JP3835652B2 - Method for determining Japanese / English of document image and recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a document image Japanese / English determination method and a recording medium for determining whether each character region in a document image is a Japanese region or an English region.
[0002]
[Prior art]
When character recognition processing is performed on a document image, it is necessary to select an appropriate language. In other words, when trying to recognize Japanese with English OCR, it is impossible to recognize anything other than alphabets and numbers. Conversely, when trying to recognize English with Japanese OCR, English OCR is used for character segmentation and language processing. The recognition rate will be lower than the case.
[0003]
Therefore, it is necessary to perform language identification before performing the character recognition process. Conventionally, various methods for identifying a character type in a document have been proposed. For example, there is a document recognition apparatus that counts the number of black and white inversions in the vertical or horizontal direction of a binarized character line and identifies the character type based on the distribution (see Japanese Patent Laid-Open No. 5-108876).
[0004]
There is also a document recognition apparatus that recognizes a read word and discriminates the language type of a recognized character based on the matching rate between the recognition result and the dictionary (see Japanese Patent Laid-Open No. 6-150061).
[0005]
[Problems to be solved by the invention]
In the former device, the number of black / white inversions is used as a feature for identifying the character type, but this feature varies greatly depending on the font and document content (ratio of kana, kanji, numbers, etc.). There is a problem that becomes low.
[0006]
On the other hand, since the latter device performs character recognition once, if the OCR performance is good, the character type can be determined with a high probability, and it is possible to accurately distinguish between Japanese and English. Become. However, OCR has a problem that it takes a long time for processing.
[0007]
The present invention has been made in consideration of the above circumstances,
An object of the present invention is to identify Japanese and English with high accuracy and at high speed, and to identify a Japanese / English language of a document image that can identify both for each character region and for each page as to a range to be identified. It is to provide a recording medium.
[0008]
[Means for Solving the Problems]
In order to achieve the above object, according to the first aspect of the present invention, there is provided a document image Japanese / English determination method for determining whether each character region in a document image is a Japanese region or an English region, A line is cut out from each character area, and the frequency number of the rectangle when the ratio of the height of each rectangle in the line to the maximum height of the rectangle in the line is high (hereinafter, the first frequency number), and the maximum number of rectangles in the line The frequency number of the rectangle when the ratio of the height of each rectangle in the row to the height is low (hereinafter, the second frequency number) is calculated, and the first frequency number / second frequency number is a predetermined first number. Each character area is determined to be a Japanese area when a threshold value of 1 is exceeded, and each character area is an English area when the first frequency number / the second frequency number is less than a predetermined second threshold value It is determined that there is, otherwise it is determined as an unknown area, and the unknown area is calculated in advance Determines that the a Japanese region when close to the characteristic values of the Japanese, are characterized by determining that the English region when close to the previously calculated characteristic values in English.
[0009]
According to a second aspect of the present invention, there is provided a computer-readable recording medium on which a program for causing a computer to implement the method for determining Japanese / English of a document image according to the first aspect is recorded.
[0010]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.
<Example 1>
FIG. 1 shows the configuration of Embodiment 1 of the present invention. In the figure, 101 is an image input means for inputting a document image, 102 is an image reduction means for reducing the input document image, 103 is a connected component extraction means for extracting a connected component from the document image, and 104 is an extracted connection. An area generating means for generating a character area by classifying and integrating components, 105 is a Japanese / English discriminating means for discriminating Japanese and English in character area units or page units, 106 is a control unit for controlling the whole, 107 is a data storage unit for storing various data such as input document image data, connected component data, and area data, 108 is a data communication path, 109 is data communication for connecting to a host or the like via a network, line, or the like. Means.
[0011]
FIG. 2 shows an overall process flowchart of the first embodiment of the present invention. The processing operation of the present invention will be described below with reference to FIG.
First, the image input unit 101 obtains a document image by reading a document (step 201). The image input means is, for example, a scanner or a fax machine, and an image may be obtained from another device via the network via the data communication means 109.
[0012]
Next, the image reduction means 102 reduces the input document image (step 202). This process is, for example, a process of reducing the OR of the input document image to about 1/8. That is, 8 × 8 pixels are reduced to one pixel. If there is at least one black pixel in 64 pixels, the reduced pixel is a black pixel.
