JP4936635B2 - Character string search device, character string search method, and program for causing computer to execute the method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
In the present invention, when a predetermined character string is accepted as a keyword, each character forming the document is converted into a character based on a distance value between each character forming the document to be searched and a standard character registered in advance. More particularly, the present invention relates to a character string search device, a character string search method, and a program for causing a computer to execute the method, which search for a character string that forms the keyword from the document based on the recognition result of the character recognition. Character search device, character search method, and program for causing computer to execute the method capable of efficiently searching for a character string corresponding to a predetermined keyword by reducing search omission and search noise even when there are characters that cannot be processed About.
[0002]
[Prior art]
2. Description of the Related Art Conventionally, electronic filing for storing and managing handwritten documents captured from a scanner or the like as image data is widely known. In electronic filing, since a large amount of image data is accumulated, it is an important issue to develop an efficient document search technique.
[0003]
For this reason, Non-Patent Document 1 discloses a technique for performing character recognition of a handwritten document stored as image data and extracting a document including a designated keyword by full-text search.
[0004]
In the technique of Non-Patent Document 1, a plurality of character recognition candidates for a character cut out from image data are expressed by a lattice structure, and the character string recognition accuracy is improved by searching using the lattice structure. However, the technique of Non-Patent Document 1 has a problem in that when there is a character that cannot be recognized in the character string to be searched, the character string becomes a search omission.
[0005]
For this reason, when searching for a character string from image data, Patent Document 1 calculates similarity by comparing the shape characteristics of a character image between characters that cannot be recognized and characters that have been created in advance. A document filing device that outputs a character string including characters that cannot be recognized as a search result when the degree is greater than a threshold value is disclosed.
[0006]
Further, in Patent Document 2, even if some characters that cannot be recognized are included in the character string of the keyword to be searched, if the ratio of the number of matched characters to the number of characters in the entire character string is equal to or greater than a threshold value, that character A filing device that outputs a column as a search result is disclosed.
[0007]
[Non-Patent Document 1]
Shuji Senda and two others, “Construction of a document image database system capable of full-text search”, Digital Library, Library and Information University, October 23, 1996, No. 8
[Patent Document 1]
JP 2000-57315 A
[Patent Document 2]
Japanese Patent Laid-Open No. 8-27213
[0008]
However, when the conventional technique of Patent Document 1 is used, the similarity is calculated by collating the character that cannot be recognized with the shape feature of the character that has been created in advance. However, the character that cannot be recognized has the character image quality in the first place. Since it is in a state where it is difficult to discriminate due to low quality, there is a problem that the degree of similarity is small and the omission of detection of character strings is not improved so much.
[0009]
Furthermore, with this conventional technology, when the character image of a handwritten character that is recognized is low quality, the character shape feature is collated forcibly. It may happen that the efficiency is deteriorated.
[0010]
Further, in the prior art of Patent Document 2, any character that could not be recognized is searched for a character string if the degree of coincidence of other parts is high. There is a problem that even a character string with no character is searched as a search noise, which confuses the user.
[0011]
For these reasons, it is important to realize a technology that can efficiently search for a character string corresponding to a given keyword while reducing search omissions and search noise even when there are characters that cannot be recognized. It is a difficult issue.
[0012]
The present invention has been made to solve the above-described problems (problems). Even when there are characters that cannot be recognized, search omissions and search noise are reduced, and a character string corresponding to a predetermined keyword is efficiently used. An object of the present invention is to provide a character search device, a character search method, and a program for causing a computer to execute the method.
[0013]
[Means for Solving the Problems]
The present invention has been made to achieve the above object, The present invention Then, when a predetermined character string is accepted as a keyword, each character forming the document is recognized based on a distance value between each character forming the document to be searched and a standard character registered in advance, A character string search device that searches a character string that forms the keyword from the document based on a recognition result of the character recognition, and uses a distance value between the predetermined character forming the document and the standard character. If the character cannot be recognized, the distance value giving means for giving a predetermined distance value to the predetermined character, and the distance value given by the distance value giving means and each character that forms the character-recognized document And a keyword search means for searching for a character string constituting the keyword existing in the document based on the distance value of the document.
[0014]
Also, The present invention A character lattice storage means for storing one or a plurality of character recognition candidates obtained by character recognition of each character forming the document as a character lattice, and the keyword search means includes the character lattice storage means. The recognition result stored as a character lattice is read, and the recognition result stored as the character lattice based on the distance value given by the distance value giving means and the distance value of each character forming the character-recognized document To search for a character string forming the keyword existing in the document.
[0015]
Also, The present invention When the predetermined character string is accepted as a keyword, each character forming the character string forming the keyword is searched in order from the first character of the character string based on the result of the character recognition, and the character string is formed. If the predetermined character to be recognized cannot be recognized using the distance value from the standard character, the search is continued by skipping the predetermined character, and the number of characters that cannot be recognized with respect to the number of characters in the character string is determined. A character string extracting unit configured to extract the character string when the ratio is equal to or lower than a predetermined value, wherein the keyword searching unit forms the character string extracted by the character string extracting unit; The extracted character based on the predetermined distance value assigned to the character by the distance value assigning means and the distance value of each character recognized to form the extracted character string There is judged whether the character string form the keyword, wherein the search for a character string which forms the keywords based on the determination.
[0016]
Also, The present invention When the first character of the character string forming the keyword cannot be recognized, the character string extracting means searches each character forming the character string forming the keyword in reverse order from the last character of the character string, and If character recognition cannot be performed for a predetermined character forming a sequence using the distance value from the standard character, the search is continued by skipping the predetermined character, and the character recognition for the number of characters in the character string is performed. The character string is extracted when the ratio of the number of characters that cannot be processed is equal to or less than a predetermined value.
[0017]
Also, The present invention , Applying dynamic programming to the character string extracted by the character string extracting means, and associating each character forming the extracted character string with each character forming the character string forming the keyword The distance value assigning means further includes a predetermined character that forms a character string to which dynamic programming is applied by the character associating means, even if a distance value between the standard character and the predetermined character is used. When the character cannot be recognized, a value relating to the character forming the character string forming the keyword associated with the predetermined character is assigned to the predetermined character by the character associating unit as the predetermined distance value. Then, the keyword search means is based on the distance value given by the distance value giving means and the distance value of each character recognized to form the character string to which the dynamic programming is applied. Issued string is determined whether a character string constituting the keyword, wherein the search for a character string which forms the keywords based on the determination.
[0018]
Also, The present invention The distance value assigning unit is unable to recognize a character using a distance value between the standard character and a predetermined character forming a character string to which dynamic programming is applied by the character matching unit. In addition, the predetermined distance value based on the dispersion or standard deviation of the distance values of the characters that form the character string that forms the keyword associated with the predetermined character by the character associating means is assigned to the predetermined character. It is characterized by giving.
[0019]
Also, The present invention A recognition rank assigning means for assigning numbers as recognition ranks to the character recognition candidates stored in the character lattice storage means as character lattices in ascending order of the distance values of the character recognition candidates. The keyword search unit is configured to determine the document based on the recognition rank given by the recognition rank granting unit, the distance value given by the distance value granting unit, and the distance value of each character forming the character-recognized document. The character string which makes the said keyword which exists in is searched.
[0020]
Also, The present invention Then, when a predetermined character string is accepted as a keyword, each character forming the document is recognized based on a distance value between each character forming the document to be searched and a standard character registered in advance, A character string search method for searching a character string forming the keyword from the document based on a recognition result of the character recognition, wherein a predetermined character forming the document is determined using a distance value from the standard character If a character cannot be recognized, a distance value giving step for giving a predetermined distance value to the predetermined character, and each character forming the document recognized as a distance value given by the distance value giving step And a keyword search step of searching for a character string forming the keyword existing in the document based on the distance value of the document.
[0021]
Also, The present invention Then, when a predetermined character string is accepted as a keyword, each character forming the document is recognized based on a distance value between each character forming the document to be searched and a standard character registered in advance, A program that causes a computer to execute a character string search method for searching for a character string that forms the keyword from the document based on the recognition result of the character recognition, wherein a predetermined character forming the document is between the standard character and When the character cannot be recognized even if the distance value is used, a distance value giving step for giving a predetermined distance value to the predetermined character, and the distance value given by the distance value giving step and the character recognized Causing a computer to execute a keyword search step of searching for a character string constituting the keyword existing in the document based on a distance value of each character forming the document. And features.
[0022]
DETAILED DESCRIPTION OF THE INVENTION
Exemplary embodiments of a character string search apparatus, a character string search method, and a program that causes a computer to execute the method will be described below in detail with reference to the accompanying drawings. In the present embodiment, a case where a predetermined search character string is searched from a horizontally written handwritten document image will be described.
[0023]
First, the configuration of the character string search device according to the present embodiment will be described. FIG. 1 is a functional block diagram showing the configuration of the character string search device according to the present embodiment. As shown in the figure, when searching for a document including a specific keyword from a plurality of handwritten document images 102 such as a form, the character string search apparatus 100 designates a keyword, and the keyword is the document image 102. It is a device that searches for whether it is included.
[0024]
In this character string search apparatus 100, even if there is a character that cannot be recognized in part, if the degree of matching between the keyword and the character string in the document image 102 as a whole is high, the character string is searched. can do.
[0025]
The character string search device 100 includes a document image storage unit 101, a document image 102, a search character string reception unit 103, a line / character cutout processing unit 104, a distance value calculation unit 105, a character recognition processing unit 106, a recognition order assignment unit 107, A character lattice storage unit 108, a character lattice 109, a character string extraction unit 110, a character association unit 111, a distance value assignment unit 112, a keyword search unit 113, and a control unit 114 are included.
[0026]
The document image storage unit 101 is a storage unit that stores and reads a plurality of handwritten document images 102, and the search character string reception unit 103 is a reception unit that receives a character string that is a keyword to be searched from the document image 102. It is. The search character string receiving unit 103 receives a character string input from a keyboard or the like, a character string transmitted from a terminal device via a network, and the like.
[0027]
The line / character cutout processing unit 104 specifies a document area in the document image 102 to be subjected to character recognition, cuts out a line from the document in the area, and extracts each character considered to constitute one character from each cut out line. Furthermore, it is a processing part to cut out.
[0028]
The specification of the document area for retrieving the character string is specifically performed according to the layout rule for each document. This specification may be performed manually. Further, the document area may be specified by another known method.
[0029]
The distance value calculation unit 105 calculates a distance value between a feature value of a standard character created by averaging characters written by various people and a feature value of each character in the cut document image 102. It is a calculation part. This standard character is stored in a storage unit (not shown). The standard character creation method is not limited to the method described here, and various methods can be used.
[0030]
The character recognition processing unit 106 is a processing unit that performs character recognition of the extracted character. Specifically, the distance value calculated by the distance value calculation unit 105 is referred to, and when the distance value is smaller than a predetermined threshold, it is recognized that the character corresponds to the compared standard character.
[0031]
The recognition rank assigning unit 107 is a grant unit that assigns a recognition rank to a plurality of character candidates obtained by performing character recognition on a certain character based on the obtained distance value of each character. is there. Specifically, the distance values of a plurality of character candidates are compared, numbers are assigned in ascending order of the distance values, and the recognition rank is set.
[0032]
The character lattice storage unit 108 is a storage unit that stores, as a character lattice 109, and reads out one or a plurality of character recognition candidates obtained by character recognition of each character forming the document image 102. Further, the distance value and recognition order of each character candidate are stored together. The character lattice 109 will be described in detail later.
[0033]
The character string extraction unit 110 is an extraction unit that extracts a character string that forms a keyword to be searched from the document image 102 using the character lattice 110. Specifically, each character forming the keyword is searched in order from the first character in the character lattice.
[0034]
At this time, if there is a character that cannot be recognized, the character is skipped and the next character is searched. When the ratio of the number of unrecognizable characters to the number of characters in the character string forming the keyword is equal to or less than a predetermined value, the character string is extracted as a keyword candidate to be searched.
[0035]
Also, if the first character of the keyword cannot be recognized, it will not be possible to start searching for the character string that makes up the keyword, as will be explained later, even if the other part matches the keyword. Therefore, the search for the character string candidates is performed in reverse order from the last character of the keyword to prevent the search omission.
[0036]
The character association unit 110 is a processing unit that performs a process of associating each character forming the extracted character string with each character of a keyword to be searched using dynamic programming.
[0037]
When there is a character that cannot be recognized in the character string associated by the character association unit 110, the distance value assigning unit 112 is based on the distribution of distance values of characters in the keyword corresponding to the character that cannot be recognized. The assigning unit assigns the obtained value as the distance value of the character that cannot be recognized. In addition, a predetermined rank is set as the recognition rank of characters that cannot be recognized.
[0038]
The keyword search unit 113 calculates an average distance value of the character string extracted by the character string extraction unit 110, and if the calculated average distance value is smaller than a predetermined threshold value, the keyword search unit 113 It is a search part which determines as corresponding to a keyword and outputs it as a search result. The average distance value is calculated based on the distance value and recognition order assigned to characters that cannot be recognized, and the distance value and recognition rank of characters that have been recognized.
[0039]
The control unit 114 is a control unit that performs overall control of the character string search device 100, and is a control unit that controls transmission and reception of various data between the functional units.
[0040]
FIG. 2 is a diagram illustrating an example of a horizontally written handwritten document to be searched by the character string search device 100. As shown in the figure, the character string search apparatus 100 can search a character string from a free-pitch document in which the character spacing between adjacent characters is free.
[0041]
Next, line and character cutout processing by the line / character cutout processing unit 104 will be described. FIG. 3 is a diagram for explaining line extraction processing by the line / character extraction processing unit 104. As shown in the figure, the line / character cutout processing unit 104 creates a histogram (line histogram) representing the frequency of black pixels in the horizontal direction, and cuts out each line using the valley portion between the line histograms as the line cutout position. To do.
[0042]
In addition, the character cutting process can be performed in the same manner. Specifically, a histogram (character histogram) representing the frequency of black pixels in the vertical direction is created for each extracted line, and each character is extracted using the valley portion between each character histogram as the character extraction position. .
[0043]
Next, the structure of the character lattice 109 shown in FIG. 1 will be described. FIG. 4 is a diagram illustrating an example of the basic segment cut out by the line / character cut-out processing unit 104. As shown in the figure, the line / character cutout processing unit 104 sets character division position candidates for each line by using a character histogram, and cuts out characters according to the division positions to create basic segments.
[0044]
The basic segment is a minimum unit that can be a character candidate divided by a character histogram. In the basic segment, there is a case where there is a part separated into the left and right in one character (for example, “Company”, “Ha”, etc.), and the character may be erroneously cut out as two characters.
[0045]
FIG. 5 is a diagram illustrating an example of candidate segments created by combining basic segments. As shown in the figure, in order to cope with a case where one character is erroneously cut out to two or more characters, candidate segments combining basic segments are created, and various character candidates are listed.
[0046]
FIG. 6 is a diagram illustrating an example of combinations of basic segments and candidate segments arranged in the order of appearance in a document. As shown in the figure, character recognition can be performed more accurately by combining basic segments and candidate segments and considering various character candidates.
[0047]
FIG. 7 is a diagram illustrating an example of the structure of the character lattice 109. This figure shows a character lattice 109 created by using a part of the basic segment and candidate segments shown in FIG. As shown in FIG. 7, the character lattice 109 represents various character candidates obtained by performing character recognition on the basic segment and candidate segment shown in FIG. 6 in a lattice structure.
[0048]
Here, it is also conceivable that a plurality of character candidates recognized as characters correspond to the same character image. In this case, for example, the character lattice 109 is generated by selecting only those whose recognition ranks are from the top to the tenth and the distance value of the candidate character is a predetermined threshold value or less.
[0049]
In addition, a segment number is assigned to each candidate character in the character lattice 109 in the order of appearance in the document. In the example of FIG. 7, “no”, “ba”, “ba”, “gu”, “b”, “t”, “age”, “fire”, “burn”, “so”, “no”, Segment numbers are assigned in the order of "Zen", "Sa", "-", "Rice", "Rate", and "Dou".
[0050]
Next, a process for calculating the distance value of the character feature amount by the character recognition processing unit 106 shown in FIG. 1 will be described. FIG. 8 is an explanatory diagram for explaining the calculation process of the distance value of the character feature amount by the character recognition processing unit 106 shown in FIG. In this process, standard characters created by averaging characters written by many people are used.
[0051]
First, as shown in FIG. 8A, processing for extracting the contour of a standard character image is performed. Subsequently, as shown in FIG. 8B, the character image from which the contour is extracted is divided into mesh regions. In this example, when a mesh area composed of two vertical and horizontal squares is overlapped with each other, half of the mesh area is divided so that 25 (= 5 × 5) are generated.
[0052]
When extracting character feature values, the number of character contours in each direction shown in FIG. 8 (c) is counted for each divided mesh region, and feature vectors are generated from the feature values. To do. In this example, since the number of directions of the contour to be extracted is eight, the dimension of the feature vector is 200 dimensions (= 5 × 5 × 8).
[0053]
Such a feature vector is created for each character cut out from the document image 102, and a distance value between the standard character and each character is calculated. Here, various distance measures such as Mahalanobis distance, Euclidean distance, or city block distance can be used as the distance value.
[0054]
Next, the segment number table created by the character lattice storage unit 108 shown in FIG. 1 will be described. FIG. 9 is a diagram showing an example of the segment number table created by the character lattice storage unit 108 shown in FIG. As shown in the figure, the character lattice storage unit 108 rearranges the generated character lattice 109 for each character and stores it as a segment number table in a form in which segment number, distance value, and recognition rank information are added.
[0055]
The segment number, distance value, and recognition order of each character are stored as a set such as (segment number, distance value, recognition order). In the example of FIG. 9, the character recognized as “I” appears three times in the document image, and the segment number, distance value, and recognition order are (64, 179, 1), (68, 214). 9) and (76, 125, 1).
[0056]
Next, character string extraction processing by the character string extraction unit 110 using the segment number table will be described. FIG. 10 is an explanatory diagram for explaining the character string extraction processing by the character string extraction unit 110 using the segment number table.
[0057]
As shown in the figure, when the keyword to be searched is “biofuel”, the character string extraction unit 110 searches the segment number table for the character “ba” that is the first character of the keyword. In FIG. 10A, it can be seen that the character “B” is in the segment number “69”.
[0058]
Subsequently, the next character “I” is searched. At this time, a predetermined number width is set, and the search is performed within the number width. This number width is appropriate so that only the “I” characters that follow the “B” character are selected in the document image, and the “I” characters that do not follow the “B” character are not included. Set to.
[0059]
For example, when the number width is 12, the character “I” whose segment number is 76 is 7 (= 76−69) since the segment number increment from the character “B” is “B”. Selected as a sequence of characters. Since the character “I” whose segment numbers are 64 and 68 is smaller than the segment number of the character “B”, it can be determined that the character has appeared before the character “B” and is not selected.
[0060]
Similarly, the character “o” with the segment number 78 is selected as a character subsequent to the character “a” because the increment of the segment number from the character “a” is 2 (= 78-76). . In the same manner, the characters up to “charge” are selected, and character string candidates corresponding to the keyword “biofuel” as shown in FIG. 10B are extracted. When the extraction is completed, the segment numbers of the first character and the last character of the extracted character string are stored.
[0061]
Here, the number width to be searched is set to 12, but the number width can be arbitrarily set as long as consecutive characters are appropriately selected. For example, the character number width of a character string included in the length of three times the height of one line of the document image may be calculated, and the character search may be performed within the number width. This is a setting method based on the prediction that consecutive characters in a character string appear within a length three times the height of one line.
[0062]
Next, character string extraction processing by the character string extraction unit 110 when there is a character that cannot be recognized will be described. FIG. 11 is an explanatory diagram illustrating character string extraction processing by the character string extraction unit 110 when there is a character that cannot be recognized. Here, a case where the character “o” with the segment number 78 cannot be recognized will be described.
[0063]
As shown in FIG. 10, the character string extraction unit 110 searches for the character “A” in the same manner as in FIG. However, since the character “o” having the segment number 78 could not be recognized, it is determined that there is no character “o”.
[0064]
In that case, the character string extraction unit 110 searches for the character “FUN”, which is the character following “O”, by expanding the number width for searching for the character. Here, the number width to be enlarged can be arbitrarily set. Here, it is set to 24 (= 12 × 2), which is twice the initial number width.
[0065]
By setting in this way, it is possible to select the character “burn” having the segment number 82. Thereafter, the subsequent characters are searched again with the normal number width.
[0066]
Here, when the first character of the keyword to be searched cannot be recognized, the segment number for starting the search is not determined, so that the search for each character forming the keyword cannot be started. Therefore, in this case, a process for searching for keywords in reverse order is performed.
[0067]
That is, when the character “B” of “Biofuel” cannot be recognized, the character “Fee” is searched, and then the character “Fuel” is searched in the reverse order. Using the example of FIG. 11, when the character “Rate” with the segment number 94 is selected, the character width “Fuel” is searched in reverse order with a number width of 12. In the same manner, the search is performed up to the character “I”.
[0068]
When there is a character that cannot be recognized, if the ratio of the number of characters that cannot be recognized to the number of characters of the keyword to be searched is equal to or less than a predetermined threshold value, a character string composed of each character searched by the character string extraction unit 110 is selected. Extracted as a character string corresponding to a keyword.
[0069]
In the example of FIG. 11, since the number of characters of the keyword “biofuel” is 5 and the number of characters that cannot be recognized is 1, the percentage of characters that cannot be recognized is 20%, and the threshold is set to 30%. If so, this character string is extracted as a character string corresponding to the keyword. Then, the segment numbers of the first character and the last character of the extracted character string are stored. This threshold value can be set arbitrarily.
[0070]
Next, a search process for searching for a character string that forms a keyword by the character association unit 111, the distance value assigning unit 112, and the keyword search unit 113 will be described. FIG. 12 is an explanatory diagram illustrating a search process for searching for a character string that forms a keyword by the character association unit 111, the distance value assigning unit 112, and the keyword search unit 113.
[0071]
First, the character associating unit 111 performs processing for associating each character forming the character string extracted by the character string extracting unit 110 with each character forming the character string forming the keyword by dynamic programming.
