JP3727422B2 - Character recognition apparatus and method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a character recognition apparatus and method for recognizing a document image in which items described in advance, such as forms and business cards, are clear.
[0002]
[Prior art]
In recent years, character recognition devices have become widespread, and it is desired to improve the accuracy of character recognition.
FIG. 16 is a block diagram of a conventional character recognition apparatus described in Japanese Patent Laid-Open No. 63-311492. This character recognition device includes an image input unit 1, an image memory 2, a sub character pattern cutout unit 3, a character pattern extraction unit 4, a recognition unit 5, a recognition dictionary 6, a word processing unit 7, a word dictionary 8 and a display unit 9.
[0003]
The sub character pattern cutout unit 3 cuts out a sub character pattern that is a constituent element of a character such as “bias” and “旁” from a document image if it is a Chinese character. The character pattern extraction unit 4 combines the cut sub-character patterns, and the recognition unit 5 recognizes the combined character pattern and creates a plurality of candidate character strings. The word processing unit 7 determines one word as a recognition result from candidate character strings created by collating the word dictionary 8.
[0004]
[Problems to be solved by the invention]
By the way, in the conventional apparatus, since character recognition is performed in consideration of the possibility of combining various sub-character patterns, it takes time to recognize characters. Furthermore, although a plurality of candidate character strings are collated with a word dictionary to determine one word, if the collation with the word dictionary can be performed, the determined word is output as a correct recognition result. The recognition accuracy is not necessarily high.
[0005]
In view of the above problems, an object of the present invention is to provide a character recognition device and a method for recognizing characters in a document in which a plurality of items are described accurately in a short time.
[0007]
[Means for Solving the Problems]
To solve the above problems The present invention is a character recognition device for recognizing characters of a document image, in which the target document image is composed of character string images corresponding to a plurality of description items. Character cutting means for cutting out a character image based on the height of the character or on the basis of the instruction, a recognition dictionary in which character characteristics are registered in the character type, and a character image cut out by the character cutting means Character candidate selection means for selecting a plurality of character candidates by collating with the entire range of the recognition dictionary or the designated range, and classifying words related to all the description items of the document image to be recognized for each attribute. The word dictionary registered in combination with the consecutive character candidates of the character image selected by the character candidate selection means and collated alone in the entire range of the word dictionary or in the designated range, Ah If there is no match, the word extraction unit extracts the character candidate as the word candidate as it is and the character candidate extracted by the word extraction unit as a keyword Description item classification means for classifying into description items in units of images, a related rule table that records a list of related attributes and description conditions of words that are description contents for each description item, and the entire range of the word dictionary by the word extraction means When it is determined that the word candidate obtained by collating with the relation candidate satisfies the contents of the description item of the relation rule table and the relation determination means does not satisfy the word candidate, An instruction means for giving a predetermined instruction to the output means, the character candidate selection means or the word extraction means, and the word candidate and the indicator determined to be satisfied by the association determination means Output means for outputting the word candidates extracted according to the instruction of the instruction, the instruction means misrecognizes the word candidates that the association determination means determines that the description content attribute or the like is not satisfied for each description item Instructing the character cutting means and the like. Therefore, recognition accuracy is improved.
[0008]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, a character recognition device according to the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram of an embodiment of a character recognition device according to the present invention. This character recognition device includes an image input unit 101, a character cutout unit 102, a recognition dictionary 103, a character recognition unit 104, a word dictionary 105, a word collation unit 106, an item classification unit 107, and an item related rule storage. Unit 108, reprocessing determination unit 109, and result output unit 110.
[0009]
The image input unit 101 is composed of a scanner or the like, and a pixel (black pixel) composed of binary data by performing L / E (Light / Electric) conversion or the like on a document whose description items prepared by an operator are determined in advance. And a white image) are notified to the character segmentation unit 102 as a document image. FIG. 2 shows a business card document image 201 notified to the character cutout unit 102.
[0010]
Upon receiving the notification of the document image from the image input unit 101, the character cutout unit 102 scans the document image, counts the number of black pixels in the vertical direction or the horizontal direction, and determines the character string from the distribution of black pixels for each scanning line. Extract images. A number is assigned to the extracted character string image, and the position, for example, the coordinates of the upper left corner point and the lower right corner point of the rectangle circumscribing the character string image are obtained, and the height of the character string image is calculated.
