JP2866920B2 - Standard pattern creation method and apparatus, and character recognition apparatus and method - Google Patents

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 creating a standard pattern and an apparatus and a method for character recognition.

[0002]

2. Description of the Related Art A character recognition device is suitable as an information processing device for digitizing a large number of printed documents, creating a database, document processing, automatic translation, and the like, and further research and development are ongoing.

[0003] Conventional character recognition systems for recognizing character images printed or handwritten on paper generally include:
(1) document image input, (2) character extraction, (3) preprocessing (smoothing, normalization, thinning, etc.), (4) feature extraction,
(5) Large classification, (6) Fine classification, (7) Post-processing, etc.

In such a character recognition system, a character image on a sheet is first read as an optical image and converted into an electric signal. The text image loaded into the system is based on the histogram of the marginal distribution, etc.
After clipping is performed for each predetermined recognition unit, for example, for each character, preprocessing is performed for efficient recognition. In the feature extraction, a feature of an input character for performing a recognition process using a structural analysis method, a pattern matching method, or the like, for example, a topological feature or a feature of each pixel divided into meshes is obtained. The large classification is used particularly when targeting kanji having a large number of character categories, and narrows down candidate categories by a simple method. In the fine classification, detailed recognition processing is performed on the narrowed candidates. Further, in the post-processing, when candidates cannot be completely determined by recognizing individual input characters, adjacent input characters are connected, and a determination is made as a character string by referring to a grammar or the like.

[0005] This type of conventional character recognition system has the following problems.

In the conventional feature extraction, a character image normalized to a certain size is scanned in units of pixels (bits), and the character is focused on by focusing on the relationship between the scanned pixel and its neighboring pixels. Features (quantified features)
Is extracted. Such features represent the overall or comprehensive features of each character, and do not necessarily highlight the salient features of each character. Therefore, there is a problem that these feature amounts are easily affected by noise.

[0007] In the conventional character recognition used for large classification and fine classification, pattern matching for measuring the degree of overall similarity between an unknown input character represented by the above-described feature and a standard pattern of a dictionary is known. (Distance calculation) is performed, and an appropriate number of candidate character types are selected in ascending order of the distance. In such a conventional method, since the distance scale is the only criterion for classification, there is a drawback that remarkable structural features of each character cannot be flexibly utilized in the classification process. Therefore, in the conventional method, it is necessary to compare the unknown input pattern with the standard patterns of all character types. For example, if the number of character types is 5000, basically 5000
One time distance calculation is required. This is the biggest obstacle to speeding up character recognition. Furthermore, it becomes difficult to check the correctness of classification and recognition results in the course of character recognition, and all checks and corrections of erroneous recognition are left to post-processing such as matching with an enormous word dictionary. This is also a major obstacle to speeding up.

In order to solve the above problem, the present inventor has disclosed a character recognition system based on a new principle in Japanese Patent Application No. 6-312049. This character recognition method generally focuses on the fact that information on the structure of a character is directly reflected in a peripheral distribution (histogram) in each direction of a character image, and obtains the information from a histogram of a cut-out input character image. The 0-1 pattern expression obtained is partially collated with each 0-1 pattern expression obtained from the histogram of each standard character in the dictionary to narrow down the candidate character types.

In this character recognition method, a set of samples of a character set to be recognized is used as a standard pattern, and a method of absorbing variations in fonts due to differences in fonts and character sizes using a unique large classification method is used. ing. If it cannot be absorbed by this method, the 0-1 pattern expression of the misrecognized character is added to the standard pattern. For this reason, if the font type of the character to be recognized is increased, the variation of the font will also increase, and if the large variation is to be absorbed by the large classification method, the candidate character type narrowed down to the unknown input character by the large classification The number inevitably increases. Even when the range of the size of the unknown input character is greatly expanded, the number of candidate character types increases in substantially the same manner. If the number of candidate character types increases, the burden on the sub-classification increases and the recognition speed decreases. Therefore, an attempt was made to avoid the increase in the number of candidate character types by adding a 0-1 pattern expression of misrecognized characters to a standard pattern. However, this time, the time required for the large classification increases with the increase in the number of standard patterns due to the addition of the standard patterns, and the recognition speed decreases. However, no method for speeding up the increase in the number of standard patterns (increase in the size of the dictionary) has not been provided. On the other hand, if a statistical standard pattern in which the variation of the font is taken into consideration is used, distance calculation with all standard patterns is required as in the case of the conventional pattern matching method, and speeding up is not possible.

[0010]

Therefore, in such a character recognition system, a new type which is completely different from the conventional one is required in order to absorb various variations of the font and further improve the recognition speed and the recognition accuracy. It is required to produce standard patterns. Even when the number of standard patterns increases (the dictionary becomes large-scale), it is required to improve the recognition speed while maintaining the recognition accuracy.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and provides a character recognition apparatus and method capable of simultaneously improving all aspects of noise resistance, recognition speed, and recognition accuracy as compared with the related art. It is another object of the present invention to provide a method and an apparatus for creating a standard pattern for creating a standard pattern capable of effectively absorbing variations in characters due to differences in fonts as a standard pattern to be registered in the recognition dictionary. I do. It is another object of the present invention to provide a character recognition device and method capable of improving the recognition speed while maintaining the recognition accuracy even when the number of standard patterns increases.

[0012]

Means for Solving the Problems The present invention (claim 1) provides:
A histogram is created from the dot matrix of the input character image, a 0-1 pattern expression extracted from the created histogram of the character image, and one or a plurality of standardized character types prepared for each character type. A 0-1 pattern expression serving as the standard pattern is created for a character recognition device that searches a candidate for a character type corresponding to the input character image by partially collating the 0-1 pattern expression serving as a pattern. A standard pattern creation method, wherein a histogram is created from a dot matrix of two or more character images having the same character type but different fonts, and a 0-1 pattern expression is created from each created histogram. Each done 0
The present invention is characterized in that a maximizing process and a minimizing process are performed based on a -1 pattern expression, and the result is used as the standard pattern.

According to the present invention (claim 8), a histogram is created from a dot matrix of an input character image, a 0-1 pattern expression extracted from the created histogram of the character image, and each character type. A character recognition device that partially matches each of one or a plurality of standard patterns of each character type prepared for each character type with a 0-1 pattern expression and searches for a character type candidate corresponding to the input character image. Means for creating a histogram from a dot matrix of two or more character images of the same character type but different fonts, respectively. Means for creating a 0-1 pattern expression from each of the generated histograms, and means for performing a maximization process and a minimization process based on each of the created 0-1 pattern expressions Characterized by comprising a.

Here, for characters that are separated vertically and horizontally (for example, “gate”, “ran”, etc.), it is also possible to prepare each part separated as the character type and recognize each separated part. .

Further, the character recognition device according to the present invention (claim 9) is a checking means for cutting out an input document image for each predetermined recognition unit, and a histogram creating means for forming a histogram from the cut out recognition unit. And the 0-1 pattern expression extracted from the created histogram of the recognition unit and the previously prepared 0-1 pattern expression of each recognition target category are partially collated to correspond to the recognition unit. Large classification means for searching for a candidate for a recognition target category to be executed.
The one-pattern expression is a predetermined maximizing process and a predetermined minimizing process based on 0-1 pattern expressions respectively created from a histogram obtained from a dot matrix of two or more character images having the same character type but different fonts. It is characterized by being obtained by performing processing. Preferably, the recognition unit is a character unit, and the recognition target category is a character type.

Further, in the character recognition method according to the present invention (claim 10), for each of a plurality of character types to be recognized, a dot matrix of two or more character images having the same character type but different fonts is respectively used. A histogram is created, a 0-1 pattern expression is created from each of the created histograms, a maximizing process and a minimizing process are performed based on each of the created 0-1 pattern expressions, and the result is used as a recognition target. Is registered in the recognition dictionary as a 0-1 pattern expression that is a standard pattern corresponding to the character type to be extracted, and an input document image is cut out for each predetermined recognition unit, and a histogram is created from the cut out recognition unit and created. The 0-1 pattern expression extracted from the histogram of the recognized recognition unit and the standard pattern corresponding to each character type to be recognized registered in the recognition dictionary. By matching the over emissions become 0-1 pattern representations respectively partially, characterized by searching the character types of candidates corresponding to the recognition unit. Preferably, the recognition unit is a character unit.

Further, according to the character recognition device of the present invention (claim 11), at least two character images of the same character type but different fonts are provided as at least one standard pattern for each recognition target category. The 0-1 pattern expression obtained by performing predetermined maximization processing and minimization processing based on the 0-1 pattern expression respectively created from the histogram obtained from the dot matrix of FIG. A recognition dictionary registered in association with information indicating a class to which the standard pattern belongs obtained by classifying based on a predetermined classification condition, and a checking unit for cutting out an input document image for each predetermined recognition unit A histogram creating means for creating a histogram from the cut out recognition units,
Preprocessing means for classifying the 0-1 pattern expression extracted from the created histogram of the recognition unit according to the predetermined classification condition to obtain a corresponding class, and extracting from the created histogram of the recognition unit. 0-
One pattern expression is partially collated with a standard pattern associated with information indicating the same class as the class obtained by the pre-processing means among the standard patterns registered in the recognition dictionary. Large classification means for searching for a candidate for a recognition target category corresponding to a recognition unit. Preferably, the recognition unit is a character unit, and the recognition target category is a character type.

