JP2001175661A - Device and method for full-text retrieval - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that a character recognition result has an error with higher probability and retrieval is not correctly performed/frequently since a key word does not match characters in a text when an index using only a 1st recognition candidate character of a character recognition result is generated from a text generated as a result of character recognition. SOLUTION: While a document of recognition candidate characters matching the key word is retrieved by referring to indexes, shape features of a character image and shape features of the characters constituting the key word are compared with each other to retrieve a document which meets retrieval conditions.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a full-text search apparatus and a full-text search apparatus for performing a full-text search by using an arbitrary keyword from document / drawing data created by identifying a character image described in a document or drawing. It relates to the search method.

[0002]

2. Description of the Related Art A method of accumulating digitized text which can be read by a computer and performing a retrieval process of the digitized text by using a keyword includes (1) a method of directly collating the text content and the keyword one by one; (2) There is a method in which a character appearing in a text and its position information are extracted in advance to create an index, and a positional relationship between a keyword and a character in the text is tested using the index at the time of search.

[0003] In the above (2), the index can be largely classified from the unit of a character string for creating an index into an index in units of consecutive N (N is an integer) characters and an index in units including grammatical elements such as words and morphemes. . Further, there are a method of describing a text number or the like from the description content of the position information, and a method of describing an appearance position of a character in the text in addition to the text number.

In the above (1), in order to collate a text with a keyword at high speed, it is necessary to develop the text in a memory. However, when the number of stored texts increases, the time required to develop the text in the memory increases. Because of the length, there is a problem that high-speed search cannot be performed. However, since it is not necessary to create an index in advance, it is convenient when registering and deleting frequently. In the above (2), since it is necessary to create an index in advance, it takes more time to register and delete than in the above (1).
Less than in (1) above. For this reason, registration and deletion are not performed very frequently, and are suitable for handling a large number of documents.

FIG. 21 shows, for example, JP-A-10-149367.
Of the related art full text search device disclosed in
FIG. 2 is a block diagram showing a conventional example 1 in which the index creation method related to the above (2) is applied. In the figure, 201 is a text storage means, 202
Is a main index registration unit, 203 is a sub index registration unit, 204 is a main index storage unit, 205 is a sub index storage unit, 206 is a sub index management unit, 207 is a main index search unit, 208 is a sub index search unit, and 209 is a sub index search unit. Keyword search control means,
210 is a keyword search result storage unit, 211 is a search condition input unit, 212 is a logical condition analysis unit, and 213 is a search result output unit.

Next, the operation will be described. The text stored by the text storage unit 201 is registered with an index of consecutive N characters by the main index registration unit 202 and stored by the main index storage unit 204.

At the time of a search, the keyword search control means 209 is used by using the search conditions obtained from the search condition input means 211.
Obtains a search result by searching the primary index and the secondary index. From the search result, the keyword search result storage unit 210 stores the number of search results (text identification number).
The sub-index creation unit 206 is activated to create a sub-index for those having a large number of words and those having a large ratio between the number of character positions in the text and the number of text identifications in the search result.

In the first conventional example, in addition to the main index of the N-character index, a sub-index is held, the sub-index is accessed first, and if a keyword does not exist in the sub-index, the main index is accessed.
The main index holds the document number and the character position number, and the sub index holds only the document number. For this reason, the secondary index is smaller in size than the main index, and requires less index test processing. When there is an N-character index of a keyword in the sub-index, there is no need to access the main index, and the search processing time is shortened. In addition, by deleting from the secondary index an index that is searched less frequently based on the search history, the size of the index can be reduced.

Next, in order to perform a search on a document image in which the document is not character-coded (no digitized text is created), character recognition processing is executed to extract a character portion from the document image. To create and save digitized text. For example, Japanese Patent Application Laid-Open No. 8-7033 discloses a technique of increasing the ratio of correct characters by holding a plurality of recognition candidate characters for each character image as a result of character recognition.

FIG. 22 is a block diagram showing a conventional full-text search device disclosed in Japanese Patent Application Laid-Open No. 8-7033 (hereinafter referred to as Conventional Example 2). In FIG.
222 is an output unit, 223 is a character recognition unit, 224 is a document search unit, 225 is a keyword input unit, 226 is image data, 227 is text information, and 228 is a search file.

Next, the operation will be described. In the second conventional example, when a document image is input from the image input unit 221, character recognition is performed using the character recognition unit 223, and the recognition candidate character is stored in the search file 228. In order to store a plurality of recognition candidate characters, the description of the search file 228 is described as [number of candidate characters] [candidate character 1] [candidate character 2] using the number of recognition candidate characters and the recognition candidate characters. .

For example, when storing a plurality of recognition candidate characters for a character image "new document filing",
[1] New [4] Jobun Girls [1] Book [1] F [1] A [1] I [1] L [1] [1] At the time of search, the document search means 224 sets the search file 228
The matching between the text and the keyword is executed, and if all of the same characters as the keyword are included in the recognition candidate characters, the matching is determined to be successful. For example, when the keyword “document” is searched for the text “new document filing”, “sentence” and “sho” are present in the recognition candidate characters [4] Therefore, the collation succeeds and the search result is output.

By combining Conventional Example 1 and Conventional Example 2, it becomes possible to create an index including recognition candidate characters and perform a search. For example, if N = 2, in the example of “New document filing” in Conventional example 2,
"New length", "New sentence", "New woman", "Shinko", "Jojo",
By creating an index using recognition candidate characters such as "document", "girl book", and "letters", it becomes possible to adapt to the first conventional example.