[0011]
Next, the character string image is scanned in a direction perpendicular to the direction in which the character string image exists, and a character image having the same width as the height of the character string image is cut out from the black pixel distribution. That is, a full-width character image is cut out from each character string image. A number is assigned to the clipped character image, and the character recognition unit 104 is notified together with the number.
3 shows character string images L1 to L5 extracted from the document image 201 shown in FIG. 2, and character images C1, C2,... Cut out for each character string image.
[0012]
When the character cutout unit 102 receives a notification of the character image cutout condition and the number (processing range) of the character image based on the content of the reprocessing condition table 801 described later from the reprocessing determination unit 109, the character cutout unit 102 A new character image is cut out using the character image in the range of the number notified in accordance with the image cut-out condition, for example, equal pitch or proportional pitch, as the processing range. The newly cut-out character image is the same as the above-described processing, and each character image is newly numbered and notified to the character recognition unit 104 together with its position and run information.
[0013]
The recognition dictionary 103 registers character codes and standard feature vectors of characters corresponding to the character codes. Here, the character code includes codes such as symbols in addition to codes of kanji, hiragana, katakana, numbers, and alphabets. In the recognition dictionary 103, character codes and the like are registered in character types such as kanji,.
When the character recognition unit 104 receives a notification of a character image or the like from the character cutout unit 102, the character recognition unit 104 extracts the feature vector, compares it with the standard feature vectors of all character codes in the recognition dictionary 103, and obtains, for example, the city distance A plurality of character codes corresponding to the standard feature vector having a small value are extracted from the top and stored as character candidates. When this character image recognition processing is performed for all character images of all character string images, the word collating unit 106 is activated.
[0014]
FIG. 4 is a diagram for explaining the result of recognition processing for the character image shown in FIG. The character images C1, C2,... Of the character string images L1 to L5 are shown in a column, and the top, first to third places of each character candidate are shown in a row. This character candidate is actually stored as a character code.
For example, the first character candidate of the character image C1 of the character string image L1 is “approval”, the second character candidate is “magazine”, and the third character candidate is “filled”.
[0015]
Further, when the character recognition unit 104 receives the notification of the character image recognition condition from the reprocessing determination unit 109, the character recognition unit 104 limits the condition of the recognition dictionary 103, for example, the character type to only numbers, English letters, numbers, symbols, etc. Similar to the above, a plurality of character codes are extracted and stored as character candidates in the collation range. When all the re-recognition processes for the re-cut character image are completed, the word collating unit 106 is activated.
[0016]
As shown in FIG. 5, the word dictionary 105 includes a word column 501 and an attribute column 502 that represents the overall contents of words registered in the word column 501, and a plurality of words are associated with each attribute. It is registered. In practice, words are indicated by character codes.
That is, the attribute representing the general content of the words “stock”, “finite”, “company”,... Is “company”, and the attribute “mail” includes “mail”, “MAIL”, “jp”. ",... Are included. The contents of the word dictionary 105 are changed depending on the document image to be recognized.
[0017]
In this embodiment, since “business card” is targeted as the document image 201, the contents of such a word dictionary 105 are obtained.
In addition, each word in the word column 501 included in the attribute column 502 from “company” to “mail” is used as a keyword for classifying a description item to be described later.
When activated by the character recognizing unit 104, the word collating unit 106 reads out consecutive character candidates from the first character candidate of each character string image extracted by the character recognizing unit 104 as word candidates, and the word column of the word dictionary 105 If there is a word that matches the registered word, it is determined as a word candidate and stored together with the attribute of the word. At this time, the word candidates are read out by giving priority to the upper character candidate, and when the word matching the first character candidate is not registered in the word dictionary 105, the first place, the second place, the first place, and the third place In this way, word candidates are read out by combining character candidates. When a word is not found even if the first, second and third character candidates are combined with the next character candidate, the first character candidate is determined as the word candidate, and the attribute is stored as “undefined”. When the word matching process is completed for all character candidates of all character string images, the item classification unit 107 is activated.
[0018]
FIG. 6 shows word candidates obtained by collating the word dictionary 105. In the case of the character string image L1 shown in FIG. 4, since the combination of the first character candidates of the character images C1 and C2 is registered as the word “recognition” in the attribute “general” of the word dictionary 105, the character image C1. , C2 is determined as the word candidate “recognition”. Subsequently, since the combination of the first character candidates of the character candidates C3 and C4 is registered as the word “Tsuyo” in the attribute “name”, the character images C3 and C4 are determined as the word candidate “Tsuyo”. Similarly, the character images C5 and C6 are determined as the word candidate “company”.
