JPH0721817B2 - Document image processing method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a method for processing a document image, and more specifically, it is suitable for character recognition in document information captured as image data or clipping of a specific character area. Document image processing method.

[Conventional technology]

In a document image processing system using a large-capacity memory such as an optical disk as a file device, it is required to automatically input index information necessary for searching stored documents. The document image is roughly divided into a photographic area that essentially takes a gray value and an area represented by binary information such as a character or a graphic.
In order to automatically input the index information, it is necessary to extract a specific character string or character in the character area in the input document image and convert the index term into a character code by character recognition.

Conventionally, as a technique for recognizing characters in a document image, for example, I E E E Proceeding 6th ICP
R), 1982, pp. 1023 to 1026, a method is reported for recognizing English characters (Roman characters) by separating character groups (words) that are in contact with each other from a character string. . Further, as a technique for separating individual characters from a character string for character recognition, for example, the Institute of Electronics and Communication Engineers, V.
ol.J67-D, No.10 (October 1984), an example of character segmentation for Japanese documents is reported. However, all of these reports are based on the assumption that the type of document constituent characters (English or Japanese) is known, and target document images in which the character type is unknown or multiple types of characters are mixed. It is not what

Meanwhile, IPSJ Transactions, Vol.25, No.2 (March 1985)
Pp. 257 to 264, there is reported a method for identifying the type of document constituent characters by visual characteristics of character patterns. In this report, for example, when a character is scanned with a small unit mesh of 3 picture elements x 3 picture elements and only one white pixel area and one black pixel area appear in the unit mesh, the pattern of the unit mesh Focusing on the fact that the types are limited to 66 patterns (legal patterns) out of 2 ⁹ = 512, the types of legal patterns obtained during character scanning and the frequency distribution are used to determine the inclination direction of the character stroke and the curvilinearity. The characteristics such as the aspect ratio of characters are extracted, and the Roman characters, Greek characters, Russian characters, Japanese characters, Hangul characters, etc. are discriminated by this.

[Problems to be solved by the invention]

However, since the character type identification method using the legal pattern described above uses the statistical characteristics of the legal pattern appearing in the document, when two or more characters are mixed in the same document, for example, a title, The important parts such as the author and abstract part cannot be applied to the academic papers which are described in English as well as Japanese. Further, in this method, since it is necessary to determine the stroke direction, shape, etc. for all black pixels, the processing is complicated, and there is a problem that a large memory capacity is required for data processing for the determination.

An object of the present invention is to provide a document image processing method particularly suitable for distinguishing a Japanese (Japanese sentence) region and an English region when two types of character regions are mixed in the same document image.

[Means for solving problems]

In order to achieve the above object, a document image processing method according to the present invention includes a step of taking document information including a character string into a memory as image data, a step of obtaining a circumscribed rectangle for each black pixel continuous area from the image data, A step of obtaining a circumscribed rectangle for each character string from the image data, and a frequency distribution regarding the vertical position of the upper and lower sides of the circumscribed rectangle for each black pixel continuous area included in the circumscribed rectangle of the character string for each of the character strings And a step of determining from the peak position of the frequency distribution whether the character type forming the character string is an English character or a Japanese character.

[Action]

In the character area in the document, the circumscribed rectangle for each black pixel continuous area indicating the stroke of each character is, for example, Kanji, Katakana,
Comparing Japanese sentences consisting of hiragana and English sentences consisting of uppercase and lowercase alphabets, there is a difference in the arrangement. For example, in the case of an English sentence, there is a space between words, but there is no space in a Japanese sentence. Also, in English, there is a clear difference in the size of the circumscribed rectangle of the black pixel continuous area between uppercase and lowercase, but there is no difference in the size of the alphabet between the Kanji and Hiragana or Katakana in Japanese.
Therefore, there is a characteristic in the adjacency relationship of these circumscribed rectangles or the variation in the vertical direction. If the degree of separation of the circumscribed rectangle of the black pixel continuous area in the character string direction is obtained as a frequency distribution, it can be determined from the distribution pattern whether each character string is a Japanese sentence or an English sentence. Similarly, it is possible to distinguish Japanese sentences from sentences such as Greek, Russian, and German.

According to the present invention, for each character string, by determining the character type of the character string from the frequency distribution regarding the position of the circumscribed rectangle of the black pixel continuous region, for example, academic conference publications written in Japanese, Extract the text area in English from the paper,
This can be cut out and an abstract created automatically. In addition, because the character recognition process can be executed after knowing the character type for a specific area, misrecognition can be reduced and conversion to character codes can be performed quickly, and these character codes are automatically used as an index. It is also possible to register.

〔Example〕

 Examples of the present invention will be described below.

FIG. 1 is a hardware configuration diagram of a document image processing system for carrying out the present invention. 1 is a data processor, 2 is a memory for storing a program required by the processor, and 3 is required for character recognition. This is a memory for storing a dictionary. In this example, a dictionary for Japanese sentences and a dictionary for English sentences are prepared. Reference numeral 4 is a memory for storing data such as processing results or intermediate results, 5 is a keyboard for inputting various commands or numerical data to the processor 1, and 6 is a document image input device. A typical example of the document image input device 6 is a scanner that reads image information from a document, but it may be a magnetic tape, a magnetic disk, or the like that reads data from a recording medium in which data has already been input. Further, 7 is a frame memory for temporarily storing the input document image data, 8 is a display device for displaying the document image data, 9 is a printer device for making a hard copy, and these elements are a bus. Connected by 10.

In the present invention, the document image data input from the image input device 6 is processed by the processor 1 as shown in FIG.

That is, for the document image 11 input to the frame memory 7, a circumscribed rectangle 12 of black pixel continuous components and a circumscribed rectangle 13 of each character string are obtained. The circumscribed rectangle of the black pixel continuous component can be extracted by a known method known in, for example, JP-A-60-129379. For the circumscribing rectangle 13 of the character string, for example, a publicly known method reported in "Technical Technical Report PRL-83-7" entitled "Individual Character Extraction Method from Vertical Writing Horizontal Writing Document" can be adopted.

In the present invention, the character type 14 forming each character string is discriminated from the two types of circumscribing rectangles 12 and 13 thus extracted from the input image, and based on this, the character cutting process 15,
Alternatively, character recognition processing 16 is performed.

FIG. 3 (A) shows an example of an input document in a Japanese sentence area 20.
FIG. 3 (B) shows a part of the input image 11 including the text area 21 and the English text area 21. FIG.
And a circumscribed rectangle of black pixel continuous components (small rectangle in each character string circumscribed rectangle). In this example, the circumscribed rectangle of the string (below,
Looking at the relationship between the character string rectangle) and the circumscribing rectangle of the black pixel continuous area contained therein (hereinafter referred to as the black pixel circumscribing rectangle), it can be seen that the Japanese language area 20 and the English language area 21 have the following differences. .

(1) In the Japanese sentence, most of the black pixel circumscribed rectangles are aligned with the bottom of the character string rectangle, but in the English sentence, “g”, “j”,
Since some characters and symbols such as “p”, “q”, “y”, and “,” have a downward protruding stroke, which determines the base of the character string rectangle, most of the black pixel circumscribing rectangles It is far from the bottom of the string.

(2) In English, most characters consist of continuous strokes, and the upper side of the black pixel circumscribed rectangle is the first level that matches the upper side of the character string rectangle and the second level that corresponds to the upper side of lowercase letters.
There are two remarkable levels, but in the case of Japanese,
The upper side of the black pixel circumscribed rectangle is concentrated on the upper side of the character string rectangle, and there is no clear second level.

(3) Regarding the adjacency relationship between the black pixel circumscribed rectangles in each character string, in the case of English sentences, there are large gaps between words, but in Japanese sentences there is no such gap, and the gaps between rectangles are also uneven. Is.

The present invention identifies the type of constituent character for each character string by paying attention to the above-mentioned feature difference in the two types of character areas.

FIG. 4 shows a display symbol of each of the character string rectangles L including a plurality of black pixel circumscribing rectangles R1 to R7. String rectangle L
The size and position of is indicated by the coordinate address (LXmin, LYmin), (LXmax, LYmax) of the upper left and lower right pixels, and similarly,
The size and position of each black pixel circumscribed rectangle is also the coordinate address (RXmin, RYmin), (RXmax, RYmax) of the upper left and lower right of each rectangle.
Indicated by. All the black pixel circumscribing rectangles R included in one character string rectangle L are To be satisfied. The coordinate values of each rectangle extracted from the input image are stored for each rectangle ID in a black pixel circumscribing rectangle table 51 and a character string rectangle table 52 prepared in the data memory 4, as shown in FIG. 5, for example. .

When distinguishing the constituent character types of each character string by paying attention to the above properties (1) and (2), for example, as shown in FIG.
Each character string rectangle Lj is divided in the height direction into n areas (8 areas a to h in this example), and the upper side RXmax and the lower side RYmin of the black pixel circumscribed rectangle R included in this character string are It suffices to find out in which divided region each of them is located and obtain its frequency distribution. As shown in FIG. 7, the frequency distributions of RXmax and RYmin show different characteristics in Japanese and English. Therefore, for example, the character strings showing high frequency in the upper and lower divided areas a and h are Japanese and the inside A character string having a high frequency in the areas b and g can be judged as an English sentence.

On the other hand, if attention is paid to properties (3), as shown in FIG. 8, determine the distribution of the distance D between the black pixel circumscribed rectangular area contained within the same string, as in FIG. 9, D ₁ If there are two prominent peaks of D ₂ English, otherwise it may be determined that Japanese.

Hereinafter, the procedure of the document image processing of the present invention using the first determination focusing on the properties (1) and (2) and the second determination focusing on the property (3) will be described with reference to a program flow chart.

FIG. 10 shows the overall procedure of document image processing, and includes the following processing steps.

Process 61: A document image is input from the image input device 6.

Process 62: The circumscribed rectangle R of the black pixel continuous component is extracted. The coordinate address of the extracted rectangle is stored row by row in the rectangle table 51 of FIG. Here, if the table format is such that the address of the diagonal point is stored but the address of the rectangle can be uniquely defined, for example, the coordinate address of one point and another display format for expressing the width and height of the rectangle are adopted. May be.

Process 63: The character string rectangle L is extracted and each address data is stored in the character string rectangle table 52.

Process 64: The first determination process for determining whether the character string is a Japanese sentence or an English sentence is performed by referring to the black pixel circumscribing rectangular table 51 and the character string rectangular table 52.

Process 65: For the character string that could not be determined in process 64, the second determination process of Japanese sentence or English sentence is performed.

Process 66: With reference to the above determination result, the dictionary memory 3 is used to perform character recognition, character segmentation, and the like.

FIG. 11 is a flow chart showing the details of the first judgment processing 64, and comprises the following steps.

Process 701: The counter UC for counting the upper side distribution frequency of the black pixel circumscribed rectangle R and the counter DC for counting the bottom side distribution frequency are cleared for each character string.

Process 702: The coordinate data of the rectangle R is read from the black pixel circumscribed rectangle table 51 in the order of the rectangle ID.

Process 703: Referring to the table 52, a character string including the rectangle R
Search for Lj. The inclusion condition is based on the above-mentioned formula (1).

Process 704: The character string Lj is divided into n pieces in the height direction. Upper limit address Amax (i) and lower limit address Amin of each divided area
(I) is defined by the following equation.

Amax (i) = LYmin + (LYmax-LYmin) -i / n Amin (i) = LYmin + (LYmax-LYmin)-(i-1) / n Process 705: Which of the above divided areas is the upper side address RYmax of the rectangle R It is checked whether it is included, and the corresponding counter UC (i) is incremented. Segment area index i and RYma
The relationship with x is obtained by Amin (i) ≦ RYmax ≦ Amax (i).

Process 706: It is checked which of the divided areas the bottom address RYmin of the rectangle R is included in and the corresponding counter DC (i) is incremented. Segmented area index and RYmin
The relationship between and is obtained by Amin (i) ≦ RYmin ≦ Amax (i).