[0013]
The connected component extracting means 103 extracts a black pixel connected component from the reduced image (step 203). The area generation unit 104 classifies the extracted connected components and integrates them to generate a character area (step 204). As this region generation method, for example, a known method described in JP-A-6-20092 may be used. At this time, information on the connected components constituting each character area is stored and held in the data storage unit 107.
[0014]
Subsequently, the Japanese-English discriminating means 105 judges whether the generated character area is Japanese or English (step 205).
[0015]
In step 202, the image is OR-reduced to fuse neighboring black pixels. Here, in English, there is a feature that there is a space between words and that there is a very narrow space between characters in the word. On the other hand, in Japanese, the character spacing does not change greatly except before and after punctuation.
[0016]
FIG. 3 shows image examples of English and Japanese sentences and their circumscribed rectangles. The circumscribed rectangle 302 represents the result of reducing the English image 301 and extracting the connected components in the circumscribed rectangle (note that the circumscribed rectangle 302 should be smaller than the image 301 because it is reduced, but here In the same size). In English images, connected components are formed by merging words.
[0017]
For the examples of the Japanese images 303 and 305, when the connected components are extracted in the same manner and expressed by their circumscribed rectangles, the circumscribed rectangles 304 and 306 are obtained, respectively.
[0018]
In the case of English sentences, since the number of characters constituting a word is constant to some extent, there is a feature that the number of circumscribed rectangles in which the aspect ratio is about 2 to 6, 7 times increases. On the other hand, the Japanese language has a feature that a long rectangle that does not appear in English as shown in the circumscribed rectangle 304 is generated, or conversely, a large number of small rectangles such as the circumscribed rectangle 306 are generated.
[0019]
Therefore, the above-described connected component rectangles are classified into three types of “short”, “medium”, and “long”, and these are totalized for each character region. FIG. 4 shows a processing flowchart of Japanese-English determination in the first embodiment. The process of FIG. 4 is performed for each character area. When the row direction is horizontal, for example, the width / height is 2 or less “short”, the width / height 2 to 6 is “medium”, and the rectangle is classified as “long” (step) (step). 401). Then, the classification results in the character area are totaled (step 402), and it is determined whether each character area is Japanese or English (step 403). Here, assuming that the number of “short” rectangles is SCNT, the number of “medium” rectangles is NCNT, and the number of “long” rectangles is LCNT, the determination of Japanese and English is as shown in FIG. 8 (detailed flowchart of step 403). Done.
[0020]
First, it is examined whether LCNT / (NCNT + SCNT)> Thl is satisfied (step 801). Th1 is a predetermined threshold value, for example, about 0.3. If this conditional expression is satisfied, it means that there are sufficiently long rectangles, and it is determined that the character area is a Japanese area (step 804).
[0021]
Next, when it is determined No in step 801, it is checked whether NCNT / (LCNT + SCNT) <Th2 is satisfied (step 802). Th2 is also a predetermined threshold value, for example, 3. If this conditional expression is satisfied, it means that there are few middle rectangles, and it is determined that the character area is a Japanese area (step 804). If neither condition is satisfied, it is determined as an English region (step 803).
[0022]
<Example 2>
In the first embodiment described above, Japanese / English is determined in units of character areas. In this case, the number of characters may be very small depending on the character area. In such a case, since the number of rectangles cannot be obtained sufficiently, it may be difficult to make a Japanese-English determination based on the ratio of the number of rectangles. The second embodiment is an embodiment that considers a case where the number of rectangles is not sufficient.
[0023]
FIG. 5 shows a process flowchart of the second embodiment. The Japanese-English discriminating means 105 checks whether or not the number of rectangles in the tabulated area is sufficient (that is, whether or not it is greater than or equal to a predetermined threshold Th) (step 501). Japanese-English discrimination using OCR described in JP-A-6-150061 is performed (step 503). In this case, since the number of characters is small, an increase in processing time can be reduced even if OCR processing is performed. If the number of rectangles is sufficient, the identification of Japanese and English by the rectangle length described in the first embodiment is performed (step 502).