In the character recognition method according to the present invention (claim 16), two or more character images of the same character type but different fonts are provided as at least one standard pattern for each recognition target category. The 0-1 pattern expression obtained by performing predetermined maximization processing and minimization processing based on the 0-1 pattern expression respectively created from the histogram obtained from the dot matrix of FIG. Information indicating the class to which the standard pattern belongs obtained by classifying based on a predetermined classification condition is registered in association with a recognition dictionary, and an input document image is cut out for each predetermined recognition unit, A histogram is created from the cut-out recognition units, and the 0-1 pattern expression extracted from the created histogram of the recognition units is expressed in accordance with the predetermined classification condition. The corresponding class is obtained by classifying the class, and the 0-1 pattern expression extracted from the created histogram of the recognition unit and the same class as the obtained class among the standard patterns registered in the recognition dictionary. Is partially collated with the standard pattern associated with the information indicating the search unit, and a candidate for a recognition target category corresponding to the recognition unit is searched. Preferably,
The recognition unit is a character unit, and the recognition target category is a character type.

Preferably, a target range of the dot matrix of the character image based on which the histogram is created is an inscribed rectangular area defined by a height and a width of the character image.

Preferably, the histogram used to create the 0-1 pattern expression is created by dividing the count value of black pixels projected from a predetermined direction on the dot matrix of the character image by the number of pixels arranged in the predetermined direction. The obtained histogram is obtained in two or more predetermined directions, and these are connected in the horizontal axis direction of the histogram.

Preferably, the predetermined direction is two directions of a vertical direction and a horizontal direction, two directions of a diagonal ± 45 degrees direction, or four directions of a vertical direction, a horizontal direction and a diagonal ± 45 degrees direction.
It is characterized by two directions.

The predetermined direction may be any direction other than the vertical direction, the horizontal direction, and the oblique direction of ± 45 degrees.

Preferably, in creating the 0-1 pattern expression, the horizontal axis of the connected histogram is L
Divided into m sections, the vertical axis is divided into m sections, and mx
L regions are formed, and in each of the m × L regions,
When a histogram passing through the region exists, 1 is assigned to the region, and when it does not exist, 0 is assigned.

[0024] Preferably, L sections obtained by dividing the horizontal axis i of the connected histogram are overlapped by a predetermined width in each of the histograms on which the connected histogram is based. Features.

Preferably, in the maximizing process, m × L regions of the 0-1 pattern expression of the standard pattern correspond to two or more character images of the same character type but different fonts. A selection is made for each of the L sections on the horizontal axis from any of the 0-1 pattern representations. In the selection, m sections are arranged in the vertical axis direction for each section on the horizontal axis. , And selects the m areas in the 0-1 pattern expression in which the number of 1 areas is the largest.

Preferably, in the minimizing process, m × L areas of the 0-1 pattern expression of the standard pattern correspond to two or more character images of the same character type but different fonts. A selection is made for each of the L sections on the horizontal axis from any of the 0-1 pattern representations. In the selection, m sections are arranged in the vertical axis direction for each section on the horizontal axis. And selecting the m areas in the 0-1 pattern expression in which the number of areas that are 1 out of the areas is the smallest.

The present invention (Claim 12) provides Claim 1
1, the histogram used to create the 0-1 pattern expression is a histogram created by dividing the count value of black pixels projected from a predetermined direction on the dot matrix of the character image by the number of pixels arranged in the predetermined direction. , 2
The above-mentioned predetermined directions are obtained, and these are connected in the horizontal axis direction of the histogram. The 0-1 pattern expression divides the horizontal axis of the connected histogram into L sections and sets the vertical axis to m sections. M × L areas are formed, and 1 is assigned to each of the m × L areas when a histogram passing through the area exists, and 0 when no histogram exists. It is characterized by being created by

The present invention (claim 13) provides claim 1
In 1 or 12, the maximizing process includes the step of converting the m × L areas of the 0-1 pattern expression of the standard pattern into two or more character images corresponding to the same character type but different fonts from each other. A selection is made for each of the L sections on the horizontal axis from any of the 0-1 pattern representations. In the selection, m sections arranged in the vertical axis direction are selected for each section on the horizontal axis. And selecting the m areas in the 0-1 pattern expression in which the number of areas that are 1 among the areas is the largest. L regions are determined for each of the L sections on the horizontal axis from any of the 0-1 pattern expressions corresponding to the two or more character images having the same character type but different fonts. To choose, In other words, for each section on the horizontal axis, the m areas in the 0-1 pattern expression in which the number of areas that are 1 are the minimum among the m areas arranged in the vertical axis direction are selected. There is a feature.

Further, the present invention (claim 14) provides claim 1
1, wherein the classification condition specifies a shape of a pattern consisting of 0 and 1 in the 0-1 pattern expression.

The present invention (claim 15) provides claim 1
2, the classification condition includes one or a plurality of different criteria, and each criteria is defined by a rule on a value in a vertical axis direction of a section on a horizontal axis in the 0-1 pattern expression and a rule on this rule. It is characterized by the provision of the section of the horizontal axis to be satisfied.

In the character recognition device and method according to the present invention,
Note that the size and the positional relationship between the vertical and horizontal line segments are directly reflected in the structure of the histogram in the vertical and horizontal directions.
Focusing on the fact that the degree is reflected in the histogram in two directions, the effect of noise is efficiently absorbed by making full use of the 0-1 pattern expression of the histogram, and the structural information of the character is effectively used. .

That is, in the present invention, an input document image is firstly cut out for each predetermined recognition unit, for example, for each character unit, and a vertical or horizontal histogram is created from the cut out recognition unit.

Then, the 0-1 pattern expression extracted from the created histogram and the previously prepared 0-1 pattern expression of each recognition target category are partially collated to correspond to the recognition unit. A candidate for a category to be recognized is searched for, and category candidates (for example, character candidates) are narrowed down. The one or more candidates for the recognition target category are output as a recognition result (or a large classification result).

In the present invention, there is no need for complicated image processing such as distance calculation by pattern matching, normalization of character images, thinning, and the like at all with simple operations such as comparison judgment and counting. There is a great advantage that a simple process of generating a histogram is the only image processing. In this way, by comparing an unknown input character with a standard pattern with a small amount of calculation, dramatic speed-up of large classification is achieved.

If partial matching based on a 0-1 pattern expression using a histogram of a character image, which is provided for the first time in the present invention, is used, the structural information of the character is directly included in the histogram as distinctive feature information. Since it is reflected, characters can be recognized with high accuracy by the simple collation processing as described above.

Now, with the method and apparatus for creating a standard pattern according to the present invention, the following can be achieved in the above-described character recognition apparatus and method.

First, in the standard pattern creating method, a histogram is created from a dot matrix of two or more character images having the same character type but different fonts, and a 0-1 pattern expression is created from each created histogram. Then, maximization processing and minimization processing are performed based on each of the created 0-1 pattern expressions, and the result is used as the standard pattern.

As a result, when the standard pattern created as described above is used, the input character image and the standard pattern
In the case of partial matching using the -1 pattern expression, if the input character image is a font used when creating the standard pattern, the input character image always matches the condition of partial matching and is therefore selected as a recognition candidate.

Therefore, given the range of font types and character sizes to be recognized, the variation of fonts is absorbed using the sample set, and the average number of candidate character types obtained in the large classification is set to a desired value. It is possible to provide a method for systematically creating a standard pattern set as follows.

Next, a description will be given of the speeding up of the recognition according to the enlargement of the dictionary according to the present invention.

In Japanese and Chinese, the number of character types to be recognized is increased (usually more than 5,000 character types) and there are many various fonts. The size of the dictionary will soon reach tens of thousands. According to the present invention, 0-1
The standard pattern is classified into an appropriate number of classes in advance based on classification conditions (classification criteria) based on a combination of salient features appearing in the pattern expression, and a corresponding class is determined for the input character image based on the classification conditions. The above-described large classification is performed using the standard pattern of the class. In this case, in practice, the scale of each class may be, for example, about several thousands. Therefore, in the present method based on the 0-1 pattern expression of the histogram of the character image, the determination of the class to which the input character image corresponds is extremely accurate and Easy to do.

Therefore, according to the present invention, even when a dictionary has a large scale, for example, about 50,000 standard patterns, the time required for the large classification is the size of the class of the standard pattern to which the unknown input corresponds. (Approximately several thousand), which is almost equal to the required time.

There is no conventional character recognition method for introducing a standard pattern classification to speed up the recognition, and the present invention provides a character recognition method utilizing character structure information via a histogram of a character image. It can be said that this is a unique speed-up method.

As described above, according to the present invention, it is possible to efficiently realize high precision and high speed printing character recognition.

[0045]

Embodiments of the present invention will be described below with reference to the drawings.

First, a histogram used in character recognition according to the present embodiment and its 0-1 pattern expression will be described. In addition,
The standard pattern creating method according to the present invention devises a method for creating a standard pattern of the 0-1 pattern expression (that is, a dictionary pattern to be registered in a dictionary), and uses one or a very small number of standard patterns with the same character type having different fonts. Can be quickly and accurately recognized. Further, the method for speeding up recognition according to the present invention enables high-speed recognition even when the number of standard patterns increases (for example, even if it increases to about tens of thousands).

The histogram used in the present embodiment is a so-called marginal distribution, which is a distribution of black pixel count values obtained by projecting a dot matrix of a character image in a vertical, horizontal or ± 45 degrees direction. However, in the present embodiment, an inscribed rectangular area defined by the height and width of the character is defined as the target range of the histogram of the character image.

FIG. 1A shows an example of a dot matrix showing a character image "book". The inscribed rectangular area is an area including 7 horizontal dots and 8 vertical dots. (B)
FIG. 1 shows an example of a vertical histogram, and FIG. 2C shows an example of a horizontal histogram. FIG. 2 shows an example of the histogram in the ± 45 degrees direction. FIGS. 3 and 4 show examples of histograms of various character images in the vertical and horizontal directions.

Next, the 0-1 pattern expression will be described. As will be described later, the 0-1 pattern expression is used for collating a character obtained from a dot matrix of input characters with a character registered in a dictionary by a classification procedure described later. The 0-1 pattern expression is created based on histograms in some predetermined directions. Here, a case where vertical and horizontal histograms are used will be described.

FIGS. 5A and 14A show examples of histograms in the vertical and horizontal directions, and FIG. 5C shows an example of 0-1 pattern expression (or a part thereof). Hereinafter, an operation performed to obtain the 0-1 pattern expression of (c) from the histograms of (a) and (b) of FIG. 5 and FIG. 14 will be described with reference to FIG.

First, as shown in FIG. 5, two vertical histograms as shown in FIG. 5A and two horizontal histograms as shown in FIG. The horizontal axis of each of the histograms (a) and (b) is divided into L1 and L2 sections, respectively. Adjacent sections are overlapped by an appropriate width δ in consideration of the change of the font.

In FIG. 5, the heights of the histograms in the vertical and horizontal directions are expressed as percentages (%) with respect to the height and width of the character image, respectively. The section [0, 100] on the vertical axis is divided into m sections, and in the m × (L1 + L2) areas, 1 is assigned to a histogram when there is a histogram passing through the area, and 0 is assigned to no such histogram. . In this way, a 0-1 pattern expression of the histogram as shown in (c) is obtained. However, if the height of the histogram is 100
When it is equal to%, it is considered that the image has passed the boundary of 100% on the vertical axis.

Each section on the horizontal axis is represented by coordinates i (i = 1, 2, 2, 3).
.., L1 + L2). The vertical axis is represented by threshold value θ, and both ends and division points are represented by θt (t = 1, 2,..., M + 1).
However, θ1 = 100% and θm + 1 = 0%. Then, each area of the 0-1 pattern representation of the histogram is
(I, θt), i = 1,..., L1 + L2, t = 1, 2,
.., M.

FIG. 5 shows an example in which L1 = 3 and L2 = 4, and FIG. 14 shows an example in which L1 = 5, L2 = 5, δ = 0,
In this example, m = 13.

It should be noted that the 0-1 pattern expression for the two-direction histograms of ± 45 degrees or the vertical and horizontal directions and ±
The 0-1 pattern expression and the like for the 45-degree four-direction histogram are configured in exactly the same manner.

Here, the duality of the 0-1 pattern expression will be described.

Since the 0-1 pattern expression represents a histogram, (A) the threshold value θt is sequentially reduced from θ1 = 100% (to θm + 1 = 0%), and t = 1, In the process of scanning as 2, 3,..., If the region (i, θt) is 1, it is 1 in all the regions (i, θt ′) where t ′ ≧ t, and the region (i, θt) ) Is 0, it is 0 in all regions (i, θt ′) where t ′ ≦ t.

On the contrary, the following can be said.

(B) The scale of the threshold value θt is 100% at the above (A), 0%, and 0% at 1%.
00%, and sequentially from θ1 = 100% (θm + 1 = 0)
%), In the process of scanning t = 1, 2, 3,..., If the region (i, θt) is 0, all the regions (i, θt ′) satisfying t ′ ≧ t Is 0 and if the region (i, θt) is 1, it is 1 in all the regions (i, θt ′) where t ′ ≦ t.

Such a relationship between (A) and (B) is referred to herein as a 0-1 pattern duality.

A classification procedure performed based on the above (A) in a large classification procedure (described later) is referred to as a classification procedure 1A, and a classification procedure based on the above (B) is referred to as a classification procedure 1B. One pattern is targeted. Also,
In the sub-classification (described later), the same classification procedure as the classification procedure 1A and the classification procedure 1B is called a classification procedure 2A and a classification procedure 2B, respectively.
-1 pattern is targeted. That is, the basic processing procedures are the same.

Next, the character recognition device of this embodiment will be described. The configuration of this character recognition device is basically the same as that of the character recognition device based on the new principle disclosed by the present inventor in Japanese Patent Application No. 6-312049. There is a feature in a standard pattern having a -1 pattern expression.

This character recognition method generally focuses on the fact that information on the structure of a character is directly reflected in a peripheral distribution (histogram) in each direction of the character image, and focuses on the extracted input character image. The partial matching is performed between the 0-1 pattern expression obtained from the histogram and the 0-1 pattern expression obtained from the histogram of each prepared standard character.

As shown in FIG. 6, the present character recognition apparatus includes a document image input unit 2, a character cutout processing unit 4, a large classification processing unit 6, a fine classification processing unit 8, and a histogram creation for creating a histogram from a character dot matrix. The unit 10 includes a 0-1 pattern expression dictionary 12 having a standard dictionary of 0-1 pattern expressions used for large classification and, if necessary, fine classification.

The 0-1 pattern expression dictionary 12 has 0-1 pattern expressions (for all m × (L1 + L2) areas) of the characters to be recognized as standard patterns. For each character to be recognized, a plurality of creation conditions (m,
L1, L2, δ, font type or size range represented by one standard pattern described later, etc.)
May be prepared in correspondence with 0-1 pattern expressions. The details of the standard pattern according to the present embodiment, a method of creating the standard pattern, and a classification procedure using the standard pattern will be described later.

The document image input section 2, character cutout processing section 4, histogram creation section 10, large classification processing section 6, and fine classification processing section 8 are the same as those disclosed in Japanese Patent Application No. 6-312049. is there. Hereinafter, the main points will be described.

The document image input unit 2 takes in a character image printed on a document as an optical image and converts it into an electric signal by using, for example, an image scanner, and then converts the density value according to a predetermined threshold value. Quantizes to black and white binary and outputs.

The character cutout processing section 4 performs a checkout process for cutting out characters from an input image by a known method based on the regularity of the arrangement of characters on a sheet.

The histogram creating section 10 creates a histogram from the character dot matrix as described above.

The large classification processing section 6 includes a histogram creation section 1
Using 0, a histogram for large classification (here, histograms in the vertical and horizontal directions) is obtained from the dot matrix of the cut-out character. Then, the 0-1 pattern expression dictionary 1
2 using the large classification process (here, a classification procedure 1A described later).
/ 1B) to narrow down candidate character types for unknown input characters.

The sub-classification processing unit 8 selects a recognition result to be output based on the set of candidate character types obtained by the large classification processing unit 6 and determines a priority order when there are a plurality of character candidates. I do. At this time, if necessary, a histogram for fine classification obtained from the dot matrix of the cut-out character (here, ±
Classification procedure 2A / 2B using the two histograms in the 45-degree direction) and the 0-1 pattern expression dictionary 12
The recognition result is determined based on the result.

The final processing result is output, for example, as a character code. If there are a plurality of character candidates, only the first one or several with the priority information are output. The output character code or the like is stored in a storage device such as a RAM or a magnetic disk, and / or displayed on a display device such as a CRT or a printer. When some candidates are output, the user is made to select as necessary.

Next, the classification procedure 1A used for the large classification will be described. The classification algorithm of the classification procedure 1A is shown in FIGS.
Shown in This process is roughly divided into four processes (a),
(B), (c), and (d).

(Process a) First, an initial value is set (step S11). The overlap width δ of the section and the constant κ as a guide for narrowing down the number of candidate character types are determined. Also, as an initial value, a set Ω of character types to be classified is set as Ω = {a set of all character types (0-1 pattern expression)}, and the vertical axis is divided into m sections. The initial value of t is t = 1, θ1 = 100%, (Example: θt + 1 = θt-5%,
t = 1, 2,..., m = 20). It is assumed that none of the coordinates i is initially fixed.

Also, (i, θt), i = 1, 2,..., L
The array of 0 and 1 (horizontal array) taken by each area of 1 + L2 is u
In the same manner, those for (i, θt + 1), (i, θt + 2) and (i, θt + 3) are represented by v, w, and x, respectively. Hereinafter, u, v, w, and x are each simply referred to as an array.

U, v, w, x are defined as 0-
A partial pattern of one pattern expression is shown (FIGS. 5C and 15C).

(Process b) With respect to the histogram of the unknown input character X belonging to Ω, the arrays u, v, w and x are obtained for the threshold values θt, θt + 1, θt + 2 and θt + 3, respectively (FIG. 5 (c), FIG. 15 (c)), (Step S12).

(Process c) When there are no coordinates other than the fixed coordinates where the value of u is 1 and there are no coordinates where the value of x is 0 (Yes in step S13),
θt = θt + 1, θt + 1 = θt + 2, θt + 2 = θt + 3, θt +
With 3 = θt + 4, the process moves to the process b (step S12) (step S14).

In other cases (No in step S13), coordinates where the value of u is 1 are i, and coordinates where the value of x is 0 are i '(step S15). From the nature of the histogram, i and i 'are different.

Further, the value of w is 1 at the coordinate i where the value of u is 1, and the value of w is 0 at the coordinate i 'where the value of x is 0. Then, in Ω, all such coordinates i (1 ≦ 1)
i ≦ L1 + L2) and i ′ (1 ≦ i ′ ≦ L1 + L2)
, A set of all character types in which the value of w is 1 and 0 is represented by Sw (i = 1, i ′ = 0). However, when one of such i and i 'does not exist, only the one existing in parentheses is described. That is, instead of Sw (i = 1, i ′ = 0), Sw (i =
1) or Sw (i ′ = 0) (Step S1)
6).

Then, 1 ≦ | Sw (i = 1, i ′ = 0)
If | ≦ κ (Yes in step S17), Sw (i = 1, i ′ = 0) is output as a set of candidate character types (step S23), and the process ends.

If | Sw (i = 1, i ′ = 0) |> κ (Yes in step S18), only i used here is set as fixed coordinates, and the process proceeds to a process d (step S30) described later. The process proceeds (step S28).

If 1 ≦ | Sw (i = 1) | ≦ κ (Yes in step S19), S is set as a set of candidate character types.
w (i = 1) is output (step S24), and the process ends.

If | Sw (i = 1) |> κ (Yes in step S20), the process proceeds to the process d (step S30) described later, with all i used here as fixed coordinates (step S29).

If 1 ≦ | Sw (i ′ = 0) | ≦ κ (Yes in step S21), Sw (i ′ = 0) is output as a set of candidate character types (step S25), and the processing ends. I do.

If | Sw (i ′ = 0) |> κ (Yes in step S22), Ω = Sw (i ′ = 0), θt = θt + 1, θt + 1 = θt + 2 , Θt + 2 = θt + 3,
With θt + 3 = θt + 4, the process moves to the process b (step S12) (step S26).

| Sw (i = 1, i ′ = 0) | = 0 or | Sw (i = 1) | = 0 or | Sw (i ′ = 0) | =
If it is 0 (No in step S22), the set Ω of the character type currently targeted as the set of candidate character types is output (step S27), and the process ends.

(Process d) Coordinate i (1 ≦ i ≦ L1 + L2)
Are fixed (Ye in step S30)
s), as a set of candidate character types, a set where i and i ′ exist is output as Sw (i = 1, i ′ = 0), and if only i is present, Sw (i = 1) is output. (Step S31),
finish.

When there is a coordinate i that is not a fixed coordinate (No in step S30), Ω = Sw (i = 1, i ′ = 0) or Ω = Sw according to the above two cases.
(I = 1), θt = θt + 1, θt + 1 = θt + 2, θ
With t + 2 = θt + 3 and θt + 3 = θt + 4, the process proceeds to the process b (step S12) (step S32).

The above is the classification algorithm of the classification procedure 1A.

Another Classification Procedure 1 Used in Major Classification
The classification algorithm of B is obtained by replacing 0 after processing b (step S12) with 1 and 1 with 0 in the classification algorithm of the classification procedure 1A.
This means that it is equivalent to applying the classification algorithm of the classification procedure 1A by replacing 0 and 1 in the 0-1 pattern expression in the classification procedure 1B. As described above, this relationship is called duality between the classification procedure 1A and the classification procedure 1B.

Next, the sub-classification processing will be briefly described.

The classification procedure 1A in the large classification can be regarded as a classification procedure based on black pixels of a character image, and the classification procedure 1B can be regarded as a classification procedure based on white pixels, and the influence of noise differs. Therefore, in the sub-classification, when the candidate character type common to both classification results is only one, it is regarded as a true candidate with sufficient reliability and is output as a recognition result.

If not, a process for determining the order of reliability is performed and a plurality of candidates are output. In this case, if necessary, the classification procedure 2A and the classification procedure 2B which are the same classification algorithms as the classification procedure 1A and the classification procedure 1B
Is used. Here, the classification procedure 2A and the classification procedure 2B use a histogram in the ± 45 degree direction as a histogram of the character image, and are particularly useful for the classification of a character having a line segment whose inclination is close to ± 45 degrees.

Further, a distance between vector patterns is calculated for prioritizing as necessary. Unknown input character X
The 0-1 pattern of the vertical, horizontal, and diagonal (± 45 degrees) histograms for
Arranged in the order of horizontal, -45 degrees, and +45 degrees, each coordinate on the horizontal axis, that is, a serial number is assigned to each section, and a variable whose value is the number of 1s in each section i is represented by yi. However,
1 ≦ i ≦ L1 + L2 + L3 + L4 = n. Then, the vector Y is defined as follows. From the definition of Y = y1 + y2 +... + Yn Y, each element yi takes a nonnegative positive value.

As the vector pattern used here, the entire standard pattern may be used. However, if necessary, the vertical and horizontal axes of the partial pattern or the histogram may be partially and finely divided for flexible use. It is effective.

FIGS. 10 to 13 show an example of the procedure of the small classification process after performing the classification procedure 1A and the classification procedure 1B in the large classification process. Here, for the unknown input character X, selection of the first-place candidate character type and error checking are performed based on the candidate character types obtained as a result of the large classification. This processing procedure is an example, and various other variations are possible.

In the embodiment described above, the histograms in the vertical and horizontal directions are used in the classification procedures 1A and 1B in the large classification, and ± 45 in the classification procedures 2A and 2B in the fine classification.
Although the histogram of the degree direction was used, the vertical,
Four histograms in the horizontal and ± 45 degrees directions can be used, and histograms in other directions such as ± 30 degrees and ± 60 degrees can be used, and can be set as appropriate.

Further, even in a configuration in which only one of the classification procedure 1A and the classification procedure 1B is performed in the large classification and the sub-classification is not performed, a sufficient effect can be obtained as compared with the related art. That is, the sub-classification is not essential to the present invention, but shows one method of post-processing.

In the following, first, 0-1 according to the present embodiment will be described.
A standard pattern (dictionary pattern) for pattern expression, a method for creating the pattern, and a classification procedure using the standard pattern will be described in detail.

The procedure of character recognition disclosed in Japanese Patent Application No. 6-312049 is as follows: character extraction → histogram generation → 0
-1 pattern expression → large classification (extraction of structural information, collation with standard pattern, selection of candidate character type) → fine classification → output (1
Rank, second place candidate, etc.). The present invention, as a standard pattern used in the above-described large classification, is completely different from the conventional one, and absorbs all font variations due to differences in fonts and character sizes appearing in the histogram of a given sample set. It provides a method for creating a standard pattern that allows the

First, a conventional standard pattern used in the classification algorithm will be described. Conventionally, when a character recognition algorithm is given, in order to realize high-precision recognition using the algorithm, a standard representing a certain finite set (represented by S) of the same character type having different fonts is used. One or more patterns must be created for each character type. In a recognition method based on a pattern matching method, a set S
, Perform statistical processing, clustering, etc., and create a small number of statistical standard patterns representing the set for each character type.

In character recognition based on a histogram, FIG.
And histograms in each direction of characters (vertical and horizontal, or vertical, horizontal and oblique ± 45 degrees, etc.) as shown in FIG.
0-1 pattern expression obtained by quantizing is equivalent to the feature space. In the 0-1 pattern expression, in the area of arbitrary coordinates i on the horizontal axis (that is, small section i), the section on the vertical axis is the section of one column (corresponding to the histogram of black pixels) extending from bottom to top. , Which extend from top to bottom (corresponding to a histogram of white pixels). The vertical axis is divided into an appropriate number (represented by m) of small sections.

As shown in FIGS. 5 and 14, the vertical axis of the 0-1 pattern expression is expressed in units of this small section.
The value of the vertical axis at the end of the section of column 1 (boundary with the section of column 0) is given by the number of 1s included in column 1. Therefore, when the number of 1s at an arbitrary coordinate i on the horizontal axis is represented by yi, the 0-1 pattern expression is an n-dimensional vector YY = (y1, y2,..., Yn), n = l ₁ + l ₂ + l ₃ + l ₄ Represented by (1). Where l ₁ , l ₂ , l ₃ , and l
₄ indicates the number of small sections on the horizontal axis of the four histograms in the vertical, horizontal, and oblique directions of ± 45 degrees, respectively. Note that the above n varies depending on the type of histogram used. For example, when only vertical and horizontal histograms are used, n = l ₁
+ L ₂ .