[0014]

Since the conventional full-text search apparatus is configured as described above, when an index is created from a text created as a result of character recognition, the first recognition candidate of the character recognition result is used. When an index using only characters is created, the probability that the character recognition result contains an error increases, and there is a problem that the keyword and the character in the text do not match, and the search is often not performed correctly.

In a search for actually collating text using recognition candidate characters as in Conventional Example 2, the probability that the correct character is included in the text is higher than when only the first-ranked recognition candidate character is held. However, the larger the amount of data, the longer the time required to load the text file into the memory, so that there was a problem that the search could not be performed at high speed.

Further, when an index is created and searched using recognition candidate characters, if all the correct characters are not included in the recognition candidate characters, the index of the correct character string cannot be created correctly. There was an issue that was not correctly searched. For example, if the recognition result of the character image "character recognition" incorrectly recognizes "characters" as "u", such as "bunu recognition", the indexes to be created are "bunu", "uho", " Recognition ", and the original" character "and" character recognition "indexes cannot be created. As a result, the keyword" character recognition "cannot be correctly searched.

Further, for example, if three recognition candidate characters are held for each character, the combinations for creating an index of two consecutive characters are 3 × 3 = 9, and the recognition candidate characters are stored one character at a time. Nine times that of holding. In a combination of three consecutive characters, 3 × 3 × 3 = 27 combinations. As the number of recognition candidate characters increases, the number of consecutive N
There is also a problem that the number of character combinations increases, and as a result, the capacity of the index becomes very large.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and has as its object to provide a full-text search device and a full-text search method capable of performing high-speed and high-precision full-text search. Another object of the present invention is to provide a full-text search device capable of reducing the capacity of an index.

[0019]

A full-text search device according to the present invention refers to an index to search for a document of a recognition candidate character that matches a keyword, and at the same time, shapes features of a character image extracted by a feature extraction unit. And a search unit for searching for a document that matches the search condition by comparing the shape characteristics of the characters constituting the keyword.

In the full-text search device according to the present invention, a concatenated character in which two or more recognition candidate characters are combined is included in an index creation target.

[0021] In the full-text search device according to the present invention, among the candidate characters output by the character recognizing means, recognition candidate characters whose accuracy is lower than the reference accuracy are excluded from the index creation target.

In the full-text search device according to the present invention, even if the accuracy of the recognition candidate character output by the character recognition means is lower than the reference accuracy, the recognition candidate character relating to the character image having no recognition candidate character with an accuracy exceeding the reference accuracy is provided. In the case of, the recognition candidate character is included in an index creation target, and an identification symbol for distinguishing the recognition candidate character from other recognition candidate characters is added.

The full-text search device according to the present invention stores the shape characteristics of the character image in the database, and stores the recognition candidate characters for the character image and the character codes of the characters that may form words in the database. It is like that.

The full-text search device according to the present invention is configured to determine each recognition candidate character and a character that may constitute a word, in consideration of linguistic information or a type of character.

A full-text search apparatus according to the present invention calculates a distance between a shape feature of a character image extracted by a feature extraction unit and a shape feature of a character constituting a keyword, and searches when the distance satisfies a predetermined criterion. It is designed to recognize that the conditions are met.

The full-text search device according to the present invention is provided with setting means for setting whether or not to execute the shape feature collation processing by the search means.

A full-text search device according to the present invention excludes a document including a recognition candidate character matching a keyword from a target of shape feature matching.

The full-text search device according to the present invention compares the shape features of the character image extracted by the feature extraction unit with the shape features of the characters constituting the keyword only when there is no recognition candidate character matching the keyword. It is like that.

In the full-text search device according to the present invention, when specifying a collation target of a shape feature with respect to a keyword, a recognition candidate character to which an identification code is added is treated as a wild card.

[0030] The full-text search device according to the present invention determines whether or not to include a concatenated character, which is a combination of two or more candidate recognition characters, in the index creation target, in consideration of the appearance probability of appearance of the concatenated character in the entire document. Is determined.

In the full-text search device according to the present invention, when each of the recognition candidate characters constituting the concatenated character is the only recognition candidate character for the character image, the number of appearances of the concatenated character is counted up to increase the appearance probability. It is intended to be updated.

In the full-text search device according to the present invention, the appearance probability is updated by counting up the number of appearances of the connected character that matches the keyword.

[0033] In the full-text search device according to the present invention, when the recognition candidate character output by the character recognition means is corrected, the appearance frequency is updated by counting up the number of occurrences of the concatenated character including the corrected recognition candidate character. It is like that.

In the full-text search method according to the present invention, while referring to an index, a document of a recognition candidate character matching a keyword is searched, and a shape characteristic of a character image is compared with a shape characteristic of a character constituting a keyword. , A document that matches the search condition is searched.

[0035]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a full-text search apparatus according to Embodiment 1 of the present invention. In the figure, reference numeral 1 denotes an image input means for inputting an image, and 2 denotes each character image included in the input image. A character recognition unit that outputs one or more recognition candidate characters for a character image and outputs the accuracy (similarity) of each recognition candidate character. Reference numeral 3 denotes a correspondence relationship between each recognition candidate character output by the character recognition unit 2 and a character position. This is an index creating means for creating an index indicating the following.

Reference numeral 4 denotes a case where, among character images included in the input image, there is a character image having no recognition candidate character having a certainty exceeding the reference certainty, a shape characteristic of the character image is extracted and linguistic information is extracted. Or, considering the type of characters,
Ambiguous text extraction means (feature extraction means) for determining a character that may constitute a word (character string) and a recognition candidate character for the character image, and extracting the character string as an ambiguous text; The search condition input means (input means, setting means) 6 for inputting a keyword as a search for the document number of the recognition candidate character matching the keyword with reference to the index, while being extracted by the vague text extraction means 4 Retrieval means for comparing the shape characteristics of the character image with the shape characteristics of the characters constituting the keyword and searching for a document number matching the document search conditions, and an output means 7 for outputting a search result of the search means 6.

Reference numeral 8 denotes a character recognition dictionary used by the character recognition means 2 for character recognition, 9 denotes a shape feature dictionary used by the search means 6 at the time of keyword search, and 10 denotes an ambiguous text extracted by the ambiguous text extraction means 4. An ambiguous text database, 11 is an index database for storing the index created by the index creation means 3, and 12 is a recognition character database for storing recognition candidate characters and the like.

Next, the operation will be described. First, FIG.
The document registration method will be described with reference to FIG. First, in step ST100, the image input unit 1 inputs a document image that can be processed by a computer. As a configuration of the image input unit 1, a scanner, a digital camera, or the like may be used, or a computer-generated image created in advance may be input via a network or the like. here,
It is assumed that the document image shown in FIG.

Next, in step ST110, the character recognizing means 2 performs a character recognizing process on the input image input from the image input means 1, and outputs a character code and a similarity indicating its certainty. Since the method of character recognition can be performed by using a known technique, details are omitted. The character recognition means 2 outputs a plurality of recognition candidate characters and respective degrees of similarity to each character image included in the input image.

FIG. 4 shows a part of the recognition result of the character recognizing means 2. Here, the recognition results of the respective character images from the first line to the second line in FIG. It shows the recognition candidate characters up to the rank and their similarity. In FIG. 4, “◆” existing in the recognition candidate character means that the corresponding character code is not stored.

Next, in step ST120, the index creating means 3 narrows down the recognition candidate characters used for the search from the recognition results shown in FIG. As a method of narrowing down the recognition candidate characters, for example, the similarity value of the recognition candidate character and the recognition candidate character are obtained in advance from the learning data, the probability that the recognition candidate character is a correct answer, the probability of the correct answer is high,
In addition, a threshold value TH1 at which sufficient narrowing can be performed is set, and recognition candidate characters having a similarity equal to or greater than the threshold value TH1 are stored.

When there is no recognition candidate character having a similarity of the threshold value TH1 or more, there is a high probability that a correct character is not included, indicating that there is a high possibility that a correct character is not included in addition to each recognition candidate character. Add the “*” symbol. Although “*” is used in this example, another character code may be assigned, or a value other than the character code may be assigned. FIG. 5 shows the result of narrowing down the recognition candidate characters.
For example, if TH1 = 80 is set, there is no recognition candidate character having a similarity of 80 or more for character position number 4 and character position number 9 (see FIG. 4). "(Reference numeral 2 in FIG. 5).
3, 24). The index creation means 3 converts the narrowed-down recognition candidate characters shown in FIG.
To save.

Next, in step ST130, the index creating means 3 creates an index. Here, an index for each character and an index for two consecutive characters are created from the recognition candidate characters shown in FIG. Here, a method of creating an index will be specifically described. FIG. 9 shows a two-character index created by the index creating means 3 from the recognition candidate characters shown in FIG. The creation method is as follows. For characters adjacent to each other in order from the first character in FIG. 5, the character code of the character before and after the two adjacent characters, the appearance position of the previous character, the recognition candidate order of the previous character Calculate and store the product of the recognition candidate rank of the subsequent character. The appearance position is described as “XY”, and means the Y-th character from the beginning of the document number X. Here, the document number of the document image in FIG. 3 is “1”.

For example, an index of “document” 25 in FIG. 9 is created from “sentence” 21 and “book” 22 in FIG. In this case, since the position information of the "sentence" 21 is the first character from the beginning of the document 1, the character position is "1-1" and the "sentence" 2
Since the recognition candidate ranks of 1 and 22 are both first,
The recognition candidate rank is 1 × 1 = 1. FIG. 10 is a table storing the positions of one-character indexes and the recognition order.
The character code, the character appearance position, and the recognition candidate rank are held. For a character determined not to include the correct character code, “*” 31 and a character position 32 are retained.

Next, in step ST140, the ambiguous text extracting means 4 extracts an ambiguous text including a character which does not include the correct character code. That is, the ambiguous text extracting means 4 creates the character shape feature from the character image of the character code with “*” from the recognition candidate characters shown in FIG. 5 and stores it in the ambiguous text database 10 together with several characters before and after the character. Store. The method of determining the characters before and after is, for example,
By performing a known morphological analysis, a character whose morphological analysis has failed from before and after the character code marked with “*” may be used, or may be in the same category (alphabet,
(Kanji, numbers, hiragana, katakana) may be continuous characters, or the number of characters may be fixed. Here, the last character is retained.

