JP2574776B2 - Character recognition method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method, and more particularly to a character recognition method suitable for the purpose of recognizing a printed character pattern whose character font is unknown with high accuracy.

[Conventional technology]

In the conventional print character recognition method, when a character font is known (this is called single-font character recognition), recognition is performed with high precision using a pattern dictionary including a standard pattern of the character font. However, if there are many fonts and it is not known which one is used (this is called multi-font recognition), a pattern dictionary consisting of standard patterns for all target fonts is used. Since it must be used, the dictionary memory increases, the recognition speed decreases, and the recognition accuracy deteriorates.

To deal with this problem, for example, Japanese Patent Publication No. 59-3793 prepares a plurality of pattern dictionaries, recognizes one line of characters using one of them, and counts the number of rejects in the recognition result. A method is described in which when the number of rejects is large, the dictionary is replaced with another dictionary and recognition is performed again. In Japanese Patent Application Laid-Open No. 53-51928, a plurality of discriminators corresponding to a plurality of fonts are prepared, all of them are used to recognize one field of characters, and the number of rejects in the recognition result by each discriminator is determined. And a method of selecting a recognition result that minimizes the number of rejects is described. In this method, a plurality of discriminators are used. However, even if one discriminator is used to sequentially switch and recognize a plurality of pattern dictionaries, the speed is reduced but the same effect is obtained in terms of the recognition rate. It is clear that

[Problems to be solved by the present invention]

In a conventional method such as Japanese Patent Publication No. 59-3793 or Japanese Patent Application Laid-Open No. 53-51928, standard patterns for all character categories are prepared corresponding to a plurality of fonts, so that only (number of character categories) × (number of fonts) is required. A standard pattern is required, and the dictionary memory is increased.

An object of the present invention is to print multi-font characters.
An object of the present invention is to provide a character recognition method that requires a relatively small dictionary memory and can obtain a recognition rate as high as that of a single font print character.

[Means for solving the problem]

In the present invention, the above problem is solved by storing a plurality of standard patterns for each character category instead of preparing a standard pattern for each character font as a dictionary, and even if fonts are different in each character category. If the patterns are the same, the problem is solved by reducing the number of standard patterns to be stored by using the same standard pattern. At that time, a separate dictionary was prepared for each font by separately preparing a memory that previously stored information specifying which standard pattern to use for each character category for each font. Has the same effect as

[Action]

As a dictionary, a standard pattern for each character font is not prepared, and a dictionary in which a plurality of standard patterns are stored for each character category is used, so even if different character fonts are similar in the same category, Memory can be saved because standard patterns for them can be shared.

At the time of character recognition, the character pattern of the unknown character font is given, and the recognition device compares each of the plurality of standard patterns belonging to the character category with the input character pattern of the unknown character font, and determines the most similar standard pattern. decide. For each character font, which standard pattern corresponds to each character category is specified in advance. After recognizing a plurality of character patterns (for example, for one line), a character font (may be plural) corresponding to the above-described most similar standard pattern is determined, and an input document is determined from the tally result. Determine the font of the character. Using the character font information determined in this way, a standard pattern corresponding to the character font in the dictionary can be selected, and if the input pattern is recognized again,
Effectively, it is equivalent to single-font character recognition, and high-precision character recognition becomes possible. After the character font is determined for a part of the document, the character recognition can be performed for the remaining part of the document assuming the character font, thereby enabling high-speed recognition.

In the present invention, since the information to be specified for each character font need only be the standard pattern number for each character category,
Compared with the conventional method of storing a pattern for each character font, the total memory amount is greatly reduced.

In addition, it is possible to cope with the character shape variation in the same font simply by adding information specifying which of the plurality of standard patterns in the dictionary corresponds to each variation pattern.

〔Example〕

 Hereinafter, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a PAD (Problem Analysis Diagram) diagram showing the flow of processing in an embodiment of the character recognition system of the present invention. FIG. 2 is a block diagram showing the configuration of an apparatus for performing the processing of FIG. Each part of the device is connected to a bus 1, and the overall operation is controlled by a control unit 2.

