JP3236732B2 - Character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device,
In particular, the present invention relates to a character recognition device that recognizes characters in an input document image and converts the characters into character codes.

[0002]

2. Description of the Related Art In recent years, a printed document is read by optical / electrical conversion or the like, and is then converted into bit information for each pixel, and characters in the image data are recognized. Research and development have been made on systems for saving input, further translating into foreign languages, and uttering voices for blind persons, visually impaired persons, children and school children, and some of them have been put into practical use.

[0003] The present invention relates to a character recognition device employed in such a system. By the way, in a character recognition device employed in such a conventional system, a black pixel is projected in a vertical direction on a character string with respect to a cut-out character string image, and a continuous image area is defined as a character image. (For example, Akiyama et al., “Putting Out Characters in Article Area of Printed Material” PRL80-70). By the way, aside from simple recognition objects such as alphabets, there are many types and numbers of characters to be recognized in Japanese, Chinese, etc. It is necessary to know the exact shape and size of the character being written. Further, even in the case of a printed document in a language of European and American languages, it is important to accurately determine the size of the character to be recognized because various types of symbols are used in scientific and technical papers. However, in the above-described character extraction technology, a character image is extracted using a continuous area of orthographic projection as a character. Therefore, when there is a contact between characters (hereinafter referred to as a contact character), for example,
Separated characters consisting of multiple character elements separated in the character string direction such as "north" and "river" in horizontal writing documents and "2" and "3" in vertical writing documents, and characters originally consisting of single character elements. However, if there is a character that has been separated into a plurality of character elements due to blurring (hereinafter, both are referred to as a separated character), accurate extraction cannot be performed.
Therefore, several techniques have been proposed for correctly extracting characters from a character string image of a document that includes contact characters and separated characters. For example, in Japanese Patent Application Laid-Open No. 5-128308 "Character recognition device", a character element having a width smaller than a predetermined character size is treated as a separated character, and a character element before and after (left, right, up and down) within a width not exceeding the character size. Connect and cut out as one character. In addition, a character element wider than a predetermined character size is treated as a contact character, and is divided by the character size.
Each of the divided parts is cut out as one character. When a character element having a width smaller than the predetermined character size and a character element having a width wider than the predetermined character size are continuous, a separation position when the character element is divided for each character size from a head position of the character element having a small width. The separation position when divided for each character size from the top position of the wide character element is set as a cutout candidate position, and all the images sandwiched between two different cutout candidate positions among these cutout candidate positions are once subjected to character Tentatively recognized, and a character having a high evaluation value of the character tentatively recognized is selected as an original character. This takes into account the possibility that the latter half element of the separation character is in contact with the next character.
On the other hand, in Japanese Patent Laid-Open No. 5-128307 "Character Recognition Apparatus", while Japanese Patent Laid-Open No. 5-128308 generates a clipping candidate position for each fixed character size, a half-width character width is applied to the character size. All the separation positions between the case where the character width is applied and the case where the full-width character width is applied are all candidates.

[0004] Also, regarding the detection of margins in a character string,
The interval width of the cut-out characters is compared with the full-width character width or the half-width character width, and a case where the margin width is larger is defined as a margin.

[0005]

However, the method disclosed in Japanese Patent Laid-Open No. 5-128308 is based on the premise that the character size is constant. For this reason, it is not possible to properly cope with a document printed in a mixture of full-width and half-width and a document with an irregular pitch. In addition, technical papers and the like have many documents in these modes. On the other hand, in the method disclosed in Japanese Patent Application Laid-Open No. 5-128307, the number of clipping candidate positions becomes extremely large, and all images sandwiched between two different clipping candidate positions among these clipping candidate positions are once recognized as characters. Therefore, the amount of calculation becomes extremely large, and the processing speed of the entire system is delayed.

In the detection of margins, a narrow half-width margin cannot be detected when a full-width character width is used as the character width. Conversely, when a half-width character width is used as a character width, for example, when a narrow-width character continues in a document printed at a full-width pitch, such as “11”, the margin width between the characters is changed to a half-width character. Since it is larger than the width, an extra margin is detected. As a result, it is difficult to properly cope with a document printed in a mixture of full-width and half-width and a document with an irregular pitch.

SUMMARY OF THE INVENTION In view of the above problems, the present invention provides a character recognition apparatus capable of quickly and accurately extracting characters and detecting margins even for documents printed in full-width and half-width and documents of irregular pitch. The purpose of this is to provide.

