein a character code corresponding to a specific combination is recorded in the path memory instead of the first character code. 8. The character recognition method according to claim 7, wherein the specific combination is a combination when the second character code is “ta” and the first character code is “cho”. A character recognition method, wherein the character code corresponding to the combination is "town". 9. The character recognition method according to claim 7, wherein the specific combination is a combination when the second character code is “tree” and the first character code is “dimension”. A character recognition method, wherein a character code corresponding to the combination is "village". 10. The character recognition method according to claim 1, wherein the extraction of the candidate character pattern includes: (a1) extracting a black block area of the character string pattern as a segment, and storing the position coordinates of the segment in the coordinate memory. And (a2) integrating the extracted segments based on their respective position coordinates to generate a new segment, and adding the position coordinates of the new segment to the coordinate memory. A character recognition method, wherein the position coordinates of the segment and the new segment recorded in the coordinate memory by executing the steps (a1) and (a2) are used as the position coordinates of the candidate character pattern. 11. The character recognition method according to claim 10, wherein the generation of the new segment includes: reading a position coordinate of the recorded segment #Sn (n is an integer) from the coordinate memory; Reading the position coordinates of a segment #Sk (k is an integer) different from Sn from the coordinate memory; and a distance value D between the segment #Sn and the segment #Sk.
_{determining nk} from each of the read position coordinates; and a constant e (e is a positive real number) of a line height L of the character string pattern.
Comparing the distance with the distance value D _nk obtained above, and when the result of the comparison is D _nk ≦ e · L, the segment #S
n, a segment #Sk, and a segment between these segments to generate the new segment. 12. An original image including a character string pattern to be recognized is read, an image input unit having an image memory for storing the original image, and a plurality of continuous candidate character patterns are extracted from the character string pattern. A candidate character pattern extraction unit that records the position information in a coordinate memory; and a character recognition unit that recognizes the extracted candidate character pattern to determine an appropriate number of character codes and records the character codes in a character code memory. Unit, arrange the recorded character code based on the recorded position information, a candidate path creation unit that creates a candidate path, from among the created candidate paths, based on the result of the recognition, In the character recognition device including a best result selection unit that selects a best path, the candidate path creation unit specifies the extracted candidate character pattern as a first candidate character pattern. A first character reading unit that reads out the character code obtained for the first candidate character pattern from the character code memory as a first character code, based on the recorded position information, A second designation unit that designates a candidate character pattern adjacent to the first candidate character pattern as a second candidate character pattern; and the character code determined for the second candidate character pattern as a second character code. A second character reading unit that reads from the code memory; and a connection rule that records the first character code in a path memory when a connection relationship between the first character code and the second character code satisfies a predetermined connection rule. A character recognition device comprising an inspection unit. 13. The character recognition device according to claim 12, wherein the connection rule is satisfied when the first character code and the second character code have the same connection type. Character recognition device. 14. The character recognition device according to claim 13, wherein the first character code is a character code belonging to a class of homomorphic characters. 15. The character recognition device according to claim 14, wherein the isomorphic characters are horizontal bar characters such as a hyphen "-", a long sound "-", and a Chinese numeral "1". apparatus. 16. The character recognition device according to claim 14, wherein the isomorphic characters are round characters such as a numeral “0” and an alphabetic character “O”. 17. The character recognition device according to claim 14, wherein the isomorphic characters are vertical bar characters such as a numeral “1” and an alphabetical character “l”. 18. The character recognition device according to claim 12, wherein the connection rule checking unit is configured to determine the first character code and the second character code.
When the character code is a specific combination, the second character code is deleted from the path memory, and the character code corresponding to the specific combination is recorded in the path memory instead of the first character code. A character recognition device, comprising: 19. The character recognition device according to claim 18, wherein the specific combination is a combination when the second character code is “ta” and the first character code is “cho”; A character recognition device, wherein the character code corresponding to the combination is "town". 20. The character recognition device according to claim 18, wherein the specific combination is a combination when the second character code is “tree” and the first character code is “dimension”, A character recognition device, wherein a character code corresponding to the combination is "village". 21. The character recognition device according to claim 12, wherein the candidate character pattern extraction unit extracts a black block area of the character string pattern as a segment, and records the position coordinates of the segment in the coordinate memory. A segment extraction unit; and a segment integration unit that integrates the extracted segments based on their respective position coordinates to generate a new segment, and adds the position coordinates of the new segment to the coordinate memory. A character recognition device for extracting position coordinates recorded in a memory as position coordinates of the candidate character pattern. 22. The character recognition device according to claim 21, wherein the segment integrating unit reads a position coordinate of the recorded segment #Sn (n is an integer) from the coordinate memory; A second step of reading the position coordinates of a segment #Sk (k is an integer) different from the segment #Sn from the coordinate memory;
Distance value D between coordinate reading unit and segment #Sn and segment #Sk
a distance value detection unit for determining _nk from the read position coordinates; and a constant e (e is a positive real number) of a line height L of the character string pattern.
A comparing unit that compares the distance with the calculated distance value D _nk , and when the result of the comparison is D _nk ≦ eL, the segment #S
n, a segment #Sk, and a segment between these segments to generate the new segment by integrating the segment, and a writing unit that additionally records the generated position information of the new segment in the coordinate memory. And a character recognition device comprising: