JP2746345B2 - Post-processing method for character recognition - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a post-processing method for character recognition.

[Conventional technology]

In general, a character recognition device recognizes a character by comparing and comparing a feature pattern extracted from a character image on an input image with a standard feature pattern registered in a dictionary in advance.

However, in such character recognition on a character-by-character basis, when there are many similar characters such as kanji, a plurality of candidates are often found for one character. Post-processing is required.

Conventionally, as such post-processing, a character string of a single semantic category, such as only an address, only a name, or only a book title, is targeted, and the character recognition in units of one character is performed by knowledge processing using word knowledge. Methods to correct the results are being considered. This is effective when the position of the character string in the document and the semantic category to which the character string belongs are known, such as when characters are entered in a preset frame, but multiple characters are included in the target document. If there is a character string belonging to the semantic category and the position in each document is ambiguous, it cannot be handled.

[Problems to be solved by the invention]

By the way, character strings belonging to two or more different semantic categories, which are continuously printed with characters of the same size, such as an address and a telephone number in a business card, are cut out as one character string block. Such a character string block should be processed by dividing it into semantic categories and output. However, according to the conventional post-processing method, one block is processed collectively.

It is an object of the present invention to provide a post-processing method for character recognition in which blocks of character strings belonging to two or more different semantic categories are divided and processed and output for each semantic category.

[Means for solving the problem]

The present invention performs character recognition on a character-by-character basis for each character string block cut out from an input image and corrects the result of the character recognition by knowledge processing using word knowledge for each semantic category. In the processing method, a plurality of knowledge processes for each semantic category are performed on one character string block, and it is determined whether a character group of a plurality of semantic categories exists in the character string block. Is divided for each semantic category. In addition, a character string of a processing result of each of the divided parts is output with a label of its semantic category added.

(Operation)

As a result of knowledge processing of a certain semantic category for one character string block of interest, if a partial character string of the character string block is unprocessed, knowledge of another semantic category for the unprocessed character string is obtained. Processing is performed, divided into two or more parts for each semantic category, and the result of knowledge processing is output for each semantic category.

〔Example〕

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

First, the contents of the processing in one embodiment of the present invention will be described using the business card image 21 shown in FIG. 2 as an example.

When a character string block is cut out from the business card image 21, rectangular areas 31 to 35 shown in FIG.
It is cut out as. This character string block is cut out by, for example, a method of integrating circumscribed rectangles of black pixels in a business card image, but may be another method.

The contents of each character string block obtained in this way are unknown, but if the target image is specified in advance as a horizontally written business card image or if it is determined at the time of cutting out the character block, Since the relative positional relationship and arrangement order of the character strings of each semantic category in are specified to some extent, the relative position and character size (such as the height of the character string blocks) of the extracted character string blocks in the image
Thus, the candidate of the semantic category can be estimated.
That is, the semantic category can be estimated such that the character string block looks like a company name and the character string block looks like a title.

Next, a character in each character string block is cut out, and character recognition is performed for each character by comparing and comparing the characteristic pattern with a standard characteristic pattern registered in the character dictionary, and a candidate character is determined. Although the number of candidate characters for each character can be set in advance, the number of candidate characters may be determined by, for example, comparing a distance difference from the first candidate with a preset threshold.

The result of the character recognition of the character string block in units of one character is, for example, as shown in FIG.

Here, the character string block is a block composed of a character string of an address and a character string of a telephone number. Since both are printed successively with characters of the same size, they are cut out as one block. Such a block frequently occurs in a business card image or the like, but it is desirable to divide a character string in the block according to a semantic category.

The candidate character strings of each character string block obtained by the above-described character recognition on a character-by-character basis are corrected from the beginning of the candidate character strings by knowledge processing by comparison and collation with a knowledge dictionary of words of each meaning category. At this time, as described above, the knowledge dictionary of the semantic category estimated based on the relative position of the character string block and the like is used. However, knowledge processing may be sequentially performed on a plurality of semantic categories, and the result of the semantic category that provides the most reliable processing result may be selected.

Taking the character string block as an example, by comparing and comparing the knowledge dictionary of the address, which is the estimated semantic category, with the candidate character string, for example, a portion of "〒223 Shinagawacho, Kohoku-ku, Yokohama-shi" is obtained as a processing result It is assumed that the character string that follows does not match the knowledge dictionary and has not been processed.

By counting the number of characters of this unprocessed character string and comparing the number of characters with a preset threshold value, "the unprocessed character string portion contains a character string of another semantic category" Determine whether or not. In this example, the unprocessed character string is 16 characters long enough to assume that a character string of another semantic category is included.