[0024]
<Example 3>
Next, a description will be given of a third embodiment that performs Japanese-English discrimination on a page basis. 6 and 7 are detailed flowcharts of step 205 according to the third embodiment. In the method shown in FIG. 6, the number of “short”, “medium”, and “long” rectangles is aggregated for the entire page, not for each character area (steps 601 and 602), and the results are used for the day by page. The determination of English is performed (step 603). This Japanese-English determination method is performed according to the processing flowchart of FIG. The threshold values Th1 and Th2 at this time may be different from those in the case of processing in units of character areas.
[0025]
In the method shown in FIG. 7, Japanese / English is determined for each character area (step 702), and based on the result, Japanese / English is determined for the page (step 703). Specifically, Jn is the number of areas determined to be Japanese, and En is the number of areas determined to be English. If Jn> En, it is determined to be a Japanese page, and if En> Jn, it is determined to be an English page. If Jn = En, it may be rejected or judged to be either English or Japanese.
[0026]
<Example 4>
A description will be given of a Japanese-English identification method using features different from the above-described embodiments. FIG. 9 shows the configuration of the fourth embodiment. The difference from the first embodiment is that a line segmentation unit 902, a block extraction unit 903, and an in-block character type discrimination unit 904 are provided. The other components are the same as those of the first embodiment. FIG. 10 shows a process flowchart of the fourth embodiment.
[0027]
First, the line cutout unit 902 cuts out lines from the character area of the document image (steps 1001 and 1002). When the technique described in Japanese Patent Laid-Open No. 6-20092 is used as the area generation process, line information is obtained at the stage of extracting the area. A method using a projection described in “Division of document image using density distribution, linear density, circumscribed rectangle feature” (Akiyama et al., August 1986, Vol. J69-D No. 8) may be used. .
[0028]
Next, the block extraction unit 903 extracts a block corresponding to a word (step 1003). As this block extraction method, the method previously proposed by the present applicant in Japanese Patent Application No. 8-34781 may be used. That is, the block extraction unit 111 detects circumscribed rectangles in the row data and collects the circumscribed rectangles into block data. A method for collecting the block data is as follows. Histograms of character rectangle intervals (one rectangle is not yet determined as one character. Therefore, in the case of Kanji characters, there are many cases where each divided into a single and a rectangle is a single rectangle) are obtained. FIG. 18 shows the extracted character rectangle and the distance between the rectangles. FIG. 19 shows a histogram of rectangular intervals.
[0029]
In this histogram, the peak with the shortest distance tends to appear in the gap between kanji characters and the composition, and the distance between characters in the same word of proportional English. Even if these are integrated, different character types rarely enter the block, so that block data is formed by integrating them. By performing this process, proportional characters and single characters are separated (that is, kanji characters composed of bias and composition) into one.
[0030]
The peak with the longest distance often appears in the distance between words and the distance between the punctuation mark and the next character. These are often used at the boundary when the character type changes (especially the distance between words), and we want to avoid having the same block. Therefore, processing is performed so that character rectangles having a distance equal to or longer than the longest peak value are not made into the same block.
[0031]
Further, when the distance (A, B) between the target rectangle and the adjacent rectangle is measured and the difference (A−B) is equal to or greater than a predetermined threshold, the longer distance rectangles are not integrated and the shorter one is not integrated. Process to merge distance rectangles. FIG. 20 is a diagram for explaining a case where rectangles are not integrated at a position where the difference between the rectangles is large. In FIG. 20, since the integration of rectangles is not performed at a position where the difference is larger than a predetermined threshold value, three blocks are formed. By performing such processing, even if the distance between words is absolutely close, such as in proportional English, there should be a difference from the distance between characters, so only one word can be integrated together. Even in the case of proportional fonts, the Japanese kanji parts are arranged at relatively equal intervals, which is convenient for collecting Japanese sentences.
[0032]
By using the block extraction method described above, in the case of English, unlike Japanese documents, words are separated by a space equivalent to half-width characters, so avoid mixing with other character types to form block data. It is done.