FIG. 8 shows a specific method for creating a shape feature. In FIG. 8, the image of a character image is divided into eight parts, and the number of black pixels in each area is obtained. For example, area 41
, The number of black pixels is 13 (see reference numeral 49),
For 2, the number of black pixels is determined as 10 (see reference numeral 50). The shape features thus created are stored together with the recognition candidate characters. FIG. 6 shows an example in which the shape features extracted from the character images of the fourth and ninth characters are stored. Further, the ambiguous text extracting means 4 stores the position of the character for which the shape characteristic has been created and its characteristic value in the recognized character database 12 (see the lower part of FIG. 5).

Next, a document search method will be described. Here, as a result of the document registration processing, the index database 1
1 and the ambiguous text database 10 only store data relating to the document of document number 1.
FIG. 11 is a flowchart showing a document search method.

First, in step ST 200, the user inputs a keyword using search condition input means 5. The search condition input means 5 can be configured using a computer keyboard or a mouse, but is not limited thereto, and voice input using a microphone, a telephone, or the like is also possible. Here, it is assumed that a keyword “character” is input. Next, in step ST210, the search means 6 divides the input keyword. Here, it is decomposed into a set of one-character and two-character concatenated character strings. That is, "sentence",
Divide into "characters" and "characters".

Next, in step ST220, the search means 6 searches for a document using the index. FIG. 12 is a flowchart showing index matching. First, in step ST221, the search unit 6
Performs a process of extracting the respective indexes of the divided “character”, “sentence”, and “character” (see reference numerals 26 and 27 and 28 in FIG. 9 and 27 and 28 in FIG. 10). Specifically, the content of each index is loaded on a memory (not shown).

Next, in step ST222, the position of the character is verified to search for the document number. That is,
The document number may be searched by verifying the character positions of “sentence” and “character”, but the document number may be searched using the index 26 of “character”. here,
A search is performed using the “character” index 26. In this case, since the character position of "character" is "1-7", document number 1 is the search result. Finally, step ST224
, The search means 6 outputs a search result of the index search.

Next, in step ST230 of FIG. 11, the search means 6 performs a search using the ambiguous text. FIG. 13 is a flowchart showing ambiguous text matching. First, in step ST231, a search target document is determined. Here, in order to eliminate waste of processing, as a result of the search by the index matching (step ST220), the document of the document number which is the output candidate is excluded from the search target.

More specifically, a document number including any one of the characters “sentence” or “character” of the keyword “character” is picked up, and a search is made for a document number excluding the document of the document number output in step ST220. Target document.
That is, the document number indicated by the index 27 of “sentence” is ORed with the document number indicated by the index 28 of “character” from FIG. 10, and the search result in step ST220 is excluded from this. In this case, the OR of the document numbers of “sentence” and “character” is “1”, and since document number 1 is output in step ST220, there is no target document except for document number 1 from document number 1. I do.

Next, in step ST232, the target document is loaded into the memory. Here, since there is no target document, it is not loaded. Subsequently, in step ST233, collation at the character code level is performed, but since there is no target document, no collation is performed. Similarly, in step ST234, collation of shape features is performed, but since there is no target document, collation is not performed. In step ST235, the process proceeds to Y. In step ST236, no result is output, and the process ends. Finally, in step ST240 of FIG. 11, each search result (document number 1) is output, and the processing ends.

Next, the user inputs "task" as a keyword.
A search in the case where is input will be described. In step ST200 of FIG.
Input "Assignment" as a keyword. Step S
At T210, the search means 6 divides the keyword.
Here, it is divided into “sections”, “titles”, and “assignments”. Next, in step ST220, the search means 6 executes a search by index matching. In step ST221 of FIG. 12, each index is extracted. Here, the index 30 of "title" exists, but the indexes of "task" and "section" do not exist. Step S
The process proceeds to T222, step ST224, and ends without any result because there is no index of "assignment".

Next, in step ST230 of FIG. 11, the search means 6 searches for an ambiguous text. First, FIG.
In step ST231 of step 3, the search target document is determined. The document number indicated by the “section” index,
OR with the document number indicated by the index of "title",
From now on, the processing excluding the search result in step ST220 is executed.

Since the document number indicated by the “title” index 30 is “1” and the search result in step ST220 is absent, the document number of the target document is “1”. Next, in step ST232, the ambiguous text of the target document is loaded into the memory. Here, document number 1 shown in FIG.
Is loaded into memory.

Next, in step ST233, the search means 6 executes collation at the character code level. Here, if even one character matches the search keyword, the vicinity of the matched character position is stored as the shape feature collation range, and the process proceeds to the next step. Specifically, the vicinity of the part where the character of either “section” or “title” of the keyword “assignment” exists is set as the shape feature matching range. Here, since “theme” 33 matches in FIG. 6, this is set as the shape feature collation range.

Next, in step ST234, the search means 6 executes collation using the shape feature. Here, the shape feature of “section” is loaded from the shape feature 34 and the shape feature dictionary 9 in FIG. In FIG. 8, regions 41 to 48 are allocated to regions 1 to 8. In the calculation of the shape feature, a difference between the features of each region is calculated as described below.

[0059]

(Equation 1)

Here, D is the distance between the shape features, X _i is the i-th shape feature of the text in the ambiguous text database 10, and Y _i is the i-th shape feature (shape feature dictionary) of the corresponding keyword character. 9).