In Japanese Patent Application No. 6-312049, any one 0-1 pattern expression in the set S is used as a standard pattern, and the variation of the font is absorbed by the above-described large classification method (that is, the classification procedure 1A). If it cannot be absorbed, the basic means is to increase the number of standard patterns. On the other hand, according to the present invention, by utilizing the characteristics of the histogram, even if any character of the set S of the same character type becomes an unknown input in the large classification method, the same character type is always a candidate character type. Can be created to be included in the standard pattern. Hereinafter, a method of creating a standard pattern according to the present invention will be described in detail.

First, the basic concept will be described.

(A) 1 of 0-1 pattern expression of character image
In the case where attention is paid to the column of, the number of 1s (corresponding to the height of the column of 1, ie, the height of the black pixel histogram) in the abscissa i of the 0-1 pattern expression of the unknown input character X is represented by xi, For the abscissa, yi ≧ xi (i = 1, 2,..., N) (2)
(1). Then, at an arbitrary abscissa i from the property of the histogram, Y (1) is also 1 in a section where X is 1.
Therefore, Y (1) is always selected as a candidate character type of X in the large classification method.

(B) 0 of 0-1 pattern expression of character image
Focusing on the column of the number of 0s on the abscissa i of the 0-1 pattern expression of the unknown input character X (0 extending from top to bottom of the 0-1 pattern expression)
Ii ′, i.e., corresponding to the height of the white pixel histogram), is represented by xi ′, and all the abscissas satisfy yi ′ ≧ xi ′ (i = 1, 2,..., N) (3). Is expressed as 0-1 pattern expression having the number yi ′ of Y
(0). Then, at an arbitrary abscissa i from the nature of the histogram, Y (0) is also 0 in a section where X is 0.
Therefore, in the large classification method, Y (0) is always selected as a candidate character type of X.

Since the number of sections on the vertical axis is represented by m, the number xi of 1 in X on the abscissa i and 1 in Y (1)
Are xi = m−xi ′ and yi = m, respectively.
−yi ′. Therefore, when these are substituted into Expression (3), Y (0) becomes a 0-1 pattern expression satisfying yi ≦ xi (i = 1, 2,..., N) (4).

Therefore, in the large classification method, Y
Y = (Y (1), Y expressed as a pair of (1) and Y (0))
If (0)) is made to correspond to one character type, when focusing on the 1st column of the 0-1 pattern expression of the unknown input character X, Expression (2) is established, and focusing on the 0th column Equation (4) in the case
Holds, the character type Y is always selected as one of the candidate character types of X.

Next, a description will be given of a method of creating a standard pattern that absorbs variations in fonts.

There are various fonts for the same character type depending on various fonts and character sizes. The method of creating a standard pattern that absorbs such a change in font is as follows.

R characters Y1, having the same character type but different fonts
Let S be a finite set consisting of Y2,..., Yr. S = {Y1, Y2,..., Yr} (5) In the following, in order to simplify the description, Y1, Y2,.
Yr represents a character and is used as a symbol representing a 0-1 pattern expression of the character.

Y1i (1), y2i (1),..., Yri
(1) are represented by 0-1 pattern expressions Y1, Y2,
.., Yr as the number of 1s on the abscissa i, and the maximum value among them is represented by yi (1). Yi (1) = max ｛y1i (1), y2i (1),. ｝ (I = 1, 2,..., N) (6)

Therefore, yi (1) ≧ y1i (1), yi (1) ≧ y2i (1),..., Yi (1) ≧ yri (1) (i = 1, 2,..., N) (7) ) Holds. The number of 1s on the abscissa i is yi (1), (i =
The 0-1 pattern expression given by 1, 2,.
Represented by (1).

For example, (a), (b), and (c) of FIG.
From the three types of 0-1 pattern expressions representing the same character type as shown in FIG.
(1) is obtained. As can be easily understood by comparing the figures, the column of 1 at the coordinate i of Y (1) in FIG. 16A is the coordinate of each figure in (a), (b), and (c) in FIG. It is equal to the largest one column in i.

The above processing is called maximization processing of 0-1 pattern expression.

On the other hand, y1i (1), y2i (1),.
When the minimum value of yri (1) is represented by yi (0), yi (0) = min {y1i (1), y2i (1),..., yri (1)} (i = 1, 2,...) , N) (8).

Therefore, yi (0) ≦ y1i (0), yi (0) ≦ y2i (0),..., Yi (0) ≦ yri (0) (i = 1, 2,..., N) (9) ) Holds. If the number of 0s on the abscissa i is yi (0), (i =
The 0-1 pattern expression given by 1, 2,.
It is represented by (0).

For example, (a), (b), and (c) of FIG.
The 0-1 pattern expression Y (0) as shown in FIG. 16B is obtained from the 0-1 pattern expression. FIG.
The row of 1 at the coordinate i of Y (0) in (b) is equal to the minimum row of 1 at the coordinate i in each of FIGS. 15 (a), (b) and (c).

The above processing is called a 0-1 pattern expression minimization processing.

Next, Y (S) is represented by a 0-1 pattern expression Y
A pair of (1) and Y (0) is represented by Y (S) = (Y (1), Y (0)) (10), where Y (S) has the same character type as set S and sets S This is a representative standard pattern. When selecting a candidate character type by focusing on the 1st column of the 0-1 pattern expression of the unknown input character X, using Y (1), and selecting a candidate character type by focusing on the 0th column Is Y (0). Assuming that the abscissa of the former is i and the abscissa of the latter is i ′ in these two cases, yi (1) ≧ xi and yi ′ for all i and i ′ selected for X, respectively.
If (0) ≦ xi ′ is satisfied, Y
(S) is always included in the narrowed down candidate character types of X.

For Y1, Y2,..., Yr belonging to the set S, since the equations (7) and (9) hold,
Whichever input is unknown, Y (S) is naturally included in the narrowed down candidate character types. Therefore, the standard pattern Y (S) that absorbs all the variations of the font of the set S is created.

As described above, according to the present embodiment, given a range of font types and character sizes to be recognized, variations in fonts can be absorbed using the sample set and obtained in the above-described large classification. It is possible to provide a method for systematically creating a standard pattern set such that the average number of candidate character types is equal to or less than a desired value.

In the large classification process, the classification process 1A mainly composed of one column of the 0-1 pattern expression is referred to as Y.
By using (1) to execute the above-described classification process 1B mainly using a sequence of 0s in the 0-1 pattern expression using Y (0), an effective candidate character type utilizing the advantages of both. Can be narrowed down.

Therefore, by using such a standard pattern, the recognition accuracy and the recognition speed of a printed character can be more effectively improved.

Here, a case will be described in which the variation of the font is large for the same character type.

A finite set S = 字 Y of the same character type in equation (5)
If it is recognized that the variation of the font in 1, Y2,..., Yr} is large, the number of candidate character types obtained in the large classification increases, and the burden of the fine classification increases. In such a case, the set S may be divided into subsets (classes) having similar fonts, and a standard pattern representing each class may be created according to the method described above. The number of such classes depends on how small the average number of candidate character types obtained in the large classification is. Therefore, the number of classes is expected to vary greatly depending on the character type. For example, the characters "T" and "river" are JIS
In the range of characters such as Level 1 and Level 2, it does not depend much on the font, printer type and character size,
The number of classes is likely to be one or at most two.

The following procedure can be considered as a method of dividing the set S of the same character type into an appropriate number of classes of characters having similar fonts.

(I) Classification based on clustering method As a most general method, a set S of the same character type is divided into an appropriate number of classes by applying a clustering method, and a standard pattern representative of each class is obtained. Is created according to the method described above.

(Ii) Verification of Classification Based on Recognition Experiment Whether the classification of the set S by the clustering method is appropriate or not is verified by a recognition experiment. That is, using a standard pattern representative of each class, a recognition experiment was performed on the font data included in S and the font data not included in S, and the average number of candidates narrowed down for the former and the average number of candidates for the latter The validity of the classification is verified based on the average number of candidates and the ratio (cumulative recognition rate) in which true candidate character types are included therein. As a result of the verification,
If there is any inconvenience, the procedure of returning to the above (i) is repeated to obtain a reasonable result.

Hereinafter, a standard pattern creation device and a character recognition device including a standard pattern creation function according to an embodiment of the present invention will be described.

FIG. 17 shows an example of the configuration of the standard pattern creating apparatus, and FIG. 18 is a flowchart showing a schematic procedure thereof.

As shown in FIG. 17, the standard pattern creating apparatus includes a character image data input section 22, a histogram creating section 2
4, a 0-1 pattern expression creating unit 26 and a standard pattern creating unit 28 are provided.

First, standard characters having the same character type but different fonts for use in creating a standard pattern are determined (step S201). In determining the standard characters, verification of (i) the classification based on the clustering method and (ii) the classification based on the recognition experiment is performed as necessary.

The character image data input unit 22 takes in standard characters printed on a document as an optical image and converts them into an electric signal using, for example, an image scanner, and then converts the density into an electric signal according to a predetermined threshold value. Quantizes the value to binary black and white. Alternatively, a dot matrix corresponding to each standard character already stored in an information storage medium such as a magnetic disk is read (step S202).