[0033]
Subsequently, the in-block character type determination unit 904 performs Japanese-English determination for each block (step 1004). The method of the application mentioned above may also be used for this. That is, the in-block character type determination unit 904 determines the character type whether the block grouped by the above process is Japanese or alphanumeric. The same character type is determined in the block. The character type is determined as follows. That is, when the number of black runs in the vertical direction of the rectangle or the number of black and white inversions is greater than or equal to a predetermined threshold with respect to the width of the rectangle in the block, it is identified as a Japanese character, and the rectangle in the extracted block in the vertical direction Identify alphabetic characters based on coordinate values. FIGS. 21A and 21B show specific examples of the number of vertical runs in the case of Japanese and English characters. In an ideal case where there is no noise with alphanumeric characters, a maximum of four runs can be made with the letter “g” (FIG. 21B). Therefore, if more than 5 runs are counted, it will be in Japanese. In the case of the character “image” shown in FIG. 21A, the number of runs in the vertical direction changes as indicated by the number below the character.
[0034]
The Japanese / English discriminating means 905 tabulates the discrimination results for each block and discriminates between Japanese and English in the area (step 1005). Here, the number of blocks determined to be Japanese is JCNT, the number of blocks determined to be English is ECNT, and the number of blocks determined to be indefinite is NCNT. FIG. 11 is a detailed flowchart of step 1005. If JCNT * Th3> ENCT, it is determined that the language is Japanese (steps 1101 and 1105). Otherwise, if ECNT> JCNT, it is determined that the language is English (1102, 1104). Otherwise, it is rejected (step 1103). The threshold value Th3 is set to 2, for example.
[0035]
<Example 5>
In Example 4 described above, Japanese / English is determined in units of character areas. In this case, the number of characters may be very small depending on the character area. In such a case, since the number of rectangles cannot be obtained sufficiently, it may be difficult to perform Japanese-English determination by the ratio of the number of block discrimination results. The fifth embodiment is an embodiment where the number of blocks is not sufficient.
[0036]
FIG. 12 shows a process flowchart of the fifth embodiment. The Japanese / English discriminating means 105 checks whether or not the total number of blocks in the character area is sufficient (that is, whether or not it is greater than or equal to a predetermined threshold Th) (step 1201). Japanese-English discrimination using the OCR described in Japanese Patent Laid-Open No. 6-150061 is performed (step 1203). In this case, since the number of characters is small, an increase in processing time can be reduced even if OCR processing is performed. If the number of blocks is sufficient, Japanese and English are identified based on the discrimination result for each block described in the fourth embodiment (step 1202).
[0037]
<Example 6>
In the sixth embodiment, the Japanese / English discrimination for each character area in the fourth embodiment is changed to Japanese / English discrimination in page units. The processing flowchart of the sixth embodiment uses FIGS.
[0038]
In the processing shown in FIG. 6, JCNT, ECNT, and NCNT are aggregated not for each character area but for the entire page, and the result is used to determine Japanese / English using the processing method shown in FIG. At this time, Th3 may be different from that in the character area unit.
[0039]
In the processing shown in FIG. 7, first, each character area is determined, and Japanese-English determination of the page is performed from the result. Specifically, Jn is the number of areas determined to be Japanese, and En is the number of areas determined to be English. If Jn> En, it is determined to be a Japanese page, and if En> Jn, it is determined to be an English page. When Jn = En, it may be rejected or it may be either Japanese or English.
[0040]
<Example 7>
In the seventh embodiment, when performing Japanese-English discrimination for each character area or page, as shown in FIG. 13, a Japanese-English discrimination process using a rectangular length (step 1301) and a date using a discrimination result for each block are used. The English / Japanese discrimination is performed by the English discrimination process (step 1302). Then, finally, discrimination is made between Japanese and English from each discrimination result (step 1303).
[0041]
If both are determined to be Japanese or English, the final result may be determined to be Japanese or English as it is. If any of them is determined to be rejected, the determination result that is not rejected is the final result.
[0042]
If both of the determination results are in Japanese and the other is in English and the results do not match, one of the following determinations is made.
(1) Reject.
(2) The degree of certainty of both is calculated, and the result with the larger value is adopted.