[0072]
The character association unit 111 searches for the basic segment that forms the character string based on the segment number of the first character and the last character of the character string extracted by the character string extraction unit 110 shown in FIG. As shown in FIG. 12B, a number is assigned to each basic segment.
[0073]
Then, as shown in FIG. 12 (c), the horizontal segments are arranged with basic segments assigned numbers, and the vertical axis is arranged with keywords to be searched, and each character forming a character string forming a keyword by dynamic programming. Each basic segment is associated with.
[0074]
In FIG. 12C, when the line advances in the upward direction as indicated by the arrow (1), it means that one basic segment corresponds to a plurality of characters. This corresponds to a case where there is a character cutout error such as two types of characters included in one basic segment, or a case where two types of character recognition results are obtained for one basic segment.
[0075]
When the line advances in an oblique direction as indicated by the arrow (2), it indicates that one basic segment corresponds to one character. Further, when the line advances in the horizontal direction as indicated by the arrow (3), it indicates that one character corresponds to a plurality of basic segments. This corresponds to a case where a kanji character composed of partial and 偏 is divided into two basic segments.
[0076]
In the example of FIG. 12 (c), the basic segments 1 to 3 are detected as corresponding to the character “B” by dynamic programming, and the basic segments 6 to 9 are detected as the characters “FUN”. Correspondingly, 10 and 11 basic segments are detected as corresponding to the characters “Rate”.
[0077]
FIG. 12D shows a character string in which characters are associated in this way. Then, the distance value assigning unit 112 newly assigns a distance value and a recognition order for the character after the association process to the characters “fuel” and “fee” whose combination of basic segments has been reviewed by the association.
[0078]
Thereafter, the keyword search unit 113 calculates an average distance value between the extracted character string and the keyword. Specifically, the distance value of each character forming the extracted character string is weighted by the recognition rank, and the average value of the weighted distance values is calculated. Here, the weight is given by 2 × N / (1 + N), where N is the value of the recognition order. When the recognition order is 1 (N = 1), this weight is 1, and when the recognition order is 10 (N = 10), this weight is about 1.82.
[0079]
In the example of FIG. 12D, the recognition ranks of “B”, “I”, “O”, “Fuel”, and “Fee” are 1, 1, 2, 1, and 1, respectively. The weights are 1, 1, 1.33, 1 and 1. Therefore, the average distance value of the character string “biofuel” is calculated as (1 × 113 + 1 × 125 + 1.33 × 204 + 1 × 164 + 1 × 196) /5=173.9. The calculation formula for calculating the average distance value is not limited to this, and may be calculated using another calculation formula.
[0080]
When the average distance value is calculated, the keyword search unit 113 compares the predetermined average distance value threshold value with the calculated average distance value, and the calculated average distance value is smaller than the threshold value. In this case, it is determined that the character string matches the keyword, and the result is output. For this threshold value, for example, the average value of 3σ of each character forming the keyword can be used.
[0081]
Next, an average distance value calculation process when there are characters that cannot be recognized will be described. FIG. 13 is an explanatory diagram for explaining the calculation process of the average distance value when there is a character that cannot be recognized.
[0082]
As shown in FIG. 13A, the character association unit 111 associates the basic segments 1 to 4 with the characters “B” and “I” in the keyword, as in FIG. 12C. . The basic segment No. 5 was unrecognizable by character recognition, and it is unclear what character it actually is. In this case, the character associating unit 111 shifts to the process of associating the basic segment corresponding to the next character “FUN”, assuming that the part corresponds to “O”. The subsequent processing is the same as that described with reference to FIG.
[0083]
FIG. 13B shows a character string in which the basic segments are associated in this way. Again, the distance value and recognition order for the revised character are newly assigned to the characters “Fuel” and “Fee” whose combination of basic segments has been revised.
[0084]
However, the distance value of the character “o” that cannot be recognized is set to 3 times the standard deviation σ (variance is the square of σ) of the distance value of “o”, that is, 3σ. In this example, 3σ is 261. The recognition order is set low, and in this example, it is set to 10th.
[0085]
In this example, the recognition ranks of “B”, “I”, “O”, “Fuel” and “Fee” are 1, 1, 10, 1 and 1, respectively. , 1.82, 1 and 1. Therefore, the average distance value of the character string “biofuel” is calculated as (1 × 113 + 1 × 125 + 1.82 × 261 + 1 × 164 + 1 × 196) /5=21.4.
[0086]
When the average distance value is calculated, the keyword search unit 113 compares the predetermined average distance value threshold value with the calculated average distance value, and the calculated average distance value is smaller than the threshold value. In this case, it is determined that the character string matches the keyword, and the result is output.
[0087]
Here, the distance value of characters that cannot be recognized can be set to an arbitrary value. However, if the distance value is set too small (for example, set to 0), if the distance value of other parts is small and the recognition order is low, the character can be recognized regardless of the character that cannot be recognized. Since a column is searched, even a character string unrelated to the keyword is searched, resulting in an increase in search noise.
[0088]
On the other hand, if the distance value of the character that cannot be recognized is set too large, the average distance value of the character string becomes large, and no matter how small the distance value of other parts, search omission is likely to occur. The same applies to the setting of the recognition order.
[0089]
Therefore, in the present invention, the above-mentioned problem is solved by setting an appropriate size value for the distance value and recognition order of unrecognizable characters. The reason why the value of 3σ is used as the distance value is that character feature information is included in 3σ.
[0090]
That is, in the example of FIG. 13A, a large value of 3σ (or variance) for the character “o” means that the variation of the character feature of “o” is large. In that case, the difference in features from other characters (for example, “age”) becomes inconspicuous, whether the character is actually “o”, and the character string containing that character matches the keyword Whether or not to do so is uncertain.
[0091]
Therefore, when the character in the character string corresponding to the character “o” in the keyword cannot be recognized, setting the distance value as 3σ makes it difficult for the character string to be selected as a search result, thus reducing search noise. Can be made. The value set as the distance value is not limited to 3σ, and may be a value based on the average value of the distance values of “e”.
[0092]
Next, the procedure of the character string search process according to this embodiment will be described. FIG. 14 is a flowchart showing the procedure of the character string search process according to the present embodiment. As shown in the figure, first, the line / character cutout processing unit 104 performs line and character cutout processing after layout analysis (step S1401).
[0093]
Then, the distance value calculation unit 105 calculates the distance value of the feature amount between each extracted character and the standard character (step S1402). Subsequently, the character recognition processing unit 106 performs character recognition of each character based on the calculated distance value (step S1403). Then, when a plurality of character recognition results are obtained for a certain character, the recognition rank assigning unit 107 sets the recognition rank in ascending order of the distance value (step S1404).
[0094]
After that, the character lattice storage unit 108 generates a character lattice, and stores the character recognition result as a segment number table together with the distance value and recognition order of each character (step S1405).
[0095]
Then, the character string extracting unit 110 extracts a character string corresponding to the designated keyword using the segment number table (step S1406), and the character associating unit 111 forms each extracted character string. A process of associating the character with each character in the keyword is performed (step S1407).
[0096]
Subsequently, the distance value assigning unit 112 checks whether there is a character that cannot be recognized in the associated character string (step S1408). If there is a character that cannot be recognized (step S1408, Yes), A value based on the dispersion of character distance values in a keyword corresponding to characters that cannot be recognized is set as a distance value of characters that cannot be recognized, and a predetermined rank is set as the recognition rank of characters that cannot be recognized ( Step S1409). If there is no character that cannot be recognized (No in step S1408), the process proceeds to step S1410.
[0097]
Thereafter, the keyword search unit 113 calculates an average distance value of the extracted character string, and when the calculated average distance value is smaller than a predetermined threshold, the extracted character string corresponds to the keyword. It judges as a thing and outputs as a search result (step S1410).
[0098]
As described above, in the present embodiment, when there is a character that cannot be recognized in the character string to be searched, a predetermined value (for example, standard deviation) is set as the distance value of the character, When a lower rank (for example, 10th rank) is set as the recognition rank, an average distance value between the extracted character string and the keyword is calculated, and the average distance value is smaller than a predetermined threshold value Since the extracted character string and the keyword are output as a match, even if there is an unrecognizable character, the average distance value of the entire character string is calculated and judged, so that search omissions and search noise You can search for keywords with few.
[0099]
The embodiments of the present invention have been described so far, but the present invention is not limited to the above-described embodiments, and can be applied in various different embodiments within the scope of the technical idea described in the claims. It may be implemented.
[0100]
For example, in the present embodiment, the case where the document image to be searched is a handwritten document image is shown, but the present invention is not limited to this, and the same applies to a document image of printed characters. Can be applied.
[0101]
In the present embodiment, the case where the present invention is applied to a horizontally written document has been described. However, the present invention is not limited to this and can be easily applied to a vertically written document.
[0102]
In this embodiment, a character string that forms a keyword is searched using a character lattice. However, the present invention is not limited to this, and a character string can be obtained by a known character string search method that does not use a character lattice. It is good also as searching for.
[0103]
Furthermore, in the present embodiment, the case where a character string corresponding to one keyword is searched from a document image such as a form is shown, but the present invention is not limited to this, and matches a word in a word dictionary. The present invention can also be applied to a case where a character string to be searched is searched from a document image, and an index of the document image is created by selecting a number of searched character strings.
[0104]
Further, in the present embodiment, a case where a character string is searched from a free-pitch document in which the character spacing between adjacent characters is free has been shown, but the present invention is not limited to this, and the character spacing is constant. The same can be applied to a document with an entry frame.
[0105]
In the present embodiment, the predetermined distance value and the recognition order are given to the characters that cannot be recognized after the dynamic programming is applied. However, the present invention is not limited to this. These may be added when it is determined that character recognition cannot be performed.
[0106]
Furthermore, in the present embodiment, the character string that forms the keyword is searched in order from the first character of the character string, and when the first character is a character that cannot be recognized, the search is performed in reverse order from the last character. The present invention is not limited to this, and when the first character is a character that cannot be recognized, it is possible to sequentially search for the next character and the subsequent characters.
[0107]
In addition, among the processes described in this embodiment, all or part of the processes described as being automatically performed can be manually performed, or the processes described as being manually performed All or a part of the above can be automatically performed by a known method. In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above-described document and drawings can be arbitrarily changed unless otherwise specified.
[0108]
Each component of the illustrated character string search apparatus is functionally conceptual and does not necessarily need to be physically configured as illustrated. That is, the specific form of the character string search device is not limited to the one shown in the figure, and all or part of the character string search device is functionally or physically distributed / integrated in arbitrary units according to various loads or usage conditions. Can be configured. Furthermore, all or any part of each processing function performed in the character string search device is realized by a CPU and a program that is analyzed and executed by the CPU, or is realized as hardware by wired logic. obtain.
[0109]
The character string search method described in this embodiment can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. This program can be distributed via a network such as the Internet. The program can also be executed by being recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, and a DVD and being read from the recording medium by the computer.
[0110]
【Effect of the invention】
As explained above, The present invention According to the above, when character recognition is not possible using a distance value between a predetermined character forming a document and a standard character, a predetermined distance value is given to the predetermined character, and the given distance value is Since it is configured to search for a character string that constitutes a keyword existing in a document based on the distance value of each character that forms a character-recognized document, even if there is a character that cannot be recognized, search omissions and search noise can be avoided. While decreasing, it becomes possible to search efficiently the character string corresponding to a predetermined keyword.