[0008]

In order to achieve the above object, according to the first aspect of the present invention, a character string image is cut out from an input document image which is converted into bit information for each pixel, and then each character image is extracted. In a character recognition device that further cuts out individual character images from a column image and recognizes the individual character images and converts them into corresponding character codes, a character element that is a block of pixels constituting a character from the character string image (Composed of a line segment or a curve having the same phase with respect to the character string direction or a combination thereof), and a full-width based on the position or size of the character element extracted by the character element extraction unit A character width calculator for calculating the width of the character and the width of the half-width character by predetermined means (including a procedure and a method); and a half-width calculated by the character width calculator among the character elements extracted by the character element extraction unit. Sentence If there is a character element narrower than the width, it is detected and used as it is as a sub character element.If there is a character element wider than the half width character width, it is detected and multiple sub The character element dividing unit that divides into character elements and the sub character element generated by the character element dividing unit are recognized as constituting a single character with the character elements or sub character elements of the preceding and following character strings. A character element image that is connected and cut out so as to be as wide as possible (a concept including the length in the vertical direction if the recognition direction is up and down) as long as the target connected width does not exceed the half-width character width. A character candidate generating unit that cuts out a character element image that is connected and cut out so that the width after connection is as wide as possible without exceeding the full-width character width as a full-width character candidate image, and A recognition unit for temporarily recognizing the half-width character candidate image and the full-width character candidate image generated by the character candidate generation unit as a single character, and calculating a corresponding character code and an evaluation value indicating accuracy; and the recognition unit. Recognition result evaluation that substitutes the evaluation values obtained for the half-width character candidate image and the full-width character candidate image into a predetermined function and compares them, and determines the character candidate image with the higher evaluation value as a correct character. And a part.

In the second aspect of the present invention, the character width calculating section includes a position or a size (including both) of a character element.
And a character printing pitch calculating means for calculating the width of full-width characters and the width of half-width characters from the calculation result. In the invention according to claim 3, the character width calculating unit calculates the height of the character string image (the size or length in the direction perpendicular to the character string of the character string image,
If it is a vertical writing, it will be a so-called width. ), A temporary character width calculating means for calculating a temporary character width from the character elements extracted by the character element extraction unit, a character element having an error between the width of the character element and the temporary character width smaller than a predetermined value. A continuous part detecting means for detecting a continuous part, and calculating a character printing pitch from a distance between the center points in the character string direction of the detected continuous character elements, and calculating a width of a full-width character and a width of a half-width character from the calculation result. And a calculating means for calculating.

According to a fourth aspect of the present invention, the character width calculation section calculates a provisional character width from the height of the character string image and a character element having a width smaller than the provisional character width. In this case, the character elements are temporarily connected within a range where the width does not exceed the provisional character width, and the error between the provisionally connected character element and the width of the original character element and the provisional character width is obtained. A continuous part detecting means for detecting a part where a smaller character element is continuous, and calculating a character printing pitch from a distance between the middle points in the character string direction of the character element of the continuous part detected by the continuous part detecting means. Calculation means for calculating the width of full-width characters and the width of half-width characters from the calculation results.

According to a fifth aspect of the present invention, the recognition unit includes a half-width character recognition dictionary used to recognize the half-width character candidate image generated by the character candidate generation unit, and a half-width character recognition dictionary generated by the character candidate generation unit. And a full-width character recognition dictionary used for recognizing full-width character candidate images. In the invention of claim 6, the recognizing unit determines whether or not the recognition result of the full-width character candidate image is a character having the same shape as one half-width character divided in half in a character string direction of a single character. A half-character judging unit; and a priority evaluation unit that gives priority to the evaluation value of the full-width character candidate image when the recognition result estimating unit evaluates the half-character judging unit as such a character. It is characterized by:

[0012]

[0013] In the invention according to claim 7 , each of the character images determined to be a single character by the recognition result evaluation unit is
A margin detection unit that detects an inter-character space based on whether each character is a full-width character or a half-width character and a midpoint interval in the character string direction of two consecutive characters, based on a detection result of the margin detection unit A margin adding section for inserting a margin code at a corresponding position in the output character code.

[0014]

According to the first aspect of the present invention, a character string image is cut out from an input document image, and then individual character images are further cut out from each character string image. The following operation is performed in the character recognition device that converts the data into the corresponding character code.