[0019]
Here, combining the 2nd and 3rd character candidates of the character images C3 and C4 matches the word “stock” of the attribute “company” in the word dictionary 105, but combines the higher character candidates. Is a word candidate.
In the case of the character images C3 and C4 of the character string image L2, the combination “Taro” of the first character candidates is not in the word dictionary 105, so the second character candidate “bold” of the character image C3 and the character image Since the word “Taro” combined with the C1 first character candidate is registered in the attribute “name” of the word dictionary 105, it is determined as a word candidate.
[0020]
Since the first to third character candidates of the character images C6 and C11 of the character string image L5 are not registered as words in the word dictionary 105, the first character candidates are attributed as word candidates “old” and “属性”. Are defined as “undefined”.
When the word collation unit 106 receives a notification of the number of the character image and the word dictionary collation condition from the reprocessing determination unit 109, the word collation unit 106 combines consecutive or single character candidates recognized by the character recognition unit 104. When the word candidate is determined, the collation range of the word dictionary 105 is determined again according to the notified attribute.
[0021]
That is, when the attributes “company” and “general” are notified as the word dictionary matching condition, “stock” (attribute) combining the second and third character candidates of the character images C3 and C4 of the character string image L1. "Company") is determined as a word candidate. At this time, “Tsuyo” of the word dictionary 105 included in the attribute “name” is excluded.
Thus, when a word candidate is re-determined, the re-determining determination unit 109 is notified of the determined word candidate together with the number of the character image.
[0022]
When activated by the word collating unit 106, the item classifying unit 107 classifies which description item is scheduled with the word candidate as a keyword in units of character string images stored in the word collating unit 106. In the character string image L1, since the word candidate “company” of the character images C5 and C6 is in the attribute “company” of the word dictionary 105, the description item is classified as “company”. At this time, even if an attempt is made to classify the name “name” by the word candidate “Tsuyo” of the attribute “name”, the remaining “recognition” and “company” are not classified because they are not in the attribute “name”. In the character string image L2, the description item is classified as “name” from the attribute “name” of the word candidates “business card” and “Taro”. The items described in the character string images L3, L4, and L5 are classified into “address”, “phone”, and “mail”, respectively. The classification of the description items is determined by referring to the layout of the character string image, but the description is omitted because it differs from the gist of the present invention.
[0023]
When the classification of the description items is completed for all the character string images, the reprocessing determination unit 109 is activated.
The item relation rule storage unit 108 stores a relation rule table 701 shown in FIG. 7 and a reprocessing condition table 801 shown in FIG.
The association rule table 701 includes an attribute 703 related to the description content for each description item 702 and a special condition 704 for the description content. From this association rule table 701, it can be seen that the related attributes of the item “company” include “company” and “general”, and other attributes are not included. That is, if a word candidate with another attribute is included, the character image is erroneously recognized.
[0024]
From the special condition 704, it can be seen that the character image described next to “〒” in the item “address” is “3 or 5 digit number”. It can also be seen that the last character string image of the entry item “telephone” is “four-digit number”.
The reprocessing condition table 801 includes a description item 802 and a reprocessing condition 803. The reprocessing condition 803 includes a processing range 804, a character image cutout condition 805, a character image recognition condition 806, and a word dictionary matching condition 807. Contains.
[0025]
For example, when the character string image is classified as the description item “company”, the re-processing is performed when the attribute of the word candidate determined by the word matching unit 106 of the character string image is not included in the related attribute 703. The processing range 804 of the condition 803 is defined as “a word candidate for a character image that does not satisfy a related attribute”, and the word dictionary matching condition 807 is defined as “limited to attribute company or general”. As a result, the collation range of the word dictionary 105 for determining word candidates in the word collating unit 106 is limited to words of the attributes “company” and “general”.
[0026]
If the character string image is classified as the entry item “telephone” and the attribute of the word candidate is not included in the related attribute 703, the processing range 804 of the reprocessing condition 803 is satisfied with the related attribute. And a word candidate other than “mail” connected to the word candidate, character image cutout condition 805 is defined as proportional pitch, and character recognition condition 806 is defined as limited to “number” and “symbol”. ing.
[0027]
When activated by the item classification unit 107, the reprocessing determination unit 109 reads the related rule table 701 stored in the item related rule storage unit 108. Next, whether the word candidate included in the description item classified for each character string image unit by the item classification unit 107 has an attribute of a word candidate contrary to the related attribute 703 in the related rule table 701, whether a special condition 704 exists. At a certain time, it is determined whether there is a word candidate that does not satisfy the condition. If any determination is negative, it is determined whether or not the determination of the reprocessing condition for all the character string images has been completed. If the determination has been completed, the result output unit 110 is activated. If not completed, the determination of whether or not the next character string image is contrary to the contents of the related rule table 701 is repeated.
[0028]
If any of the determinations are affirmative, there is an error in the recognition of the character string image. Therefore, the description item of the character string image classified by the item classification unit 107 is “mail”, “address” or “phone”. , “Company” or “Name”. When the description item is determined, the reprocessing condition 803 of the description item 802 is read from the reprocessing condition table 801 stored in the item-related rule storage unit 108.
[0029]
When it is determined to be “mail”, the word candidate that does not satisfy the related attribute, the character image number of the word candidate other than “mail” that is connected to the word candidate, and the character image cutout at the proportional pitch are cut out. The word recognition unit 103 is notified that the alphabet, number, and symbol of the recognition dictionary 103 are to be collated, and the word dictionary 106 includes words, letters, numbers, and symbols that have the word collation range. To the word collating unit 107.
[0030]
When it is determined that the description item is “address”, the character segmentation unit 102 is notified that the character image is to be segmented by “3” or “5” characters having the same pitch as the number of the character image of the word candidate that does not satisfy the special condition. The character recognition unit 104 is notified that the numbers in the recognition dictionary 103 are to be collated.
When it is determined that the description item is “telephone”, the character extraction unit 102 is notified that the character image number of the word candidate that does not satisfy the related attribute and the character image at the proportional pitch are to be extracted, The character recognition unit 104 is notified that the symbol is to be collated.
[0031]
When it is determined that the description item is “company”, the word collation unit 106 is notified that the character image number of the word candidate that does not satisfy the related attribute and the attribute of the word dictionary 105 are the company and the general word is to be collated. To do.
When it is determined that the description item is “name”, the word collation unit 106 is notified that the number of the character image of the word candidate that does not satisfy the related attribute and the attribute of the word dictionary 105 are the name and the word of the title. To do.
[0032]
For example, in the case of the character string image L1, the reprocessing determination unit 109 determines that the item to be described is “company” and recognizes the character images C3 and C4 as “Tsuyo” that does not satisfy the related attributes. Therefore, the word collating unit 106 is notified so that the collation range of the word dictionary 105 for the character candidates of the character images C3 and C4 is limited to the company and the general public. As a result, the word collation unit 106 combines the second character candidate “stock” of the character image C3 recognized by the character recognition unit 104 with the third character candidate “formula” of the character image C4. Re-recognize “stock” as a word candidate.
[0033]
In the case of the character image L3, the description item is determined to be “address”, and the character image C2 is recognized as the letter “M” contrary to the special condition (FIG. 6). The character image C2 (FIG. 9A) is cut out by the character cutout unit 102 as three character images C21, C22, and 23 as shown in FIG. 9B according to the character image cutout condition 805. As a result, the character recognition unit 104 collates with the recognition dictionary 103 and extracts character candidates as shown in FIG. The word collating unit 106 collates the word dictionary 105 with the character candidate, and determines the first character candidates of the character images C21, C22, and C23 as word candidates “1”, “2”, and “1”, respectively.
[0034]
In the case of the character string image L5, the description item is determined to be “mail”, and character images C4 to C11 other than the character images C1 to C3 are set as a reprocessing range of character recognition (FIG. 10A). The character cutout unit 102 cuts out character images C4 to C14 in accordance with the proportional pitch notified from the reprocessing determination unit 109, as shown in FIG. The character recognition unit 104 collates with the recognition dictionary 103 in the range of English sentences, numbers, and symbols notified from the reprocessing determination unit 109, and extracts character candidates as shown in FIG. As shown in FIG. 10D, the word collation unit 106 collates with the word dictionary 105 by combining character candidates with the attributes notified from the reprocessing determination unit 109 in the range of mail, alphabetic characters, numbers, and symbols. Determine word candidates.
[0035]
Since the combination “lp” of the first character candidate of the character images C13 and C14 does not exist in the attribute “mail”, the second character candidate “j” of the character image C13 and the first character candidate “p” of the character image C14. The word candidate “jp” is determined from the above.
When the reprocessing determination unit 109 receives the notification of the result of the reprocessing from the word matching unit 106, the reprocessing determination unit 109 sends the number of the character string image to the result output unit 110 together with the previous recognition result outside the processing range of the reprocessing, The recognition item and the character code that is the recognition result are notified.
