JP3669626B2 - Search device, recording medium, and program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a search device, a recording medium, and a program for searching for a keyword from a character recognition result obtained by recognizing a character in an original document using an index table.
[0002]
[Prior art]
In recent years, with the spread of the Internet, search technology that extracts necessary information from a large amount of information existing on a network is regarded as important. In particular, many systems for searching for specific keywords from text data have already been provided. In such a search, it is required to perform an accurate and high-speed search from a large amount of text documents.
[0003]
In order to perform a high-speed search, a technique for searching for a specific keyword from text data using an index table is known. The index table defines an index character string including a predetermined number of characters (for example, two characters) and a position of a portion in text data that matches the character string.
[0004]
When searching for a keyword from a set of character codes (character recognition results) obtained by character recognition of characters in an original document (paper document), character recognition errors (false recognition) must be considered. I must. This is because if there is an error in character recognition, the character represented by the character code can be different from the character written in the original document. Misrecognition means that characters written in the original document are not correctly converted into character codes. Such misrecognition occurs due to, for example, fading, tilting, and dirt of characters printed on paper.
[0005]
For example, if a character string “Inu” exists at a position in the original document and the character “nu” in the character string is erroneously recognized as “su”, the character recognition result corresponding to the character string “Inu” A character string “chair” exists at the position of the inner part. As a result, the index character string “chair” and its position are registered in the index table created from the character recognition result. Therefore, even if the keyword “dog” is searched using this index table, the keyword cannot be detected at that position in the character recognition result. As described above, there is a problem of “missing search” in which a keyword cannot be detected at a certain position even though the keyword exists at a certain position in the original document.
[0006]
As a conventional technique for dealing with the search omission problem, a plurality of candidate characters may be prepared as a character recognition result for one character in the original document, and may exist in the original document based on the plurality of candidate characters. A technique for registering a plurality of character strings as an index character string in an index table is known. The keyword search is performed using this index table. Such a technique is disclosed in, for example, “Information Processing Method and Apparatus” in Japanese Patent Laid-Open No. 9-16619.
[0007]
FIG. 11 shows an example of an index table 1901 in which a plurality of character strings that may exist in an original document are registered as index character strings according to the prior art. In the example shown in FIG. 11, the index table 1901 is obtained by character recognition of an original document including a character string “using an index... In the index table 1901, the index character string “I” and the index character string “In” are both registered as existing at the character position “1”, which is the same position in the character recognition result (rows 1911 and 1911). Line 1912).
[0008]
By using the index table 1901 shown in FIG. 11, the keyword “index” can be detected. Hereinafter, processing for searching for the keyword “index” using the index table 1901 shown in FIG.
[0009]
First, a character string composed of two adjacent characters included in a keyword is generated. From the keyword “index”, five character strings “IN”, “NDE”, “DET”, “Cuck”, and “X” are generated.
[0010]
Next, these character strings are searched from the index table 1901. The character strings “IN”, “NDE”, “DET”, “Cook”, and “CUS” have character positions “1”, “2”, “3”, “4” and “4” in the character recognition result respectively. “5” is indicated (line 1912, line 1919, line 1915, line 1914 and line 1913).
[0011]
From the positional relationship of these character positions, it is determined that the keyword “index” exists in the character recognition result.
[0012]
In this way, according to the conventional technique using the index table in which a plurality of character strings that may exist in the original document are registered as index character strings, the problem of search omission can be avoided.
[0013]
[Problems to be solved by the invention]
According to such a conventional technique, there is a problem that search noise increases. Search noise means that a keyword is detected even though the keyword does not exist in the original document. For example, when keywords “denwa” and “hook” are searched using the index table 1901 shown in FIG. 11, these keywords are detected at the character position “3”. In order to determine whether the search result is valid, the user must compare the original document with the search result.
[0014]
In order to prevent the problem of search omission, as the number of candidate characters obtained as a character recognition result for one character increases, such search noise increases, and the burden on the user to determine whether the search result is valid or not. Will increase.
[0015]
The present invention has been made in view of such problems, and provides a search device, a recording medium, and a program that can perform a high-speed search and can easily determine the validity of a search result. The purpose is to do.
[0016]
[Means for Solving the Problems]
The search device of the present invention recognizes each of the characters in the original document, and uses a result of character recognition of the original document having at least one candidate character corresponding to each character to use a plurality of characters. The index table is an index configured by a character string combining the candidate characters corresponding to each of a plurality of characters constituting the character string existing in the original document. A certainty defined in advance as the probability that each candidate character exists in the original document for each character string, the position of the index character string in the original sentence, and each candidate character included in the index character string Before the same number of characters as the index string It is determined whether a keyword character string exists in the index table, and if it exists, based on the position of the index character string that matches the keyword character string, the keyword character string in the original sentence A keyword certainty factor defined as a probability that the keyword exists at the identified position based on a position identifying unit that identifies the position of the keyword and a certainty factor of an index character string that matches the character string of the keyword A calculation unit; A determination unit that determines validity of a search result based on the keyword certainty factor; Thus, the above object is achieved.
[0018]
The determination unit may determine that the search result is valid when the keyword certainty factor is equal to or greater than a predetermined value.
[0019]
The predetermined value may be set according to at least one of the number of characters included in the keyword and the type of character included in the keyword.
[0020]
Said An index table creation unit that creates an index table further includes the index table creation unit, When a plurality of candidate characters are generated based on the character recognition result, each of the candidate characters generated corresponding to each character in the character string of the original document The index character string may be generated by combining.
[0021]
Said An index table creation unit that creates an index table further includes the index table creation unit, One candidate character is generated for each character based on the character recognition result, and a similar character similar to each generated candidate character is generated based on the certainty factor, and a character string of the original document The combination of the candidate character corresponding to each character and the similar character, or the combination of the similar characters The index character string may be generated.
[0022]
The recording medium of the present invention uses a character recognition result of the original document having at least one candidate character corresponding to each character by recognizing each character in the original document, using an index table, and a plurality of characters. A computer-readable recording medium recording a program for causing a computer to execute a search process for searching for a keyword comprising the index table, wherein the index table includes a plurality of characters constituting a character string existing in the original document. For each candidate character included in the index character string, an index character string composed of a character string that combines the corresponding candidate characters, a position of the index character string in the original sentence, and the original document Each said candidate character in it And a certainty factor that is defined in advance as a probability of existence, and the search process determines whether a character string of the keyword having the same number of characters as the index character string exists in the index table. Determining the position of the keyword character string in the original document based on the position of the index character string that matches the keyword character string; and the confidence of the index character string that matches the keyword character string. Calculating a keyword certainty factor defined as a probability that the keyword exists at the specified position based on the degree, and determining a validity of a search result based on the keyword certainty factor Records the program to be This achieves the above object.
[0023]
The program of the present invention recognizes each of the characters in the original document by using an index table from the character recognition result of the original document having at least one candidate character corresponding to each character. Search process to search for By computer An index character string composed of a character string obtained by combining the candidate characters corresponding to each of a plurality of characters constituting the character string existing in the original document. A position of the index character string in the original sentence and a certainty factor defined in advance as a probability that each candidate character exists in the original document for each candidate character included in the index character string. And the search process determines whether a character string of the keyword having the same number of characters as the index character string exists in the index table, and if it exists, an index character string that matches the character string of the keyword is determined. Based on the position, the keyword in the original sentence. A keyword certainty factor defined as a probability that the keyword exists at the identified position based on the step of identifying the position of the character string and the certainty factor of the index character string that matches the character string of the keyword And steps to Determining the validity of a search result based on the keyword certainty factor; This achieves the above object.
[0024]
DETAILED DESCRIPTION OF THE INVENTION
In this specification, a character is not limited to a character used in a specific language system, and refers to any symbol including numbers and symbols (for example, “)” and “◎”). Such a symbol is assigned a code (character code) for electronically expressing the symbol.
[0025]
FIG. 1 shows a configuration of a search device 1 of the present invention. The search device 1 searches for a keyword from a character recognition result obtained by recognizing a character in the original document.
[0026]
The search device 1 includes, as its constituent elements, a terminal 100, a CPU 100 that executes document registration processing and document search processing, an image input device 120 that inputs a document as image data, a work memory 180, and a hard disk (HDD) 170. With. These components are connected to each other via an internal bus 1101. Alternatively, these components may be connected to each other via any type of network.
[0027]
The terminal 100 is an input / output device including a keyboard and a CRT, for example. The terminal 100 is used, for example, for a user to specify a process executed by the search device 1 or to display a search process result executed by the search device 1 to the user.
[0028]
The HDD 170 stores a document registration program 1103, a document search program 1104, a character recognition pattern dictionary 160, a certainty factor table 150, and document data 1102. Any type of memory may be used as the HDD 170.
[0029]
All or part of the document registration program 1103 and the document search program 1104 may be provided to the search device 1 via any type of communication line (not shown) or broadcast, or any type of computer-readable program. The search apparatus 1 may be provided in a form recorded on a simple recording medium. Such a recording medium is, for example, a DVD-ROM, a CD-ROM, a flexible disk, or the like. The document registration program 1103 and the document search program 1104 recorded on such a recording medium can be installed in the search device 1 by a reading device such as a disk drive.
[0030]
FIG. 2 shows a flow of processing executed by the search device 1 in order to search for a keyword from a character recognition result obtained by recognizing a character in the original document.
[0031]
When the user instructs the start of document registration processing from the terminal 100 (FIG. 1), the document registration program 1103 stored in the HDD 170 is loaded into the work memory 180. The CPU 110 can access the work memory 180 at high speed. When the CPU 110 executes the document registration program 1103, document registration processing is performed.
[0032]
The document registration process includes a character recognition process and an index table creation process. The character recognition process and the index table creation process are respectively performed by the CPU 110 executing a character recognition program (not shown) and an index table creation program (not shown) that are part of the document registration program. .
[0033]
In the character recognition process, the original document is read by the image input device 120, and image data (document image data) 130 of the original document is generated. The document image data 130 is stored in the HDD 170 (FIG. 1). Character recognition processing is performed based on the similarity between the shape represented by the partial area in the document image data 130 and the character shape registered in the character recognition pattern dictionary 160 (FIG. 1). A result of the character recognition process is stored in the HDD 170 as a character recognition result 140.
[0034]
Next, in the index table creation process, an index table 190 is created from the character recognition result 140. The index table 190 is stored in the HDD 170. In the index table creation process, the certainty factor table 150 (FIG. 1) can be referred to.
[0035]
Document image data 130, character recognition result 140, and index table 190 stored in HDD 170 constitute at least a part of document data 1102 (FIG. 1).
[0036]
When the user inputs a keyword from the terminal 100 and instructs the start of the document search process, the document search program 1104 stored in the HDD 170 is loaded into the work memory 180. When the CPU 110 executes the document search program 1104, document search processing is performed. In the document search process, a keyword is searched from the character recognition result using the index table 190.
[0037]
Once the index table 190 is generated by the document registration process, the keyword search is performed with reference to the index table 190. Even if the keyword to be searched changes, it is not necessary to create a new index table 190.
[0038]
Note that it is not essential that all processing shown in FIG. For example, the document registration process may be performed by a device different from the search apparatus 1, and only the document search process using the generated index table 190 may be performed by the search apparatus 1.
[0039]
FIG. 3 shows an example of the original document 1310. The original document 1310 includes a character string “search method using an index. From document data”. The original document is, for example, a document in the form of paper on which a character string is printed. The original document may be a document written in a sign, a signboard, a bulletin board, or the like.
[0040]
FIG. 4 shows an example of a character recognition result 140 obtained by performing character recognition processing on the original document 1310. The character recognition result 140 includes a character position 1042 and a candidate character 1043. In FIG. 4, the numbers in parentheses attached to each candidate character indicate the reliability Rr for each candidate character. The character recognition result 140 is the largest character recognition result of each character “I”, “N”, “De”, “T”, “K”, “S” included in the original document 1310 (FIG. 3). The candidate characters for which the reliability Rr is obtained are “i”, “shi”, “te”, “so”, “ta”, and “su”, respectively.