Through these processes, which of the divided areas a to h in FIG. 6 where the upper side and the lower side of the rectangle R are located is stored as the frequency in the counters UC and DC for each character string.

Process 707: It is checked whether the determination process has been completed for all the rectangles in the table 51. If NO, the process returns to process 702, YES.
If so, the process proceeds to step 708.

Process 708: By referring to the counters UC and DC for each character string, it is determined whether or not the respective maximum count values are generated in the counters corresponding to the divided areas at the upper end or the lower end of the character string rectangle. That is, whether max [UC (i)] = UC (h), (i = a to h) max [DC (i)] = DC (a), (i = a to h) are both satisfied. Find out. If satisfied, this character string can be determined to be a Japanese sentence.

Process 709: In order to indicate that the character string is a Japanese sentence, for example, a Japanese sentence display code is recorded in the column F indicating the determination result of the table 52.

Process 710: When it is not possible to determine a Japanese sentence in process 708, the maximum count value of the counter UC is other than the highest divided region h, and the maximum count value of the counter DC is below the lowest divided region a. If so, it is determined that this character string is an English sentence.

That is, if max [UC (i)] ≠ UC (h), (i = a to h) max [DC (i)] ≠ DC (a), (i = a to h) are both satisfied, the English sentence and Determine and process step 711
Otherwise, go to step 712.

Process 711: A code indicating an English sentence is recorded in the lower column of the character string rectangular table 52.

Process 712: A display symbol indicating that determination is impossible is recorded in the F column of the character string rectangular table 52.

Process 713: It is determined whether or not the character type determination has been completed for all the character strings, and if there is an unprocessed character string, the process returns to process step 708, and if not, this routine ends.

FIG. 12 is a flow chart showing details of the second judgment processing 65.

Process 800: The lower column of the character string table 52 is checked to find a character string that was judged as undecidable in the previous first judgment.
If not, this routine is terminated, and if there is, the process proceeds to the next step 80 1.

Process 801: Read the character string rectangle Lj data.

Process 802: Search all black pixel circumscribing rectangles included in the character string rectangle Lj. The search condition in this case is the expression (1), and the correspondence between each rectangle and Lj has been saved (for example, the rectangle ID of Lj corresponding to the table 51 is stored) in the already performed first determination processing 64. If you like, you may use it.

Process 803: Initialize a counter LC for counting the distribution of the distance D between adjacent black pixel circumscribing rectangles.

Process 804: 1 of the black pixel circumscribed rectangle searched in step 802
For each R _i , the distance D in the X direction from the adjacent rectangle R _{i + 1} is obtained using each X coordinate.

Process 805: The counter LC (n) corresponding to the distance D is incremented. The index n of the counter is obtained by discretizing the distance D by a predetermined value.

Process 806: For all black pixel circumscribing rectangles included in the character string rectangle Lj, it is checked whether or not the distance determination is completed. If there is an unprocessed rectangle, the process returns to the processing step 804, and if not, the processing step. Proceed to 807.

Process 807: Check the value of the counter LC, and as shown in FIG.
It is checked whether or not the peak values are at two places separated from each other. If there is, it is determined as an English sentence and the process proceeds to step 808, and if not, the process proceeds to step 809. For the determination of the peak value, a counter LC (i) satisfying LC (i) -LC (i-1)> K LC (i) -LC (i + 1)> K with respect to a predetermined slope K indicating the peak point There are two (i = I ₁ , I ₂ ) and | I ₁ −I ₂ |> p However, it is sufficient to determine whether or not p is a value indicating a predetermined distance.

Process 808: The code indicating that the character string rectangle Lj is an English sentence is recorded in the column F of the table 52.

Process 809: The variance of each count value of the counter LC is compared with a predetermined threshold value, and if larger than this, the process proceeds to a process step 810, and if not, the process proceeds to a process step 811.

Process 810: The character string rectangle Lj is determined to be a Japanese sentence, and the display code to that effect is recorded in the column F of the table 52.

