JP2008077454A - Title extraction device, image reading device, title extraction method, and title extraction program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To precisely extract a character string having no exterior features on a document image as a title. <P>SOLUTION: A title extraction device includes an extraction requirement acquisition part 31 for acquiring a keyword character string described near a title character string on a document image and relative location information of the title character string corresponding to the keyword character string, a character recognition part 33 for recognizing each character in a region including the title character string and the keyword character string in the document image, a keyword retrieving part 34 for retrieving the keyword character string from the recognition result to acquire the location, a title location acquisition part 35 for acquiring a location of the title character string from the acquired location of the keyword character string and the relative location of the title character string corresponding to the keyword character string, and a title output part 36 for outputting title character string data from the acquired location of the title character string. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a title extraction apparatus, an image reading apparatus, a title extraction method, and a title extraction program that extract a title from a document image.

  If a title indicating the contents of a document can be automatically extracted from a document image acquired by reading a document, the document can be easily managed by using the title as a file name of the document. Therefore, convenience is improved.

In order to extract a title from a document image in this way, conventionally, the appearance feature is applied by focusing on the appearance feature that is given to make the title portion such as the position on the document and the underline and frame line stand out. There is known a technique for taking out the recorded portion as a title (see, for example, Patent Document 1).
Japanese Patent Laid-Open No. 9-134406

  However, in a document image, an appearance feature is not necessarily applied to an appropriate character string for identifying the document image. In the conventional technique, an appearance feature is applied to an appropriate character string as a title. If not, the character string cannot be extracted as a title.

  For example, in the case of the document image shown in FIG. 25, according to the conventional technique, the most characteristic character string “invoice” is extracted as the title character string. It is inappropriate as a title to identify. In this example, the character string “IV01234567” indicating the invoice number, which is document-specific information, is appropriate as the title character string. However, this character string has no significant characteristics and is extracted as a title in the conventional method. I can't.

  In addition, character recognition is used to extract a required character string from a document image as text data. However, this character recognition cannot avoid misrecognition, and this misrecognition reduces the accuracy of title extraction, or title extraction. It is desirable to configure so that inconveniences that cannot be avoided can be avoided.

  The present invention has been devised to solve such problems of the prior art, and its main purpose is to accurately extract even a character string having no external features on a document image as a title. It is an object of the present invention to provide a title extraction device, an image reading device, a title extraction method, and a title extraction program that are configured to be capable of performing the above. Another object of the present invention is to make it possible to avoid a reduction in title extraction accuracy and inability to extract due to erroneous recognition in character recognition.

  The present invention provides a keyword character string described in the vicinity of a title character string on a document image, an extraction condition acquisition unit that acquires relative position information of the title character string with respect to the keyword character string, Character recognition means for recognizing at least an area including the title character string and the keyword character string, and a keyword character string acquired by the extraction condition acquisition means is searched from the recognition result of the character recognition means. Based on the keyword search means for acquiring the position, the position of the keyword character string acquired by the keyword search means, and the relative position of the title character string with respect to the keyword character string acquired by the extraction condition acquisition means Title position acquisition means for acquiring the position of the column and title sentence acquired by the title position acquisition means Based on the position of the column, a configuration having a title output means for outputting the data of the title string.

  According to the present invention, since the position of the title character string is obtained with reference to the keyword character string described in the vicinity of the title character string, even a character string having no appearance feature can be reliably extracted as a title. Furthermore, by setting extraction conditions for each document format, it is possible to accurately extract titles from various document images.

  A title extraction device according to a first aspect of the present invention acquires a keyword character string described in the vicinity of a title character string on a document image, and relative position information of the title character string with respect to the keyword character string. Extraction condition acquisition means, character recognition means for recognizing at least an area including the title character string and the keyword character string in the document image, and the extraction condition acquisition means from the recognition result of the character recognition means The keyword search means for searching the keyword character string acquired in step (b) and acquiring the position thereof, the position of the keyword character string acquired by the keyword search means, and the title character string for the keyword character string acquired by the extraction condition acquisition means Title position acquisition means for acquiring the position of the title character string based on the relative position, and this title Based on the position of the title character string acquired by the position acquisition means, and has a title output means for outputting the data of the title string.

  According to this, since the position of the title character string is obtained with reference to the keyword character string described in the vicinity of the title character string, even a character string having no appearance feature can be reliably extracted as a title. By setting extraction conditions for each document format, it is possible to accurately extract titles from various document images.

  In this case, the extraction condition acquisition unit acquires information regarding the extraction condition input by the user's input operation using the input operation unit (keyboard or the like) of the own device or another device.

  The title extraction device according to a second aspect of the present invention is the title extraction device according to the first aspect, wherein the extraction condition acquisition unit includes position information of a title inclusion region including the title character string and the keyword character string. Is acquired by the user's input operation, and the character recognition means performs character recognition for the title inclusion area acquired by the extraction condition acquisition means.

  According to this, since the character recognition range is limited, the accuracy and processing speed of title extraction can be increased.

  The title extraction device according to a third aspect of the present invention is the title extraction device according to the first aspect, wherein the keyword search means includes a character string dividing means for dividing the keyword character string into a plurality of partial character strings; The partial character string obtained by the character string dividing means is searched for in the character recognition result character string of the character recognizing means, and the position of each partial character string is detected. The position of the keyword character string based on the position of the partial character string acquired by the partial character string position detecting means by comparing the word order based on the position of each partial character string acquired by the character string position detecting means with the normal word order Evaluation means for evaluating whether or not is appropriate.

  According to this, even when the recognition result of the character recognition means includes erroneous recognition, the position of the keyword character string can be acquired with high accuracy.

  In this case, the character recognition means obtains a plurality of candidate characters for each character of the recognition result character string, and the partial character string position detection means compares the plurality of candidate characters with respect to each character of the partial character string. A configuration for detecting the position of the character string is preferable. According to this, it is only necessary that the partial character string and the recognition result character string match with the lower candidate characters, and it is possible to avoid the disadvantage that the partial character string search unnecessarily fails due to erroneous recognition.

  Further, the character string dividing means may divide the keyword character string at a boundary where there is a relatively large space between each character. According to this, since the keyword character string is divided at a blank portion where misrecognition due to noise, in particular, illegal character insertion is likely to occur, the influence of misrecognition due to noise can be suppressed low.

  Also, the evaluation means obtains the interval between the partial character strings from the position of each partial character string for each candidate when a plurality of suitable candidates for the position of the keyword character string remain due to the evaluation of the partial character strings in the word order. Thus, it is preferable to select a candidate whose interval between the partial character strings is closest to the regular interval. According to this, a candidate can be narrowed down to the most appropriate one.

  The title extraction device according to a fourth aspect of the present invention is the title extraction device according to the first aspect, wherein the keyword search unit is configured to determine the difference between the comparison target character string in the recognition result character string and the keyword character string. Based on the recognition result character string, the first and second position detecting means for detecting the position of the keyword character string, and the first position detecting means determines the length of the comparison target character string as the keyword character. The approximate position of the keyword character string on the recognition result character string is detected based on the degree of difference obtained by setting the same as that of the column, and the second position detecting means is acquired by the first position detecting means. Based on the degree of difference obtained by limiting the comparison target within the comparison target range based on the approximate position of the keyword string, and increasing or decreasing the length of the comparison target string within the required range Key on the row And it detects detailed positions of de string.

  According to this, even when the recognition result of the character recognition means includes erroneous recognition, the position of the keyword character string can be acquired with high accuracy. In particular, in the method of obtaining the position of the keyword character string by dividing the keyword character string into partial character strings, when an illegal character is inserted in a portion corresponding to the partial character string in the recognition result character string due to noise, Although it is difficult to detect a keyword, this method enables keyword detection without being affected by character insertion.

  In the second position detecting means, the length of the comparison target character string is increased or decreased within a required range, so that the number of comparisons increases. However, the comparison target range, that is, the comparison target character string to be compared with the keyword character string By limiting the range from which the character string is extracted from the recognition result character string, the processing speed can be increased.

  In this case, the degree of difference is preferably the Levenshtein distance (edit distance) that represents how much the two character strings are different by quantifying them according to the number of character replacements, deletions, and insertions (number of characters). .

  In addition, the character recognition unit may rank and acquire a plurality of candidate characters for each character of the recognition result character string, and the position detection unit may obtain a degree of difference according to the rank of matching candidate characters. Here, the degree of difference when matching with the lower candidate character is set higher than when matching with the highest candidate character and lower than in the case of character replacement, and further increases as the rank of the candidate character is lowered. Is set as follows. Thereby, the determination according to the recognition reliability of a candidate character can be performed.

  The comparison target range may be set by adding an expansion width according to the Levenshtein distance acquired by the first position detecting means before and after the approximate position of the keyword character string. According to this, the comparison target range can be set to a necessary and sufficient range in which the keyword character string is assumed to exist.

  Also, the minimum length of the comparison target string is obtained by subtracting the Levenshtein distance from the length of the keyword string, and the maximum length of the comparison target string is calculated by subtracting the length of the keyword string. It is recommended to add the Stein distance. According to this, the length of the comparison target character string can be increased or decreased within a necessary and sufficient range that is assumed to vary due to the deletion or insertion of characters due to erroneous recognition.

  The title extraction device according to a fifth aspect of the present invention is the title extraction device according to the first aspect, wherein the title output means is a text of the title character string in the recognition result character string acquired by the character recognition means. Data is output.

  According to this, it is easy to manage the document by adding the text data of the acquired title character string as a title name or the like to the document file storing the document image or registering it as a keyword or the like in the electronic filing system. Can be.

  A title extraction device according to a sixth aspect of the present invention is the title extraction device according to the first aspect, wherein the title output means extracts image data created by cutting out a title character string region in a document image. Output.

  According to this, it is possible to avoid outputting an incorrect title character string due to erroneous recognition in character recognition.

  An image reading apparatus according to a seventh aspect of the present invention includes the title extraction apparatus according to any one of the first to sixth aspects.

  According to this, since the position of the title character string is obtained with reference to the keyword character string described in the vicinity of the title character string, even a character string having no appearance feature can be reliably extracted as a title. By setting extraction conditions for each document format, it is possible to accurately extract titles from various document images.

  Since the title of the read document can be acquired, management of the digitized document is facilitated, and convenience is improved.

  In this case, in addition to the title extraction device, the image reading apparatus controls the title extraction apparatus to perform an image reading unit that reads an image of a document and a title extraction process for a document image acquired by the image reading unit. Means.

  A title extraction method according to an eighth aspect of the present invention acquires a keyword character string described in the vicinity of a title character string on a document image and relative position information of the title character string with respect to the keyword character string. An extraction condition acquisition step; a character recognition step for performing character recognition on an area including at least the title character string and the keyword character string in the document image; and from the recognition result of the character recognition step, the extraction condition acquisition step The keyword search step of searching for the keyword character string acquired in step (b) and acquiring the position thereof, the position of the keyword character string acquired in the keyword search step, and the title character string for the keyword character string acquired in the extraction condition acquisition step Tight to get title string position based on relative position A position obtaining step, based on the position of the title character string acquired by the title position acquisition step, and has a title output step of outputting the data of the title string.

  According to this, since the position of the title character string is obtained with reference to the keyword character string described in the vicinity of the title character string, even a character string having no appearance feature can be reliably extracted as a title. By setting extraction conditions for each document format, it is possible to accurately extract titles from various document images.

  A title extraction program according to a ninth aspect of the present invention causes a computer to execute each step in the title extraction method according to the eighth aspect.

  According to this, since the position of the title character string is obtained with reference to the keyword character string described in the vicinity of the title character string, even a character string having no appearance feature can be reliably extracted as a title. By setting extraction conditions for each document format, it is possible to accurately extract titles from various document images.

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

  FIG. 1 is a block diagram showing a schematic configuration of an image processing apparatus to which the present invention is applied. The image processing apparatus includes an image input unit 1 that extracts image data of a document image from a document file acquired from an external image reading device and the like, and an image memory 2 that stores image data of the document image acquired by the image input unit 1. A title extraction unit (title extraction device) 3 that extracts a title from the document image, and a control unit 4 that controls the operation of each unit.

  This image processing apparatus can be realized by introducing an application program for title extraction that operates on a predetermined OS into an information processing apparatus such as a PC. The application program for title extraction is provided by being stored in a computer-readable storage medium such as a CD-ROM.

  The title extracted by the title extraction unit 3 is sent to an electronic filing system via a required interface (LAN interface or the like) 5 together with image data of a document image, for example, and registered therein as a search keyword or the like.

  FIG. 2 is a block diagram showing a schematic configuration of an image reading apparatus to which the present invention is applied. The image reading apparatus optically reads an image of a document recorded on a recording medium such as paper and generates image data, and the image data of the document image acquired by the image reading unit 7. It has an image memory 2 to be stored, a title extraction unit (title extraction device) 3 that extracts titles from document images, and a control unit 8 that controls the operation of these units.

  The title extracted by the title extraction unit 3 is added as a title name or the like to a document file storing image data of a document image. The document file to which the title is added is a required interface (LAN interface or the like) 9. Or to a PC or the like via a recording medium.

  FIG. 3 shows an outline of processing in the title extraction unit 3 shown in FIGS. In various tasks, a large number of documents to be created are standardized, and in the document format, a specific character string that describes the contents of the character string is described near the appropriate character string as a title for identifying the document. In the example of the invoice shown in the figure, the character string “IV01234567” indicating the invoice number is appropriate as a title for identifying the document, and the character string “invoice number:” written on the left side of the character string is invoiced. It must be written in the document format.

  Therefore, the title extraction unit 3 specifies the position of the character string “invoice number:” as the keyword character string, and then determines the position of the title character string “IV01234567” from the relative positional relationship with the keyword character string. Identify and extract the title string.