[0041]
The character recognition process can be performed according to an arbitrary algorithm. The character recognition process can be executed, for example, according to an algorithm in which the document image data 130 is cut out in character units, and the cut out image data (partial area) in character units is converted into character codes.
[0042]
When converting from a partial area to a character code, the shape represented by the partial area is compared with the shape of the character registered in the character recognition pattern dictionary 160 (FIG. 1). Characters determined to be similar in shape based on a predetermined determination criterion are obtained as candidate characters. In this way, the partial area is converted into the character code of the candidate character. A plurality of candidate characters corresponding to one partial region may be obtained.
[0043]
Candidate characters are considered to be highly likely to match the characters written in the part corresponding to the partial area of the original document because the shape and the shape represented by the partial area are similar. Means.
[0044]
Each column (for example, the column 1044) in the character recognition result 140 corresponds to a partial region in the document image data 130. That is, it corresponds to a portion of the original document 1310 (for example, the portion 1311 shown in FIG. 3). Further, since the partial area in the document image data 130 is cut out from the document image data in units of one character, the partial area is one character of the original document 1310 (for example, a character written in the part 1311 shown in FIG. 3). "Ku").
[0045]
Candidate characters “T”, “U”, “W”, and “K” shown in column 1044 can match the characters written in the corresponding portion of original document 1310 (portion 1311 shown in FIG. 3). It is a character that can be regarded as having high character.
[0046]
A character position 1042 indicates the position of the candidate character in the character recognition result 140. For example, the character position “5” shown in the column 1044 indicates that the position of the column 1044 (part in the character recognition result) in the character recognition result 140 is the “fifth” position.
[0047]
As a method for expressing the character position 1042, any method may be used as long as the portion in the original document 1310 corresponding to the candidate character can be specified. As described above, each column in the character recognition result corresponds to a partial area of the document image data 130. Therefore, the character position 1042 may be represented by the position in the character recognition result of the column including the candidate character, or may be represented by the position in the document image data 130 of the partial area of the document image data 130 corresponding to the column. May be.
[0048]
For example, the character position 1042 may be represented by the document name of the original document, the page number, the line number, and the position (number of characters) from the beginning of the line, or the coordinates in the document image data. Or may be represented by an address.
[0049]
The reliability Rr indicates the probability of character recognition, that is, the correct probability. The reliability Rr takes a value from 0 to 1, and the greater the value, the greater the probability. For character recognition, for example, a neural network, vector quantization, or template matching can be employed.
[0050]
When a neural network method is adopted for character recognition, characters corresponding to at least one neuron whose output value is equal to or greater than a certain standard among characters registered in the character recognition pattern dictionary 160 are obtained as candidate characters. . The correspondence relationship between the neuron output value and the correct answer probability is obtained in advance, and the reliability Rr can be obtained from the neuron output value corresponding to each candidate character based on the correspondence relationship.
[0051]
In both vector quantization and template matching methods, the distance in the feature amount space between the shape represented by the partial region of the document image data 130 and the character shape registered in the character recognition pattern dictionary 160 is obtained. This is a technique for performing character recognition. One shape is represented as one representative point in the feature amount space. When these methods are employed, at least one character whose distance in the feature amount space is equal to or less than a reference among characters registered in the character recognition pattern dictionary 160 is obtained as a candidate character. A correspondence relationship between the distance in the feature amount space and the correct answer probability is obtained in advance, and the reliability Rr can be obtained from the distance in the feature amount space corresponding to each candidate character based on the correspondence relationship.
[0052]
Regardless of which method is used for character recognition, the reliability Rr reflects the similarity between the shape represented by the partial area of the document image data 130 and the character shape registered in the character recognition pattern dictionary 160. To do.
[0053]
Information other than shape similarity may be considered as the reliability Rr. For example, the deviation SR of the size of the partial area that is the target of character recognition in the document image data 130, the deviation LR of the relative position of the partial area in the line, and the like may be considered.
[0054]
The deviation SR of the size of the partial area is defined as, for example, a deviation of the size of the partial area from the average value of the sizes of all partial areas (each corresponding to one character) in the document image data 130. Can be done. By obtaining a correspondence relationship between the deviation SR and the correct recognition probability of character recognition in advance, the reliability Rr can be corrected so that the reliability Rr becomes small when the deviation SR is large.
[0055]
The relative position deviation LR of the partial areas is obtained by, for example, obtaining the average value of the positions in the direction perpendicular to the lines for all the partial areas (each corresponding to one character) in the same line in the document image data. It can be defined as the deviation from this average value of the position in the direction perpendicular to the rows of the partial areas. By obtaining a correspondence relationship between the deviation LR and the correct probability of character recognition in advance, the reliability Rr can be corrected so that the reliability Rr becomes small when the deviation LR is large.
[0056]
Thus, the reliability Rr can be set more appropriately by correcting the reliability Rr according to the deviation SR and / or the deviation LR.
[0057]
An index table 190 is created from the character recognition result 140 shown in FIG. 4 (index table creation processing).
[0058]
FIG. 5A shows a procedure of index table creation processing. Hereinafter, the procedure of the index table creation process will be described in detail.
[0059]
Step S401: It is determined whether or not the reliability Rr of the candidate character of interest in the character recognition result 140 is greater than or equal to a reference value. The reference value may be “0.05”, for example. If the determination result in step S401 is “Yes”, the process proceeds to step S402. If the determination result in step S401 is “No”, the process proceeds to step S404.
[0060]
When the character recognition result 140 (FIG. 4) is obtained by the character recognition process, only the candidate characters whose reliability Rr is equal to or higher than the reference value may be included in the character recognition result 140. In that case, the process in step S401 may be omitted.
[0061]
Step S402: The certainty factor Cr of the candidate character is calculated. The certainty factor Cr is calculated by, for example, (Equation 1) based on the reliability Rr for each candidate character.