[0036]
The result output unit 110 holds bitmap data for displaying a character code or the like on the display screen. When the result output unit 110 receives a notification of the character code or the like as a recognition result from the reprocessing determination unit 109, the result screen 110 is shown in FIG. The recognition result of the document image is displayed and the recognition result is stored.
Next, the operation of the present embodiment will be described with reference to the flowcharts of FIGS.
[0037]
The image input unit 101 receives an input of a document original, for example, “business card” from an operator, and converts it into a binarized document image (S1202).
The character cutout unit 102 extracts a character string image from the document image (S1204). A character image is cut out as a full-width character from the extracted full-character string image (S1206, S1208).
[0038]
Next, the character recognizing unit 104 extracts character candidates by collating character images cut out using all character types in the recognition dictionary 103 as a collation range, and repeats this collation for all character images of all character string images (S1210). , S1212).
The word matching unit 106 performs word matching on the obtained consecutive character candidates for words of all attributes in the word dictionary 105, and prioritizes combinations of higher-order character candidates to extract all word candidates and their attributes. It repeats until it completes about all the character images of a character string image (S1214, S1216). Since word matching can be performed if a matching word is found in the word dictionary 105 as a result of word matching by combining upper character candidates without combining all character candidates, the time for word matching is shortened. This recognition error of the word candidate can be corrected by determining the attribute or the like later.
[0039]
Next, the item classification unit 107 classifies the description items of the character string image with respect to all character string images using word candidates as keywords (S1218, S1220).
The re-recognition processing unit 109 determines whether or not the re-recognition has been completed for all the character string images (S1222). If the re-recognition processing unit 109 has not completed the re-recognition processing unit 109, the re-recognition processing unit 109 stores the related rule table 701 stored in the item-related rule storage unit 108. Whether or not the content is contrary is determined for each entry item (S1224). If not, the process returns to S1222, and the related attribute 703 in the related rule table 701 does not match the attribute of the word candidate, or the special condition 704 is violated. If so, the process proceeds to S1302.
[0040]
In step S1302, the reprocessing determination unit 109 determines whether the item described in the character string image is “mail” (S1302). If it is not “mail”, the process advances to step S1316. If it is “mail”, the character extraction unit 102 is notified of the number of the character image described in the “mail” processing range 804 of the description item 802 of the reprocessing condition table 801 ( S1304).
[0041]
The character cutout unit 102 cuts out the character image again according to the notified character image and the proportional pitch of the character image cutout condition 805 (S1306).
The character recognizing unit 104 recognizes a character candidate using the alphabet, number, and symbol of the character type in the recognition dictionary 103 of the character image recognition condition 806 as a collation range, and repeats the process until there is no character image cut out again (S1308, S1310). .
[0042]
Next, the word collation unit 106 determines a single character candidate as a word candidate by combining consecutive character candidates with the mail, English letters, numbers, and symbols of the attribute of the word dictionary 105 of the word dictionary collation condition 807 as a collation range. . This process is repeated until there are no re-recognized word candidates (S1312, S1314), and the process returns to S1222.
In step S1316, the reprocessing determination unit 109 determines whether the description item of the character string image is “address” or “phone” (S1316). If there is no item, the process proceeds to S1326. When “address” or “telephone”, the character cutout unit 102 is notified of the contents of the processing range 804 of the reprocessing condition table 801. When the described item is “address”, three or five characters of equal pitch are used. When the entry item is “telephone” as the image cutout condition 805, the proportional pitch character image cutout condition 805 is notified (S1318).
[0043]
The character cutout unit 102 cuts out the character image again according to the processing range and the cutout conditions notified from the reprocessing determination unit 109 (S1320).
The character recognition unit 104 re-cuts the character image by limiting the character type number of the recognition dictionary 103 to the collation range when “address” and limiting the character type to a number and the symbol as the collation range when “phone”. Is recognized as a character candidate, and it is repeated until the recognition of the recut character image is completed (S1322, S1324), and the process returns to S1222.
[0044]
In S1326, the word matching unit 106 defines the processing range 804 of the reprocessing condition table 801 when the description items are “company” and “name” until all word matching of the character candidates to be re-recognized is completed. When the word dictionary matching condition based on the combination of character candidates for the character image to be set is “Company”, the word attribute is determined to be “Company” and “General”. (S1328). Thereafter, the process returns to S1222.