[0111]
Also, The present invention According to the above, one or more character recognition candidates obtained by character recognition of each character forming the document are stored as a character lattice, the recognition result stored as the character lattice is read, and characters that cannot be recognized are read. The character string that constitutes the keyword existing in the document is searched from the recognition result stored as the character lattice based on the distance value assigned to each character and the distance value of each character recognized. A character string can be searched from various character candidates, and a character string corresponding to a predetermined keyword can be efficiently searched.
[0112]
Also, The present invention According to the above, when a predetermined character string is accepted as a keyword, each character forming the character string forming the keyword is searched in order from the first character of the character string based on the character recognition result, and the character string is formed. If the character cannot be recognized even if the distance between the predetermined character and the standard character is used, the search is continued by skipping the predetermined character, and the ratio of the number of characters that cannot be recognized to the number of characters in the character string is predetermined. If the character string is less than or equal to the value, the character string is extracted, and a predetermined distance value assigned to a predetermined character that does not recognize the character that forms the extracted character string and the character that forms the extracted character string are recognized. The character string extracted based on the distance value of each character is determined whether it is a character string forming a keyword, and the character string forming the keyword is searched based on the determination. Even if there is a character that can not be shaped recognized, it is possible to efficiently search for a string forming the keyword.
[0113]
Also, The present invention According to the above, when the first character of the character string forming the keyword cannot be recognized, each character forming the character string forming the keyword is searched in reverse order from the last character of the character string, and the predetermined character string forming the character string is determined. If the character cannot be recognized using the distance between the standard character and the character, the search is continued by skipping the predetermined character, and the ratio of the number of characters that cannot be recognized to the number of characters in the character string is less than the predetermined value. Since the character string is extracted in such a case, the character string forming the keyword can be efficiently searched even when the first character of the character string forming the keyword cannot be recognized.
[0114]
Also, The present invention According to the above, by applying dynamic programming to the extracted character string, each character forming the extracted character string is associated with each character forming the character string forming the keyword, and dynamic planning is performed. When a character that cannot be recognized using a distance value between a standard character and a predetermined character that forms a character string to which the law is applied, a character string that forms a keyword associated with the predetermined character is formed. A value related to the character to be assigned as a predetermined distance value to the predetermined character, the given distance value, and the distance value of each character recognized to form a character string to which dynamic programming is applied, Based on the above, it is determined whether the extracted character string is a character string that forms a keyword, and the character string that forms the keyword is searched based on the determination, so even if there are characters that cannot be recognized, Proper distance with dynamic programming Can impart value to the character, it is possible to reduce the search omission or search noise.
[0115]
Also, The present invention According to the above, when a predetermined character forming a character string to which dynamic programming is applied cannot be recognized using a distance value between the standard character and the predetermined character, the predetermined character is associated. Since the predetermined distance value based on the dispersion or standard deviation of the distance value of the characters forming the character string forming the keyword is assigned to the predetermined character, even if there is a character that cannot be recognized, An appropriate distance value including character feature information can be given to the character, and search omission and search noise can be reduced.
[0116]
Also, The present invention According to the above, for the one or a plurality of character recognition candidates stored as a character lattice, numbers are assigned as recognition ranks in ascending order of distance values of the character recognition candidates, and the given recognition rank and character recognition cannot be performed. Since it is configured to search for a character string that forms a keyword existing in the document based on the distance value assigned to the character and the distance value of each character recognized, Search noise can be further reduced.
[0117]
Also, The present invention According to the above, when character recognition is not possible using a distance value between a predetermined character forming a document and a standard character, a predetermined distance value is given to the predetermined character, and the given distance value is Since it is configured to search for a character string that constitutes a keyword existing in a document based on the distance value of each character that forms a character-recognized document, even if there is a character that cannot be recognized, search omissions and search noise can be avoided. While decreasing, it becomes possible to search efficiently the character string corresponding to a predetermined keyword.
[0118]
Also, The present invention According to the above, when character recognition is not possible using a distance value between a predetermined character forming a document and a standard character, a predetermined distance value is given to the predetermined character, and the given distance value is Since it is configured to search for a character string that constitutes a keyword existing in a document based on the distance value of each character that forms a character-recognized document, even if there is a character that cannot be recognized, search omissions and search noise can be avoided. While decreasing, it becomes possible to search efficiently the character string corresponding to a predetermined keyword.
[Brief description of the drawings]
FIG. 1 is a functional block diagram showing a configuration of a character string search device according to an embodiment.
FIG. 2 is a diagram illustrating an example of a horizontally written handwritten document to be searched by a character string search device;
FIG. 3 is a diagram for explaining line cut-out processing by a line / character cut-out processing unit;
FIG. 4 is a diagram illustrating an example of a basic segment cut out by a line / character cut-out processing unit.
FIG. 5 is a diagram showing an example of candidate segments created by combining basic segments.
FIG. 6 is a diagram illustrating an example of combinations of basic segments and candidate segments arranged in the order of appearance in a document image.
FIG. 7 is a diagram illustrating an example of a structure of a character lattice.
8 is an explanatory diagram for explaining a calculation process of a distance value of a character feature amount by a character recognition processing unit shown in FIG. 1; FIG.
9 is a diagram showing an example of a segment number table created by the character lattice storage unit shown in FIG. 1. FIG.
FIG. 10 is an explanatory diagram illustrating a character string extraction process by a character string extraction unit using a segment number table.
FIG. 11 is an explanatory diagram illustrating a character string extraction process performed by a character string extraction unit when there is a character that cannot be recognized.