[0062]
[Expression 1]
Confidence level Cr = reliability level Rr for candidate character × characteristic coefficient Kr
The character-specific coefficient Kr is defined in advance for each character (for example, “I”). The character-specific coefficient Kr depends on the appearance probability of the character in a normal document. Each character has a different appearance probability in a normal document. For example, in a general Japanese document, the character “ゐ” has a lower appearance probability than the character “ru”. As described above, the character-specific coefficient Kr is set low for characters having a low appearance probability. On the contrary, the character-specific coefficient Kr is set to be high for characters having a high appearance probability. The appearance probability for each character can be statistically obtained in advance for a large number of general documents.
[0063]
The certainty factor Cr for each candidate character is obtained by multiplying the reliability Rr for the candidate character by the character-specific coefficient Kr for the same character as the candidate character (character with the matching character code). The certainty factor Cr calculated in this way indicates the probability that the same character as the candidate character exists in a specific part in the original document. Such a specific portion is a portion in the original document corresponding to the column (the portion in the character recognition result) including the candidate character in the character recognition result 140 (FIG. 4).
[0064]
However, the certainty factor Cr is not necessarily a statistical probability itself. The certainty factor Cr may be a value obtained by normalizing a statistical probability according to a predetermined criterion. Such a predetermined criterion may be any criterion as long as the certainty factor Cr of the candidate character retains the property that the same character as the candidate character indicates the probability that the character exists in a specific part in the original document. The certainty factor Cr may be expressed by an integer expression instead of a real number expression. Alternatively, the certainty factor Cr may be represented by a symbol that indicates the level of the certainty factor Cr stepwise (for example, ◯: high, Δ: medium, x: low).
[0065]
When the appearance probability for each character is unknown, the character-specific coefficient Kr may be constant for all characters. Further, a character-specific coefficient Kr may be set for each character type (kanji, katakana, hiragana).
[0066]
Step S403: The candidate character and the certainty factor Cr obtained in step S402 are registered in the candidate character-confidence table.
[0067]
FIG. 5B shows an example of the candidate character-credibility table 1501. The candidate character and the certainty factor Cr are registered in the candidate character-confidence table 1501 for each character position 1042 (FIG. 4).
[0068]
Referring to FIG. 5A again, the description of the index table creation process will be continued.
[0069]
Step S404: It is determined whether or not the processing of steps S401 to S403 has been performed for all candidate characters at all character positions. If the determination result in step S404 is “Yes”, the process proceeds to step S405. If the determination result in step S404 is “No”, the processing from step S401 is performed on other candidate characters.
[0070]
Step S405: An index table is created from candidate characters registered at adjacent character positions in the candidate character-confidence table 1501 (FIG. 5B). The index table defines an index character string, a character position, and a certainty factor Cr.
[0071]
The index character string is generated by combining candidate characters registered at adjacent character positions in the candidate character-confidence table 1501 (FIG. 5B). For example, by combining the candidate character “I” registered at the character position “1” of the candidate character-confidence table 1501 with the candidate character “B” registered at the adjacent character position “2”, the index character The column “Ishi” is generated.
[0072]
FIG. 6 shows an example of the index table created by the index table creation process. A column 1610 of the index table 190 shows an index character string. A column 1611 indicates the character position of the first candidate character included in the index character string. A column 1612 shows a set of certainty factors Cr defined for each of the candidate characters included in the index character string.
[0073]
The row 1602 included in the index table 190 has a position of “1” in the character recognition result 140 matching the index character string “Ishi”, and is defined for the character “I” of the index character string “Ishi”. The certainty factor Cr is 0.9, and the certainty factor Cr defined for the character “I” in the index character string “Ishi” is 0.8.
[0074]
A set of certainty factors Cr defined for each candidate character included in the index character string “Ishi” (line 1602) is obtained as a set of certainty factors Cr calculated in step S402 (FIG. 5A) for each candidate character. . As a set of certainty factors Cr, a set of values obtained by multiplying the certainty factor Cr calculated in step S402 (FIG. 5A) for each candidate character by a coefficient for each index character string may be used. The coefficient for each index character string can be set low, for example, for an index character string that has a low probability of appearing in a general document. For example, the character string “Wo” and the character string “ヰ ヰ” have a low probability of appearing in a general Japanese document. For such an index character string, the coefficient for each index character string can be set low.
[0075]
Generating an index character string by combining candidate characters registered at adjacent character positions in the candidate character-confidence table 1501 (FIG. 5B) indicates that a plurality of fields shown in the character recognition result 140 (FIG. 4) are displayed. Of these, it is equivalent to combining candidate characters included in each of a plurality of adjacent (continuous) columns (for example, the column 1045 and the column 1046).
[0076]
Thus, in step S401 to step S405 shown in FIG. 5A, the CPU 110 (FIG. 1) functions as an index table creation unit that creates the index table 190.
[0077]
The index table 190 is created by registering all combinations of candidate characters registered at adjacent character positions in the candidate character-confidence table 1501 shown in FIG. 5B as index character strings.
[0078]
However, when there are overlapping combinations in all combinations of candidate characters registered at adjacent character positions in the candidate character-confidence table 1501, the index table 190 includes a plurality of character positions for one index character string. And a set of certainty factors Cr are registered. For example, the index character string “NW” is generated from the candidate characters “N” and “W” registered in the character positions “2” and “3” of the candidate character-confidence table 1501, and the character positions “4” and “4” The index character string “NW” is also generated from the candidate characters “N” and “W” registered in “5”. In this case, for one index character string “NW”, the character position 2, confidence factor Cr (0.7, 0.2) and character position 4, confidence factor Cr (0.1, 0.2) are index tables. 190 is registered (line 1604).
[0079]
The number of characters included in the index character string is determined in advance. In the example shown in FIG. 6, the number of characters included in the index character string is “2”. The number of characters included in the index character string can be any natural number. However, in general, the number of characters included in the index character string is preferably 2 or more. This is because if the number of characters included in the index character string is 1, the number of character positions and certainty factors Cr registered for one index character string increases, and the search cannot be performed at high speed.
[0080]
The index character strings in the index table 190 are preferably ordered according to a predetermined order to facilitate searching.