When the distance D is equal to or less than a threshold value THR, it is determined that the shape feature has been successfully collated, and this document is output as a search result. Now, the feature values of the “section” areas 1 to 8 in the shape feature dictionary 9 are “10”, “7”, and “12”, respectively.
Assuming that “12”, “10”, “5”, “10” and “9”, FIG.
The distance to the shape feature 34 is D = 30. Therefore, T
Since HR ≧ D holds, the matching between the features succeeds,
Document number 1 is output as a search result. Finally, in step ST240, the search result, document number 1
Is output.

In the first embodiment, the index is set to 1
Although the description has been made of the case of a character and two characters, the present invention is not limited to this, and an index of three consecutive characters may be used or more. In the first embodiment, the search is performed using both the index and the ambiguous text. However, the present invention is not limited to this, and the search result is output without performing the ambiguous text collation as shown in FIG. You may. By not using the ambiguous text, it is not possible to search for a part that failed in character recognition, but it is possible to speed up the result output. In addition, since a high-precision search can be performed by using an ambiguous text, when a search condition is input to the search condition input unit 5, whether or not to perform a search using an ambiguous text is specified, thereby improving the search accuracy. Priority or priority of search speed can be freely specified.

Although FIG. 6 is used for the ambiguous text, as shown in FIG. 7, a table is prepared which indicates the start position and the end position of the document number having the ambiguous text and the character code of the ambiguous text included in which document. May be. The operation in this case will be described. At the time of registration, the ambiguous text extracting means 4 determines the character string of several characters before and after including the character whose similarity is equal to or less than TH1 as the ambiguous text, and determines the start character position, the end character position, and the document number as described above. Hold. Now, as described with “*” 23 in FIG. 5, the last character including this character is assumed to be an ambiguous text. In FIG. 7, a start character position 4 (see reference numeral 500), an end character position 5 (see reference numeral 501), and a document number 1 (see reference numeral 502) are held.

Further, the ambiguous text extracting means 4 is provided in FIG.
A table of characters in which the ambiguous text shown in (B) appears is created. Now, the document number 1 is held for all of the recognition candidate characters existing at the start character position 4 and the end character position 5. From FIG. 5, in this example, “intelligence” 503 and “Kun” in FIG.
504, “fill” 505, “word” 506, “talk” 507,
Document number 1 is held for “title” 508.

The search process is performed in step ST220 of FIG.
Up to this, it is the same as the first embodiment. Step ST2
At 30, for the keyword “assignment”, the retrieval means 6 loads the “section” and “title” indexes from the table of FIG. 7B and determines the corresponding document. Here, "section"
Does not exist, and the document number of the document including “title” is “1”. Therefore, the document number 1 is searched using the shape feature. In FIG. 7A, 4 to 5 of document number 1
For the characters and the ninth to tenth characters, the characters and shape features are loaded from the recognized character database 12 in FIG. The following is the same as the first embodiment. As a result, double storage of the recognition character database 12 and the ambiguous text database 10 is prevented, and
It is possible to reduce the capacity for holding data.

As is clear from the above, the first embodiment
According to the method, the document number of the recognition candidate character that matches the keyword is searched by referring to the index, and the shape feature of the character image is matched with the shape feature of the character constituting the keyword to match the search condition of the document. Since the configuration is such that the document number to be searched is searched, there is an effect that high-speed and high-precision full-text search can be performed.

Embodiment 2 In the first embodiment, when all the character codes do not match, a case is described in which the document number is searched using the shape feature. However, the search may be performed using only the index file without using the shape feature. . Since the document registration method is the same as that in the first embodiment, a document search method will be described.

First, in step ST200 of FIG. 11, a keyword “task” is input. next,
In step ST210, keyword division is performed. Here, “section”, “title”, and “assignment” are created.
Next, in step ST220, a search by index collation is performed. The flowchart of index collation uses FIG. In step ST221, the search means 6 executes a process of extracting the index of each divided keyword character string. Since there is no index for “task” and “section” but an index for only “title”, the index 30 for “title” is extracted from FIG.

Next, in step ST222, collation of the character position is performed. Here, since there is no index of “assignment”, the collated document is not applicable, and the process proceeds to step ST223. In step ST223,
A collation using a “*” symbol is performed for character positions that are partially mismatched. In this search, it is possible to collate even the character strings of “* title” and “section *” even if they do not completely match the keyword like “assignment”. The processing procedure is "section",
The character position is detected from the index of “section” or “title” using the index of “title”. An index does not exist for “section”, but an index 30 exists for “title”.

Next, the index 31 of the "*" character is loaded. It is verified whether or not an index 31 of “*” is connected to the index 30 of “title”.
The character position "1-4" 32 at the beginning of "*" is 1- of "title".
The condition is satisfied because the character is one character before 5. Since there is no other character position of "title", the search result (document number 1) is output in step ST224, and the process ends. FIG.
In step 1, without performing the ambiguous text collation in step ST230, the process proceeds to step ST240, in which the search result (document number 1) is output, and the process ends.

In the second embodiment, a character "*" is added to a recognition candidate character for a character for which there is no correct answer as a recognition candidate character, and a search is performed assuming that this character matches any character. . However, if no correct character is included such as “**”, the result is not a success. Thereby, there is an effect that search omission due to erroneous recognition can be reduced.