If a histogram or 0-1 pattern expression corresponding to a standard character already exists in a magnetic disk or the like, it may be read. Needless to say, when the histogram is read, the processing of the subsequent step S203 is unnecessary for the standard character, and 0-
When the one-pattern expression is read, the subsequent steps S203 and S204 are not necessary for the standard character.

Next, the histogram creating section 24 has basically the same function as the histogram creating section 10 of FIG. 6, and creates respective histograms from the dot matrix of each standard character as described above (step S1). S20
3).

Next, the 0-1 pattern expression creating section 26
As described above, the respective 0-1 pattern expressions are created from the dot matrix of each standard character (step S20).
4).

Finally, the standard pattern creating section 28 performs the above-described maximization processing and minimization processing for each coordinate i for each of the obtained 0-1 pattern expressions corresponding to the respective standard characters. To obtain a standard pattern (step S20)
5).

The standard pattern created in this way is registered in, for example, the 0-1 pattern expression dictionary 12 in FIG. In addition, if necessary, the above-described (i) classification based on the clustering method and (ii) verification of the classification based on the recognition experiment are performed.
The standard pattern registered in the -1 pattern expression dictionary 12 is updated to a more appropriate one.

FIG. 19 shows an example of a character recognition device including the standard pattern creation function. This character recognition device is shown in FIG.
17 and FIG. 17 except that the histogram creating unit 24 in FIG. 17 is also used as the histogram creating unit 10 in FIG. Has the same configuration.

Although the recognition unit has been described as a character unit in the above description, characters separated vertically and horizontally (for example, “gate”)
It is also possible to prepare each part separated as the character type for "random" and compare and recognize the input image and the standard pattern for each separated part.

By the way, the number of character types to be subjected to character recognition usually exceeds several thousand in Japanese and Chinese. further,
Considering that fonts vary depending on various fonts, printers, character sizes, etc., the total number of standard patterns required even using the standard pattern creation method described above,
That is, the size of the character recognition dictionary is expected to exceed tens of thousands. In the conventional method, the time required for character recognition increases almost in proportion to the size of the dictionary.

In such a case, the standard pattern (0-1 pattern expression) created as described above is
-1 is classified into several classes by a method based on the salient features appearing in the pattern expression, and before the large classification process, a corresponding class is obtained from the 0-1 pattern expression of the unknown input character image, and belongs to this class. When the large classification process is performed by focusing on the standard pattern, even if the dictionary for character recognition is enlarged, it is possible to speed up the recognition process while maintaining high recognition accuracy.

In the following, a description will be given of a method for speeding up the recognition in accordance with the enlargement of the character recognition dictionary.

This method is based on the above-mentioned Japanese Patent Application No. Hei 6-31204.
By applying the basic idea of the large classification of character recognition disclosed in No. 9 in advance, standard patterns are classified in advance into an appropriate number, for example, a dozen or more classes, and the 0-1 pattern expression of an unknown input character image is obtained. This makes it possible to determine the class of the standard pattern to be collated based on accurate and simple comparison and judgment processing. The method generally comprises:
(1) Classification condition setting, (2) Classification of standard patterns according to the set classification conditions, (3) Standard pattern assignment to other classes, which is performed as necessary to avoid erroneous classification in the large classification process Addition, (4) Step 2 or Steps 2 and 3 are performed for all standard patterns.

The specific method will be described below.

First, the setting of the standard pattern classification condition will be described.

Vertical, horizontal and oblique (± 45 degrees) of the character image
When classifying the 0-1 pattern expression of the directional histogram, various classification conditions can be set. The basic concept will be described below using an example.

Here, the classification conditions are represented using the following four symbols.

A _i (u): In the section where the scale on the vertical axis of the 0-1 pattern expression exceeds u%, all columns are 0.

B _i (v): All sections where the height of the 0-1 pattern expression is less than or equal to v% are 1 column.

C _i (w, p): In the 0-1 pattern expression, there are p rows of 1 or less with a height of w% or more.
However, p ≧ 1.

D (y): In the 0-1 pattern expression,
In the section at one or both ends of the horizontal axis, the height of y% or more 1
Column exists.

Here, each symbol has a subscript i = 1, 2, 2,
According to 3 and 4, respectively, the vertical of the target character image,
Regarding the 0-1 pattern expression of the histogram in the horizontal and diagonal directions (± 45 degrees), 1 (True) is established when the proposition shown in the right sentence is satisfied, and 0 (Fa) when the proposition is not satisfied.
1se).

For example, A ₁ (u) = 1 indicates 0 in the 0-1 pattern expression of the vertical histogram of the target character image in all sections where the vertical axis scale exceeds u%.
And that A ₁
(U) = 0 indicates that it is not established.

Similarly, C ₂ (w, p) = 1 indicates that p (p ≧ 1) or less with a height of w% or more in a 0-1 pattern expression of a horizontal histogram of a target character image. Of 1
It is shown that the existence of the sequence of

A ₁ (u) = 1 and C ₂ (w, p)
= 1 or A ₁ (u) = 1 and C ₂ (w, p) =
Various classification conditions necessary for the classification of the standard pattern can be represented by various combinations of the four conditions, such as 0, and the way of giving parameters such as u, v, w, p.

Next, a specific classification method based on such classification conditions will be described.

By selecting and using classification conditions suitable for the character set to be recognized according to the characteristics of the character set to be recognized, tens of thousands of standard patterns in the dictionary are composed of standard patterns of which the size of each class is not more than several thousand. It is possible to classify into an appropriate number of classes. At this time, with respect to a standard pattern that may fall into a plurality of classes in consideration of variations in fonts, noise, and the like, each class is configured to belong to the corresponding plurality of classes.

For example, when A ₁ (u) is used as the classification condition, the standard pattern is a class in which A ₁ (u) is 1, that is, in a section where the scale on the vertical axis exceeds u%, the column from the column of all 0s. The standard pattern is classified into a class of the standard pattern and a class of A ₁ (u) of 0, that is, a class of the standard pattern in which a column of 1 also exists in a section where the scale of the vertical axis exceeds u%. However, there may be a standard pattern in the former class with one row of height not exceeding u% but at or near u%. Similarly, in the latter class there may be standard patterns with more than u% but one row of heights close to u%. Therefore, when character images corresponding to those standard patterns are input, there is a misclassification that the former corresponding to the former is classified into the latter or the latter corresponding to the former due to fluctuations in the font and noise. It is possible.

In order to avoid such misclassification, 0-
A region of u ± δ is provided around u% of the scale on the vertical axis of one pattern expression. Here, δ> 0. And A
₁ reference pattern becomes a 1 in (u) the A ₁ (u-δ) at 0, that is the height of one of the columns is less than u% (u-δ) all standard patterns exceeding% A ₁ (u)
Is added to the class where is 0. On the other hand, in A ₁ (u), 0
However, A ₁ (u) is 1 for all the standard patterns that are _{1 in} A ₁ (u + δ), that is, the standard patterns in which the height of one row exceeds u% but is not more than (u + δ)%. Add to class. However, it is necessary to determine the magnitude of δ to an appropriate value in consideration of the variation of the font and the magnitude of noise.

In the case of C _i (w, p), similar considerations are required for w by providing an area of w ± α as in the case of u. Here, α> 0. p is the number of one row having a height of not less than w%, and the variation of the number of one row is mainly due to the horizontal axis when the histogram of the columnar black pixels is converted to the 0-1 pattern expression. , One column is formed in each of the sections on both sides of the boundary line, which is converted into a total of two one columns. In a character image, the width of such a columnar portion is usually considerably smaller than the section width of the horizontal axis, so that the columnar portion is converted into one column unless it is on the boundary line.

For example, if p = 2, C _i (w, p)
Is a standard pattern having one or two rows of 1 having a height of not less than w%, and a class having a value of C _i (w, p) of 0 is 1 having a height of not less than w%. Is a standard pattern having three or more columns. Therefore, it is necessary to add to the former a standard pattern that may move from the latter class to the former in consideration of font variations and noise (however, in the sense that those standard patterns are excluded from the latter) Absent). In the case of this example, it is sufficient if the standard patterns to be added are the following standard patterns. That is, a standard pattern in which three rows of 1 having a height of w% or more are three and two of them are adjacent (there is a possibility that these two become one), and the height is w% Above 1
Is a standard pattern in which two columns are adjacent to each other (each of two adjacent pairs may be one). On the other hand, as a standard pattern that may move from the former to the latter, 1
Is only two standard patterns (if at least one of the two becomes two single columns, it will move to the latter class), and these may be added to the latter. .

As described above, if the standard patterns are classified in advance using various classification conditions, the class to which the 0-1 pattern expression of the unknown input character image corresponds can be extremely determined based on the classification conditions. It is possible to select candidate character types at high speed and with high accuracy based on a large classification method as disclosed in Japanese Patent Application No. 6-312049.