[0045]
In this manner, after character recognition is performed by cutting out character images according to the rough character image cutout conditions without considering the written items in the first place, the written items corresponding to the written contents are cut out by classifying the written items and character candidates are extracted. Recognizing and determining word candidates having attributes according to the contents of description, the accuracy and efficiency of character recognition are dramatically improved.
When the reprocessing determination unit 109 determines in S1222 that the re-recognition of all the character string images has been completed, the result output unit 110 describes the word candidates recognized from the input document image for each character string image. It displays on a display screen with an item (S1226), and complete | finishes a process.
[0046]
In this embodiment, the character recognition apparatus according to the present invention is realized with the configuration shown in FIG. 1, but the present invention is realized by a program, which is recorded on a recording medium such as a floppy disk and transferred. By doing so, it can be easily implemented by another independent computer system. FIG. 14 is a diagram for explaining a case where this is implemented with a floppy disk.
[0047]
The physical format of the floppy disk 1401 which is a recording medium body is such that tracks 1, 2,..., 80 are created concentrically from the outer periphery to the inner periphery and divided into 16 sectors in the angular direction. The program is recorded according to the allocated area.
The floppy disk 1401 is housed in a case 1402, which protects the disk from dust and external impact and can be safely transported.
[0048]
FIG. 15 is a diagram for explaining the recording / reproduction of a program on the floppy disk 1401. By connecting a floppy disk drive 1502 to the computer system 1501 as shown, a program can be recorded on and reproduced from the disk 1401. The disk 1401 is inserted into and removed from the floppy disk drive 1502 via the insertion port 1503. In the case of recording, the program is recorded on the disk 1401 by the floppy disk drive 1502 from the computer system 1501. For reproduction, the floppy disk drive 1502 reads the program from the disk 1401 and transfers it to the computer system 1501.
[0050]
【The invention's effect】
As mentioned above, According to the present invention, the target document image is composed of character string images corresponding to a plurality of description items, and is a character recognition device for recognizing characters of such a document image. Based on the height of the row image or when an instruction is received, a character cutout means that cuts out a character image based on the instruction, a recognition dictionary in which character characteristics are registered in the character type, and the character cutout means Character candidate selection means for selecting a plurality of character candidates by collating the character image with the entire range of the recognition dictionary or the designated range, and words related to all the description items of the document image to be recognized for each attribute The word dictionary classified and registered and the character candidates of the continuous character images selected by the character candidate selection means are combined and matched in the whole range of the word dictionary or the designated range, and they match. single If there is a word candidate, it is extracted along with its attribute, and if there is no match, the attribute is undefined and the character candidate is extracted as it is as a word candidate, and the word candidate extracted by the word extraction means is used as a keyword. Description item classification means for classifying into description items in character string image units, a related rule table in which a list of related attributes and description conditions of words that are description contents for each description item is recorded, and the word extraction means When it is determined that the word candidate obtained by collating in the entire range satisfies the contents of the description item of the related rule table, and when it is determined not to be satisfied by the related determination unit, An instruction means for giving a predetermined instruction to the character extraction means, the character candidate selection means or the word extraction means; and the word candidates determined to be satisfied by the association determination means; Output means for outputting the word candidates extracted in accordance with the instructions of the indicating means, so that word candidates with attributes that do not contradict the contents of the description items are recognized as correct recognition, so that recognition is not performed again. The time required is shortened, and the word candidates having attributes that are contrary to the contents of the description items are re-recognized as erroneous recognition, so that the recognition accuracy is improved.
[0051]
Further, according to the present invention, when the instruction unit determines that the relation determination unit does not satisfy the instruction, the character extraction of the character string image including the character image of the word candidate is performed by the character extraction unit. Since it includes a condition change instruction unit that gives an instruction to change the condition, the character candidate collating condition in the character candidate selecting means, and the word candidate collating condition in the word extracting means to a predetermined condition, It is possible to correctly recognize word candidates that have been erroneously recognized.
[0052]
Further, according to the present invention, the instruction means records a processing range to be recognized for each description item, a character image clipping condition, a character candidate character type, and a word candidate attribute. A processing condition table; and a reading unit that reads out the processing condition of the character string image including the word candidate for each description item from the processing condition table when it is determined that the relation determination unit does not satisfy the condition determination unit. The change instructing unit includes a first condition instructing unit that instructs the character extracting unit to specify a character image included in the processing range read by the reading unit and a character image clipping condition, and the character candidate selecting unit to the reading unit. A second condition designating unit that designates a character type that limits the collation range of the recognition dictionary read by the user, and an attribute that designates an attribute that limits the collation range of the word dictionary read by the reading unit to the word extracting unit. Since the possible and a condition instructing section, pre-reprocessing conditions for each described item becomes clear, it increases the recognition efficiency of the document image.
