JP3243389B2 - Document identification method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for identifying a document type, for example, a Japanese document and an English document with high accuracy.

[0002]

2. Description of the Related Art With the advance of computerization of information, document data is stored, searched, browsed, and reused in a wide range of fields including offices. As one method for converting document content into data, the importance of character recognition processing for converting document content printed out on paper into text data is increasing. In this character recognition processing, it is better to separate the processing into English sentences and Japanese sentences,
This is advantageous in terms of performance because the language-specific processing method can be used.

As a conventional technique for identifying such a document / character type, there is, for example, an apparatus described in Japanese Patent Application Laid-Open Nos. 4-346188 and 4-346189. In the former device, a vertical line and a horizontal line having a length equal to or more than a certain value are extracted from image data of a character line cut out by region division, and a closed region (a rectangle in a character) surrounded by the vertical and horizontal lines is extracted. Is extracted, and whether the character line is Japanese or English is recognized based on the number.
Further, the latter apparatus scans image data of a character line cut out by area division, obtains a distance between characters, and identifies English / Japanese sentences based on a distribution of the distance between characters.

[0004]

The above-described conventional apparatus employs a local method of determining the character type on a line-by-line basis in a character region cut out by region division, and as a preprocessing immediately before performing character recognition. Is convenient,
It cannot be applied to the stage where the character area before the area division is not determined (that is, it cannot be used as preprocessing of the area division). Also, if the whole document or a certain large area is to be used for the purpose of globally discriminating between English and Japanese, it is necessary to perform all the processing from the area division, so a long processing time and a large memory capacity are required. Cost. Furthermore, since the document and the character type are identified using the distribution of the character size and the distance between the characters, there is a problem that the influence of the difference in the font and the size increases.

An object of the present invention is to provide a document identification method for accurately identifying a document without being affected by differences in fonts and character sizes.

[0006]

According to a first aspect of the present invention, there is provided a document identification method for identifying a type of an input document image or a partial area of the document image. The cut-out pattern is compared with the registered template. If the pattern does not match the template and is a new pattern, the pattern is registered as a new template and a new symbol is added, and similar to the already registered template. When the pattern is replaced, the pattern is replaced with a symbol given to the template, the number of the symbols is updated, and the appearance frequency distribution of the symbol in the document image or a partial region of the document image is calculated. It is characterized in that the type of the document or the partial area is identified based on the appearance frequency distribution.

According to a second aspect of the present invention, when calculating the appearance frequency distribution of the symbols, the calculation is performed excluding a pattern of a predetermined size or less.

According to a third aspect of the present invention, the type of the document or the partial area is identified based on the average of the number of patterns represented by one template.

According to a fourth aspect of the present invention, the type of the document or the partial area is identified based on the number of types of the template.

The invention according to claim 5 is characterized in that the type of the document or the partial area is identified based on the number of types of templates having a high appearance frequency.

The invention according to claim 6 is characterized in that the type of the document or the partial area is identified based on the number of types of templates having a high appearance frequency and the total number of patterns.

[0012]

A document read from a scanner or the like is stored as document image data in a memory. A connected component is cut out from the document image data stored in the memory and stored in the pattern memory. The cut-out pattern is compared with a pre-registered template, and if it is a new pattern, it is registered as a template, and a new symbol is added to the new template and stored in the symbol string storage memory. If the similarity with the registered template is high, the number of symbols assigned to the known template is updated and stored in the symbol string storage memory. Then, the frequency distribution of the symbols is calculated with reference to the symbol string storage memory, and the Japanese sentence and the English sentence are identified based on the symbol frequency distribution. In another embodiment, when calculating the symbol frequency distribution, the calculation is performed after excluding small patterns, and the document including the picture region is identified with high accuracy.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be specifically described below with reference to the drawings. FIG. 1 is a configuration diagram common to each embodiment of the present invention. Hereinafter, for simplicity, a case in which processing is performed in units of documents will be described. However, the present invention can also be applied to an area that has been divided in advance by area identification.

A document is input as image data by the image input unit 101 and stored in the image memory 106. Here, the image input unit is, for example, a scanner,
Alternatively, data may be input from an image data file or a fax. The connected component extraction unit 102 collectively cuts out connected components from the image data stored in the image memory 106, and stores the extracted data in the pattern memory 107. In addition, as such a clipping method, for example, by obtaining the distribution (projection) of black pixels in the vertical and horizontal directions, a character is cut out or a black pixel component to be connected is extracted. What is necessary is just to use the method of cutting out by integrating only those that are close to each other.

The pattern matching unit 103 matches the extracted pattern with a template registered in the template memory 108. If the pattern is not matched and is new, the pattern matching unit 103 registers the extracted pattern as a new template in the template memory 108, and A symbol (for example, a number) is attached to the template to store the symbol string storage memory 109.
To memorize.

If the similarity with the already registered template is high, the frequency of the symbol assigned to the registered template is updated and stored in the symbol string storage memory 109. That is, in the symbol row storage memory 109, different patterns that have appeared are stored (symbolized) as symbols, and the frequency (that is, the number of similar patterns) is also stored for each symbol. The frequency distribution calculation unit 104 calculates the frequency distribution of the symbols stored in the symbol string storage memory 109. The document identification unit 105 determines a Japanese sentence and an English sentence based on the obtained frequency distribution.

Before describing each embodiment of the present invention, a conventional image compression method used in the present invention will be described. In this method, connected components are cut out from a document image, one connected component is regarded as one pattern, registered as a template, and a similar pattern is replaced with a template in the cut out pattern, thereby reducing image information. (See US Pat. No. 5,303,313). As a method of determining the similarity, the test method described in paragraph 9 of the publication is used. In other words, in this method, the patterns are superimposed while correcting the positions, and it is determined whether or not the pixels are different from each other based on the appearance position of the pixel having a difference and the pattern of the appearing pixel. .

FIG. 10 shows an example in which a similar pattern is detected from a cut-out pattern and a symbol is assigned. 80
1 is a document image composed of English characters, 802 is a template,
803 is a symbol (template number). A connected component (character pattern) is cut out from the document image 801. In the example of the figure, first, a connected component “H” is cut out, and this is subjected to pattern matching with a pattern registered in the template 802. In this case, since nothing is registered, no matching is performed, and the connected component “H” is newly registered as a template. Also, the connected component “H” is represented by the symbol “1” (template number) and stored in the memory.

Next, the connected component "e" is cut out and does not match the pattern registered in the template, so that it is newly registered as a template and is represented by the symbol "2". Hereinafter, the connected components “t”, “o”,
“L”, “d”, and “m” are processed in the same manner and registered as templates, and symbols “3” to “7” are assigned to each and stored in the memory.

Subsequently, when "e" next to "m" is cut out, the connected component "e" matches the pattern "e" registered in the template and is not newly registered as a template. However, a pattern is created by taking an average of the registered pattern and the matched pattern (or a representative value of both patterns), and the template is updated. Thereby, the similar pattern is updated so that the representative pattern is registered.

The symbol "2" which has already been determined is assigned to the symbol. Hereinafter, in the same manner, when the symbols up to the number “9” are used, the characters “h”,
Symbols “a” and “b” are assigned to “a”, respectively.

As described above, the document image is decomposed into a symbol sequence (a sequence of template numbers) and pattern information of each template. Although not directly related to the present invention, the image is compressed at a high compression rate by predictively coding these pieces of information in the predictive coding unit.

Embodiment 1 FIG. 2 is a processing flowchart according to Embodiment 1 of the present invention. An image is input from the image input unit 101 (Step 2)
01), the connected component extraction unit 102 extracts connected components from the input image data (Step 202), and outputs the extracted connected components to the pattern matching unit 103.

The pattern matching section 103 performs pattern matching between the extracted connected component and the template registered in the template memory 108 (step 203), and determines whether the extracted connected component is a new pattern. (Step 204), and if the pattern is a new pattern, a new symbol (for example, a number) is assigned and registered in the template memory 108, and the symbol is registered in the symbol row memory 109 (step 20).
6, 205).

If a match is found with the existing template, "1" is added to the number of symbols of the existing template registered in the symbol column memory 109, and then the symbol is registered in the symbol column memory 109 (step 20).
5). The pattern matching unit 103 performs the above-described pattern matching processing on all connected components in the input image (Steps 203 to 207). If all connected components have not been processed (NO in step 207), step 20
3 when all connected components have been processed (step 2
07, (YES), and proceed to step 208.

The frequency distribution calculation section 104 calculates the frequency distribution of the symbols stored in the symbol string storage memory 109 (step 208), and the document identification section 105 generates a Japanese sentence and an English language based on the calculated frequency distribution. The sentence is determined (step 209).

FIG. 3 shows a specific example of the pattern frequency distribution of a Japanese document and an English document. In the case of a Japanese document, many types of templates are distributed at a low frequency, whereas in an English document, few types of templates are distributed at a high frequency . In other words, English documents are
And it is configured from degrees about 80 types of characters generally words
Have been. On the other hand, Japanese is only 10 katakana and hiragana
If you count 0 and add Kanji, it will be over 1000 kinds
Consists of characters. Therefore, the appearance frequency of the symbol is
Japanese and English are very different.
Since it is an effective feature amount, it is possible to distinguish between Japanese and English , for example, by calculating the variance of the frequency distribution.

In the case where a picture is included in the document,
Because of the dither processing, there are many small patterns in the picture area. Based on these many small patterns, it is possible to identify a document as to whether or not a picture is included in the document.

Second Embodiment FIG. 4 is a processing flowchart according to a second embodiment of the present invention. The processing up to step 307 is the same as that described in the first embodiment. In the second embodiment, when calculating the frequency distribution, the frequency distribution calculation unit 104 checks the size of the template, and excludes small templates that are equal to or smaller than a predetermined threshold value before calculating (step 308). A document image often includes a picture area, and there are many small patterns in the picture area. Since such patterns are very large in number, they greatly affect the distribution of frequency distribution and the like.

In the second embodiment, when calculating the frequency distribution, small patterns are excluded, so that the above-mentioned effect is eliminated.
It is possible to more accurately distinguish between Japanese and English.

Third Embodiment FIG. 5 is a processing flowchart according to a third embodiment of the present invention. The processing up to step 408 is the same as that described in the first embodiment. Document identification unit 105 of the third embodiment
Then, a document is identified using the average of the number of patterns represented by one template as a feature amount (step 409).
The number of patterns per template can be determined by referring to the symbol string storage memory 109 in which the frequency (number of patterns) is written for each symbol (template). In the case of English, a character having a higher appearance frequency has a larger number of patterns per template. On the other hand, Japanese has a smaller number of patterns per template than English, so it is easy to distinguish between Japanese and English.

Fourth Embodiment FIG. 6 is a processing flowchart according to a fourth embodiment of the present invention. The processing up to step 508 is the same as that described in the first embodiment. The document identification unit 105 according to the fourth embodiment identifies a document using the number of template types (step 509). The number of template types can be determined by referring to the symbol string storage memory 109 as the number of symbol types. Since English has a sentence composed of about 80 characters, the number of required template types is much smaller than in Japanese, and it is easy to distinguish between Japanese and English by judging the number. Can be.

Embodiment 5 FIG. 7 is a processing flowchart according to Embodiment 5 of the present invention. The processing up to step 608 is the same as that described in the first embodiment. In the fifth embodiment, a document is identified using the number of template types having a high appearance frequency. The appearance frequency can be determined by referring to the frequency in the symbol string storage memory 109. In English documents, the appearance frequency of symbols is biased. For example, there are some patterns in which the appearance frequency exceeds 2%, whereas in Japanese, this is rare. Therefore, by counting the number of templates having a high appearance frequency (step 609), Japanese and English can be easily identified.

Embodiment 6 FIG. 8 is a processing flowchart according to Embodiment 6 of the present invention. The processing up to step 708 is the same as that described in the first embodiment. In the sixth embodiment, a document is identified using the total number of patterns in addition to the number of types of templates having a high appearance frequency (step 709). The total number of patterns is obtained from the sum of the frequencies of each symbol in the symbol string storage memory 109.

In the case of identifying a partial region of a document, the number of templates having a high appearance frequency increases because the total number of patterns is small, and the identification method of the fifth embodiment makes identification difficult.
Therefore, normalization is performed by multiplying the number of templates having a high appearance frequency by the total number of patterns, thereby removing the influence of the case where the total number of patterns is small, thereby enabling more accurate Japanese and English discrimination.

FIG. 9 is a diagram in which the feature amounts of the third and sixth embodiments are combined and plotted. The vertical axis is appearance frequency 2%
The above number of templates × the total number of patterns, and the horizontal axis is the average of the number of patterns represented by one template. From this figure, it can be seen that Japanese and English can be easily separated linearly.

In the above-described embodiment, a configuration is provided in which a dedicated processing unit for executing each function is provided. However, the present invention is not limited to this.
The configuration can be changed so as to be incorporated in an OM or the like and operated and processed on a general-purpose processor.

[0038]

As described above, according to the first aspect of the present invention, a pattern is cut out from a document image and made into a template, and the pattern in the image is replaced with the template to be converted into a symbol, and the pattern in the document or a partial area is formed. Calculates the appearance frequency distribution of symbols inside and uses the frequency distribution to identify the type of document or partial area, so it is possible to identify documents with high accuracy without being affected by differences in fonts and sizes be able to.

According to the second aspect of the present invention, small patterns are excluded from the frequency distribution when calculating the appearance frequency distribution. Therefore, even if a picture area is included in the document, the document can be accurately calculated. Can be identified.

According to the third aspect of the present invention, since the identification is performed using the average of the number of patterns per template, it is simpler than the first aspect of the invention without being affected by differences in fonts and sizes. The calculation can identify the document with high accuracy.

According to the fourth aspect of the invention, by using a simple index of the number of template types, a document can be identified with higher accuracy without being affected by differences in fonts and sizes.

According to the fifth aspect of the present invention, the template is identified using the number of types of templates having a high appearance frequency.
Documents can be identified with higher accuracy without being affected by differences in fonts and sizes.

According to the sixth aspect of the present invention, since the total number of patterns is further taken into account, the document is not affected by differences in fonts and sizes, and even if the amount of characters is relatively small, the document can be accurately obtained. Can be identified.

[Brief description of the drawings]

FIG. 1 is a configuration diagram common to each embodiment of the present invention.

FIG. 2 is a processing flowchart according to the first embodiment of the present invention.

FIG. 3 shows a specific example of a pattern frequency distribution of a Japanese document and an English document.

FIG. 4 is a processing flowchart according to a second embodiment of the present invention.

FIG. 5 is a processing flowchart according to a third embodiment of the present invention.

FIG. 6 is a processing flowchart according to a fourth embodiment of the present invention.

FIG. 7 is a processing flowchart according to a fifth embodiment of the present invention.

FIG. 8 is a processing flowchart according to a sixth embodiment of the present invention.

FIG. 9 is a diagram in which feature amounts of the third embodiment and the seventh embodiment are combined and plotted.

FIG. 10 shows a specific example in which a similar pattern is detected from a cut-out pattern and a symbol is assigned.

[Explanation of symbols]

 Reference Signs List 101 Image input unit 102 Connected component extraction unit 103 Pattern matching unit 104 Frequency distribution calculation unit 105 Document identification unit 106 Image memory 107 Pattern memory 108 Template memory 109 Symbol string storage memory

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06K 9/20 G06K 9/62 G06T 7/00-7/60

Claims (6)

(57) [Claims] In a document identification method for identifying the type of an input document image or a partial region of the document image, a pattern cut out from the document image is compared with a registered template, and the pattern is identified as a template. If it is a new pattern without matching, it is registered as a new template and a new symbol is added. If the pattern is similar to an already registered template, the pattern is replaced with the symbol assigned to the template and Updating the number of symbols, calculating the appearance frequency distribution of the symbol in the document image or the partial area of the document image, and identifying the type of the document or the partial area based on the appearance frequency distribution. Document identification method to use. 2. The document identification method according to claim 1, wherein when calculating the appearance frequency distribution of the symbols, the calculation is performed excluding a pattern of a predetermined size or less. 3. The document identification method according to claim 1, wherein the type of the document or the partial area is identified based on an average of the number of patterns represented by one template. 4. The document identification method according to claim 1, wherein the type of the document or the partial area is identified based on the number of types of the template. 5. The document identification method according to claim 1, wherein the type of the document or the partial area is identified based on the number of types of templates having a high appearance frequency. 6. The number of types of templates having a high appearance frequency,
2. The document identification method according to claim 1, wherein the type of the document or the partial area is identified based on the total number of patterns.