A character element extraction unit extracts a character element which is a block of pixels constituting a character from a character string image. The character width calculating unit calculates the width of the full-width character and the width of the half-width character by predetermined means from the position or size of the character element extracted by the character element extraction unit. If the character element division unit has a character element having a width smaller than the half-width character width calculated by the character width calculation unit among the character elements extracted by the character element extraction unit, the normal projection shadow of the pixel in the character string direction is not detected. Detects whether there is a continuous part, etc., and makes it as a sub character element as it is, and if there is a character element wider than half width character width, detects this and divides it into multiple sub character elements whose individual width is less than half width character width I do. The character candidate generation unit cuts out the connected character elements formed by connecting the sub character elements as much as possible within a range not exceeding the half-width character width for the sub character elements generated by the character element division unit. The image to be extracted is a half-width character candidate image, and an image cut out by a connected character element formed by connecting sub-character elements as far as possible within a range that does not exceed the full-width character width is cut out as a full-width character candidate image. The recognition unit temporarily recognizes the half-width character candidate image and the full-width character candidate image generated by the character candidate generation unit as a single character, and then evaluates the corresponding character code and accuracy by pattern matching or the like. Calculate the value. The recognition result evaluation unit substitutes the evaluation values obtained for the half-width character candidate image and the full-width character candidate image in the recognition unit into a predetermined function and compares them, and determines the character candidate image with the higher evaluation value as a result. Judge as correct characters.

According to the second aspect of the present invention, the character printing pitch calculating means in the character width calculating unit calculates the mutual positional relationship of the character elements or the size (width) in the character string direction and the size (width) in the direction orthogonal thereto. The character printing pitch is calculated from the height and the relative size with respect to the line spacing, and the width of the full-width character and the width of the half-width character are calculated from the calculation result. According to the third aspect of the present invention, the provisional character width calculation means in the character width calculation unit calculates the provisional character width from the height of the character string image. Similarly, the continuous part detecting means detects a part of the character elements extracted by the character element extraction part in which the character element having an error between the width of the character element and the temporary character width smaller than a predetermined value is continuous. Similarly, the calculating means calculates the character printing pitch from the detected distance between the middle points of the consecutive character elements in the character string direction, and calculates the width of the full-width character and the width of the half-width character from the calculation result.

In a fourth aspect of the present invention, the provisional character width calculation means in the character width calculation unit calculates the provisional character width from the height of the character string image. Similarly, in the case where the continuous part detecting means includes a series of character elements having a width smaller than the calculated temporary character width, the character elements are temporarily connected within a range that does not exceed the temporary character width, and the temporarily connected character elements are connected. Then, an error between the width of the original character element and the provisional character width is obtained, and a portion where the character elements whose values are smaller than a predetermined value continues is detected. Similarly, the calculating means calculates the character printing pitch from the distance between the middle points in the character string direction of the character element of the continuous part detected by the continuous part detecting means, and calculates the width of the full-width character and the width of the half-width character from the calculation result. I do.

In the fifth aspect of the present invention, the half-width character recognition dictionary in the recognition unit is used for recognizing the half-width character candidate image generated by the character candidate generation unit by pattern recognition or the like. Similarly, a full-width character recognition dictionary is used to recognize a full-width character candidate image generated by the character candidate generation unit. In the invention according to claim 6, the half-character judging means in the recognizing unit determines whether or not the recognition result of the full-width character candidate image is a character having the same shape as one half-width one character divided in the character string direction of a single character. to decide. Similarly, if the priority evaluation unit determines that the character is such a character by the half character determination unit, the evaluation value of the full-width character candidate image is prioritized in the evaluation by the recognition result evaluation unit.

[0019]

According to the seventh aspect of the present invention, for each character image determined by the margin detection unit to be a single character in the recognition result evaluation unit, it is determined whether each character is a full-width character or a half-width character. The inter-character margin is detected from the midpoint interval of the character in the character string direction. Similarly, the margin adding unit inserts a margin code at a corresponding position in the output character code based on the detection result of the margin detection unit.