  Further, here, a title inclusion area 21 including a title character string and a keyword character string in the document image is designated, and character recognition is performed on the title inclusion area 21. In a document created based on a predetermined document format, the area where the title character string and the keyword character string exist is determined, and by limiting the character recognition range here, the accuracy and processing speed of title extraction can be improved. Can do.

  In this example, in order to specify the position of the title inclusion area 21 on the document image, the area start position of the title inclusion area 21, that is, the coordinates (x, y) of the upper left corner of the title inclusion area 21, the width W of the area, Height H is specified.

  The title inclusion area 21 only needs to include at least both the title character string and the keyword character string. If the image expands or contracts, tilts, or misaligns due to the performance of the image reading apparatus, the title inclusion area 21 It is necessary to enlarge and set the title inclusion area 21 appropriately so that the character string and the keyword character string do not deviate from the title inclusion area 21. It is also possible to perform character recognition for the entire document image without setting the title inclusion area.

  FIG. 4 shows an example of a method for specifying the relative positional relationship of the title character string with respect to the keyword character string shown in FIG. FIG. 4A is an example in which the relative positional relationship of the title character string with respect to the keyword character string is designated by the direction. Here, since the title character string “IV01234567” is present on the right side of the keyword character string “invoice number:”, the information indicating the relative position of the title character string is “right side”.

  FIG. 4B is an example in which the relative positional relationship of the title character string with respect to the keyword character string is designated by coordinates. In this example, the coordinates of the representative point of the rectangular area 23 circumscribing the title character string “IV01234567”, where the center of the rectangular area 22 circumscribing the keyword character string “invoice number:” is a reference point, here the reference point is the origin The relative positional relationship is specified by the coordinates (x0, y0) of the upper left vertex P of the rectangular area 23 and the width w0 and height h0 of the rectangular area 23 when O (0, 0) is set. In this case, a required numerical value may be acquired by actually measuring a dimension with a document or the like that is a source of a document image, and the numerical value may be input using a keyboard or the like.

  FIG. 5 is a block diagram showing the title extraction unit 3 shown in FIGS. The title extraction unit 3 includes an extraction condition acquisition unit (extraction condition acquisition unit) that acquires a keyword character string described in the vicinity of the title character string on the document image and relative position information of the title character string with respect to the keyword character string. ) 31, an extraction condition storage unit 32 that stores the extraction conditions acquired by the extraction condition acquisition unit 31, and a character recognition unit (character recognition means) that acquires character information by analyzing an image of a predetermined character recognition target area 33, a keyword search unit (keyword search means) 34 for searching the keyword character string extracted from the extraction condition storage unit 32 from the recognition result character string acquired by the character recognition unit 33 and acquiring the position thereof, and the keyword The position of the keyword character string acquired by the search unit 34 and the matching of the title character string with respect to the keyword character string extracted from the extraction condition storage unit 32 Title position acquisition unit (title position acquisition means) 35 for acquiring the position of the title character string based on the specific position, and the position of the title character string based on the position of the title character string acquired by the title position acquisition unit 35 And a title output unit (title output means) 36 for outputting data.

  The extraction condition acquisition unit 31 acquires an extraction condition necessary for the title extraction process by a user's input operation, and the keyword character string and the position information of the title character string relative to the keyword character string. In addition, information on the title inclusion area including the title character string and the keyword character string in the document image is acquired by the user's input operation. The user's input operation is performed using a keyboard or the like in the case of the image processing apparatus including the PC shown in FIG. 1, and the operation panel of the own apparatus is used in the case of the image reading apparatus shown in FIG. It may be performed using a PC keyboard connected via a network or the like.

  The character recognition unit 33 specifies a character by a known analysis method such as pattern matching and outputs text data of the character as a recognition result. For this text data, a general character code such as ASCII, Shift-JIS, UNICODE or the like may be used.

  FIG. 6 shows an outline of character recognition performed by the character recognition unit 33 shown in FIG. In the character recognition unit 33, as shown in FIG. 6A, the position on the document image for each character, for example, the position on the document image of the rectangular area circumscribing each character, that is, the start coordinate ( x, y), information on the width w and height h of the rectangular area is acquired.

  Further, in the character recognition unit 33, as shown in FIG. 6B, recognition reliability is given to the candidate characters for each character of the recognition result character string, and the candidate characters are ranked in descending order of the recognition reliability. Then, candidate characters up to a predetermined rank (for example, the third rank) are output.

  Furthermore, as shown in FIG. 6C, the character recognition unit 33 outputs the recognition result so that it can be divided in units of lines. When an image of a plurality of lines including a line in which the title character string exists is acquired as a recognition target due to a position shift in the title inclusion area 21 illustrated in FIG. 3 due to skew or the like at the time of image reading. The character recognition is performed for all the lines in the line, and the recognition result character string is taken out line by line based on the information indicating the line break. In this example, the second line as well as the first line in which the keyword character string “invoice number:” and the title character string “IV01234567” exist are recognized, and a recognition result character string for each line is output.

  Next, the keyword search unit 34, title position acquisition unit 35, and title output unit 36 shown in FIG. 5 will be described. First, the keyword search unit 34 performs keyword search using one or both of the two methods of the word order determination method and the Levenshtein distance method, details of which will be described later. Further, as shown in FIG. 4, the title position acquisition unit 35 specifies the position of the title character string on the recognition result character string acquired by the character recognition unit 33 from the relative positional relationship with the keyword character string.

  The title output unit 36 acquires the text data of the title character string from the recognition result character string based on the position of the title character string acquired by the title position acquisition unit 35 and outputs it. In the example of FIG. 6C, the character string from the character next to the keyword character string to the end of the line in the recognition result character string on the first line where the keyword character string exists is the title character string. It should be noted that the title character string is specified by comparison with the number of characters of the title character string and the format, for example, in the example of FIG. Also good.

  The title output unit 36 may output the title character string as image data. Also in this case, as shown in FIG. 4, the position of the title character string area on the document image is specified from the relative positional relationship with the keyword character string, and the title character string in the image data of the document image is identified. Cut out the area and get the image data.

  FIG. 7 is a flowchart showing a procedure of processes performed in the image processing apparatus and the image reading apparatus shown in FIGS. First, the extraction conditions, that is, the keyword character string, the relative position information of the title character string, and the position information of the title inclusion area are acquired by the extraction condition acquisition unit 31 of the title extraction unit 3 by the user's input operation. Is stored in the extraction condition storage unit 32 (step 101).

  Then, a document image is acquired, that is, if the image processing apparatus shown in FIG. 1 is used, the image input unit 1 extracts the image data of the document image from the document file. If the image reading apparatus shown in FIG. A process of reading an image of a document is performed by the unit 7 (step 102), and image data of a document image is stored in the image memory 2.