[0053]
According to the present invention, when the word extracting unit first extracts a word candidate, the word candidate is given priority to a combination of upper character candidates among the plurality of character candidates selected by the character candidate selecting unit. When the word candidate is extracted with priority given to the collation range when the instruction is received from the third condition instruction unit, the time required for the recognition process can be shortened.
[0054]
According to the present invention, there is also provided a character recognition method for recognizing characters of a document image, in which the target document image is composed of character string images corresponding to a plurality of description items, and accepts input of the document image. A first character cutting step for cutting out a character image based on the height of the character string image, and the entire range of the recognition dictionary in which character features are registered for the character type of the character image cut out in the first character cutting step A document image to be recognized by combining the first character candidate selection step for selecting a plurality of character candidates by collating with the character candidates for the consecutive character images selected in the first character candidate selection step Words related to all the description items are collated in the entire range of the registered word dictionary classified by attribute, and if there is a matching word, it is extracted as a word candidate with the attribute, A first word extraction step for extracting a character candidate as a word candidate with the attribute undefined, and a description item for classifying the word candidate extracted in the first word extraction step into a description item for each character string image as a keyword. Whether the word candidates extracted in the classification step and the first word extraction step satisfy the content of the description item in the related rule table in which a list of related attributes and description conditions of the word as the description content is recorded for each description item An association determination step for determining whether or not, a first instruction step for instructing a character image cut-out condition when it is determined that the relation determination step is not satisfied, and a character image in accordance with the cut-out condition in the first instruction step The collation range of the recognition dictionary of the character image cut out by the second character cutting out step and the second character cutting out step A second instruction step to indicate, a second character candidate selection step for selecting a character candidate in accordance with the instruction in the second instruction step, and a collation range of the word dictionary of the character candidate selected in the second character candidate selection step A third instruction step that performs extraction, a second word candidate extraction step that extracts word candidates in accordance with instructions in the third instruction step, a word candidate that is determined to be satisfied in the association determination step, and an extraction in the second word candidate extraction step Therefore, it is possible to obtain the same effect as that of the character recognition device.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of an embodiment of a character recognition device according to the present invention.
FIG. 2 is a diagram illustrating an example of a document image converted by an image input unit according to the embodiment.
FIG. 3 is a diagram illustrating an example of a character image cut out by a character cutout unit according to the embodiment.
4 is a diagram for explaining character candidates recognized from the character image shown in FIG. 3 by the character recognition unit of the embodiment. FIG.
FIG. 5 is an explanatory diagram of an example of the word dictionary of the embodiment.
6 is a diagram for explaining word candidates and their attributes collated from the character candidates in FIG. 4 by the word collating unit of the embodiment. FIG.
FIG. 7 is a diagram showing the contents of a related rule table stored in the item related rule storage unit of the embodiment.
FIG. 8 is a diagram showing the contents of a reprocessing condition table stored in the item-related rule storage unit of the embodiment.
FIG. 9 is a diagram for explaining re-recognition processing in the character image C2 of the character string image L3 in the embodiment.
FIG. 10 is a diagram for explaining re-recognition processing in character images C4 to C11 of the character string image L5 in the embodiment.
FIG. 11 is a diagram illustrating an example of a recognition result of a document image output and displayed by the result output unit of the embodiment.
FIG. 12 is a flowchart illustrating the operation of the embodiment.
FIG. 13 is a flowchart illustrating the operation of the embodiment.
FIG. 14 is an explanatory diagram of a recording medium on which the character recognition method described in the above embodiment is recorded.
FIG. 15 is a diagram for explaining attachment of the recording medium to a computer system.
FIG. 16 is a configuration diagram of a conventional character recognition device.