[0021]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a character recognition device according to the present invention will be described based on embodiments. In the following embodiment, characters such as portrait, width and width, and a normalization unit that corrects large characters such as characters in the heading column, which are full-width but large as characters in the heading column, to standard characters, and outputs recognized characters The output unit that performs the translation and utterance based on the output result as a whole, such as the utterance unit, etc., the user performs various operations, and inputs the conditions that are known in advance. Needless to say, it has an input operation unit for performing the operation. However, since these are not directly related to the gist of the present invention, illustration and description are omitted. (First Embodiment) FIG. 1 is a configuration diagram of a first embodiment of the present invention. In FIG. 1, reference numeral 1 denotes an image input unit into which an image is input. Reference numeral 2 denotes a character string extracting unit that detects the position of the character string from the document image read by the image input unit 1 and cuts out the character string image. Reference numeral 3 denotes a character element extraction unit that extracts a character element that is a block of pixels constituting a character from the character string image cut out by the character string extraction unit 2. Reference numeral 4 denotes a character width calculation unit that calculates the width of full-width characters and the width of half-width characters from the position or size of the character element extracted by the character element extraction unit 3. Reference numeral 5 denotes a character element having a width smaller than the half-width character width calculated by the character width calculation section 4 among the character elements extracted by the character-element extraction section 3 and detects the character element and sets it as a sub-character element as it is. This is a character element division unit that detects a character element having a width wider than the width and divides the character element into a plurality of sub-character elements each having a width equal to or less than a half-width character width. Reference numeral 6 denotes an image cut out by a connected character element connected to a sub-character element as far as possible within a range that does not exceed a half-width character width for the sub-character element generated by the character element dividing unit 5. A character candidate generating unit that cuts out, as a full-width character candidate image, an image that is cut out as a candidate image and that is cut by a connected character element to which sub-character elements are connected as far as possible within a range that does not exceed a full-width character width. Reference numeral 7 denotes a recognition unit that temporarily recognizes the half-width character candidate image and the full-width character candidate image generated by the character candidate generation unit 6 as candidate characters, and calculates a corresponding character code and an evaluation value that is a probability of correctness of recognition. . A recognition result evaluation unit 8 compares the evaluation values obtained for the half-width character candidate image and the full-width character candidate image by the recognition unit based on a predetermined function, and determines the character candidate image with the higher comparison result as a correct character. Department. 9 is the correct 1 in the recognition result evaluation unit 8.
For each character image determined to be a character, detecting whether each character is a full-width character or a half-width character, and detecting a margin existing between the characters from the midpoint interval in the character string direction of two consecutive characters, This is a margin adding section for inserting a margin code at a corresponding position in the output character code. A recognition result output unit 10 outputs a recognition result.

Hereinafter, the operation of the character recognition apparatus having the above-described configuration will be described with reference to an example of a horizontally written input image as shown in FIG. In the figure, the lower two circles represent some characters that are not directly used for describing the operation of the present embodiment. The image input from the image input unit 1 has a pixel forming a character as 1 and a background pixel other than a character as 0.
Is stored in an image memory (not shown) as binary data. The character string extraction unit 2 first obtains the normal projection shadows of the pixels forming the characters in the vertical and horizontal directions with respect to the character string from the document image stored in the image memory. The width of the normal projection shadow and the interval between the normal projection shadows are compared to determine whether the input document image is a vertical writing document or a horizontal writing document (for this technology, see Japanese Patent Application No. 60-7763, for example).
No. 3). Next, the document image is divided into blocks parallel to the character string direction, the character part projection in the character string direction is obtained for each block, and a rectangular area sandwiched by the boundary coordinates of the block and the start and end coordinates of the projection is defined as the block. By examining the character string of each block for overlap in the direction perpendicular to the character string of the block, by connecting the character strings of each block and obtaining the coordinates of the character string,
Cut out a character string image. This is shown in FIG. 3A (this technique is disclosed, for example, in Japanese Patent Application No. 60-106404, "Character Recognition Apparatus"). The character element extraction unit 3 extracts a character element from the character string image (FIG. 3A) cut out by the character string extraction unit 2. As shown in FIG. 3 (b), the character string of the character string image is projected in the vertical direction, and a continuous portion of the regular projection shadow of the character part is defined as a character string element. Find rectangle coordinates. this,
S10 shown in FIG. 3C. The character width calculation unit 4 calculates the width of full-width characters and the width of half-width characters from the coordinates of the character elements extracted by the character element extraction unit 3. Here, there are several methods for calculating the character width in the character width calculation unit 4, and three methods adopted in the present embodiment will be described below.

The first method will be described with reference to FIG. The height H of the character string is multiplied by a predetermined constant α to obtain a temporary character width W.
* Is calculated. Generally, in a printed Japanese document, since the aspect ratio of many characters is almost 1, α takes a value close to about 1. However, it is needless to say that there is a slight difference depending on whether the correction is performed or the contents of the normalization processing are performed in accordance with the printed font. Next, since there is a vertical writing "one (the beginning of kanji and one of the kanji numerals)" and a horizontal writing "1 (the kanji of the kanji and one of the kanji numbers)", the character whose width is almost equal to the provisional character width W * Elements (s3, s5, s9, s10 in the figure)
Width W of the widest character element (s3 in the figure)
A value obtained by multiplying max by a predetermined constant β is defined as a full-width character width Wz. That is, the width of each character element is set to Wi (i = 1, 2,...)
Then, the full-width character width Wz is represented by the following equation.