[0081]
The index table 190 is created using a plurality of candidate characters for one character position in the character recognition result 140 shown in FIG. As a result, the index table 190 includes a plurality of index character strings corresponding to the same character position. Therefore, a plurality of index character strings can match one part of the character recognition result. For example, the index character string “Ishi” shown in the row 1602 of the index table 190 and the index character string “in” shown in the row 1603 of the index table 190 are both character recognitions indicated by the character position “1”. It matches the portion in the result 140 (the portion including the column 1045 and the column 1046). This can reduce search omissions.
[0082]
In this way, the index character string and the character recognition result portion match that each character included in the index character string is one of the consecutive portions of the character recognition result (adjacent column shown in FIG. 4). Includes the concept that it is the same as one of at least one candidate character included in (character code is equal).
[0083]
An index table including a plurality of index character strings corresponding to the same character position, such as the index table 190, can also be created from a character recognition result in which there is one candidate character for one character position.
[0084]
FIG. 7 shows an example of a character recognition result 140a in which there is one candidate character for one character position. The character recognition result 140a differs from the character recognition result 140 shown in FIG. 4 in that there is one candidate character for one character position. The character recognition result 140a indicates that the characters “I”, “N”, “De”, “T”, “K”, “S” included in the original document 1310 (FIG. 3) are “I”, “ This indicates that “SHI”, “TE”, “SO”, “TA”, and “SU” are recognized. The characters “n”, “de”, “t”, and “ku” included in the original document 1310 are erroneously recognized.
[0085]
FIG. 8A shows a procedure of processing for creating an index table (index table creation processing) from the character recognition result 140a shown in FIG.
[0086]
Step S501: With reference to the certainty factor table, a pair of similar characters and certainty factor Cr is obtained from one set of candidate characters and reliability Rr for one character position of character recognition result 140a. A plurality of pairs of similar characters and certainty factors Cr may be obtained. The process of step S501 is performed for each character position. The certainty factor table will be described later with reference to FIG.
[0087]
Step S502: It is determined whether or not the certainty factor Cr of the similar character is greater than or equal to a predetermined reference value. The predetermined reference value is, for example, 0.05. If the determination result in step S502 is “Yes”, the process proceeds to step S503. If the determination result in step S502 is “No”, the process proceeds to step S504.
[0088]
Step S503: Similar characters and the certainty factor Cr obtained in step S501 are registered in the similar character-confidence table.
[0089]
FIG. 8B shows an example of the similar character-credibility table 1801. The similar character and the certainty factor Cr are registered in the similar character-confidence table 1801 for each character position 1042 (FIG. 7).
[0090]
With reference to FIG. 8A again, the description of the index table creation processing will be continued.
[0091]
Step S504: It is determined whether or not the processes in steps S502 to S503 have been performed for candidate characters at all character positions. If the determination result in step S504 is “Yes”, the process proceeds to step S505. If the determination result in step S504 is “No”, the processing from step S502 is performed on other similar characters.
[0092]
Step S505: An index table is created from similar characters registered at adjacent character positions in the similar character-confidence table 1801 (FIG. 8B). The index character string is generated by combining similar characters registered at adjacent character positions in the similar character-credibility table 1801 (FIG. 8B). This process is the same as the process of creating the index table 190 (FIG. 6) from the candidate character-confidence table 1501 (FIG. 5B) in step S405 shown in FIG. 5A.
[0093]
The generated index table is the same as the index table 190 shown in FIG. For example, in the row 1602 of the index table 190, a column 1611 indicates the character position of the first similar character “I” included in the index character string “Ishi”. A column 1612 shows a set (0.9, 0.8) of certainty factors Cr defined for each of the similar characters included in the index character string “Ishi”.
[0094]
FIG. 9 shows an example of the certainty factor table 150. FIG. 9 shows only the portion related to the candidate character “si” in the certainty factor table 150.
[0095]
For example, when one candidate character “shi” and reliability Rr “0.9” are obtained as a character recognition result, the certainty factor table 150 is similar character “n” and certainty factor Cr “0.2”. , The similar character “shi” and the certainty factor Cr “0.8” are obtained. The similar character “n” and the similar character “si” are characters having the same or the same character shape as the candidate character “si”.
[0096]
The fact that the similar characters for the candidate character “si” are “n” and “si” indicates that when one candidate character “si” is obtained as a character recognition result, the similar character “ "Or a similar character" shi "is likely to be written.
[0097]
The certainty factor table 150 performs character recognition on an original document in which a large number of characters are written in advance, and compares the character recognition result and reliability Rr obtained thereby with characters actually existing in the original document. Can be created. For example, “0.2” of the certainty factor Cr shown in the portion 1811 of the certainty factor table 150 is obtained when character recognition is performed for a character “N” written in various fonts and various print qualities. It can be obtained from the probability that the candidate character “shi” and the reliability Rr0.9 are obtained.
[0098]
The certainty factor table 150 is prepared for all combinations of characters. However, similar characters whose certainty factor Cr is smaller than a predetermined reference need not be registered in the certainty factor table 150. Therefore, the number of similar characters obtained for one candidate character can be limited.
[0099]
When the reliability Rr obtained by character recognition does not match the reliability Rr defined in the certainty table 150 shown in FIG. 9 (for example, the reliability Rr obtained by character recognition is 0.8). The certainty factor Cr of similar characters is calculated by an appropriate method. For example, when the reliability Rr obtained by character recognition is smaller than 0.5, the row of the reliability Rr “0.5” in the certainty factor table 150 is referred to. When the reliability Rr obtained by character recognition is greater than 0.9, the row of the reliability Rr “0.9” in the certainty factor table 150 is referred to. When the reliability Rr obtained by the character recognition is a value between the two reliability levels Rr defined in the certainty degree table 150, the character recognition is performed among the two reliability levels Rr defined in the certainty degree table 150. A row having a value close to the reliability Rr obtained by is referred to.
[0100]
Note that the structure of the certainty factor table 150 is not limited to the structure shown in FIG. The certainty level table 150 may have an arbitrary structure as long as at least one similar character and certainty level Cr group is obtained from the set of candidate characters and the reliability level Rr. For example, assuming that the distribution of the reliability Cr with respect to the reliability Rr is a uniform distribution, the distribution range is expressed by the upper and lower limits of the reliability and the upper and lower limits of the reliability Rr. May be defined in the certainty factor table 150. Alternatively, assuming that the distribution of the reliability Cr with respect to the reliability Rr is a Gaussian distribution, the average value and the variance of the distribution may be defined in the reliability table 150.