As a certainty factor of the discrimination method using the rectangular length, for example, when LCNT / (NCNT + SCNT)> Thl, and Thl = 0.3, a value of LCNT / (NCNT + SCNT) * 2.5 (however, the upper limit is 1) )
When NCNT / (LCNT + SCNT) <Th2 and Th2 = 3, the value of (LCNT + SCNT) /NCNT*2.5 (however, the upper limit is 1)
When NCNT / (LCNT + SCNT)> Th2 and Th2 = 3, a value of NCNT / (LCNT + SCNT) * 0.33 (the upper limit is 1)
And
[0043]
For example, when JCNT * Th3> ECNT and Th3 = 2, the value of JCTN / (ECNT * 3) is set (however, the upper limit is 1).
In the case of ECNT> JCNT, the value of ECNT / JCNT * 0.7 (however, the upper limit is 1)
And
[0044]
<Example 8>
FIG. 14 shows a configuration of the eighth embodiment. FIG. 15 shows a process flowchart of the eighth embodiment. In this embodiment, with respect to the entire page of the input document, the Japanese / English discriminating unit 1412 uses the method of Embodiments 3 and 6 described above to identify whether the page is in Japanese or English. The selection unit 1403 selects the English document recognition unit 1404 or the Japanese document recognition unit 1405 based on the determination result, and performs the document recognition processing of the selected language (step 1504, 1502). 1505), and outputs the recognition result to an output unit such as a display (step 1506).
[0045]
Since Japanese and English have different attributes, it may be better to switch between region division processing and font identification processing. Therefore, the document recognition unit of the present embodiment includes not only the character recognition processing but also the above-described region division processing and font identification processing.
[0046]
<Example 9>
FIG. 16 shows the configuration of the ninth embodiment, and FIG. 17 shows a process flowchart of the ninth embodiment. The difference from the eighth embodiment is that Japanese-English identification is performed for each character area. For this purpose, the area dividing unit 1602 divides the input document into character areas (steps 1701 and 1702). Here, the area dividing unit uses an area dividing method that can be applied to both Japanese and English. After the division processing, the Japanese-English discriminating unit 1603 performs Japanese-English discriminating processing for each character area using, for example, the method of the first embodiment described above (step 1704). The document recognition unit 1605 or the Japanese document recognition unit 1606 is selected, document recognition processing of the selected language is performed (steps 1705 and 1706), and the recognition result is output to the output unit 1607 such as a display (step 1707). Note that the document recognition unit of the ninth embodiment performs font identification processing in addition to document recognition processing.
[0047]
<Example 10>
In each of the above-described embodiments, Japanese and English are determined using a black pixel connected component and a rectangular length as a feature amount. However, the determination method using the black pixel connected component takes a long processing time, and the method using the rectangular length may increase the generation of rejection. In addition, there is a method of discriminating whether it is Japanese or English based on the peak position of the frequency distribution of the relative position in the upper and lower rows of the circumscribed rectangle (see Japanese Patent Publication No. 7-21817), but there is a document with an inclination. Is input, there is a problem that the frequency distribution changes greatly and the identification accuracy decreases.
[0048]
Therefore, in the present embodiment, Japanese and English are accurately and quickly identified for each region of the document image by identifying Japanese and English using the histogram of the height of the circumscribed rectangle in the row with respect to the row height. To identify. For areas that cannot be identified by the above-described Japanese-English identification method, Japanese-English identification is performed using another method.
[0049]
FIG. 22 shows a configuration of the tenth embodiment. FIG. 23 is an overall process flowchart of the tenth embodiment. First, the image input unit 2201 obtains a document image by reading a document (step 2301). The image input means is, for example, a scanner or a fax machine, and an image may be obtained from another device via the network via the data communication means 2207.
[0050]
Next, the area generation unit 2202 generates a character area (step 2302). As this region generation method, for example, a method described in Japanese Patent Laid-Open No. 6-20092 may be used. Next, the line cutout means 2203 cuts out a line for character recognition from the character area. In other words, the circumscribed rectangles of the characters are obtained and integrated to generate a line (step 2303). The Japanese-English identification means 2204 performs Japanese-English identification on the generated character area (step 2304).
[0051]
Japanese and English are identified as follows. FIG. 27 is a flowchart showing details of Japanese-English identification (step 2304). FIG. 24 shows an example of a cut out line and a circumscribed rectangle in the line. First, the frequency distribution of the ratio of the circumscribed rectangle height in the row to the row height is calculated (steps 2701 and 2702). The line height is set to line height, and the rectangular height is set to height. Let the ratio be highrate = height * 100 / lineheight. In addition, in the case of a document with an inclination as shown in FIG. 25, the maximum value of the height of the rectangle of the line may be used as lineheight instead of the line height in order to identify Japanese and English more accurately. That is, for an input document with a slope, Japanese and English are identified based on a histogram of the ratio of each circumscribed rectangle height in the line to the maximum height of the in-line rectangle.
[0052]
For example, the number of rectangles when the ratio heightrate is 80 or more is lcnt, the number of rectangles when the heightrate is 70 or more and less than 80 is ncnt, and the number of rectangles when the heightrate is 40 or more and less than 70 is scnt. Lcnt, ncnt, scnt are obtained for all rectangles in the character area.
[0053]
FIG. 26 shows an example of the number of rectangles examined for Japanese and English documents. In general, Japanese has a large lcnt, and English tends to have a large scnt. Therefore, predetermined thresholds thJ and thE are set. When lcnt / scnt> thJ, it is determined that the language is Japanese (step 2703), and when lcnt / scnt <thE, it is determined that the language is English (step 2704). Otherwise, it is set as an unknown area (step 2705).
[0054]
The above unknown area can be identified using a statistical method. FIG. 28 is a detailed process flowchart for the unknown area. For example, the feature values lcnt, ncnt, and scnt of the Japanese region and English region are normalized in advance, and the average value and the inverse matrix of the covariance matrix are obtained for Japanese and English, respectively. Then, using the average value and the inverse matrix of the covariance matrix, the Mahalanobis distance is obtained for each of Japanese and English (steps 2801 and 2802).
[0055]
When the Mahalanobis distance in Japanese is Dj and the Mahalanobis distance in English is De, if the predetermined threshold is Me and Mj, it is judged as English when Dj / De> Me (step 2803), and Dj / De <Mj It is determined that the language is Japanese (step 2804). If neither condition is satisfied, it is determined as an unknown area (step 2805). In place of the Mahalanobis distance described above, an Euclidean distance from the average value or a city block distance may be used.
[0056]
Furthermore, Japanese-English identification is performed on the area determined to be unknown using the certainty of English recognition. FIG. 29 is a detailed process flowchart of step 2805. A certainty factor is calculated by English recognition (step 2901). Next, regarding the calculated certainty factor, for example, the number of words having a certainty factor of 60% or more is Good, the number of words having a certainty factor of less than 60% and a non-certainty factor is Bad, and the number of words having a certainty factor of zero is Zero. (Step 2902).
[0057]
When the judgment value for Japanese-English identification is Value, Value = Good / (Good + Bad + Zelo)
(Step 2903), if the value exceeds a predetermined threshold value th eocr (step 2904), it is determined as English, and if it is less than that, it is determined as Japanese.
[0058]
Note that Zelo may be weighted. For example, if Zelo is three pieces of Bad, Value is
Since Bad = Bad + Zelo × 3 Value = Good / (Good + Bad)
Thus, it can be determined that the value is English if the value exceeds the threshold value th eocr, and that the value is Japanese if the value is less than the threshold value. As described above, even in an area where the number of characters for Japanese-English discrimination determination is small, Japanese-English identification is performed with a certainty factor based on English recognition, so that Japanese-English identification is performed in units of areas with high accuracy.
[0059]
<Example 11>
In this embodiment, a circumscribed rectangle is generated from an image obtained by reducing an input document image, and the generated rectangles are appropriately integrated, and a Japanese-English identification is made more accurate by using a histogram of the rectangular length aspect ratio after the integration. This is a well-executed embodiment.
[0060]
FIG. 30 shows the configuration of the eleventh embodiment. FIG. 31 is an overall process flowchart of the eleventh embodiment. The document image input by the image input unit 3001 is reduced by the image reduction unit 3002 in the same manner as in the above-described embodiment (Steps 3101 and 3102). In this process, for example, the document image is OR-compressed to about ¼ (4 × 4 pixels are reduced to one pixel, and if there is at least one black pixel in 16 pixels, the reduced image is black).
[0061]
Next, the area generation means 3003 generates a character area (step 3103). As this region generation method, for example, a method described in Japanese Patent Laid-Open No. 6-20092 may be used. Subsequently, the rectangle integration means 3004 integrates the rectangles so that the characteristics of Japanese and English are well expressed (step 3104). For example, as shown in FIG. 32, when the vertical coordinates of the y-coordinates (vertical direction) of the rectangles 1 and 2 are close and the x-coordinates of the adjacent rectangles 1 and 2 are very close (for example, the horizontal distance between the rectangles). If the is less than the distance corresponding to the English space), merge the rectangles. Also, for example, as shown in FIG. 33, when the left rectangle 1 is in a positional relationship including the right rectangle 2 with the y coordinate and the x coordinates of the adjacent rectangles 1 and 2 are very close (for example, If the horizontal distance between the rectangles is less than the distance corresponding to the English space), merge the rectangles.
[0062]
Then, using the rectangular aspect ratio (rectangular long aspect / rectangular long aspect), it is divided into four feature amounts of a long rectangle, a medium rectangle, a small rectangle, and a minimal rectangle (FIG. 34). In general, Japanese has a high ratio of long rectangles, and English has a high ratio of medium rectangles. Using this difference in characteristics, the Japanese-English discriminating means 3005 creates an identification judgment formula and performs Japanese-English discrimination (step 3105). FIG. 35 is a detailed flowchart of the Japanese-English identification process.
[0063]
For example, the number of long rectangular areas lcnt in the area
Number of middle rectangular areas in the area ncnt
The number of small rectangular areas in the area scnt
The number of small rectangular areas sscnt (in many cases of noise) in the area is calculated (step 3501), and the ratio of long rectangles in the area ratio1 = lcnt / (ncnt + scnt) is calculated (step 3502). The ratio of the middle rectangle at ratio2 = ncnt / (lcnt + scnt) is calculated (step 3503). When calculating the above ratio, sscnt was ignored as noise.
[0064]
Then, the ratio is x-coordinate and ratio2 is the y-coordinate, and the classification is divided into a Japanese area, an English area, and a rejection area so that misidentification is reduced as much as possible, and a portion overlapping Japanese and English is rejected. For example, if ratio2 / ratiool> thE, it is determined as an English region (step 3504), if ratio2 / ratiool <thJ, it is determined as a Japanese region (step 3505), and other regions are determined as Japanese / English unknown (step 3506). ). Here, thE and thJ are predetermined threshold values.
[0065]
Similar to the tenth embodiment, a region determined to be unknown to English is identified using Japanese statistical methods. For example, the feature values lcnt, ncnt, and scnt of the Japanese region and English region are normalized in advance, and the average value and the inverse matrix of the covariance matrix are obtained in Japanese and English, respectively. Find the Mahalanobis distance in Japanese and English using the mean and the inverse of the covariance matrix. When the Mahalanobis distance in Japanese is Dj and the Mahalanobis distance in English is De, the predetermined threshold is Me and Mj. If Dj / De> Me, English is determined, and if Dj / De <Mj, Japanese is determined. If neither condition is satisfied, it is determined as unknown. In place of the Mahalanobis distance, an Euclidean distance from the average value or a city block distance may be used.
[0066]
<Example 12>
The present invention is not limited to the above-described embodiments, and can be realized by software. When the present invention is realized by software, as shown in FIG. 36, a computer system including a CPU, a memory, a display device, a hard disk, a keyboard, a CD-ROM drive, a scanner, etc. is prepared, and a computer such as a CD-ROM is prepared. On the readable recording medium, a program for realizing the Japanese / English determination function and the document recognition function of the present invention is recorded. A document image or the like input from image input means such as a scanner is temporarily stored in a hard disk or the like. When the program is started, the temporarily stored document image data is read, Japanese / English determination processing and document recognition processing are executed, and the results are output to a display or the like.
[0067]
【The invention's effect】
As described above, according to the present invention, since a plurality of determination methods are used in combination, Japanese and English can be determined with high accuracy. Further, Japanese and English can be accurately distinguished for each character area in the document image, and Japanese and English can be accurately distinguished for each page of the document image. Furthermore, since an appropriate document recognition process is performed on the document image determined to be Japanese or English, a highly accurate recognition result can be obtained.
[Brief description of the drawings]
FIG. 1 shows a configuration of Embodiment 1 of the present invention.
FIG. 2 shows an overall process flowchart of Embodiment 1 of the present invention.
FIG. 3 shows examples of English and Japanese images and their circumscribed rectangles.
FIG. 4 shows a processing flowchart of Japanese-English determination in Example 1;
FIG. 5 shows a processing flowchart of Embodiment 2.
FIG. 6 shows a first detailed flowchart of step 205 according to Embodiment 3;
FIG. 7 shows a second detailed flowchart of step 205 according to Embodiment 3;
FIG. 8 shows a detailed flowchart of step 403;
9 shows a configuration of Example 4. FIG.
FIG. 10 is a flowchart illustrating a process according to the fourth embodiment.
FIG. 11 is a detailed flowchart of step 1005;
12 shows a process flowchart of Embodiment 5. FIG.
FIG. 13 is a flowchart illustrating a process according to the seventh embodiment.
14 shows a configuration of Example 8. FIG.
FIG. 15 is a flowchart illustrating a process according to the eighth embodiment.
16 shows the configuration of Example 9. FIG.
FIG. 17 shows a process flowchart of Embodiment 9.
FIG. 18 shows extracted character rectangles and the distance between the rectangles.
FIG. 19 shows a histogram of rectangular intervals.
FIG. 20 is a diagram illustrating a case where rectangles are not merged at a position where the difference in spacing between rectangles is large.
FIGS. 21A and 21B show specific examples of the number of vertical runs in the case of Japanese and English characters.
22 shows the configuration of Example 10. FIG.
FIG. 23 is an overall process flowchart of Embodiment 10;
FIG. 24 shows an example of a cut line and a circumscribed rectangle in the line.
FIG. 25 shows an example of a line when a document is tilted and a circumscribed rectangle in the line.
FIG. 26 shows an example of the number of rectangles examined for a Japanese document and an English document.
FIG. 27 is a detailed processing flowchart of Japanese-English identification (step 2304).
FIG. 28 is a detailed process flowchart for an unknown area;
FIG. 29 is a detailed process flowchart of step 2805;
30 shows the structure of Example 11. FIG.
FIG. 31 is an overall process flowchart of Embodiment 11;
FIG. 32 shows an example of integrating rectangles.
FIG. 33 shows another example of integrating rectangles.
FIG. 34 shows rectangles classified into four types.
FIG. 35 is a detailed processing flowchart of Japanese-English identification processing according to the eleventh embodiment;
36 shows the configuration of Example 12. FIG.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 101 Image input means 102 Image reduction means 103 Connected component extraction means 104 Area | region production | generation means 105 Japanese-English discrimination means 106 Control part 107 Data storage part 108 Data communication path 109 Data communication means

Claims (2)

A document image Japanese / English determination method for determining whether each character region in a document image is a Japanese region or an English region, wherein a line is cut out from each character region, and the maximum height of a rectangle in the line When the ratio of the height of each rectangle in the line to the rectangle is high (hereinafter referred to as the first frequency number) and the ratio of the height of each rectangle in the line to the maximum height of the rectangle in the line is low A rectangular frequency number (hereinafter referred to as a second frequency number) is calculated, and each character area is a Japanese area when the first frequency number / the second frequency number exceeds a predetermined first threshold value. And when each of the first frequency number / second frequency number is less than a predetermined second threshold, each character area is determined to be an English area, and otherwise, it is determined to be an unknown area. The unknown area is determined to be a Japanese area when it is close to a pre-calculated Japanese characteristic value. And, Japanese English determination method of the document image, wherein the determining that the English region when close to the previously calculated characteristic values in English. A computer-readable recording medium recording a program for causing a computer to implement the method for determining Japanese / English of a document image according to claim 1 .

Images