Wmax = max {Wi | Wi / W * 〜1} Wz = β1 · Wmax (where β1 to 1) where max {U} represents the maximum value of the set U. A to B in A to B indicate that A and B have substantially the same value. The half-width character width Wh is calculated by multiplying the full-width character width Wz by a predetermined constant γ.

Wh = γ · Wz (where,
γ to 0.5), the second method will be described with reference to FIG. In the same manner as in the first method, the provisional character width W * is calculated, and a plurality of character elements having a width substantially equal to the provisional character width W * are present close to each other (in the figure, (s9, s10) The average value or the maximum value of the distance between the middle points (d1 in the figure) in the character string direction is set as the character printing pitch P, and the full-width character width Wz and the half-width character width Wh are calculated by the following equations. . This is due to the fact that as many characters as possible, which are as long as possible, are to be evaluated, and a change in the character printing pitch is not flexibly handled.

Wz = β2 · P (here, β2-1 and β2 <1) Wh = γ · Wz (here, γ-0.5). The third method will be described with reference to FIG. Similar to the first method, the provisional character width W * is calculated, and the provisional character width W * is calculated.
When a character element having a smaller width continues (s6,
In s7), character elements are tentatively connected to each other as long as the width does not exceed the provisional character width W * (s' in the figure), so that the characters are treated as a single character, and a character whose width is substantially equal to the provisional character width W * The average of the distances between the midpoints in the character string direction (d1, d2 in the figure) in the case where a plurality of elements or temporarily connected character elements exist close to each other ( s5 , s' and s9, s10 in the figure) The value or the maximum value is set as the character printing pitch P, and the full-width character width Wz and the half-width character width Wh are respectively calculated by the following equations. This is also for the purpose of evaluating as long and as many characters as possible.

Wz = β2 · P (where,
β2−1 and β2 <1) Wh = γ · Wz (here, γ〜0.5) It should be noted that which of the above three methods is adopted depends on the form of the original document to be processed, ie, Whether the character spacing is described relatively closely as in newspapers, whether the character spacing is relatively large as in the specification of the patent application,
It is needless to say that full-width kanji and half-width numbers and alphabets are used together as in an academic paper, and furthermore, it is determined whether or not unequal intervals are used. Further, as a matter of course, optimal values of the constants such as γ and β2 are appropriately selected according to the document to be recognized.

The operation performed by the character element division unit 5 for each character element si of the character elements extracted by the character element extraction unit 3 will be described with reference to FIG. The half-width character width calculated by the character width calculation unit 4 is represented by Wh, and the character element s
Let the width of i be Wi. First, the number of divisions N is determined. Division number N = [Wi / Wh + 0.5]. here,
[X] is a Gaussian symbol and represents a maximum integer value not exceeding X.

Next, s4, s5 in FIG.
When the number of divisions N> 1 as in s7 and s10, the character element si of interest as the recognition target is divided into N equal parts. S1, s2, s3, s6, s in FIG.
If the number of divisions N = 1 as in 8, s9, nothing is done.

Next, referring to FIG. 6, the character candidate generation unit 6, the recognition unit 7, the recognition result The details of the processing performed by the evaluation unit 8 to determine the recognition result will be described. First, the character candidate generation unit 6 focuses on the sub-character element s1 at the head of the character string, and as much as possible cuts out the connected character elements connected to the sub-character elements as long as the width after connection does not exceed the half-width character width Wh. A character candidate image c11 and a full-width character candidate image c21 cut out by a connected character element connected to sub-character elements as far as possible within a range that does not exceed the full-width character width Wz after connection are generated. Next, the recognizing unit 7 recognizes the characters after setting the two character candidate images as a single character image, and calculates a character code and an evaluation value. (Note that this evaluation technique is disclosed, for example, in Japanese Patent Application No. 63-310288, "Character Recognition Method.") In this case, the half-width character candidate image is changed to a half-width character candidate image in order to improve the recognition speed and accuracy. The recognition is performed using the recognition dictionary, and the full-width character candidate image is recognized using the full-width character recognition dictionary. In the figure, the character code for the half-width character candidate image c11 is “8 (Arabic numerals)”, and the evaluation value = 37.
The character code for the full-width character candidate image c21 is “Akira”, and the evaluation value is = 61. These results are compared by the recognition result evaluation unit 8, and the character candidate image with the higher evaluation is determined to be a character.