[0101]
By executing the index table creation processing shown in FIG. 8A using such a certainty factor table 150, the same character recognition result 140a (FIG. 7) having one candidate character for one character position is the same. An index table 190 (FIG. 6) including a plurality of index character strings corresponding to character positions can be created.
[0102]
The certainty factor table 150 may be omitted when the search device 1 executes the index table creation process according to the procedure shown in FIG. 5A.
[0103]
In this manner, the index table 190 can also be created using a plurality of similar characters that are associated in advance with one candidate character for one character position in the character recognition result 140a shown in FIG. As a result, the index table 190 includes a plurality of index character strings corresponding to the same character position. Therefore, a plurality of index character strings can match one part of the character recognition result. As a result, it is possible to reduce omissions even when erroneous recognition occurs in the character recognition process.
[0104]
However, each character included in the index character string is not necessarily the same as one candidate character included in one of consecutive portions of character recognition results (adjacent columns shown in FIG. 7). For example, the character “N” included in the index character string “IN” shown in the row 1603 of the index table 190 is not the same as one candidate character “SE” included in the column 1046a of the character recognition result 140a (FIG. 7). Absent. However, the character “n” included in the index character string “in” is the same as the similar character “n” previously associated with the candidate character “si” by the certainty factor table 150 (FIG. 9).
[0105]
In this way, the index character string and the character recognition result portion match that each character included in the index character string is one of the consecutive portions of the character recognition result (adjacent column shown in FIG. 4). Including the concept of being the same as one of at least one character previously associated with one candidate character included in the character (equal character code).
[0106]
Next, a process (document search process) for searching for a keyword from a character recognition result using the index table 190 (FIG. 6) will be described.
[0107]
FIG. 10 shows the procedure of the document search process. Hereinafter, each step of the document search process will be described in detail.
[0108]
Step S301: A keyword is input. Hereinafter, a case where the keyword is a character string “index” will be described as an example.
[0109]
Step S302: A set of two consecutive characters (a character string having a length of 2) is extracted from the keyword. In this example, a pair of two characters “IN”, “NDE”, “DET”, “KOK”, and “KUS” are extracted. Note that the length of the extracted character string is set to be equal to the length of the index character string defined in the index table. Therefore, when the length of the index character string is n (n is a natural number), a set of n characters (a character string having a length of n) is extracted from the keyword. In the following description, it is assumed that n = 2.
[0110]
The extracted sets of two characters overlap each other in part. However, a set of two characters may be extracted from the keyword so as not to overlap. For example, a set of two characters “IN”, “DET”, and “KUS” may be extracted from the keyword “index”. However, a set of two characters is extracted from the keyword so that each character included in the keyword is included in at least one of the two sets of extracted characters.
[0111]
Step S303: Referring to the index table 190 (FIG. 6), the character position and the certainty factor Cr corresponding to the set of two characters are extracted. In this example,
A character position “1” corresponding to a two-character set “IN”, a set (0.9, 0.7) of confidence factor Cr (line 1603),
A character position “2” corresponding to a two-character set “Nde”, a set (0.7, 0.8) of confidence Cr (line 1605),
A character position “3” corresponding to a two-character set “de”, a set of confidence factors Cr (0.8, 0.3) (line 1606),
A character position “4” corresponding to a two-character set “Cook”, a set of confidence factors Cr (0.3, 0.1) (line 1607), and
A set (0.1, 0.9) (line 1608) of the character position “5” and the certainty factor Cr corresponding to the two-character set “kus” is obtained.
[0112]
In order to efficiently extract the character position and the certainty factor Cr corresponding to each character string from the index table 190 shown in FIG. 6, the character code of the character included in the index character string and the line including the index character string are displayed. A correspondence table with stored addresses (for example, addresses on the HDD 170) may be used. Such an address may be obtained using a binary tree search method.
[0113]
Step S304: It is determined whether or not the process of step S303 has been performed for all sets of two characters. If the determination result in step S304 is “No”, the process of step S303 is performed for the other two character sets. If the determination result in step S304 is “Yes”, the process proceeds to step S305.
[0114]
Step S305: It is determined whether or not all sets of two characters are arranged in a predetermined order. This determination is made based on the character position obtained for each two-character group in step S304. Specifically, the character position m (m is a natural number) obtained for a set of two characters starting from the k-th character (k is a natural number) of the keyword is “m−k = If the relationship of “constant” is satisfied, it is determined that all sets of two characters are arranged in a predetermined order.
[0115]
The fact that all two character sets are arranged in a predetermined order indicates that the keyword matches a specific part in the character recognition result. The specific portion is a portion including a portion in the character recognition result in which each character included in the keyword matches.
[0116]
In this example, the keyword “index” matches the portion 1047 of the character recognition result 140 (FIG. 4) or the portion 1047a of the character recognition result 140a (FIG. 7).
[0117]
The position of the part 1047 or the part 1047a is specified as the character position “1” in the head column of the part.
[0118]
In this example, it is determined that all sets of two characters extracted from the keyword “index” are “in a predetermined order” in order to satisfy the relationship described above.
[0119]
If the determination in step S305 is “Yes”, the process proceeds to step S306. If the determination in step S305 is “No”, the process proceeds to step S308.
[0120]
As described above, in steps S302 to S305, the CPU 110 (FIG. 1) is based on the index character string included in the index table 190 (FIG. 6) and the position of the portion in the character recognition result that matches the index character string. Then, it is determined whether or not the keyword matches the part in the character recognition result, and if it matches, it functions as a position specifying unit that specifies the position of the part in the character recognition result that matches the keyword.