  Next, based on the position information of the title inclusion area extracted from the extraction condition storage unit 32, the character recognition unit 33 performs processing for extracting the image of the title inclusion area from the document image read from the image memory 2 and recognizing the character ( Step 103) Next, the keyword search unit 34 performs processing for searching the keyword character string extracted from the extraction condition storage unit 32 from the recognition result character string acquired by the character recognition unit 33 and acquiring the position (Step 103). 104).

  Then, when the keyword character string position can be specified by the keyword search in the keyword search unit 34 (step 105), information on the acquired keyword character string position and the title extracted from the extraction condition storage unit 32 Based on the relative position information of the character string, processing for acquiring the position of the title character string is performed in the title position acquisition unit 35 (step 106), and then the title character string based on the position of the title character string acquired here is obtained. Data is output from the title output unit 36 (step 107).

  On the other hand, when the position of the keyword character string cannot be specified by the keyword search in the keyword search unit 34 (step 105), it is determined that the title extraction has failed (step 108), and the process ends.

  FIG. 8 shows an example of erroneous recognition caused by character recognition in the character recognition unit 33 shown in FIG. In the example of FIG. 8A, an illegal character “,” is inserted into the third and seventh recognition result character strings due to noise in the image data. In the example of FIG. 8B, the first and second recognition result character strings are replaced with illegal characters “da” and “晝”. In the example of FIG. 8C, each misrecognition of the insertion of FIG. 8A and the replacement of FIG. 8B occurs in combination.

  In the example of FIG. 8D, as in FIG. 8A, the second illegal character “,” is inserted in the recognition result character string. An invalid character is inserted in.

  Such misrecognition can be reduced to some extent by performing preprocessing such as noise removal on the image data, but it cannot be completely eliminated. Even if such misrecognition occurs, the recognition result characters In order to accurately obtain the position of the keyword character string on the column, keyword search is performed using one or both of the following two methods: the method based on word order determination and the method based on Levenshtein distance. Is called.

  FIG. 9 is a block diagram showing the keyword search unit 34 shown in FIG. The keyword search unit 34 includes a character string dividing unit (character string dividing unit) 41 that divides the keyword character string into a plurality of partial character strings, and character recognition units that obtain the partial character strings obtained by the character string dividing unit 41. A partial character string position detecting unit (partial character string position detecting means) 42 for detecting the position of each partial character string by searching in the recognition result character string acquired in 33, and an evaluation value for obtaining an evaluation value of the partial character string A candidate registration unit 44 in which the position of each partial character string acquired by the calculation unit 43, the partial character string position detection unit 42 and the evaluation value acquired by the evaluation value calculation unit 43 are registered, and the partial character string position detection unit 42 The word order based on the position of the acquired partial character string is compared with the normal word order to evaluate whether the position of the keyword character string based on the position of the partial character string acquired by the partial character string position detection unit 42 is appropriate. Candidate evaluation unit (evaluation means) 45 The has.

  FIG. 10 shows an example of character string division performed by the character string division unit 41 shown in FIG. In the character recognition by the character recognition unit 33, misrecognition due to noise is likely to occur in a portion where there is a relatively large blank between each character, and the keyword is separated from the blank portion so that the influence of this misrecognition is reduced. The string is split. In this example, the keyword character string “invoice number:” is divided into partial character strings “invoice”, “number”, and “:” with a blank portion as a boundary.

  FIG. 11 shows the point of partial character string position detection performed by the partial character string position detection unit 42 shown in FIG. The partial character string position detection unit 42 compares the recognition result character string acquired by the character recognition for the title inclusion region with the partial character string, and at this time, the comparison position of the partial character string with respect to the recognition result character string is set to the beginning of the line. While shifting from character to character toward the end of the line, it is determined whether or not a partial character string exists at the comparison position, and the position of the partial character string on the recognition result character string is specified.

  The presence / absence of a partial character string is determined by comparing each character in order from the first character of the partial character string with the candidate characters in the corresponding recognition result character string. If not, the process moves to comparison with the next candidate character. If there is a match among the candidate characters, the process proceeds to the comparison of the next character of the partial character string with the candidate character, the comparison with the candidate character is performed for all characters in the partial character string, and the partial character If there is a match among all candidate characters for all characters in the column, it is determined that a partial character string exists at the comparison position.

  On the other hand, if there is no match among the candidate characters, it is determined that there is no partial character string at the comparison position, and the recognition result character string and the partial character string are compared at the comparison position shifted by one character toward the end of the line. When it is determined that the partial character string does not exist at all of the possible comparison positions, the keyword search is terminated assuming that the partial character string does not exist in the recognition result character string.

  In the evaluation value calculation unit 43, if there is a match among the candidate characters for all characters in the partial character string, the recognition reliability assigned by the character recognition unit 33 for each of the matched candidate characters Is added to obtain the evaluation value of the partial character string.

  In this example, the candidate characters are set up to the third place, but other modes are possible.

  FIG. 12 shows information registered in the candidate registration unit 44 shown in FIG. The candidate registration unit 44 includes the position (start position and end position) of each partial character string acquired by the partial character string position detection unit 42 and the evaluation value for each partial character string acquired by the evaluation value calculation unit 43. be registered. When the recognition result character strings of a plurality of lines are acquired by character recognition for the title inclusion area in the character recognition unit 33, the parts shown in FIGS. 12A to 12C with each line as one candidate A table of character string positions and evaluation values is created for each line, and evaluation by the candidate evaluation unit 45 is performed on a line-by-line basis.

  The example of FIG. 12A is a case where the recognition result character string shown in FIG. 11 is targeted. In this case, for example, in the partial character string “invoice”, the reliability of the first candidate that matches the first character “confirmation” is 1.0, and the reliability of the second candidate that matches the second character “request”. The degree of reliability is 0.8, the reliability of the third candidate that matches the third character “Cho” is 0.7, and the evaluation value of the partial character string “Bill” is the recognition reliability for each character. Is added to 2.5.

  In the example of FIG. 12B, as shown in FIG. 6C, the first character line in which the title character string exists due to the positional displacement in the title inclusion area due to skew or the like at the time of image reading. In addition, when the second line is also recognized, a keyword search is performed on the recognition result character string of the second line, and the character string “invoice” does not exist in the recognition result character string. .

  The example of FIG. 12C is a case of erroneous recognition in which an illegal character is inserted as shown in FIGS. 8A and 8C, and the second character string “number” and 3 The position of the second character string “:” is shifted.

  Next, the candidate evaluation unit 45 shown in FIG. 9 will be described. In this candidate evaluation unit 45, for each candidate registered in the candidate registration unit 44, the word order is obtained from the position of each partial character string, and this word order is compared with the normal word order to determine whether it is appropriate as the keyword character string position. It is determined whether or not. In the example of FIG. 11, it is determined whether or not the partial character strings “invoice”, “number”, and “:” are arranged in this order. If no candidate registered in the candidate registration unit 44 has the correct word order, the keyword search has failed.

  Further, in the candidate evaluation unit 45, for each candidate registered in the candidate registration unit 44, the interval between the partial character strings, that is, the number of characters or spaces existing between the partial character strings is obtained from the position of each partial character string. The one whose interval between the partial character strings is closest to the regular interval is selected as the final candidate. The selection based on the partial character string interval is performed when a plurality of candidates are determined to be appropriate by the above-described word order determination, whereby the final candidates can be narrowed down to one.

  In the example of FIG. 11, the interval (number of characters) between the end position of the first partial character string “invoice” and the start position of the second partial character string “number”, and the second partial character string “number” An interval (number of characters) between the end position and the start position of the third partial character string “No.” is calculated, and these intervals (number of characters) are closest to the regular interval, that is, the partial character string “invoice” and The difference between the “number” and the regular interval 1 is selected as the final candidate, and the difference between the partial character strings “number” and “:” between the regular interval 0 and the minimum is selected as the final candidate. .

  Further, when the evaluation value acquired by the evaluation value calculation unit 43 is abnormally small or when the interval (number of characters) between the partial character strings is abnormally large, the position of the keyword character string is inappropriate, and the candidate evaluation unit 45 The validity of the position of the keyword character string is determined by comparing the evaluation value and the partial character string interval with a predetermined threshold. If it is determined that the keyword character string is not valid, the keyword search is assumed to have failed.

  FIG. 13 is a flowchart showing a keyword search procedure performed by the keyword search unit 34 shown in FIG. This is executed by the keyword search (step 104) shown in FIG.

  Here, the character string dividing unit 41 first reads the keyword character string from the extraction condition storage unit 32 and divides it into partial character strings (step 201). Then, one partial character string is extracted from the partial character strings acquired here (step 202). As shown in FIG. 11, the recognition result character string and partial characters of the title inclusion area acquired by the character recognition unit 33 are obtained. The partial character string position detection unit 42 performs processing for comparing the columns and detecting the position of each partial character string (step 203). When a partial character string is detected in the recognition result character string (step 204), the start position and end position of the partial character string are registered in the candidate registration unit 44 (step 205). The evaluation value calculation unit 43 calculates the evaluation value of the partial character string, and the evaluation value is registered in the candidate registration unit 44.

  Next, it is determined whether or not all the partial character string comparison processes have been completed (step 206). If not, the process proceeds to the next partial character string comparison process, and the comparison process for all the partial character strings is performed. If completed, the candidate evaluation unit 45 determines whether the candidate is registered in the candidate registration unit 44, that is, whether the candidate is an appropriate candidate based on the word order (step 207). Here, when there are a plurality of lines of recognition result character strings, the above processing (steps 202 to 206) is repeated for each line, and when there are a plurality of candidates determined to be appropriate in the word order determination, The candidate evaluation unit 45 performs a process of selecting a final candidate by narrowing down based on the character string interval.

  Then, a determination is made as to whether or not a keyword position has been detected (step 208). If a keyword position is detected, information on the keyword position is output (step 209), while no keyword position is detected. If it is found that the keyword search has failed (step 210), the process ends.

  FIG. 14 is a block diagram showing another example of the keyword search unit 34 shown in FIG. Here, the keyword search unit 34 detects the approximate position of the keyword character string on the recognition result character string based on the Levenshtein distance (difference) between the comparison target character string and the keyword character string in the recognition result character string. An approximate position detector (first position detector) 51 and a detailed position detector (second position detector) 52 are provided.

  FIG. 15 shows a procedure for calculating the Levenshtein distance shown in FIG. The Levenshtein distance is a numerical representation of how different two character strings are. For a character string that serves as a reference for comparison, the conditions for replacement, deletion, and insertion are as follows: When this is the case, the Levenshtein distance is 1, and the Levenshtein distance is 0 if the two character strings completely match.

  In the comparison character string 1, the Levenshtein distance is 0 because it matches the correct character string “ABCDEF” that is a reference for comparison. In the comparison character string 2, “C” is replaced by “X”, in the comparison character string 3, “C” is deleted, and in the comparison character string 4, “C” and “D” are “ X "is inserted, and in the comparison character strings 2 to 4, the Levenshtein distance is 1 because the replacement, deletion, and insertion conditions of one character are applicable.

  In addition, when each of the replacement, deletion, and insertion conditions is applied multiple times (multiple characters), or when the replacement, deletion, and insertion conditions are combined and applied, depending on the number of times (number of characters) 1 is added to the value of the Levenshtein distance. In the comparison character string 5, “B” is deleted, E is replaced with “Y”, and “X” is inserted between “C” and “D”. The Levenshtein distance is 3 because of the replacement and the insertion of one character.

  FIG. 16 is a flowchart showing a keyword search procedure performed by the keyword search unit 34 shown in FIG. This is executed by the keyword search (step 104) shown in FIG.

  Here, the process of detecting the approximate position of the keyword is first performed by the approximate position detection unit 51 (step 301), and the keyword approximate position candidate obtained here is registered in the candidate registration unit 56. Next, a process for detecting the detailed position of the keyword based on the keyword approximate position candidates registered in the candidate registration unit 56 is performed by the detailed position detection unit 52 (step 302).

  Then, it is determined whether or not the keyword position has been detected (step 303). If the keyword position is detected, information on the keyword position is output (step 304), while the keyword position is detected. If not, it is determined that the keyword search has failed (step 305), and the process ends.

  Next, the approximate position detector 51 shown in FIG. 14 will be described. The approximate position detection unit 51 sets the length of the comparison target character string in the recognition result character string to be compared with the keyword character string to be the same as the keyword character string and shifts the comparison position while shifting the comparison position. A distance calculation unit 53 that calculates a Levenshtein distance (dissimilarity) with a column, a provisional candidate registration unit 54 in which the Levenstein distance for each comparison position acquired by the distance calculation unit 53 is registered as a provisional candidate, and this provisional A candidate evaluation unit 55 that selects a candidate with a minimum Levenshtein distance as a candidate for a keyword approximate position among provisional candidates registered in the candidate registration unit 54. The keyword position acquired by the candidate evaluation unit 55 Are registered in the candidate registration unit 56.

  FIG. 17 shows the outline of keyword approximate position detection performed by the approximate position detection unit 51 shown in FIG. The distance calculation unit 53 of the approximate position detection unit 51 compares the recognition result character string acquired by the character recognition for the title inclusion region with the keyword character string, and at this time, compares the keyword character string with the recognition result character string. While shifting the position one character at a time from the beginning of the line to the end of the line, the Levenshtein distance between the comparison target character string and the keyword character string in the recognition result character string at the comparison position is obtained.

  In this example, the length (number of characters) of the keyword character string “invoice number:” is 7 characters when the space is counted as one character. First, a character string of 7 characters from the first character of the recognition result character string. Is compared with the keyword character string as a comparison target character string. Next, a character string of 7 characters from the second character of the recognition result character string is compared with the keyword character string as a comparison target character string, and thereafter, the comparison between the comparison target character string and the keyword character string is performed as a comparison target character string. Is repeated until the end of the line of the recognition result character string is reached.

  Furthermore, if one character in the comparison target character string is replaced with a different character, that is, if there is no character that matches the keyword character string in the first candidate, it is compared with a lower candidate. A value corresponding to the character order is added to the Levenshtein distance. The value added to the Levenshtein distance when matching with the lower candidate character is set to be higher than (0) when matching with the highest candidate character and lower than (1) in the case of character replacement. Is set to increase as the value decreases.

  Here, when compared with the candidate characters up to the third place and matches the second or third place candidate characters, the values added to the Levenshtein distance are 0.1 and 0.2, respectively. If there is no matching character by the third candidate, 1 is added to the Levenshtein distance as described above.

  In the example of FIG. 17, when a character string for 7 characters from the first character of the recognition result character string is compared with the keyword character string as a comparison target character string, Since “Request” matches with the second candidate and “Call” matches with the third candidate, the Levenshtein distance is 0.1 + 0.2 = 0.3.

  FIG. 18 shows an example of a Levenstein distance calculation result in the distance calculation unit 53 shown in FIG. This is an example in which characters are replaced by misrecognition as in the examples of FIGS. 17 and 8B. In the candidate evaluation unit 55, a candidate having the minimum Levenshtein distance is selected as a candidate, and if there are a plurality of candidates having a minimum Levenshtein distance, all of them are selected as candidates. In this example, since the Levenshtein distance is minimum at 0.3 at the comparison position where the starting point is 0 as described above, this is selected as a candidate.

  FIG. 19 shows another example of the Levenstein distance calculation result in the distance calculation unit 53 shown in FIG. This is an example in which an invalid character “,” is inserted into the recognition result character string due to noise, as in the example of FIG. 8D. In this example, the Levenshtein is at the comparison position where the starting point is 0. Since the Levenshtein distance is 2 even at the comparison position where the distance is 2 and the starting point is 1, both of these are selected as candidates because the Levenshtein distance is the minimum.

  FIG. 20 is a flowchart showing a procedure for keyword approximate position detection performed by the approximate position detection unit 51 shown in FIG. This is executed by the keyword approximate position detection (step 301) shown in FIG.

  Here, the length (number of characters) of the keyword character string is first counted (step 401), and then a comparison target character having a length corresponding to the keyword character string from the recognition result character string acquired by the character recognition for the title inclusion area. The column is extracted (step 402), and the distance calculating unit 53 performs processing for calculating the Levenshtein distance between the comparison target character string and the keyword character string (step 403), and the result is provisional candidate registration. Registered in the unit 54. Here, the initial value of the comparison position of the keyword character string with respect to the recognition result character string is the beginning of the line of the recognition result character string.

  Next, it is determined whether or not the calculation of the Levenshtein distance at all the comparison positions has been completed (step 404). If not, the character is moved by one character in the line end direction as shown in FIG. Proceeding to the calculation of the Levenshtein distance at the comparison position, and when the calculation of the Levenshtein distance at all the comparison positions is completed, the value is selected from the Levenshtein distance for each comparison position registered in the provisional candidate registration unit 54. The candidate evaluation unit 55 performs a process of selecting the smallest candidate for the keyword approximate position (step 405), and the candidate is registered in the candidate registration unit 56.

  Next, the detailed position detection unit 52 shown in FIG. 14 will be described. The detailed position detection unit 52 includes a comparison target range setting unit 57 that sets a comparison target range in the recognition result character string based on the candidate keyword approximate position and the Levenshtein distance registered in the candidate registration unit 56, and a candidate A comparison target character string length setting unit 58 for setting a length range of the comparison target character string in the recognition result character string to be compared with the keyword character string based on the candidate Levenshtein distance registered in the registration unit 56; A distance calculation unit 59 that calculates a Levenshtein distance based on the comparison target character string length range set by the comparison target character string length setting unit 58 and the comparison target range set by the comparison target range setting unit 57; A provisional candidate registration unit 60 in which the Levenshtein distance for each comparison position and comparison target character string length acquired by the distance calculation unit 59 is registered as a provisional candidate; And a candidate evaluating section 61 Levenshtein distance is selected to the candidate what the smallest among those registered in the provisional candidate registration unit 60.

  FIG. 21 shows a procedure for keyword detailed position detection performed by the detailed position detection unit 52 shown in FIG. Due to insertion or deletion of characters due to erroneous recognition in character recognition, the length of the character string corresponding to the keyword character string in the recognition result character string may be different from the length of the regular keyword character string. In this example, since the illegal character “,” due to noise is inserted in the second of the recognition result character string, the length of the character string corresponding to the keyword character string in the recognition result character string is the normal keyword character. Unlike the 7 characters in the column “Invoice No .:”, 1 character is added to 8 characters, and the exact position of the keyword character string in the recognition result character string is in the range from 0 to 7 in the character order. Become.

  As described above, when the length of the character string corresponding to the keyword character string in the recognition result character string is different from the normal length, the above-described method for the approximate position detection, that is, the recognition result character string to be compared with the keyword character string. In the method in which the comparison target character string is set to the same number as the keyword character string and the keyword position is detected, the exact position of the keyword character string in the recognition result character string cannot be acquired.

  Therefore, the detailed position detection unit 52 enlarges / reduces the length of the comparison target character string in the recognition result character string to be compared with the keyword character string within a required range, and sets the Leven for each length of the comparison target character string. The Stein distance is obtained, and the one with the minimum Levenshtein distance is determined as the exact position of the keyword string.

  Here, first, the comparison target range setting unit 57 enlarges the candidate keyword approximate position acquired by the approximate position detection unit 51 forward and backward by a predetermined enlargement width, and compares the comparison target range in the recognition result character string, that is, the comparison target. The range to extract the character string is set. This expansion width is the Levenshtein distance D of the keyword approximate position candidate. In this example, the approximate keyword position is the character order 0 to 6, and the Levenstein distance D = 2 is added before and after this, so that the comparison target range in the recognition result character string is the character order -2 to 8. .

When the Levenshtein distance D acquired by the approximate position detection unit 51 is not an integer, it is made an integer by calculation such as rounding up. It is also possible to use D2 obtained according to the following equation instead of D using the error recognition rate E (0 ≦ E ≦ 1.0) acquired by the character recognition unit 33.
D2 = D + D × E.

  In the comparison target character string length setting unit 58, based on the Levenshtein distance D of the keyword approximate position candidate acquired by the approximate position detection unit 51, the comparison target character string in the recognition result character string to be compared with the keyword character string. The range of the length ML, that is, the maximum value MLmax and the minimum value MLmin are set. The minimum value MLmin of the comparison target character string length is obtained by subtracting the Levenshtein distance D from the keyword character string length KL, and the maximum value MLmax of the comparison target character string length is obtained by adding the Levenshtein distance D to the keyword character string length KL. It is obtained by adding. In this example, since the keyword character string length KL = 7 and the Levenstein distance D = 2 of the keyword approximate position candidate, the minimum value MLmin = KL−D = 7−2 = 5 of the comparison target character string length is obtained. The maximum value of the target character string length is MLmax = KL + D = 7 + 2 = 9.

  The distance calculation unit 59 calculates the Levenshtein distance LD between the comparison target character string and the keyword character string while shifting the comparison position one character at a time within the comparison target range set by the comparison target range setting unit 57. Further, at each comparison position, the comparison target character string and the keyword character string are changed while changing the length of the comparison target character string one character at a time within the range of the comparison target character string length set by the comparison target character string length setting unit 58. Levenshtein distance LD between and is calculated.

  Here, the starting point of the comparison target range is set to the initial value of the comparison position MS, that is, the comparison start position, and the comparison position MS is shifted character by character from the comparison start position toward the end point of the comparison target range. In this example, the initial value of the comparison position MS is a value ST−D = −2 obtained by subtracting the Levenshtein distance D from the starting point ST of the keyword approximate position. Further, the minimum value MLmin of the comparison target character string length ML is set as the initial value of the comparison target character string length ML, and is increased by one character until the comparison target character string length ML reaches the maximum value MLmax. In this example, the initial value of the comparison target character string length ML is the minimum value MLmin = KL−D = 5.

  FIG. 22 shows a provisional candidate list registered in the provisional candidate registration unit 60 shown in FIG. In this temporary candidate list, the comparison position (starting point) MS acquired by the distance calculation unit 59 and the Levenshtein distance LD for each comparison target character string length ML are registered with candidate management numbers as temporary candidates for the keyword detailed position. Is done.

  Based on this temporary candidate list, the candidate evaluation unit 61 selects a candidate having the minimum Levenshtein distance LD as a final candidate. In this example, the candidate with the candidate management number 13 is the final candidate. According to this final candidate, the detailed position of the keyword character string is 0, which is the value of the comparison position (starting point) MS, and the keyword in the recognition result character string. The length of the character string is 8, which is the value of the comparison target character string length ML, and the position of the keyword character string on the recognition result character string is in the range of 0 to 7 in the character order shown in FIG.

  Further, the candidate evaluation unit 61 determines whether or not the final candidate is valid as the position of the keyword character string. When the Levenstein distance LD is abnormally large, the position of the keyword character string is inappropriate. When the Levenstein distance LD of the final candidate is compared with a predetermined threshold TH and the Levenstein distance LD exceeds the threshold TH, It is assumed that the keyword search has failed because the final candidate is determined to be inappropriate as the position of the keyword character string.

  When a plurality of keyword approximate position candidates are acquired by the approximate position detection unit 51, the comparison condition, that is, the comparison position MS and the comparison target character string length ML are set between the temporary keyword detail position candidates acquired for each candidate. In some cases, the duplication candidates having the same comparison condition also have the same Levenshtein distance LD. Therefore, the Levenshtein distance calculation of the duplication candidate is omitted and is not registered in the temporary candidate list.

  FIG. 23 is a flowchart showing a keyword detail position detection procedure performed by the detail position detection unit 52 shown in FIG. This is executed by the keyword detailed position detection (step 302) in FIG.

  Here, one candidate of approximate keyword position acquired by the approximate position detector 51 and registered in the candidate registration unit 56 is first extracted (step 501), and the Levenstein distance D of the candidate is extracted (step 502). . Next, the keyword character string length KL is calculated (step 503). Further, a process for setting the comparison target range in the recognition result character string is performed by the comparison target range setting unit 57 (step 504), and the starting point of the comparison target range is set as the initial value of the comparison position MS. The comparison target character string length setting unit 58 performs a process of setting the length range of the comparison target character string in the recognition result character string, that is, the maximum value MLmax and the minimum value MLmin (step 505). The minimum value MLmin of the column length ML is set to the initial value.

  Next, a process for calculating the Levenshtein distance between the comparison target character string in the recognition result character string and the keyword character string according to the set comparison position MS and the comparison target character string length ML is a distance calculation unit. 59 (step 506), and the Levenstein distance acquired here is registered in the provisional candidate list (see FIG. 22) of the provisional candidate registration unit 60 (step 507).

  Then, it is determined whether or not the comparison target character string length ML has reached the maximum value MLmax (step 508). If the comparison target character string length ML has not reached the maximum value MLmax, the comparison target character string length is determined. ML is incremented by 1 (step 509), the Levenshtein distance is calculated and registered with the comparison target character string length ML increased by one character, and this processing is continued until the comparison target character string length ML reaches the maximum value MLmax. Repeated.

  When the comparison target character string length ML reaches the maximum value MLmax, whether or not the comparison position MS has reached the value obtained by adding the Levenshtein distance D to the starting point ST of the keyword approximate position, that is, the comparison position It is determined whether or not the MS has reached the final position within the comparison target range (step 510). If the comparative position MS has not reached the final position, the comparative position MS is incremented by 1 (step 511). ) The Levenshtein distance is calculated and registered at the comparison position MS shifted by one character, and this process is repeated until the comparison position MS reaches the final position.

  When the comparison position MS has reached the final position, it is determined whether or not all of the keyword approximate position candidates registered in the candidate registration unit 56 have been completed (step 512). Proceed to the next candidate process. When all of the keyword approximate position candidates are completed, the candidate evaluation unit 61 performs a process of selecting, as the final candidate, the one with the minimum Levenshtein distance D by referring to the provisional candidate list of the provisional candidate registration unit 60. Further, the candidate evaluation unit 61 determines whether or not the final candidate is valid as the position of the keyword character string by comparing the Levenshtein distance LD with a predetermined threshold TH (step 513). In the case, the final candidate keyword position is output.

  FIG. 24 is a flowchart showing another example of a keyword search procedure performed by the keyword search unit 34 shown in FIG. This uses both the keyword search based on the word order determination shown in FIG. 13 and the keyword search based on the Levenshtein distance shown in FIG. In this case, the keyword search unit 34 has both the units shown in FIG. 9 and the units shown in FIG.

  First, keyword search based on word order determination is performed by the keyword search unit 34 (step 601), and then it is determined whether or not the keyword position is detected (step 602). If the keyword position is detected, the keyword search is performed. The position information is output (step 603).

  On the other hand, when the keyword position is not detected by the keyword search by the word order determination, the keyword search by the Levenshtein distance is performed by the keyword search unit 34 (step 604), and then it is determined whether or not the keyword position is detected. If the keyword position is detected, information on the keyword position is output (step 606). If the keyword position is not detected, it is determined that the keyword search has failed (step 606). 607), the process ends.

  Keyword search based on word order determination is performed prior to keyword search based on Levenshtein distance because the processing procedure is relatively simple and the keyword position can be detected quickly, and the processing procedure is relatively complicated and takes time to process. The In the keyword search based on the word order determination, as shown in the example of FIG. 8D, in the case of erroneous recognition in which an illegal character is inserted in a portion corresponding to the partial character string in the recognition result character string, the keyword position is changed. Although it cannot be detected, the keyword search based on the Levenshtein distance is not easily affected by such misrecognition, and the keyword position can be accurately detected. By carrying out when the keyword position cannot be detected by the search, the keyword position can be reliably detected.

  The title extraction device, the image reading device, the title extraction method, and the title extraction program according to the present invention have an effect of accurately extracting a character string having no appearance feature on a document image as a title. The present invention is useful as a title extraction device, an image reading device, a title extraction method, and a title extraction program that extract a title from an image.

1 is a block diagram showing a schematic configuration of an image processing apparatus to which the present invention is applied. 1 is a block diagram showing a schematic configuration of an image reading apparatus to which the present invention is applied. The figure which shows the outline | summary of the process in the title extraction part shown in FIG. 1 and FIG. The figure which shows the example of the designation | designated method of the relative positional relationship of the title character string with respect to the keyword character string shown in FIG. Block diagram showing the title extraction unit shown in FIG. 1 and FIG. The figure which shows the outline | summary of the character recognition performed in the character recognition part shown in FIG. FIG. 1 is a flowchart showing a procedure of processing performed in the image processing apparatus and the image reading apparatus shown in FIG. The figure which shows the example of the misrecognition which arises by the character recognition in the character recognition part shown in FIG. The block diagram which shows the keyword search part shown in FIG. The figure which shows an example of the character string division | segmentation performed by the character string division | segmentation part shown in FIG. The figure which shows the point of the partial character string position detection performed by the partial character string position detection part shown in FIG. The figure which shows the information registered into the candidate registration part shown in FIG. The flowchart which shows the procedure of the keyword search performed in the keyword search part shown in FIG. The block diagram which shows another example of the keyword search part shown in FIG. The figure which shows the point of the Levenshtein distance calculation shown in FIG. The flowchart which shows the procedure of the keyword search performed in the keyword search part shown in FIG. The figure which shows the point of the keyword approximate position detection performed by the approximate position detection part shown in FIG. The figure which shows an example of the Levenshtein distance calculation result in the distance calculation part shown in FIG. The figure which shows another example of the Levenstein distance calculation result in the distance calculation part shown in FIG. FIG. 14 is a flowchart showing a procedure for keyword approximate position detection performed by the approximate position detection unit shown in FIG. The figure which shows the point of the keyword detailed position detection performed by the detailed position detection part shown in FIG. The figure which shows the temporary candidate list registered into the temporary candidate registration part shown in FIG. The flowchart which shows the procedure of the keyword detailed position detection performed by the detailed position detection part shown in FIG. The flowchart which shows another example of the procedure of the keyword search performed in the keyword search part shown in FIG. Figure showing an example of a document image

Explanation of symbols

3 Title Extraction Unit (Title Extraction Device)
21 Title inclusion area 22/23 Rectangular area 31 Extraction condition acquisition unit (extraction condition acquisition means)
32 Extraction condition storage unit 33 Character recognition unit (character recognition means)
34 Keyword search part (keyword search means)
35 Title position acquisition unit (title position acquisition means)
36 Title output section (title output means)
41 Character string dividing unit (character string dividing means)
42 Partial character string position detection unit (partial character string position detection means)
43 Evaluation Value Calculation Unit 44 Candidate Registration Unit 45 Candidate Evaluation Unit (Evaluation Means)
51 Approximate position detector (first position detector)
52 Detailed Position Detection Unit (Second Position Detection Unit)
53 distance calculation unit 54 provisional candidate registration unit 55 candidate evaluation unit 56 candidate registration unit 57 comparison target range setting unit 58 comparison target character string length setting unit 59 distance calculation unit 60 temporary candidate registration unit 61 candidate evaluation unit

Claims (9)

An extraction condition acquisition means for acquiring a keyword character string described in the vicinity of the title character string on the document image, and relative position information of the title character string with respect to the keyword character string;
Character recognition means for performing character recognition on an area including at least the title character string and the keyword character string in a document image;
From the recognition result of the character recognition means, keyword search means for searching the keyword character string acquired by the extraction condition acquisition means and acquiring its position;
Title position acquisition means for acquiring the position of the title character string based on the position of the keyword character string acquired by the keyword search means and the relative position of the title character string with respect to the keyword character string acquired by the extraction condition acquisition means When,
A title extraction apparatus comprising: a title output unit that outputs data of a title character string based on the position of the title character string acquired by the title position acquisition unit. The extraction condition acquisition means acquires position information of a title inclusion area including the title character string and the keyword character string by a user input operation,
The title extracting apparatus according to claim 1, wherein the character recognizing unit performs character recognition on the title inclusion area acquired by the extraction condition acquiring unit. The keyword search means
Character string dividing means for dividing the keyword character string into a plurality of partial character strings;
Each partial character string obtained by this character string dividing means is searched in the recognition result character string of the character recognition means, and a partial character string position detecting means for detecting the position of each partial character string;
The keyword character string based on the position of the partial character string obtained by the partial character string position detecting means by comparing the word order based on the position of each partial character string obtained by the partial character string position detecting means with the normal word order The title extracting apparatus according to claim 1, further comprising: an evaluation unit that evaluates whether or not the position of the file is appropriate. The keyword search means
First and second position detecting means for detecting the position of the keyword character string on the recognition result character string based on the difference between the comparison target character string and the keyword character string in the recognition result character string;
The first position detecting means detects the approximate position of the keyword character string on the recognition result character string based on the degree of difference obtained by setting the length of the comparison target character string to be the same as the keyword character string. ,
The second position detection means limits the comparison target within a comparison target range based on the approximate position of the keyword character string acquired by the first position detection means, and requires the length of the comparison target character string. The title extraction device according to claim 1, wherein the detailed position of the keyword character string on the recognition result character string is detected based on the degree of difference obtained by increasing or decreasing within the range.   The title extraction device according to claim 1, wherein the title output unit outputs text data of a title character string in the recognition result character string acquired by the character recognition unit.   2. The title extracting apparatus according to claim 1, wherein the title output means outputs image data created by cutting out a title character string area in a document image.   An image reading apparatus comprising the title extracting apparatus according to claim 1. An extraction condition acquisition step of acquiring a keyword character string described in the vicinity of the title character string on the document image, and relative position information of the title character string with respect to the keyword character string;
A character recognition step for performing character recognition on an area including at least the title character string and the keyword character string in a document image;
From the recognition result of this character recognition step, the keyword search step of searching the keyword character string acquired in the extraction condition acquisition step and acquiring its position;
Title position acquisition step for acquiring the position of the title character string based on the position of the keyword character string acquired in the keyword search step and the relative position of the title character string with respect to the keyword character string acquired in the extraction condition acquisition step When,
A title extraction method comprising: a title output step for outputting data of a title character string based on the position of the title character string acquired in the title position acquisition step.   A title extraction program for causing a computer to execute each step in the title extraction method according to claim 8.

Images