Embodiment 3 FIG. 15 is a block diagram showing a full-text search apparatus according to Embodiment 3 of the present invention. In the figure, the same reference numerals as those in FIG. 1 denote the same or corresponding parts, and a description thereof will be omitted. 13 is a recognition result correcting means for correcting the recognition result of the character recognition means 2, 14 is a character chain appearance probability dictionary updating means (appearance probability updating means) for changing the character chain appearance probability, and 15 is a character chain occurrence probability. When creating an index, the character chain appearance probability dictionary 16 refers to the character chain appearance probability dictionary 15 to determine whether or not to include a concatenated character in which two or more recognition candidate characters are combined in the index creation target. It is an index creation means.

Next, the operation will be described. Here, a method of creating an index using the character chain appearance probability dictionary 15 and a method of updating the character chain appearance probability dictionary 15 will be described. In the document registration process, step ST12 in FIG.
Processing up to 0 is performed in the same manner as in the first embodiment.

In step ST130 of FIG. 2, the index creating means 16 narrows down the recognition candidate characters as in the first embodiment, and creates an index from the recognition candidate characters shown in FIG. At this time, using the character chain appearance probability dictionary 15, it is determined whether to create an index for a combination of recognition candidate characters. FIG.
6 shows an example of the character chain appearance probability dictionary 15. In “probability” in FIG. 15, the number of occurrences of combinations of N consecutive characters in a document is calculated in advance from many learning documents, and the Find the probability. The total number is the number of combinations that actually appear in the learning document. The sum of the probabilities of the groups in which the first characters of the combination characters (joint characters) are the same is “1”. For example, the sum of the probabilities of combinations starting from “sentence” such as “letter”, “literature”, and “sentence” is “1”.

The following equation is defined, E is calculated from the combination of the recognition candidate characters in FIG. 5, and it is determined whether or not to create an index based on the value of E.

[0076]

(Equation 2)

Here, R represents the similarity in character recognition,
R _ij indicates the similarity of the j-th recognition candidate character at the i-th character position from the beginning of the sentence. Similarly, R _{(i + 1) k} indicates the similarity of the k-th recognition candidate character at the (i + 1) -th character position from the beginning of the sentence. P _{ij (i + 1) k} is the probability that the k-th recognition candidate character at the (i + 1) -th character position from the beginning of the sentence follows the j-th recognition candidate character at the i-th character position from the beginning of the sentence. Is shown. α and β are constants.

Specifically, in FIG. 5, for example, i =
In the case of 7, "bunyu", "letters", "literature", "jou"
E is calculated for the six types of “Jo” and “Jo”, and if each value exceeds a certain threshold, the combination is created in the index. Do not leave. Now, α = 0.5, β = 300
Then, E (bunyu) = 0.5 × (90 + 86) + (1
−0.5) × 300 × 0.001 = 88.15.
Calculate in the same way, E (character) = 102, E (literature) = 8
6.5, E (length) = 78.15, E (length) = 77.
15, E (length) = 75.15. Therefore, E
When storing a character set of> 85 or more as an index, only a combination of “character”, “bun”, and “literature” is registered. At this time, in the two-character index of FIG.
E is assigned in descending order of the value of E. Here, "character" is held as 1, "bun" is held as 2, and "literature" is held as 3.

The document search method is the same as in the first embodiment. By calculating the value using the similarity used for character recognition and the probability that combinations of characters will appear consecutively in the document, the possibility of a correct answer as a character is low, or as a character string in the document By eliminating combinations having a low probability of existing characters, it is possible to create an index for search compactly and to reduce erroneous deletion of correct characters.

Embodiment 4 Next, a method of changing the character chain appearance probability dictionary 15 will be described. In documents having the same or similar contents and fields, important words appearing in each document are similar and appear relatively frequently. Therefore, by learning the combination of the characters that appear and updating the character chain appearance probability dictionary 15 of the document for each field, it is possible to make the index compact without reducing the accuracy of the search so much. In the fourth embodiment, an example will be described in which the number of appearances is counted for a combination of characters considered to be correct from the character recognition result, and this value is reflected in the character chain appearance probability dictionary 15.

FIG. 17 is a flowchart showing a document registration method. The documents used for document registration are the same as those in the first embodiment. Until step ST120, processing is performed in the same manner as in the first embodiment. In step ST135, an index is created as in the first embodiment. Thereafter, the character chain appearance probability dictionary updating unit 14 counts the number of appearances of a combination of characters in which the number of candidates is one and continues from among the recognition candidate characters shown in FIG.

In FIG. 5, the number of appearances is counted for a combination of “document”, “intelligence”, “performance”, “noh”, “direction”, and “improvement”. The character chain appearance probability dictionary updating means 14 holds each combination and the number thereof in a buffer (not shown), and updates the character chain appearance probability dictionary 15 of FIG. 16 at a certain timing, for example, once every several document registrations. . Alternatively, the update may be performed by a user issuing an update instruction. Hereinafter, in step ST140, an ambiguous text is created as in the first embodiment, and the process ends.

When the user corrects a character recognition error using the recognition result correcting means 13 for the recognition candidate character, the number of corrected character combinations is counted and the character chain appearance probability dictionary 15 is updated. It is also possible. FIG.
9 is a flowchart showing a document registration method. FIG.
In step 9, up to step ST120, the same processing as in the first embodiment is performed.