FIG. 12 is an explanatory diagram illustrating search processing for searching for a character string that forms a keyword by a character association unit, a distance value assigning unit, and a keyword search unit;
FIG. 13 is an explanatory diagram illustrating an average distance value calculation process when there are characters that cannot be recognized.
FIG. 14 is a flowchart showing a processing procedure of a character string search process according to the present embodiment.
[Explanation of symbols]
100 Character string search device
101 Document image storage unit
102 Document image
103 Search string acceptance section
104 Line / character cutout
105 Distance value calculator
106 Character recognition processing section
107 recognition rank assigning unit
108 character lattice storage
109 character lattice
110 Character string extractor
111 Character mapping part
112 Distance value giving unit
113 Keyword search part
114 Control unit

Claims (9)

When a predetermined character string is accepted as a keyword, based on a distance value between each character candidate forming a document image to be searched and a standard character created by averaging characters written by many people A character string search device for performing character recognition of the document image and searching the character string forming the keyword from the document image based on the recognition result of the character recognition,
When the distance value between the standard character corresponding to each character forming the character string and the character candidate forming the document image is less than a predetermined threshold, the distance value is associated with the character candidate. Character recognition means that the character recognition was successful in
When the character recognition means fails in character recognition for at least one of the characters forming the character string, the large number used when creating the standard character for the standard character corresponding to the character dispersion of the distance values between each character and said standard characters written by a person, the standard deviation or average the calculated the dispersion, the position corresponding to the character value calculated in advance based on the standard deviation or the average as the distance value A distance value giving means for giving the character candidate existing in
The distance value assigned by the distance value assigning unit to the character candidate for which the character recognition unit has failed to recognize the character, and the distance value associated with the character candidate for which the character recognition unit has successfully recognized the character. A character string search device comprising: keyword search means for searching for the keyword based on the above. Character lattice storage means for storing, as a character lattice, one or a plurality of the character candidates obtained when the character recognition means has succeeded in character recognition.
The keyword detecting means includes
The character string search device according to claim 1, wherein the keyword is detected for the character candidates included in the character lattice stored in the character lattice storage unit. The keyword search means includes:
The keyword candidate is a set of the character candidates in which a value obtained by dividing the number of characters that the character recognition unit has failed in the character recognition divided by the total number of characters of the character string is equal to or less than a predetermined threshold. The character string search device according to claim 1 or 2. The character recognition means includes
When character search fails for the first character among the characters forming the character string with the search direction of the character candidates forming the document image in the normal order, the search direction is set in the reverse order. The character string search apparatus according to claim 1, 2, or 3, wherein character recognition is performed from the last character among the characters forming the character string. The character recognition means includes
The character recognition is performed on the character candidate after generating the character candidate by combining one or more basic segments by dynamic programming. The character string search device described in 1. The keyword search means includes:
A distance value assigned by the distance value assigning unit to the character candidate for which the character recognition unit has failed in character recognition, and a distance value associated with the character candidate for which the character recognition unit has succeeded in character recognition. The character string search apparatus according to claim 1, wherein a set of the character candidates whose average distance value calculated by use is equal to or less than a predetermined threshold is set as the keyword candidate. The character recognition means includes
Corresponding recognition ranks in ascending order of distance values from the standard characters to the character candidates that have been successfully recognized.
The distance value giving means is
Giving the recognition rank which is not at least the first place to the character candidate given the distance value,
The keyword search means includes:
The distance value given by the distance value giving unit to the character candidate for which the character recognition unit has failed to recognize the character and the recognition rank, and the character recognition unit is associated with the character candidate that has succeeded in character recognition. The character string search device according to claim 1, wherein the keyword is searched based on a distance value and the recognition order. A character string search method for searching a keyword consisting of a predetermined character string from a document image,
A distance value between each character candidate forming the document image and a standard character created by averaging characters written by a large number of people is calculated, and the standard character corresponding to each character forming the character string A character recognition step in which when the distance value with the character candidate forming the document image is less than a predetermined threshold, the character candidate is successfully recognized after associating the distance value with the character candidate;
When the character recognition process fails in character recognition for at least one character among the characters forming the character string, the large number used when creating the standard character for the standard character corresponding to the character dispersion of the distance values between each character and said standard characters written by a person, the standard deviation or average the calculated the dispersion, the position corresponding to the character value calculated in advance based on the standard deviation or the average as the distance value A distance value giving step to be given to the character candidate existing in
The distance value assigned by the distance value assigning step to the character candidate for which the character recognition step has failed in character recognition, and the distance value associated with the character candidate for which the character recognition step has succeeded in character recognition. A character string search method comprising: a keyword search step for searching for the keyword based on the above. When a predetermined character string is accepted as a keyword, based on a distance value between each character candidate forming a document image to be searched and a standard character created by averaging characters written by many people A character string search program for performing character recognition of the document image and searching the character string forming the keyword from the document image based on the recognition result of the character recognition,
When the distance value between the standard character corresponding to each character forming the character string and the character candidate forming the document image is less than a predetermined threshold, the distance value is associated with the character candidate. A character recognition procedure that assumes that character recognition succeeded in
When the character recognition procedure fails in character recognition for at least one character among the characters forming the character string, the large number used in creating the standard character for the standard character corresponding to the character dispersion of the distance values between each character and said standard characters written by a person, the standard deviation or average the calculated the dispersion, the position corresponding to the character value calculated in advance based on the standard deviation or the average as the distance value A distance value giving procedure to be given to the character candidate existing in
The distance value assigned by the distance value assigning procedure to the character candidate for which the character recognition procedure has failed in character recognition, and the distance value associated with the character candidate for which the character recognition procedure has succeeded in character recognition. A character string search program that causes a computer to execute a keyword search procedure for searching for the keyword based on the above.

Images