Here, a description will be given of a character recognition system to which the above-described method for increasing the size of the character recognition dictionary, that is, a method for using a standard pattern classified into classes, is applied. This character recognition system is shown in FIGS.
9 can be realized by making appropriate modifications to the character recognition system of FIG. Except for the portion to be corrected, it is basically the same as that described above, and thus only the portion to be corrected will be described.

First, the above-described class classification processing for the standard pattern is performed by the class classification unit provided after the standard pattern generation unit 28 in the standard pattern generation device in FIG. 17 or the character recognition system in FIG.

Then, the standard pattern classified into the class is 0 in the character recognition system shown in FIG. 6 or FIG.
-1 is registered in the pattern expression dictionary 12. That is, for each standard pattern, the data of the 0-1 pattern expression and the information indicating the class to which it belongs are stored in association with each other.

When an unknown input character is input to the character recognition system, as a pre-process of the large classification processing by the large classification processing unit 6, a standard pattern is first added to the 0-1 pattern expression obtained from the unknown input character image. Is subjected to a class classification process under the same conditions as when the class was classified, and a class corresponding to the unknown input character is obtained. Of course, the determination of the class of the unknown input character is extremely accurate and easy.

In the large classification process, the large classification process is performed focusing on standard patterns belonging to the same class as the class obtained for the unknown input character.

As described above, high-speed and high-precision character recognition can be easily realized even when the dictionary is enlarged.

Hereinafter, effects obtained by the present embodiment will be described.

In Japanese Patent Application No. 6-312049, a standard pattern prepared in advance as a dictionary is character data (sample) for one set of character sets to be recognized. Then, the variation of the font for each character type is dealt with by the large classification method, and a character that cannot be dealt with and is erroneously recognized is added to the dictionary as a standard pattern.
Therefore, for a recognition target composed of a plurality of fonts and characters having different sizes, a learning experiment for large classification is performed using a learning set, and characters that have been erroneously recognized are added to a dictionary prepared in advance as a standard pattern. The actual recognition is performed using the dictionary reinforced by performing such an operation.

On the other hand, the standard pattern (dictionary) according to the present invention
According to the method of making, the following is possible.

(1) Creation of Standard Pattern (Dictionary) for Absorbing Variations of Various Fonts As described above, given the font type and character size of the character to be recognized, based on one set thereof, A near-minimum number of standard patterns (dictionaries) that completely absorb the variation of the font of each character type can be created immediately.

A specific example will be described below in this regard.

The size is 6, 7, 8, 10,.
JIS first-level kanji characters 2965 for 5,11,12,14,20 points of Ishii Mincho style, and 9 points of Ishii Tai Mincho style, L Mincho style, and M Mincho style
A standard pattern was created for the character type by the method according to the present invention, and a large classification experiment was performed by the above-described large classification method. However, the horizontal axis of the vertical and horizontal histograms of the character image is divided into six, and the horizontal division of the ± 45-degree histogram is divided into three, that is, 6, 6, 3, 3. (Of course, if the number of divisions on the horizontal axis is increased, the number of candidate character types in the large classification is reduced to a smaller number.)

In this case, there are nine types of character sizes and four types of fonts, and there are 11 samples having different sizes or fonts for each character type (corresponding to the above-described set S). A set of 11 characters having the same character type but different fonts is represented by one standard pattern (Y (S) described above) for each character type, and a standard pattern is created for 2965 kanji characters. A large classification experiment was performed using all the captured characters as unknown input.

As a result, in the case of Mincho Ishii, the average number of candidates was about 4 for any size. Among them, 100% of true candidates were included. In this case, when the horizontal axis is divided into 7, 7, 3, and 3, the average number of candidates is about 1.5 in the case of Minami Ishii.

Further, as for the 2965 kanji characters, a large classification experiment of 22 points of Akashi Ishii which was not included in the creation of the standard pattern showed that the average candidate was 3.8.
The percentage of individuals that contained true candidates was 98.2%.

If an attempt is made to absorb variations in fonts without applying the standard pattern according to the present invention to the above-mentioned recognition target, not only a considerable amount of time and effort is required for a learning experiment, but also a large classification is required. It is expected that the average number of candidate character types obtained will be close to two digits. Furthermore, while the present invention makes it relatively easy to determine the number of valid standard patterns for each character type in a large classification, a technique that relies solely on a large classification method and a learning experiment requires such a determination. Considerable difficulties are expected.

(2) Correction and Addition of Standard Pattern Based on Misrecognition An input character X having an error is compared with a standard pattern (generally a plurality) of the same character type as the input character X.
By utilizing the conditions of Equations (2) and (4) at each coordinate on the horizontal axis of the 0-1 pattern expression, there are standard patterns whose number of coordinate axes that do not satisfy the conditions is equal to or less than a predetermined number (threshold). Then, the standard pattern is corrected by a method of creating a new standard pattern representing two of X and its standard pattern. If such a standard pattern does not exist, X
Correction / addition of the standard pattern efficiently by giving a method of adding X to the standard pattern, or giving a method of creating a new standard pattern by performing clustering again by including X in the existing standard pattern of that character type Can be.

(3) Application to Recognition of Collapsed Characters Very small characters with a character size of 6 points or less or complicated Chinese characters may be collapsed by an image scanner of an input device, or collapsed during printing. Often occur. For this reason, it can be said that measures against broken characters are one of the most important practical issues in terms of improving the accuracy of character recognition. The present invention is expected to function effectively as a countermeasure.

That is, character crushing is regarded as a special case in which the variation of the font is large, and the clustering of the 0-1 pattern expression in various cases where crushing occurs is performed for characters that are likely to be crushed. Create a representative standard pattern. Then, when it is determined that the possibility of crushing is large based on information at the time of character extraction (for example, when a small letter of 6 points or less is used in a footnote, etc., or when printing is bad and there is crushing, etc.). ),
It is expected that the method of using a collapsed standard pattern in accordance with the situation will contribute to higher recognition accuracy.

(4) When Considering Variations of Fonts over a Wide Range When considering variations of fonts over a wide range such as differences in fonts and sizes and various printers, fonts, fonts, A font, size, and printer in a vector space that represents a vector obtained by concatenating vectors of all character types targeted for each font, each size, and each printer in the same order, in a level in units of size and printer. A simple method can be considered in which clustering is collectively performed to create a standard pattern for each obtained class. This method also provides a highly practical standard pattern.

The present invention is not limited to the above-described embodiments, but can be implemented with various modifications within the technical scope.

[0188]

According to the present invention, input characters and standard patterns of character types are partially collated by a 0-1 pattern expression based on, for example, a vertical or horizontal histogram of those dot matrices. In a character recognition device that searches for a character type candidate for a character, based on 0-1 pattern expressions respectively created from a histogram obtained from a dot matrix of two or more character images of the same character type but different fonts from each other. In this case, a standard pattern is created so that each font conforms to partial matching, so it is possible to create a standard pattern that can effectively absorb variations in characters due to differences in fonts, and create this standard pattern By registering a dictionary and recognizing characters, even if characters with the same character type but different fonts are input, character type candidates can be obtained quickly and accurately. It is.

According to the present invention, the standard pattern (0
-1 pattern expression) are classified into several classes in advance according to classification conditions based on salient features appearing in the 0-1 pattern expression.
Since the corresponding class is obtained from one pattern expression and the large classification process is performed by focusing on the standard patterns belonging to this class, the recognition process can be performed while maintaining high recognition accuracy even if the character recognition dictionary is enlarged. It is possible to increase the speed.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating an example of a dot matrix of a character image “book” and an example of histograms in the vertical and horizontal directions thereof.

FIG. 2 is a diagram illustrating an example of a dot matrix of a character image “book” and an example of a histogram in a ± 45-degree direction thereof;

FIG. 3 is a diagram showing an example of vertical and horizontal histograms of various character images.

FIG. 4 is a diagram showing an example of vertical and horizontal histograms of various character images.

FIG. 5 is a diagram for explaining vertical and horizontal histograms and 0-1 pattern expression;

FIG. 6 is a block diagram showing a configuration of a character recognition system according to an embodiment of the present invention.

FIG. 7 is a flowchart showing a classification algorithm of a classification procedure 1A.

FIG. 8 is a flowchart showing a classification algorithm of a classification procedure 1A.

FIG. 9 is a flowchart showing a classification algorithm of a classification procedure 1A.

FIG. 10 is a flowchart illustrating an example of a sub-classification procedure.

FIG. 11 is a flowchart illustrating an example of a sub-classification procedure.

FIG. 12 is a flowchart illustrating an example of a sub-classification procedure.

FIG. 13 is a flowchart showing an example of a sub-classification procedure.

FIG. 14 is a diagram for explaining histograms in the vertical and horizontal directions and 0-1 pattern expression;

FIG. 15 is a diagram showing three types of 0-1 pattern expressions representing the same character type.

FIG. 16 is a diagram for explaining a standard pattern according to the present invention.

FIG. 17 is a block diagram showing a configuration of a standard pattern creation device according to an embodiment of the present invention.

FIG. 18 is a flowchart illustrating an example of a standard pattern creation procedure.

FIG. 19 is a block diagram showing a configuration of a character recognition system according to an embodiment of the present invention.

[Explanation of symbols]

 2 Document image input unit 4 Character extraction processing unit 6 Large classification processing unit 8 Fine classification processing unit 10 Histogram creation unit 12 0-1 pattern expression dictionary 14 Standard character histogram dictionary 22 Character image data input unit 24 ... Histogram creation unit 26 ... 0-1 pattern expression creation unit 28 ... Standard pattern creation unit

Claims (16)

(57) [Claims] 1. A histogram is created from a dot matrix of an input character image, and a 0-1 pattern expression extracted from the created histogram of the character image, and one or a plurality of patterns are prepared for each character type. The standard pattern 0-1 as the standard pattern for the character recognition device for partially comparing the character pattern 0-1 pattern expression as the standard pattern of each character type and searching for a character type candidate corresponding to the input character image. A standard pattern creation method for creating a pattern expression, wherein a histogram is created from a dot matrix of two or more character images having the same character type but different fonts, and a 0-1 pattern expression is created from each created histogram. And performing maximization processing and minimization processing based on each of the generated 0-1 pattern expressions, and using the result as the standard pattern. Standard pattern creation how. 2. The method according to claim 1, wherein the target range of the dot matrix of the character image based on which the histogram is created is an inscribed rectangular area defined by the height and width of the character image. The standard pattern creation method described. 3. A histogram used to create the 0-1 pattern expression is created by dividing a count value of black pixels projected from a predetermined direction on the dot matrix of the character image by the number of pixels arranged in the predetermined direction. Histogram
3. The standard pattern creating method according to claim 1, wherein the predetermined directions are obtained and connected in the horizontal direction of the histogram. 4. The predetermined direction is two directions of a vertical direction and a horizontal direction, two directions of a diagonal ± 45 degrees direction, or four directions of a vertical direction, a horizontal direction and a diagonal ± 45 degrees direction. 4. The standard pattern creating method according to claim 3, wherein: 5. When creating the 0-1 pattern expression, the horizontal axis of the connected histogram is divided into L sections, and the vertical axis is divided into m sections, and m × L areas are divided. 5. The standard according to claim 3, wherein, in each of the m × L regions, 1 is assigned to the region when a histogram passing through the region exists, and 0 is assigned to the region when no histogram exists. Pattern creation method. 6. The maximizing process includes the step of:
The mxL area of the -1 pattern expression is represented by any one of the 0-1 pattern expressions corresponding to the two or more character images having the same character type but different fonts from each other. The selection is performed for each of the L sections. In the selection, for each of the sections on the horizontal axis, the number of 0-areas, which is the largest among the m areas arranged in the vertical axis direction, is the largest. 6. The standard pattern creating method according to claim 5, wherein said m regions in one pattern expression are selected. 7. The minimizing process includes the step of:
The mxL area of the -1 pattern expression is represented by any one of the 0-1 pattern expressions corresponding to the two or more character images having the same character type but different fonts from each other. The selection is made for each of the L sections, and in the selection, the number of the 0-areas, which is the smallest among the m areas arranged in the vertical axis direction for each of the sections on the horizontal axis, is the smallest. 6. The standard pattern creating method according to claim 5, wherein said m regions in one pattern expression are selected. 8. A histogram is created from a dot matrix of an input character image, and a 0-1 pattern expression extracted from the created histogram of the character image and one or a plurality of patterns are prepared for each character type. The standard pattern 0-1 as the standard pattern for the character recognition device for partially comparing the character pattern 0-1 pattern expression as the standard pattern of each character type and searching for a character type candidate corresponding to the input character image. A standard pattern creating apparatus for creating a pattern expression, comprising: means for creating histograms from dot matrices of two or more character images of the same character type but having different fonts; 0-1 from each of the created histograms; Means for creating a pattern expression; and means for performing a maximization process and a minimization process based on each of the created 0-1 pattern expressions. Quasi-pattern forming apparatus. 9. An inspection means for cutting out an input document image for each predetermined recognition unit, a histogram making means for making a histogram from the cut out recognition unit, and a histogram extracted from the created histogram of the recognition unit. By partially comparing the 0-1 pattern expression with the 0-1 pattern expression of each of the recognition target categories prepared in advance,
Large classification means for searching for a candidate of a recognition target category corresponding to the recognition unit, wherein the 0-1 pattern expression of each of the previously prepared recognition target categories has the same character type but different fonts. Character recognition characterized by being obtained by performing predetermined maximization processing and predetermined minimization processing based on 0-1 pattern expressions respectively created from a histogram obtained from a dot matrix of a character image of apparatus. 10. For each of a plurality of character types to be recognized, a histogram is created from a dot matrix of two or more character images having the same character type but different fonts, and a histogram is created from each of the created histograms. One pattern expression is created, maximization processing and minimization processing are performed based on each created 0-1 pattern expression, and the result is a 0-1 pattern that is a standard pattern corresponding to the character type for which recognition is to be performed. Registered as an expression in the recognition dictionary, the input document image is cut out for each predetermined recognition unit, a histogram is created from the cut out recognition unit, and 0-1 extracted from the created histogram of the recognition unit is created.
The pattern expression is partially collated with the 0-1 pattern expression that is the standard pattern corresponding to each character type to be recognized registered in the recognition dictionary, and the character type corresponding to the recognition unit is compared. A character recognition method characterized by searching for candidates. 11. A 0-bit histogram prepared from a dot matrix of two or more character images having the same character type but different fonts, which are prepared as at least one standard pattern for each recognition target category. A 0-1 pattern expression obtained by performing predetermined maximization and minimization processing based on one pattern expression, and a class obtained by classifying the 0-1 pattern expression based on predetermined classification conditions. A recognition dictionary in which information indicating a class to which the standard pattern belongs is registered, a check means for cutting out an input document image for each predetermined recognition unit, and a histogram creation for forming a histogram from the cut out recognition units Means, classifying the 0-1 pattern expression extracted from the created histogram of the recognition unit according to the predetermined classification condition, Preprocessing means for obtaining a corresponding class; 0-1 pattern expression extracted from the created histogram of the recognition unit; and a class obtained by the preprocessing means among standard patterns registered in the recognition dictionary. A large classifying means for partially collating a standard pattern associated with information indicating the same class and searching for a candidate for a recognition target category corresponding to the recognition unit. Recognition device. 12. A histogram used to create the 0-1 pattern expression is created by dividing a count value of black pixels projected from a predetermined direction on the dot matrix of the character image by the number of pixels arranged in the predetermined direction. The histogram
Two or more predetermined directions are obtained, and these are connected in the horizontal axis direction of the histogram. In the 0-1 pattern expression, the horizontal axis of the connected histogram is divided into L sections, and the vertical axis is It is divided into m sections to form m × L areas, and in each of the m × L areas, 1 is set to the area when there is a histogram passing through the area, and 0 is set to 0 when no histogram exists. The character recognition device according to claim 11, wherein the character recognition device is created by assigning. 13. The maximizing process includes the step of converting m × L regions of the 0-1 pattern expression of the standard pattern into two or more character images corresponding to the same character type but different fonts from each other. From any of the 0-1 pattern representations,
The selection is performed for each of the L sections on the horizontal axis. In the selection, the number of areas that are 1 out of m areas arranged in the vertical axis direction is maximum for each section on the horizontal axis. The m areas in the 0-1 pattern expression are selected, and the minimization processing is performed by setting the m × L areas in the 0-1 pattern expression of the standard pattern to the same character type. Corresponds to each of two or more character images having different fonts.
A selection is made for each of the L sections on the horizontal axis from any one of the one-pattern representations. In the selection, for each of the sections on the horizontal axis, m areas in the vertical axis direction are selected. 13. The character recognition device according to claim 11, wherein the m regions in the 0-1 pattern expression in which the number of regions that are 1 are the smallest are selected. 14. The character recognition apparatus according to claim 11, wherein said classification condition specifies a shape of a pattern consisting of 0 and 1 of said 0-1 pattern expression. 15. The classification condition comprises one or more sets of different criteria, each criterion defining a value in the vertical axis direction of a section of the horizontal axis in the 0-1 pattern expression. 13. The character recognition device according to claim 12, wherein the character recognition apparatus includes a rule for a section on the horizontal axis that satisfies the rule. 16. At least one of the standard patterns prepared for each of the recognition target categories, each of which is created from a histogram obtained from a dot matrix of two or more character images having the same character type but different fonts from each other. A 0-1 pattern expression obtained by performing predetermined maximization and minimization processing based on one pattern expression, and a class obtained by classifying the 0-1 pattern expression based on predetermined classification conditions. Information indicating the class to which the standard pattern belongs is registered in the recognition dictionary in association with it, the input document image is cut out for each predetermined recognition unit, and a histogram is created from the cut out recognition units. The 0-1 pattern expression extracted from the histogram of the recognition unit is classified into classes according to the predetermined classification condition to obtain a corresponding class. The 0-1 pattern expression extracted from the histogram of the recognition unit and a standard pattern associated with information indicating the same class as the obtained class among the standard patterns registered in the recognition dictionary. A character recognition method characterized by partially collating and searching for a candidate for a recognition target category corresponding to the recognition unit.