In step ST125, the character is corrected using the recognition result correcting means 13. For example, the user corrects the character positions 8 and 9 in FIG. 5 as shown by 60 and 61 in FIG. Next, in step ST133, the index creating means 16 creates an index from the recognition candidate characters shown in FIG. Next, in step ST143, the character chain appearance frequency is counted. The character chain appearance probability dictionary updating means 14 counts the number of combinations in which the recognition candidate character is one character including before and after the corrected character. Here, the number of combinations is counted for "character recognition" and "recognition" in FIG. The character chain appearance probability dictionary 15 is updated at a certain timing, for example, after a certain number of corrections.

The frequency of occurrence of a character chain is counted from the keyword used in the search, not limited to the correction of the misrecognized character.
By reflecting it in the character chain appearance probability dictionary 15, it is possible to leave the character string used as a keyword at the time of registration more accurately.

[0086]

As described above, according to the present invention, while referring to the index, the document of the recognition candidate character matching the keyword is searched, while the shape feature of the character image extracted by the feature extracting means and the keyword are searched. Is configured to provide a search unit that searches for documents that match the search conditions by comparing the shape characteristics of the characters that constitute the character string. Therefore, there is an effect that high-speed and high-precision full-text search can be performed.

According to the present invention, since a concatenated character in which two or more recognition candidate characters are combined is included in the index creation target, there is an effect that high-speed and high-precision full-text search can be performed. .

According to the present invention, among the recognition candidate characters output by the character recognition means, recognition candidate characters having lower accuracy than the reference accuracy are excluded from the index creation target, so that the search accuracy is deteriorated. Therefore, there is an effect that the capacity of the index can be reduced without inducing.

According to the present invention, even when the accuracy of the recognition candidate character output by the character recognition means is lower than the reference accuracy, in the case of a recognition candidate character relating to a character image having no recognition candidate character having an accuracy exceeding the reference accuracy, The recognition candidate character is included in the index creation target, and an identification symbol for distinguishing from the other recognition candidate character is added to the recognition candidate character. There is an effect that a search using only the database becomes possible.

According to the present invention, the shape characteristics of the character image are stored in the database, and the recognition candidate characters for the character image and the character codes of the characters that may form words are stored in the database. Therefore, there is an effect that search accuracy can be improved.

According to the present invention, each recognition candidate character and a character that may form a word are determined in consideration of linguistic information or a type of character, so that search accuracy is improved. There is.

According to the present invention, the distance between the shape feature of the character image extracted by the feature extraction means and the shape feature of the character constituting the keyword is calculated, and when the distance satisfies a predetermined criterion, the search condition is satisfied. , The effect of being able to customize the shape feature dictionary is provided.

According to the present invention, since the setting means for setting whether or not to execute the shape feature matching processing by the search means is provided, the processing in the search processing is considered in consideration of the importance of the search speed and the search accuracy. There is an effect that the priority of the type can be set.

According to the present invention, since the document including the recognition candidate character matching the keyword is excluded from the target of the shape feature comparison, it is possible to reduce the useless search at the time of matching the shape feature. effective.

According to the present invention, the configuration is such that the shape feature of the character image extracted by the feature extraction means is compared with the shape feature of the character constituting the keyword only when there is no recognition candidate character matching the keyword. Therefore, there is an effect that the search speed can be increased.

According to the present invention, when specifying the collation target of the shape feature with respect to the keyword, the recognition candidate character to which the identification code is added is treated as a wild card, so that the search using only the index database can be performed. There is an effect that can be implemented.

According to the present invention, it is determined whether or not the connected character is to be included in the index creation target in consideration of the appearance probability that the connected character in which two or more recognition candidate characters are combined appears in the entire document. With such a configuration, there is an effect that the capacity of the index can be efficiently reduced.

According to the present invention, when each recognition candidate character constituting a connected character is the only recognition candidate character for the character image, the number of appearances of the connected character is counted up to update the appearance probability. , There is an effect that the probability that an important keyword is not searched can be reduced.

According to the present invention, since the appearance frequency is updated by counting up the number of appearances of the concatenated character matching the keyword, the priority of the important character is increased, and the important character is not searched. There is an effect that the probability can be reduced.

According to the present invention, when the recognition candidate character output by the character recognition means is corrected, the appearance probability is updated by counting up the number of occurrences of the connected character including the corrected recognition candidate character. Therefore, the priority of the important character is raised, and the probability that the important character is not searched can be reduced.

According to the present invention, a document of a recognition candidate character that matches a keyword is searched for by referring to an index, and a shape characteristic of a character image is compared with a shape characteristic of a character constituting a keyword, and a search condition is determined. Is configured to be searched for a document that matches with, so that there is an effect that a high-speed and high-precision full-text search can be performed.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing a full-text search device according to a first embodiment of the present invention.

FIG. 2 is a flowchart illustrating a document registration method.

FIG. 3 is an explanatory diagram showing an input image.

FIG. 4 is an explanatory diagram showing a recognition result of a character recognition unit.

FIG. 5 is an explanatory diagram showing a result of narrowing down recognition candidate characters.

FIG. 6 is an explanatory diagram showing an example in which shape features extracted from a character image are stored.

FIG. 7 is an explanatory diagram showing a start position and the like of a document number having an ambiguous text;

FIG. 8 is an explanatory view showing a specific method of creating a shape feature.

FIG. 9 is an explanatory diagram showing an example of a two-character index.

FIG. 10 is an explanatory diagram showing a table storing a position of a one-character index and a recognition order.

FIG. 11 is a flowchart illustrating a document search method.

FIG. 12 is a flowchart illustrating index matching.