The processing contents of FIG. 1 will be described. FIG. 1 (A), (B)
Are connected at points in the figure. FIG. 1A shows 101 to 104, 138 and character cutout portions related to the overall control.
FIG. 1 (B) is composed of portions 127 to 136 for recognizing characters using the determined font. Reference numeral 101 denotes initialization of various variables used during processing. This processing includes reading the pattern dictionary from the file 3 into the memory 61 of FIG. 2 and reading the standard pattern-character font number correspondence table into the memory 62. The memories 61 and 62 form a part of the memory 6 together with 63 to 65 described later. The method of creating the pattern dictionary and the standard pattern-character font number correspondence table will be described later. 101 includes processing for resetting a switch SW to be described later to 0.

Reference numeral 102 in FIG. 1 denotes a document image input. The information (document image) on the document 4 shown in FIG. 2 is scanned by the photoelectric conversion device 5, digitized, and stored in the memory 63 via the bus 1. Corresponding to

Before describing the processing, the configuration of the dictionary and how to look it up will be described. FIG. 3 shows an example of the data arrangement of the dictionary 61 in the memory. As can be seen from FIG. 3, there is a category pointer table 301 in the dictionary, and there are entries for the total number J of categories, and the contents are the category names D (j).
And standard pattern tables 302-1 to 3 for each category
It consists of a pointer B (j) pointing to the address 02-J. The standard pattern table 302-j for each category has entries for the number M (j) of standard patterns registered for each category, and each entry indicates the serial number of the standard pattern. That is, the i-th pattern P (j, i) of category j
Indicates that it has a serial number w. Reference numeral 303 denotes a pattern address table, the contents of which indicate an address A (w) occupied in the dictionary main unit 304 by the w-th standard pattern main unit. The dictionary main unit 304 simply arranges pattern data of a total of W standard patterns or feature values obtained by feature extraction. If the length of each pattern is constant, the above address A (w) does not need to be specified, but if the length is indefinite, it is necessary to specify at least the start address. The end address can be calculated, for example, as A (w + 1) -1. As will be understood from the above description, if the category j is specified, the main body of each of the M (j) standard patterns belonging to j can be reached in the order of 301 → 302 → 303 → 304.

Next, the configuration of the standard pattern-character font number correspondence table and how to look it up will be described. FIG. 4 shows an example of the data arrangement in the memory of the standard pattern-character font number correspondence table. As can be seen from FIG. 4, there is a character font pointer table 401 having entries for the total number K of character fonts, and the contents indicate the addresses of the standard pattern correspondence table groups 402-1 to 402-K for each character font. The internal structure of the standard pattern correspondence table 402-k is
This is shown in the enlarged view 403. As can be seen from the enlarged view, the number q of patterns registered for the character font k is the first record, and the following q combinations of the category Ca and the pattern number Pa continue. This means that this font
Indicates that the pattern number Pa is used as the font of the character of the category Ca. A plurality of pattern numbers may be registered for the same category, in which case they represent deformations within the same font. By referring to the standard pattern-character font number correspondence table 62 described above, it is apparent that information on which standard pattern number each character font uses for each character category can be obtained.

Returning to the description of FIG. 1, reference numeral 103 in FIG. 1 is a process for extracting a character line from a document image. Here, the unit of repetition is the character line on the document image, but instead of the character line,
A character field, that is, a band-shaped area in a character line, or a character line area, that is, an area configured by a plurality of character lines may be used. A method of extracting a character line or a character region from a document image is described in, for example, IEICE Transactions No. 66-D.
Volume 1, Issue 111, by Teruo Akiyama and Isao Masuda, "Paper Constituent Elements Extraction Method without Format Specification Information" (Showa
The method introduced in Jan. '58 may be used.

104 in FIG. 1 is a character line f (f = 1 to F) on the document image
Is a loop control that repeats the processing of 105 to 113.
Here, F is the number of character lines on the document.

Reference numeral 105 denotes a process for extracting a character pattern group from the character line (or character field or character area). Since this processing is publicly known, its details are omitted. The extracted character pattern is stored in the memory 64 of FIG.
N represents the number of extracted character patterns.