[0121]
Step S306: The keyword certainty factor Kc is calculated. The keyword certainty factor Kc is, for example, a certainty factor Cr corresponding to each of the two-character sets “IN”, “NDE”, “DET”, “Cuck”, and “CUS” extracted from the keyword “index”. It is obtained as an arithmetic average of the value on the left side of the set and the value on the right side of the set of certainty factors Cr corresponding to the last two character set “kus” constituting the keyword “index”. This is equivalent to obtaining an arithmetic average of the certainty levels defined for each character included in the keyword. In this example, the keyword certainty factor Kc = (0.9 + 0.7 + 0.8 + 0.3 + 0.1 + 0.9) /6=0.61.
[0122]
The keyword certainty factor Kc may be calculated by a geometric mean, a median value, or a mode value. The keyword certainty factor Kc may be calculated using only the smaller value of the certainty factor Cr pairs corresponding to the two character pairs. When the certainty factor Cr is less than a predetermined reference value, the certainty factor Cr may not be used for calculating the keyword certainty factor Kc.
[0123]
Thus, the keyword certainty factor Kc is calculated based on the certainty factor Cr defined for each character included in the keyword.
[0124]
The calculated keyword certainty factor Kc indicates the probability that the keyword exists at a position in the original document corresponding to the character position in the character recognition result.
[0125]
For example, in step S305, it is determined that the keyword “index” matches the portion 1047 of the character recognition result 140 (FIG. 4) or the portion 1047a of the character recognition result 140a (FIG. 7). When the position is specified as the character position “1”, the probability that the keyword “index” exists at the position in the original document corresponding to the character position “1” (that is, at the beginning of the original document) is 0.61. .
[0126]
As described above, in step S306, the CPU 110 determines the keyword at the position in the original document corresponding to the position of the portion in the character recognition result that matches the keyword based on the certainty factor Cr included in the index table 190 (FIG. 6). It functions as a calculation unit that calculates the keyword certainty factor Kc indicating the probability of the presence of.
[0127]
Step S307: It is determined whether or not the keyword certainty factor Kc is greater than or equal to a reference value (predetermined value). The reference value may be 0.5, for example. This reference value may be a fixed value or set according to a keyword. For example, the reference value may be changed according to the number of characters of the keyword.
[0128]
If this reference value is increased, search noise can be reduced. However, if it is too high, search omission is likely to occur. In general, when the number of characters in a keyword is large, it is preferable to reduce the omission of search even when misrecognition is frequent by setting the reference value low. This is because when the number of characters in the keyword (the number of characters included in the keyword) is large, the search noise does not increase so much even if the reference value is set low.
[0129]
The reference value may be changed according to the type of characters (character type) included in the keyword. For example, if each character of the keyword is katakana, kanji, mixed kanji and katakana, or hiragana, the reference value is set optimally for more effective search. It can be performed.
[0130]
Alternatively, this reference value may be specified by the user. The user can set an appropriate reference value according to the purpose of preventing search omission or reducing search noise.
[0131]
If the determination in step S307 is “Yes”, the process proceeds to step S309. If the determination in step S307 is “No”, the process proceeds to step S308.
[0132]
As described above, after a search result is obtained that the keyword matches the character recognition result 140 or the character recognition result 140a in step S305, step S307 is executed to determine whether or not the keyword exists in the actual original document ( Whether the search result is valid) is determined based on the keyword certainty factor Kc. In step S307, the CPU 110 functions as a determination unit that determines the validity of the search result based on the keyword certainty factor Kc.
[0133]
Step S308: It is determined that the keyword does not exist in the original document. For example, a message “Keywords could not be found” is displayed on the display of the terminal 100 (FIG. 1).
[0134]
Step S309: It is determined that a keyword has been detected. As a search result, the character position indicating the detected location and the keyword certainty factor Kc are obtained. When there are a plurality of detection locations, a plurality of sets of character positions and keyword certainty factors Kc are obtained.
[0135]
In the example described above in which the keyword is “index”, the character position “1” and the keyword certainty factor Kc “0.61” are obtained as search results.
[0136]
The search result is displayed on the terminal 100, for example. For example, the search device 1 displays the document image data 130 (and / or the character recognition results 140 and 140a) stored in the HDD 170 on the display of the terminal 100, and the document image data 130 (and / or / displayed on the display). Alternatively, the area corresponding to the keyword is highlighted in the areas of the character recognition results 140 and 140a). The highlighting is performed, for example, by changing the attributes of the displayed characters (for example, the character color and density, the character background color and density, the character size, the character thickness, the font, etc.). Such attributes may be changed according to the keyword certainty factor Kc. For example, the keyword certainty factor Kc between 0.5 and 1.0 may be divided with a step size of 0.1, and different attributes may be set for each division to perform highlighting. In this case, since the user can visually grasp the magnitude of the keyword certainty factor Kc, there is an advantage that the user can easily make a further determination of the validity of the search result visually. .
[0137]
Or you may display the area | region corresponding to a keyword in an order from the detection location with high keyword reliability Kc.
[0138]
In this manner, when the user further determines the validity of the search result, the reference value used in step S307 may be set low.
[0139]
Alternatively, step S307 may be omitted. In this case, all of the determination of the validity of the search result is performed by the user. The user can easily determine the validity of the search result based on the keyword certainty factor Kc.
[0140]
Hereinafter, an example in which search noise is suppressed by the document search process shown in FIG. 10 will be described.
[0141]
When the keyword “wax” is specified and the document search process shown in FIG. 10 is performed using the index table 190 shown in FIG. 6, the determination in step S305 is “Yes”, and the character position “3” is specified. Is done.
[0142]
In step S306, the keyword certainty factor Kc = (0.2 + 0.3 + 0.1 + 0.9) /4=0.38 is calculated.
[0143]
Since the keyword certainty factor Kc is smaller than the reference value 0.5, it is determined that no keyword exists.
[0144]
When the keyword “denwa” is designated, the determination in step S305 is “Yes”, and the character position “3” is specified.
[0145]
In step S306, the keyword certainty factor Kc = (0.8 + 0.1 + 0.2) /3=0.37 is calculated.
[0146]
Since the keyword certainty factor Kc is smaller than the reference value 0.5, it is determined that no keyword exists.
[0147]
Similarly, when the keyword “hook” is designated, the determination in step S305 is “Yes”, and the character position “3” is specified.