FIG. 13 is a flowchart showing ambiguous text matching.

FIG. 14 is a flowchart illustrating index matching.

FIG. 15 is a configuration diagram showing a full-text search device according to a third embodiment of the present invention.

FIG. 16 is an explanatory diagram showing a character chain appearance probability dictionary.

FIG. 17 is a flowchart illustrating a document registration method.

FIG. 18 is an explanatory diagram showing correction contents of a recognition result.

FIG. 19 is a flowchart illustrating a document registration method.

FIG. 20 is a flowchart illustrating a document search method.

FIG. 21 is a configuration diagram showing a conventional full-text search device (conventional example 1).

FIG. 22 is a configuration diagram showing a conventional full-text search device (conventional example 2).

[Explanation of symbols]

1 image input means, 2 character recognition means, 3 index creation means, 4 ambiguous text extraction means (feature extraction means), 5 search condition input means (input means, setting means),
6 search means, 7 output means, 8 character recognition dictionary, 9
Shape feature dictionary, 10 ambiguous text database, 11
Index database, 12 recognized character database, 13 recognition result correcting means, 14 character chain appearance probability dictionary updating means (appearance probability updating means), 15 character chain appearance probability dictionary, 16 index creation means.

Claims (16)

[Claims] 1. A character recognition means for identifying each character image included in an input image, outputting one or more recognition candidate characters for each character image, and outputting the accuracy of each recognition candidate character; Index creation means for creating an index indicating the correspondence between each recognition candidate character output by the means and the document, and a character image having no recognition candidate character with a certainty exceeding the reference certainty among the character images included in the input image Exists, a feature extracting unit for extracting a shape feature of the character image, an input unit for inputting a keyword as a document search condition, and referring to the index,
While searching for a document of the recognition candidate character that matches the keyword, the shape feature of the character image extracted by the feature extraction unit is compared with the shape feature of the character constituting the keyword, and a document that matches the search condition is searched. A full-text search device including a search unit for searching. 2. The full-text search device according to claim 1, wherein the index creation unit includes a connected character in which two or more recognition candidate characters are combined in an index creation target. 3. The index creation unit according to claim 1, wherein, among the recognition candidate characters output by the character recognition unit, recognition candidate characters having a lower accuracy than the reference accuracy are excluded from an index creation target. Full text search device. 4. The method according to claim 1, wherein the index generating means outputs the recognition candidate character output from the character recognizing means even if the accuracy of the recognition candidate character is lower than the reference accuracy. 4. The full-text search device according to claim 3, wherein the recognition candidate character is included in an index creation target, and an identification symbol for distinguishing the recognition candidate character from other recognition candidate characters is added. 5. The feature extracting means stores shape features of a character image in a database, and stores, in the database, recognition candidate characters for the character image and character codes of characters that may form words. The full-text search device according to any one of claims 1 to 4, characterized in that: 6. The full-text according to claim 5, wherein the feature extracting unit determines each recognition candidate character and a character that may form a word in consideration of linguistic information or a type of character. Search device. 7. A search means calculates a distance between a shape feature of a character image extracted by the feature extraction means and a shape feature of a character constituting a keyword, and when the distance satisfies a predetermined criterion, matches a search condition. 7. The full-text search device according to claim 1, wherein the full text search device is certified. 8. The full-text search apparatus according to claim 1, further comprising setting means for setting whether or not to execute the shape feature matching process by the search means. 9. The full-text according to claim 1, wherein the search unit excludes a document including a recognition candidate character that matches the keyword from a target of shape feature matching. Search device. 10. A search unit checks a shape feature of a character image extracted by a feature extraction unit against a shape feature of a character constituting the keyword only when there is no recognition candidate character matching the keyword. The full-text search device according to any one of claims 1 to 7, characterized in that: 11. The full-text search device according to claim 4, wherein the search means handles a recognition candidate character to which an identification code is added as a wild card when specifying a matching target of the shape feature with respect to the keyword. 12. The index creation means determines whether or not to include the connected character in an index creation target in consideration of an appearance probability that a connected character in which two or more recognition candidate characters are combined appears in the entire document. 3. The full-text search device according to claim 2, wherein: 13. In the case where each of the recognition candidate characters constituting the concatenated character is the only recognition candidate character for the character image, an appearance probability updating means for counting up the number of appearances of the concatenated character and updating the appearance probability is provided. 13. The full-text search device according to claim 12, wherein the device is provided. 14. The full-text search device according to claim 12, further comprising an appearance probability updating unit that counts up the number of appearances of the concatenated character that matches the keyword and updates the appearance probability. 15. Appearance probability updating means for updating the appearance probability by counting up the number of appearances of a connected character including the corrected recognition candidate character when the recognition candidate character output by the character recognition means is corrected. 13. The method according to claim 12, wherein
Description full-text search device. 16. Identifying each character image included in the input image, outputting one or more recognition candidate characters for each character image and the accuracy of each recognition candidate character, and indicating the correspondence between each recognition candidate character and a document. In addition to creating an index, among the character images included in the input image, extracting the shape features of character images that do not have recognition candidate characters with a degree of accuracy exceeding the reference accuracy, and entering a keyword as a document search condition, By referring to the index, a search is made for a document of a candidate character that matches the keyword, and at the same time, the shape characteristics of the character image are compared with the shape characteristics of the characters constituting the keyword to search for a document that matches the search condition. Full-text search method to do.