Therefore, the relative position of the character string block in the business card image, the appearance information of the semantic category of the already processed character string (for example, the company name and name have already been found, so they no longer appear), the length of the character string Thus, the semantic category to which the unprocessed character string belongs is estimated. In this example, the semantic category is "phone number" or "address attached"
Is determined to be high, a keyword indicating the telephone number information is searched from the character string.

In this example, since the keyword "telephone" is found, the character string after "telephone" is processed as a character string indicating a telephone number. Then, as shown in FIG. 5, as a processing result of the character string block, the character string is divided before and after by a "phone" part, and the processing result character string of each part is added with a label of its semantic category. And output (write to result memory).

If a part of the character string after "telephone" remains as an unprocessed character string and the number of characters is large, the same knowledge processing is repeated for the unprocessed character string.

Although the example described here is a case where the address and the telephone number are included in one character string block, other combinations are processed in the same procedure.

As described above, in the present embodiment, the knowledge processing of one semantic category is performed on the character recognition result of the character string block, and when an unprocessed character string remains, the character processing is performed based on the number of characters and the possibility of the semantic category. When it is determined that the character string of the semantic category of remains in the block,
The operation of performing knowledge processing of another semantic category for the unprocessed character string is repeated, and the result of the knowledge processing for each semantic category is output.

An example of a character recognition post-processing device that performs the above-described processing will be described with reference to FIG.

In FIG. 1, a document (business card, document) to be processed is optically read by a scanner 1 and stored in an image memory 2 as a binary image. A character string block is cut out from the input image by a character string block cutout unit 3, and the image is stored in a block image memory 4. At this time, information for estimating the semantic category such as the relative position of each character string block is also extracted. Note that information on a document such as a horizontally written business card may be specified in advance, or may be determined automatically based on information such as the relative position and size of a character string block.

From the image of each character string block, character extraction is performed by the character extraction recognition unit 5 to extract a characteristic pattern.
Character recognition in units of one character is performed by comparison and collation with the character dictionary 6.

The above-described post-processing is performed on the candidate character string obtained by performing character recognition on a character-by-character basis for each character string block.

That is, in the post-processing / knowledge dictionary matching unit 7, knowledge processing is performed for each character string block by comparing and matching the candidate character string with the knowledge dictionary 8 of words of each meaning category. The meaning category at this time is estimated as described above. Then, the result of the knowledge processing is output to the different category character string existence determining unit 9 together with the label of the semantic category.

In the different category character string existence determining unit 9, the processing result sent from the post-processing
By comparing the character string with the candidate character string by character recognition on a character-by-character basis, it is determined whether or not the unprocessed character string includes a character string of another category based on the number of characters or the like as described above.
If it is determined that a string of another category is included,
The post-processing / knowledge dictionary matching unit 7 performs knowledge processing of another semantic category for the candidate character string of the unprocessed part, and for the processed character string part, the character string of the processing result is stored in the result memory together with the label of the semantic category. Write to 10.

 Reference numeral 11 denotes a control unit that controls the flow of the above processing.

Note that the present invention is also applicable to speech recognition correction processing.

〔The invention's effect〕

As is apparent from the above description, according to the present invention, when one or more character string blocks cut out from an input image include character strings of two or more different semantic categories, the character string block is It is possible to output the processing result by dividing by the semantic category.

[Brief description of the drawings]

FIG. 1 is a schematic block diagram showing an example of a character recognition device embodying the present invention, and FIGS. 2 to 5 are views for explaining processing in one embodiment of the present invention. FIG. 3 is a view showing an example of an image, FIG. 3 is a view showing a character string block cutout result of a business card image, and FIG. 4 is a view showing a result of character recognition of a character string block in FIG. FIG. 5 is a diagram showing the result of knowledge processing on the character string block in FIG. DESCRIPTION OF SYMBOLS 1 ... scanner, 2 ... image memory, 3 ... character string block extraction part, 4 ... block image memory, 5 ... character extraction / recognition part, 6 ... character dictionary, 7 ... post-processing / knowledge dictionary Matching unit, 8: Knowledge dictionary, 9: Different category character string existence determining unit, 10: Result memory.

Claims (2)

(57) [Claims] 1. A character recognition method for performing character recognition on a character-by-character basis for each character string block cut out from an input image and correcting the character recognition result by knowledge processing using word knowledge for each semantic category. In the post-processing method, a plurality of knowledge processes for each semantic category are performed on one character string block, and it is determined whether or not a character string group having a plurality of semantic categories exists in the character string block. A post-processing method for character recognition, wherein a group is divided into semantic categories. 2. A post-processing method for character recognition according to claim 1, wherein a character string as a processing result of each of the divided parts is
A character recognition post-processing method characterized by adding a label of the semantic category and outputting.