[0148]
In step S306, the keyword certainty factor Kc = (0.2 + 0.3 + 0.1) /3=0.2 is calculated.
[0149]
Since the keyword certainty factor Kc is smaller than the reference value 0.5, it is determined that no keyword exists.
[0150]
As described above, according to the search device 1 of the present invention, it is possible to suppress the detection of the keyword even when the keyword does not exist in the original document, that is, to suppress the search noise.
[0151]
The document search process of the present invention is not limited to being realized by software on a computer. The document search processing of the present invention may be realized by hardware, or may be realized by a combination of software and hardware.
[0152]
In the embodiment described above, a Japanese document has been described as an example. However, the application of the present invention is not limited to Japanese documents. The present invention can also be applied to other arbitrary documents (for example, Chinese documents, English documents, Korean documents).
[0153]
【The invention's effect】
According to the present invention, based on the certainty factor included in the index table, the keyword certainty factor indicating the probability that the keyword exists at the position in the original document corresponding to the position of the part in the character recognition result that matches the keyword is calculated. Is done. Therefore, it becomes easy to determine the validity of the search result based on the keyword certainty factor.
[0154]
Since the search device of the present invention uses an index table, it is possible to perform a high-speed search.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of a search device 1 according to the present invention.
FIG. 2 is a diagram illustrating a flow of processing executed by the search device 1 to search for a keyword from a character recognition result obtained by recognizing a character in an original document.
FIG. 3 is a diagram showing an example of an original document 1310
4 is a diagram showing an example of a character recognition result 140 obtained by performing character recognition processing on an original document 1310. FIG.
FIG. 5A is a flowchart showing the procedure of index table creation processing;
FIG. 5B is a diagram showing an example of a candidate character-confidence table 1501
FIG. 6 is a view showing an example of an index table 190 created by index table creation processing;
FIG. 7 is a diagram showing an example of a character recognition result 140a in which there is one candidate character for one character position.
FIG. 8A is a flowchart showing a processing procedure for creating an index table from the character recognition result 140a shown in FIG.
FIG. 8B is a diagram showing an example of a similar character-confidence table 1801
FIG. 9 is a diagram illustrating an example of a certainty factor table 150;
FIG. 10 is a flowchart showing a procedure of document search processing.
FIG. 11 is a diagram showing an example of an index table 1901 in which a plurality of character strings that may exist in an original document are registered as index character strings according to a conventional technique.
[Explanation of symbols]
1 Search device
100 terminals
110 CPU
120 Image input device
130 Document image data
140 Character recognition result
170 HDD
180 Work memory
190 Index table

Claims (7)

Retrieval for recognizing a keyword consisting of a plurality of characters from the character recognition result of the original document having at least one candidate character corresponding to each character by recognizing each character in the original document using an index table A device,
The index table includes an index character string composed of a character string combining the candidate characters corresponding to each of a plurality of characters constituting the character string existing in the original document, and the index character in the original sentence. A column position and, for each candidate character included in the index character string, including a certainty factor defined in advance as a probability that each candidate character exists in the original document,
It is determined whether a character string of the keyword having the same number of characters as the index character string exists in the index table, and if present, based on the position of the index character string that matches the character string of the keyword, A position specifying unit for specifying the position of the character string of the keyword in the original sentence;
A calculation unit that calculates a keyword certainty factor defined as a probability that the keyword exists at the specified position based on a certainty factor of an index character string that matches the character string of the keyword;
And a determination unit for determining validity of the search results based on the keyword credibility, retrieval device.   The search device according to claim 1, wherein the determination unit determines that the search result is valid when the keyword certainty factor is equal to or greater than a predetermined value.   The search device according to claim 2, wherein the predetermined value is set according to at least one of the number of characters included in the keyword and the type of character included in the keyword. An index table creating unit for creating the index table;
The index table creation unit, when a plurality of candidate characters are generated based on the character recognition result, by combining each of the candidate characters generated corresponding to each character in the character string of the original document The search device according to claim 1, wherein the index character string is generated. An index table creating unit for creating the index table;
The index table creation unit generates one candidate character for each character based on the character recognition result, and generates a similar character similar to each generated candidate character based on the certainty factor. The search device according to claim 1, wherein the index character string is also generated by a combination of the candidate character corresponding to each character in the character string of the original document and the similar character or a combination of the similar characters. . Retrieval for recognizing a keyword consisting of a plurality of characters from the character recognition result of the original document having at least one candidate character corresponding to each character by recognizing each character in the original document using an index table A computer-readable recording medium recording a program for causing a computer to execute processing,
The index table includes an index character string composed of a character string combining the candidate characters corresponding to each of a plurality of characters constituting the character string existing in the original document, and the index character in the original sentence. A column position and, for each candidate character included in the index character string, including a certainty factor defined in advance as a probability that each candidate character exists in the original document,
The search process
It is determined whether a character string of the keyword having the same number of characters as the index character string exists in the index table, and if present, based on the position of the index character string that matches the character string of the keyword, Identifying the position of the keyword string in the original document;
Calculating a keyword certainty factor defined as a probability that the keyword exists at the specified position based on a certainty factor of an index character string that matches the keyword character string;
Including determining the validity of the search results based on the keyword credibility, recording medium recording a program. Retrieval for recognizing a keyword consisting of a plurality of characters from the character recognition result of the original document having at least one candidate character corresponding to each character by recognizing each character in the original document using an index table A program for causing a computer to execute processing,
The index table includes an index character string composed of a character string combining the candidate characters corresponding to each of a plurality of characters constituting the character string existing in the original document, and the index character in the original sentence. A column position and, for each candidate character included in the index character string, including a certainty factor defined in advance as a probability that each candidate character exists in the original document,
The search process
It is determined whether a character string of the keyword having the same number of characters as the index character string exists in the index table, and if present, based on the position of the index character string that matches the character string of the keyword, Identifying the position of the keyword string in the original sentence;
Calculating a keyword certainty factor defined as a probability that the keyword exists at the specified position based on a certainty factor of an index character string that matches the keyword character string;
Determining the validity of the search result based on the keyword certainty factor.

Images