106 is a portion that branches depending on the value of the switch SW.
Is 0, which corresponds to the case where the font of the input document has not been determined, and the character font is determined by the processing of 107 to 126.

107 is for each of the cut out character patterns.
This is a loop control for repeating the processing of 114, where n indicates a character pattern number. That is, the processing of 108 to 114 represents the recognition processing of one character. Reference numeral 108 denotes initialization processing in which the similarity maximum value R0 is set to 0 prior to the following processing. 109
Is a loop control for repeating the processing of 110 to 114 for each character category in the dictionary 61, j is the category number, and J is the total number of categories.

At 110, the number P (j) of standard patterns belonging to the j-th character category is extracted. 111 is a loop control for repeating the processing of 112 to 114 for the P (j) standard patterns, and i is a standard pattern number in the character category j. As shown in FIG. 3, the corresponding standard pattern serial number w is determined from i and j. Reference numeral 112 denotes a portion for comparing the n-th cut-out character pattern with the w-th standard pattern to obtain a similarity R. If the standard pattern has been subjected to feature extraction, it is needless to say that the input pattern also needs to be feature extracted prior to similarity calculation. Although the degree of similarity is determined by the software of the control unit 2 in FIG. 2, dedicated hardware may be used. The definition of the degree of similarity may use an appropriate method among the methods used in ordinary character recognition. A determination unit 113 compares the maximum similarity R0 with the current similarity R. If the result of the determination is that R is greater, R0 is updated at 114, and the character number j and standard pattern serial number w at that time are X
(1, n) and X (2, n) are stored in the memory 65 of FIG.

As will be understood from the above description, when the loop of 107 is completed, X (1, n) and X (2, n) in the memory 65 are set.
In (n = 1 to N), a character category most similar to the character pattern n and a standard pattern serial number are entered.

Reference numerals 115 to 124 denote character font identification processing. 115
Then, the initial value of the maximum matching degree S0 is reset to 0. Reference numeral 116 denotes a loop control for repeating the processing of steps 117 to 121 for M character fonts, and m indicates a character font number. 117
Then, the initial value of the coincidence S is reset to 0. Reference numeral 118 denotes loop control for repeating the processing of steps 119 to 121 for the cut out N character patterns, and n indicates a character pattern number. At 119, the character category j = X (1, n) and the standard pattern serial number w = X (2, n) are extracted. In 120, the font is m
Is used as key information to refer to the pattern-character font number correspondence table 62 to determine whether a combination of a character font m, a character category j, and a standard pattern serial number w exists. If the result of the determination is YES, the degree of coincidence S is increased by 1 at 121. After the loop 118 is completed, the maximum matching score S0 and the matching score S for the font m at 122 are reached.
And if S0 <S, the maximum matching score S0 is updated at 123 and the character font number m is stored as m0.

After the loop of 116 is completed for M character fonts, the maximum matching degree S0 is compared with the threshold value θ · N at 124. Here, θ is a parameter satisfying 0 <θ ≦ 1. When S0 is large, 1 is assigned to the switch SW at 125 to indicate that the character font has been determined. At that time, m0 is the character font of the inputted N character patterns.

If S0 is smaller than the threshold value at 124, the switch SW remains 0, indicating that the character font could not be determined. At this time, at 126, the character category X (1, n) (n = 1 to
N) is output as the recognition result of the N character patterns.

127 branches according to the value of the switch SW, and in the case of 1, the font of the input document is determined.
As recognition.

128 is a loop control for repeating the processing of 129 to 136 for N character patterns, and n indicates a character pattern number. 129-136 are the same as 108-114 above, but 133
Is different. 133 is the same as the determination 120.Referring to the pattern-character font number correspondence table 62 using the character font m0 as key information, it is determined whether or not a combination of the character font m0, the character category j, and the standard pattern serial number w exists. Is determined. Judgment result is YE
Only in the case of S, the processes of 134 to 136, that is, the similarity calculation and the update of the maximum similarity are performed. At the end of the loop of 128, X (1,1) for character patterns n = 1 to N
The character category of the recognition result and its standard pattern number are obtained for n) and X (2, n). In this case, only the former character category is significant. Character category X at 137
(1, n) (n = 1 to N) is output to the output device 7 as a recognition result. Reference numeral 138 denotes an end process, for example, a process of closing a data set created in the output device 7 for an input document.

Note that 101 to 138 are created as one subroutine,
It is preferable that the document is started when the document is set in the photoelectric conversion device 4 or in response to an instruction from an operator.

In addition, the operator uses the console 8 to interact with the control unit 2 in order to operate the entire process. However, since the content is a known process, the detailed description is omitted.

FIG. 5 is a PAD diagram of an example of processing for creating a dictionary as shown in the above embodiment. This processing is performed offline in advance. First, at 501, a learning character pattern with a total of I character fonts known is a category code C (i) and a character font code m (i).
(I = 1 to I). At 502, dictionaries that have been learned in the past are read from a learning file (not shown). Let W be the total number of standard patterns in the dictionary.
Here, the internal representation of the dictionary is different from that of FIG. 3 and is as shown in FIG. That is, the dictionary includes the standard pattern control table 601 and the main body 602. The standard pattern control table 601 has W entries 601-1 to 601-W, each of which is composed of a character category code C (w) and an address pointer A (w). The address pointer indicates the head address of each standard pattern main body in the main body 602. The character font code tables 603 and 604 are the same as 401 and 402 in FIG. M is the total number of registered fonts as described above.

FIG. 503 is a loop control for performing the processing of 504 to 514 for I input characters. i is the serial number of the character pattern. In step 504, the maximum similarity value R0 is reset to zero. 505 is 506 to W for the registered standard patterns.
This is a loop control for performing the process of 509. w is the serial number of the standard pattern. At 506, the category C (i) of the input pattern i is compared with the category D (w) of the standard pattern w, and the processes of 507 to 509 are performed only when they match. 507 is
This is a part for calculating the similarity R by comparing the input character pattern i with the standard pattern w. The calculation formula and parameters of the similarity need to be the same as those used in character recognition. 50
8 compares the maximum similarity value R0 up to that point with R, and R becomes R
If it exceeds 0, R0 is updated at 509 and w is stored as wx. After the loop of 505 for the W standard patterns is completed, the maximum value R0 is compared with the threshold α at 510, and if R0 is larger than α, the total number of registered patterns W is increased by 1 at 511, and the dictionary main body 602 at 512. Of the standard pattern control table 601 and the (W + 1) th entry 601- (W + 1). 513: (W + 1) wx
To be stored.

FIG. 514 is a portion for updating the character font code tables 603 and 604. The character font code m (i) of the input character is searched in the table, and if not registered, the entry of the character font is added. And the like, and the standard pattern correspondence table is updated using the information of the category C (i) of the input pattern i and the standard pattern number wx, but the details are omitted. I
After learning is completed for the input patterns, the contents of the memory are output to a learning file.

As can be understood from the above description, the present learning process is assumed to perform additional learning to a dictionary to which learning has progressed to some extent. In that case, 0 may be written to W of the learning file by another process. After the learning is completed as described above, in order to use this dictionary for character font recognition, the internal representation of the dictionary (FIG. 6) in the learning file is converted as shown in FIG. , But its realization is easy, so the description is omitted.

Although the embodiment of the present invention has been described above, it goes without saying that various changes can be made without affecting the essence of the invention. For example, in the embodiment, one standard pattern most similar to each input character pattern is selected. However, a plurality of standard patterns having high similarity are selected, and in the determination of 120, any one of the standard patterns is selected. If they match, the determination result may be set to YES. Also, instead of simply adding the combination of a character font, a character category, and a standard pattern as a binary value of 1,0 as the character font matching degree S, The similarity with respect to the standard pattern, the rank of the similarity, and the like may be added.

Further, in the embodiment, one font is determined for a character line. However, in the process of 123, not only the font having the maximum matching degree but also the font having the second or third or lower matching degree is stored. Then, the determination may be made using a plurality of upper fonts.

In the embodiment, the font m0 is fixed to 128 to 1 for the N patterns used to determine the character font.
Although re-recognition is performed at step 36, only those included in the font m0 are extracted from the recognition results using all fonts, and the speedup can be achieved by omitting the re-recognition processing. Good.

In the embodiment, when the SW is 1 in the determination of 127, the font is fixed and the subsequent recognition is performed. When the recognition result is obtained in 137, the likelihood of the recognition result is evaluated using the similarity. When the evaluation value is low, the recognition may be performed again using all the fonts.

Further, in the embodiment, the character category of the standard pattern that gives the maximum similarity is used as the recognition result.However, the likelihood of the recognition result is evaluated using the magnitude of the maximum similarity and the difference value with the next-largest similarity. When the evaluation value is low, a reject may be made.

In the embodiment, a table indicating which standard pattern is used for each character category is provided for each character font. Instead of this table, information of the character font using the standard pattern is used for each standard pattern. May be provided.

In the embodiment, when there are similar standard patterns between different categories, the standard patterns are separately provided for each category. However, these can be shared. For example, the numeral 1 (Ichi) of the character font A and the lowercase letter l (L) of the character font B can be shared to form the same standard pattern, and each address pointer can point to the same standard pattern. One way to do this is to eliminate decision 506 in FIG.

Further, in the embodiment, the matching degree for each font is evaluated by directly using the recognition result for the character pattern for one line, but language processing or arithmetic operation is performed on the recognition result for the character pattern for one line, Only for standard patterns whose linguistic or numerical validity is correct, the degree of coincidence of each font may be evaluated.

〔The invention's effect〕

As described above, according to the present invention, a character recognition method that can prevent a memory capacity of a dictionary for character font recognition from increasing and achieve a recognition rate similar to a single font for a large number of character fonts is realized.

[Brief description of the drawings]

FIG. 1 is a diagram showing the flow of processing in the character recognition system of the present invention, FIG. 2 is a block diagram showing the configuration of an apparatus for executing the processing contents of FIG. 1, and FIGS. 3, 4, and 6 are FIG. FIG. 5 is a view for explaining an example of the internal structure of a dictionary used in the processing of FIG. 5, and FIG. 5 is a view showing a flow of processing for creating a dictionary used in the processing of FIG. DESCRIPTION OF SYMBOLS 1 ... Bus, 2 ... Control part, 3 ... File, 4 ... Document, 5 ... Photoelectric conversion device, 6 ... Memory, 7 ... Output device, 8 ... Console.

 ──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-58-78276 (JP, A) JP-A-59-109978 (JP, A) JP-A-60-237584 (JP, A)

Claims (3)

(57) [Claims] 1. A pattern dictionary for storing a plurality of standard patterns for each character category in advance, and a standard for storing for each character category in the pattern dictionary which of the plurality of standard patterns a character font corresponds to. What is claimed is: 1. A character recognition method for performing character recognition having a pattern character font correspondence means, wherein: (a) a document image is scanned and digitized by a photoelectric conversion device, and a character pattern whose character font is undetermined is cut out from the document image (B) sequentially comparing the extracted character pattern whose character font is undetermined with a plurality of standard patterns of the character category in the pattern dictionary to determine a standard pattern similar to the character pattern whose character font is undetermined And (c) a character phone corresponding to the similar standard pattern determined above. Is obtained from the standard pattern character font correspondence means. (D) The above (b) and (c) are sequentially performed on each of the plurality of cut-out character patterns for which the character fonts are undetermined, and A character font having the highest matching degree among character fonts obtained for a character pattern whose character font has not been determined, is determined as the character font of the document image. 2. The standard pattern character font correspondence means according to claim 1, wherein a standard pattern of a character category corresponding to a character font determined as a character font of said document image is obtained by said standard pattern character font correspondence means. A character recognition method characterized by sequentially comparing a cut-out character pattern with a standard pattern of each character category corresponding to the determined character font, and performing character recognition on the document image. 3. The character font extracted from the document image according to claim 2, wherein the character font of the character pattern cut out from the document image is not registered in the standard pattern character font correspondence means. A character recognition method, characterized by updating the contents of the standard pattern character font correspondence means with respect to which of the plurality of standard patterns in the pattern dictionary corresponds to which standard pattern.