[Explanation of symbols]
101 Image input unit
102 character cutout
103 recognition dictionary
104 Character recognition part
105 word dictionary
106 Word verification unit
107 Item classification
108 Item-related rule storage unit
109 Reprocessing determination unit
110 Result output section
1401 Floppy disk
1501 Computer system
1502 Floppy disk drive

Claims (5)

A target document image consists of character string images corresponding to a plurality of description items, and is a character recognition device for character recognition of such a document image,
A character cutout means for cutting out a character image based on the height of the character string image when receiving an input of a document image or based on the instruction when receiving an instruction;
A recognition dictionary in which character features are registered as character types;
Character candidate selection means for selecting a plurality of character candidates by collating the character image clipped by the character cutout means with the entire range of the recognition dictionary or the designated range;
A word dictionary in which words related to all description items of a document image to be recognized are classified and registered for each attribute;
Combining consecutive character candidates of the character image selected by the character candidate selecting means and collating them alone in the entire range of the word dictionary or the designated range, and if there is a matching word, it is extracted along with its attributes as a word candidate If there is no match, the word extraction means for extracting the character candidate as it is as the word candidate without defining the attribute,
A description item classification means for classifying the word candidates extracted by the word extraction means into a description item in character string image units as keywords;
A related rule table that records a list of related attributes and description conditions of words that are description contents for each description item,
Relevance determining means for determining whether or not word candidates obtained by collating with the entire range of the word dictionary by the word extracting means satisfy the contents of the description item of the related rule table;
An instruction unit that gives a predetermined instruction to the character extraction unit, the character candidate selection unit, or the word extraction unit when it is determined that the relation determination unit does not satisfy the condition determination unit;
A character recognition apparatus comprising: an output unit that outputs a word candidate determined to be satisfied by the association determination unit and a word candidate extracted according to an instruction from the instruction unit. The instruction means includes
When it is determined that the relation determination unit does not satisfy the condition, the character image cut-out condition in the character cut-out unit of the character string image including the character image of the word candidate and the character candidate selection unit in the character candidate selection unit and matching condition, the character recognition system according to claim 1, characterized in that it comprises a condition change instruction unit that gives an instruction to change the matching condition of word candidates in the word extraction means on a predetermined condition. The instruction means includes
A processing condition table that records processing conditions to be recognized for each description item, character image clipping conditions, character candidate character types, and word candidate attributes;
A reading unit that reads out the processing condition of the character string image including the word candidate for each description item from the processing condition table when it is determined not to be satisfied by the association determination unit;
The condition change instruction unit
A first condition instructing unit for instructing the character extracting unit to specify a character image included in the processing range read by the reading unit and a character image extracting condition;
A second condition instructing unit that instructs the character candidate selecting unit to specify a character type that limits a collation range of the recognition dictionary read by the reading unit;
3. The character recognition device according to claim 2, further comprising: a third condition indicating unit that instructs the word extracting unit to specify an attribute that limits a collation range of the word dictionary read by the reading unit. The word extracting means includes
When extracting word candidates for the first time, a word candidate is extracted by giving priority to a combination of upper character candidates among a plurality of character candidates selected by the character candidate selecting means,
Wherein when receiving an instruction from the third condition instructing section, the character recognition apparatus according to claim 3, wherein the extracting the word candidate in favor of the comparison range. A character recognition method in which a target document image consists of character string images corresponding to a plurality of description items, and character recognition of such a document image is performed,
A first character extraction step of extracting a character image based on the height of the character string image upon accepting an input of a document image;
A first character candidate selection step of selecting a plurality of character candidates by collating the character image extracted in the first character extraction step with the entire range of the recognition dictionary in which character characteristics are registered for the character type;
A word dictionary in which words related to all description items of a document image to be recognized are classified and registered for each attribute in combination with and independently of the consecutive character candidates selected in the first character candidate selection step. A first word extraction step for collating with all the ranges, extracting a word candidate with its attribute if there is a matching word, and extracting a character candidate as a word candidate with the attribute undefined if there is no matching word;
A description item classification step for classifying the word candidates extracted in the first word extraction step into description items in character string image units as keywords;
It is determined whether or not the word candidate extracted in the first word extraction step satisfies the content of the description item in the related rule table in which a list of related attributes and description conditions of the word as the description content is recorded for each description item. Relevance determination step,
A first instruction step for instructing a character image cut-out condition when it is determined that the relation determination step is not satisfied;
A second character cutting step of cutting out a character image in accordance with the cutting conditions in the first instruction step;
A second instruction step for instructing the collation range of the recognition dictionary of the character image extracted in the second character extraction step;
A second character candidate selection step of selecting a character candidate according to the instruction in the second instruction step;
A third instruction step for instructing a collation range of the word dictionary of the character candidates selected in the second character candidate selection step;
A second word candidate extraction step for extracting word candidates in accordance with the instructions in the third instruction step;
A character recognition method comprising: an output step of outputting as a recognition result the word candidate determined to be satisfied in the association determination step and the word candidate extracted in the second word candidate extraction step .

Images