This comparison may be made simply based on the magnitude of the evaluation value (recognition evaluation value) of the recognition result. However, in this embodiment, in order to obtain a more accurate result, the evaluation value is subject to conditions. The following function operation is performed, and the final comparison is performed based on the magnitude of the evaluation value (final evaluation value) obtained as a result. Now,
The recognition evaluation value of the half-width character candidate image is represented by v1, the recognition evaluation value of the full-width character candidate image is represented by v2, the final evaluation value of the half-width character candidate image is represented by V1, and the final evaluation value of the full-width character candidate image is represented by V2.

First, the recognition result of the full-width character candidate image is
The first half of one character, for example, the horizontal character "Kai" and "Hachi" have the same shape as "I" and "No" in the left half, respectively. If the character has the same shape as one half-width character, the value obtained by multiplying the recognition evaluation value v2 of the full-width character candidate image by a constant δ1 (here, δ1> 1.0) is used as the final value of the full-width character candidate image. Evaluation value V2
And

That is, V2 = δ1 × v2.
Next, in other cases, a value obtained by multiplying the recognition evaluation value v2 of the full-width character candidate image by a constant δ2 (here, 1.0 <δ2 <δ1) is set as a final evaluation value V2. That is, V2 = δ2 × v23.

In the half-width character candidate image, the recognition evaluation value v1 is used as it is as the final evaluation value V1. That is, V1 = v
14. If V1 <V2, the full-width character candidate image is determined to be a character. If V1 ≧ V2, the half-width character candidate image is determined to be a character. This is because weighting is performed so that the longest possible pixel is evaluated and recognized.

In the case of FIG. 6, the final evaluation value V1 of the half-width character candidate image c11 is 37, the final evaluation value V2 of the full-width character candidate image c21 is 67, and V1 <V2. And the character code "Akira" is the recognition result. When the recognition result of one character is determined as described above, the same process is sequentially repeated for the adjacent sub-character elements, and as a result of the recognition, "1960" indicated by a bold frame in the character code column in the figure is displayed. can get. This is shown in FIG. Note that the character candidate generation unit 6
In, when the half-width character candidate image c11 and the full-width character candidate image c21 match, the full-width character is processed without performing other recognition processing.

When the recognition result is determined for all the characters in the character string, the margin adding unit 9 determines whether or not there is a margin,
If a margin is detected, the margin is inserted at a corresponding position in the output character code. This processing will be described with reference to FIG. 3A and FIG. 3A, FIG.
(B) is added for explanation. Each character candidate image ci (i = 1, 2,...) Converted into a character by the recognition result evaluation unit 8 is represented by a double-byte character. The margin adding unit 9 sequentially calculates the distance di between the middle points in the character string direction with respect to the next character candidate image ci + 1 for each of the character candidate images ci that have been converted into characters by the recognition result evaluation unit 8, and calculates the distance di between the middle points. The presence or absence of a margin is determined as follows. Here, P is a character printing pitch, ε1, ε
2 and ε3 are constants of about 1.0, 0.75 and 0.5, respectively. However, other specific values are adopted according to the print content.

1. When both characters ci and ci + 1 are full-width characters, if di> ε1 · P, there is a blank space between characters ci and ci + 1. 2. When one of the characters ci and ci + 1 is a full-width character and the other is a half-width character, if di> ε2 · P, the characters ci and ci +
There is a blank space between one. 3. If both characters ci and ci + 1 are half-width characters, di> ε
If it is 3 · P, there is a blank space between the characters ci and ci + 1.

4. In other cases, there is no margin between ci and ci + 1. In FIG. 7A, since the distance d5 between the midpoints of the full-width character c5 * and the half-width character c6 is d5> ε2 × P, a blank code is inserted between the fifth and sixth characters of the recognition result. On the other hand, in FIG. 7B, the width of the margin between the characters c1 * and c2 * is wide, but since both characters are full-width characters, the midpoint distance d1 <ε1
× P, and is determined not to be a margin. (Second Embodiment) The basic configuration of the second embodiment of the present invention is the same as that of the first embodiment. For this reason, in the present embodiment, the description of the configuration, operation, and the like of each unit based on the entire configuration diagram is omitted, and the principle, purpose, configuration, operation, and effect of only the unique portion will be described. About the principle.

In a normally printed document, only one of full-width and half-width characters is usually used.
Even if both types of characters are used in combination, one of two-byte characters and one-byte characters is mainly used, and the other type of characters is almost always used exceptionally. Also, when the other type of character is used in exceptional cases, the exceptional character is almost always used continuously. Therefore, when determining whether a character is a two-byte or one-byte character, if there is a character already determined before and after the character, this can be used for the determination. About purpose.

In addition to improving the character recognition speed and utilizing the half-width character recognition dictionary and the full-width character recognition dictionary more effectively, portions printed with different types of characters are
The performance of the system as a whole is improved by subjecting it to translation or the like separately or independently of other parts. About the configuration.

Initially or at the beginning of a sentence, the character type is determined by the constituent elements (requirements) of the first embodiment according to the first to sixth aspects of the present invention. Indicates that the following character is temporarily determined to be the same type as the already determined character, and character recognition is performed using only one of the full-width character recognition dictionary and the half-width character recognition dictionary. If an unrecognizable character appears, character recognition is performed using the other dictionary. If still unrecognizable, the character is determined using the same components (requirements) as in the first embodiment. together,
The user is alerted to this effect. About the effect.

This is the reverse of the purpose, and it not only determines the character type and recognizes the character, but also improves the performance of the entire system. As described above, the present invention has been described based on the embodiments,
The present invention is not limited to the above embodiment. That is, the following is also included in the present invention.

(1) A plurality of components (requirements, steps) described in the claims are provided for convenience of manufacture and the like. Conversely, one is set to a plurality. Alternatively, these are appropriately combined. (2) The image extracting means has a built-in storage unit such as a detachable and mountable floppy disk and is separate from other means.

Further, another means separately takes out the storage section and makes it mountable. Similarly, the output unit outputs to a floppy disk or the like. Thereby, the portability is improved, the expensive character recognition unit main body and the printing unit are effectively used, and the device is used in combination with other devices. (3) Characters to be recognized are not limited to kanji and Arabic numerals, but may be kana characters (kana), Hangul characters, alphabets, or characters consisting of these characters and kanji, or symbols such as "=". Of course, various symbols are included.

(4) The order in which characters are written is not limited to any one of left-to-right writing, right-to-left writing, and up-down direction, and has a function that allows the user to input this fact in advance. In the case of the vertical direction, there are two types, left to right and right to left. The same applies to this.

(5) Each constant such as β1, β2, ζ1, ζ2, and γ is determined by the user who has visually recognized the characters and contents of the processed document, and the type of the document such as a newspaper, a collection of papers, and a specification of a patent application. Can be pre-input with rough guesswork
In addition, a function is added that enables the user to change or correct the temporary recognition result. (6) Similarly, by putting underlines, upper and lower lines, and the like with predetermined markers on corresponding portions of the characters on the sheet before being read, inputting the position of half-width characters and special symbols on the device side. A function to enable it is added. This may be useful in newspaper reading systems for people with visual impairments or school children, and in translation systems for foreign language articles in areas where special symbols are frequently used.

(7) The whole system, such as a translation system, detects that the translation cannot be performed properly, or the user knows that there are many errors in the character recognition result, and based on this, the character recognition device A function of correcting and learning the recognition process by changing various constants automatically or by an instruction is added. (8) The evaluation method or the recognition means itself in character recognition is not limited to pattern recognition with a dictionary, but may use other methods such as the decision tree method (Japanese Patent Application No. 5-68586 “decision tree type character recognition device”). Has adopted.

(9) When it is determined that one of the divided character elements in the full-width character does not constitute the half-width character from the result of the collation as the half-width character or the previous projection shadow length (for example, FIG. In (a) (S1, S2)) in c), a function for determining a full-width character is added. vice versa,
A function to determine that a half-width character is a full-width character when it is determined that the character is not a half-width character based on the result of collation as a half-width character or an orthographic projection length before that (for example, day (S10) in FIG. 3C). Is added.

(10) In European and American documents, a margin is provided between words, and the margin is adjusted so that characters are placed at both ends of a line. Therefore, a function to deal with the margin between words instead of between characters may be added.

[0050]

As described above, according to the present invention, all of the extracted character elements are once converted into sub-widths smaller than half-width character width.
A character element dividing unit for dividing the character elements, and connect the divided character element, and character candidate generation unit for cutting out single-byte character candidate image and the full-width character candidate image, recognizes each byte character candidate image and the full-width character candidate image And a recognition result evaluation unit for selecting a higher evaluation value as a correct character. For this reason, for each character, recognition of a half-width character candidate image and recognition of a full-width character candidate image are performed at most twice, so that a document printed in a mixture of full-width and half-width characters and a document with an unfixed pitch can be obtained. Character extraction can be performed accurately . Further , since it is possible to determine whether the character is a full-width character or a half-width character, it is possible to accurately detect a margin.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is an example of an input document image.

FIG. 3 is a diagram for explaining a character element extraction process.

FIG. 4 is a diagram for explaining a character width calculation process.

FIG. 5 is a diagram for explaining a character element division process.

FIG. 6 is a diagram for explaining candidate character image generation processing, recognition processing, and recognition result evaluation processing.

FIG. 7 is a diagram for explaining margin addition processing.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input part 2 Character string extraction part 3 Character element extraction part 4 Character width calculation part 5 Character element division part 6 Candidate character image generation part 7 Recognition part 8 Recognition result evaluation part 9 Margin addition part 10 Recognition result output part

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Satoshi Emura 1006 Oaza Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (56) References JP-A-4-282789 (JP, A) JP-A-5 -16060 (JP, A) JP-A-63-83887 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/62 620 G06K 9/20 340 G06K 9/34 JICST file (JOIS)

Claims (7)

(57) [Claims] 1. A character string image is cut out from an input document image, then individual character images are further cut out from each character string image, and each character image is recognized and converted into a corresponding character code. In the character recognition device, a character element extraction unit that extracts a character element that is a lump of pixels constituting a character from the character string image, and a position or a size of the character element extracted by the character element extraction unit A character width calculating unit that calculates the width and the width of the half-width character by predetermined means; and a character element narrower than the half-width character width calculated by the character width calculating unit among the character elements extracted by the character element extracting unit. and as it is the sub-character element to detect, if any, to detect, if any wide character element width than single-byte character width
Divided into multiple sub-character elements whose individual width is less than half-width character width
, And replace all character elements with narrow sub-width characters
A character element dividing unit to be a character element, and all the sub character elements generated by the character element dividing unit are regarded as constituting a single character with the character element or the sub character element of the preceding and following character strings. with the character element image width after the recognition target and the connection is cut connected so widely made as much as possible without exceeding the byte character width and half-width character candidate image, width after connecting said A character candidate generation unit that cuts out a character element image that is connected and cut out as wide as possible within a range that does not exceed the full-width character width as a full-width character candidate image, and a half-width character candidate image generated by the character candidate generation unit; A recognition unit for temporarily recognizing the full-width character candidate image as a single character, and calculating a corresponding character code and an evaluation value indicating accuracy; a half-width character candidate image and a full-width The evaluation values obtained for character candidate image comparison by substituting a predetermined function,
A character recognition device comprising: a recognition result evaluation unit that determines a character candidate image having a higher evaluation value as a result as a correct character. 2. The character width calculating section has a character printing pitch calculating means for calculating a character printing pitch from a position or a size of a character element and calculating a width of a full-width character and a width of a half-width character from the calculation result. The character recognition device according to claim 1, wherein the character recognition is performed. 3. The character width calculation unit calculates a provisional character width from a height of a character string image, and a width of a character element among the character elements extracted by the character element extraction unit. Part detecting means for detecting a part in which a character element in which the error between the character and the provisional character width is smaller than a predetermined value, and a printing pitch of the character based on the distance between the middle points in the character string direction of the detected continuous character element 2. The character recognition device according to claim 1, further comprising: calculation means for calculating a width of a full-width character and a width of a half-width character from the calculation result. 4. A character width calculating unit for calculating a temporary character width from the height of a character string image; and a character element having a width smaller than the calculated temporary character width. The character elements are temporarily connected within a range where the width does not exceed the provisional character width, and an error between the width of the provisionally connected character element and the original character element and the provisional character width is determined. A continuous portion detecting means for detecting a portion where a smaller character element is continuous, and calculating the character printing pitch from the distance between the middle points in the character string direction of the character element of the continuous portion detected by the continuous portion detecting means, 2. The character recognition device according to claim 1, further comprising a calculating unit configured to calculate a width of a full-width character and a width of a half-width character from the calculation result. 5. The recognition unit includes: a half-width character recognition dictionary used for recognizing a half-width character candidate image generated by the character candidate generation unit; and a full-width character candidate image generated by the character candidate generation unit. 5. The character recognition apparatus according to claim 1, further comprising a full-width character recognition dictionary used for recognition. 6. A half-character determination unit that determines whether a recognition result of a full-width character candidate image is a character having the same shape as one half-width character when half of a single character divided in the character string direction is determined. And priority evaluation means for giving priority to the evaluation value of a full-width character candidate image when the recognition result evaluation unit evaluates the character if the half-character determination means determines that the character is such a character. The character recognition device according to claim 1, 2, 3, 4, or 5, wherein 7. For each character image determined to be a single character by the recognition result evaluation unit, the midpoint interval in the character string direction of two characters that are continuous with each other as to whether each character is a full-width character or a half-width character And a margin detection unit that detects a margin between characters from the above, and a margin addition unit that inserts a margin code at a corresponding position in the output character code based on the detection result of the margin detection unit. 7. The character recognition device according to claim 1 , wherein the character recognition device is a character recognition device.