Process 811: It is determined whether the process has been completed for all character strings, and if there is an unprocessed character string, the process returns to processing step 800,
If not, this routine ends.

Although the determination processes 64 and 65 have been described above, these processes may be interchanged with each other, or only one of them may be used. In addition, the determination process 64 or
In 65, if the character type cannot be determined, the type of the unknown character string may be estimated from the determination results of the preceding and following character strings.

Incidentally, referring to the English area 21 in FIG. 3, in the majority of the alphabet, the strokes of each character are continuous, and the black pixel circumscribed rectangular area is separated in the vertical direction (for example, a lowercase letter "i", "J") is extremely rare. That is, in the case of an English sentence, since the variances of the respective values of RYmin and RYmax are small, this can be used to distinguish it from a Japanese sentence. For example, if the number of black pixel circumscribing rectangles included in one character string is m, the variance Smin of RYmin is Then, the variance Smax of RYmax can be similarly obtained. Therefore, the Japanese sentence can be judged by Smin> Sminφ and Smax> Smaxφ with respect to the dispersion thresholds Smaxφ and Sminφ. The values of Smaxφ and Sminφ can be changed according to the size of the character by defining them as a function of the height of the character string.

This determination can be added to the flowchart of FIG. 10 instead of the first and second determinations, or as a third determination that supplements these.

Since the character type for each character string can be automatically discriminated by the above method, as shown in FIG. 2, character cutting or character recognition can be performed based on these character types.

〔The invention's effect〕

According to the present invention, in the character area in the document image data, the type of the constituent characters can be automatically discriminated for each character string. Therefore, the discrimination result is used to extract (cut out) the specific character area and perform the character recognition. It is effective in a document image processing system in which the results can be registered and the result is automatically registered as an index of an image file or auxiliary information such as an abstract.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram showing an example of a document image processing system for carrying out the present invention, FIG. 2 is a diagram for explaining a schematic procedure of document image processing according to the present invention, and FIG. 3 (A). ，
FIG. 4B is a diagram showing a concrete example of the input document image and the circumscribed rectangle of the black pixel continuous area and the character string circumscribed rectangle extracted therefrom, and FIG. 4 is a display of the character string circumscribed rectangle and the black pixel continuous area circumscribed rectangle. FIG. 5 is an explanatory diagram of data, FIG. 5 is a diagram showing a configuration of a storage table of each of the above rectangles, and FIGS. 6 and 7 are explanations of frequency distribution count and frequency distribution characteristics in the first character type determination method. FIG. 8 and FIG. 9 are diagrams for explaining the characteristics of the count and frequency distribution of the area distribution in the second character type determination method, and FIG. 10 is the entire image processing according to the present invention. Flow chart, 11th
The figure is a flow chart for implementing the first determination method,
FIG. 12 is a flowchart for carrying out the second determination method. Code distribution 1: Processor, 2: Program memory 3: Dictionary memory, 4: Data memory 5: Keyboard, 6: Image input device 7: Frame memory, 8: Display device, 20: Japanese area, 21: English area L : Character string circumscribing rectangle R: Black rectangle continuous circumscribing rectangle.

Claims (2)

[Claims] 1. A step of fetching document information including a character string into a memory as image data, a step of obtaining a circumscribed rectangle for each black pixel continuous area from the image data, and a step of obtaining a circumscribed rectangle for each character string from the image data. Step, for each of the character string, a step of obtaining a frequency distribution regarding the vertical relative position of the upper side and the lower side of the circumscribing rectangle for each black pixel continuous region included in the circumscribing rectangle of the character string,
A document image processing method comprising the step of determining whether the character type forming the character string is an English character or a Japanese character from the peak position of the frequency distribution. 2. The document image processing method according to claim 1,
Further, a step of obtaining the frequency distribution of the interval in the left-right direction between adjacent black pixel continuous circumscribed rectangles, and determining whether the character type forming the character string is an English character or a Japanese character from the number of peaks of the frequency distribution of the interval A method for processing a document image, comprising the steps of: