JP2918666B2 - Text image extraction method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a method of extracting a character image from a document image input as a binary image in a character recognition system.

[Conventional technology]

In a character recognition system, a line is cut out from a document image input as a binary image by a scanner or the like, and a character image is cut out for each line to perform character recognition.

The method of extracting a character image is roughly classified into a method using projection and a method using a circumscribed rectangle of connected black pixels. However, in any method, a block of projection or one circumscribed rectangle is not always the data of one character, so a technique for obtaining one character data by combining or dividing those data is known. is necessary.

Therefore, conventionally, a circumscribed rectangle of a connected black pixel (character element) is extracted, and a width of a single circumscribed rectangle or a combination of a plurality of circumscribed rectangles is compared with a preset standard character width, and a width that can be regarded as one character A method of cutting out a character image using a circumscribed rectangle or a combination area thereof as a character frame (Japanese Patent Laid-Open No. 61-17670) is known.

Also, a method is known in which a single circumscribed rectangle or a combination of regions is regarded as a character image and character recognition is performed, and a character image is determined by evaluating the certainty of the result (see 1988 patent by the same applicant). Application No. 133424).

[Problems to be solved by the invention]

Characters consisting only of vertical line segments, such as the kanji numbers "1, 2, 3" in vertical lines and the "Kanji" in horizontal lines, may not have the width of one character sentence. If they appear many times or continuously, no matter where they are separated, a reasonable recognition result can be obtained.

Therefore, in the case where such characters appear continuously, etc., the cutout error is likely to occur in the above-described conventional method, even if the width of the front and rear spaces is considered.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a character image extracting method capable of extracting an accurate character image even when a character image such as a Chinese numeral consisting only of line segments perpendicular to the line direction appears continuously.

[Means for solving the problem]

The present invention relates to a character image extracting method for extracting a character image from a line image by obtaining a character frame by integrating a character element circumscribed rectangle extracted from a line image, focusing on a character element circumscribed rectangle having a width smaller than a predetermined value. Performing a condition determination on the length of the focused vowel circumscribed rectangle and the succeeding vowel circumscribed rectangle and the width at the time of integration, and according to the result of the condition determination, the focused vowel circumscribed rectangle and the subsequent vowel circumscribed rectangle By controlling the integration of the characters, a character frame of a specific character consisting of only line segments perpendicular to the line direction is obtained.

(Operation)

In the case of a specific character consisting of only line segments perpendicular to the line direction, such as the vertical Chinese character "123", each line segment constituting each character is extracted as a character element, but its circumscribed rectangle is The width is sufficiently small, and the length of each line segment has a ratio relationship unique to the character. The width of the entire character is also unique to the character.

Therefore, according to the above-described present invention, when the character string “123” in the vertical writing line is taken as an example, it is possible to cut out exactly as follows.

Attention is first paid to the fact that the width of the character circumscribed rectangle of “one” is sufficiently small. The width integrated with the next character circumscribed rectangle (corresponding to the upper line segment of “two”) is the space between characters Since the space between the characters is too large for a standard character, the target circumscribed rectangle is determined to constitute a separate character from the next morpheme circumscribed rectangle, and is not integrated. The image of “one” can be cut out using the circumscribed rectangle alone as a character frame.

Next, we focus on the character circumscribed rectangle corresponding to the upper line segment of "2". Not as overkill. Furthermore, the length (length of a line segment) of the character circumscribed rectangle is shorter than the length of the next character circumscribed rectangle. Under these conditions, the two circumscribed rectangles are "two"
Are determined, and the images are integrated, and the image of “two” can be cut out using the integrated rectangle as a character frame.

Next, attention is paid to the circumscribed rectangle of the character element corresponding to the uppermost line segment of “3”. The width integrated with the next morpheme circumscribed rectangle is not too large as a standard character width, but the relationship of length is opposite to "two". The width integrated up to the next circumscribed rectangle is
The standard character width is not excessive, and the length of the last character circumscribed rectangle is larger than the length of the central character circumscribed rectangle. From such conditions, it is possible to determine that the three circumscribed rectangles constitute "three" and integrate them, and to cut out the image of "three" using the integrated rectangle as a character frame.

〔Example〕

FIG. 1 is a block diagram of a character recognition system according to one embodiment of the present invention.

A document image to be processed is input as binary image data from a scanner or the like and stored in the original image memory 1.
The character image cutout processing unit 2 performs a process of cutting out the character image of the document image in the original image memory 1 and stores the cutout character image in the character image memory 3. Character recognition processing unit 4
Performs normalization of the character image, extraction of the feature amount, and comparison with the dictionary to recognize the character, and stores the recognition result in the recognition result memory 5.

In the character image cutout processing unit 2, the line cutout unit 6 extracts one line of character elements from the document image and stores the character element data in the character element data memory 7. For example, a connection of black pixels on a document image is searched for, a block of connected black pixels is extracted as a character element (a group of black pixels forming a character or a part of a character), and coordinates of a circumscribed rectangle of the character element (for example, (Coordinates of diagonal vertices) are extracted as character element data. Further, the horizontal and vertical distances between the character circumscribed rectangles are obtained, and a set of character circumscribed rectangles whose distance in both directions is smaller than a certain threshold is cut out as one line, and the data is stored in the character element data memory 7.
To be stored. Such line segmentation has a stronger skew of a document than line segmentation by line direction irradiation.

The parameter setting unit 8 overlooks the entire character element data in the character element data memory 7, sets a relative parameter for use in character extraction (character element integration and character pattern synthesis), and stores the parameter in the parameter memory 9. The character element integration is performed a plurality of times. Each time the processing is completed once, the parameters are reset based on the corrected character element data in the character element data memory 7.

 The parameters to be set are as follows.

(A) line height (the distance from the bottom to the top of a horizontal line,
(Distance from the left edge to the right edge of the vertical writing line) This is the maximum height of the character circumscribed rectangle constituting the line. However, the height of the character circumscribed rectangle is the distance from the lower end in the case of horizontal writing to the farthest side of the character circumscribed rectangle from the left side in the case of vertical writing.

(B) Standard character width (width is the size in the line direction) A value slightly larger than the line height, for example, [line height] +
(Dot).

(C) Minimum character spacing The minimum spacing between adjacent character circumscribed rectangles. However,
It is determined only for a character circumscribed rectangle having a certain width, for example, half or more of [standard character width].

Before the first character element integration processing, the following initial value is set.

[Initial value] =-[line height] / 30 (Negative value means overlapping with the character circumscribed rectangle.) (D) Maximum character width The character circumscribed rectangle constituting the line has the maximum width. I do.

(E) Knth This is the maximum width of a character circumscribed rectangle that can be regarded as a character consisting of only one line segment in the vertical direction on the line, and is set to, for example, [line height] / 8.

The character element integration unit 10 integrates character element circumscribed rectangles that are regarded as the same character in order from the character element at the head of the line by a condition determination using the parameters in the parameter memory 9,
The character data in the character data memory 7 is corrected. This character element integration processing includes a normal processing mode and an exception processing mode.

First, the contents of the exception handling mode will be described. Second
The figure is a flowchart for explaining the exception handling mode. However, W1 is the character width of the current character element, W2 is the character width of the next character element of the current character element, W3 is the character width of the next character element of W2, and L1 is the current character. The length of the element, L2 is the length of the next element of the current element, L3 is the length of the element next to the element W2, and W1.2 is the element of the element of interest and the next element. W1.2.3 is the character width when the unified character element and the next two character elements are integrated.

Certain characters that are separated into one line segment perpendicular to the line direction, such as "kawa" in horizontal lines and "123" in vertical lines, are separated into specific If the interval is sufficiently wider than the character element intervals of other characters, it is possible to cut out in the normal processing mode, but such a condition is rarely satisfied in a few documents. This exception handling mode is
This is a mode for performing a character element integration process for accurately extracting such a specific character.

A comparison is made between the character width w1 (the size of the circumscribed rectangle in the line direction) of the focused pixel 1 and the parameter Knth (step). If w1 <Knth, the focused pixel 1 is perpendicular to the line direction. Assuming that the line consists of only one line segment, the system enters the exception handling mode.

In the exception handling mode, first, a comparison is made between the character width 1.2 when the target character element 1 and the next character element 2 are integrated and the standard character width (step), and if w1.2 ≧ standard character width The target character element 1 is determined as a single character element. For example, in the character string of "123" in the vertical writing line as shown in FIG. 3, when focusing on the first character element of "1", it becomes w1.2 standard character width, so it is determined as a single character element. I do.

If w1.2 <standard character width in the step, the length of the character element 1 of interest (the size in the direction perpendicular to the line direction of the circumscribed rectangle) L1 and the length L2 of the next character element 2 (L)>L1> Leme 2 if L2> L1
Is integrated into one character. For example, if the character element of the line segment above “two” shown in FIG.
Since the condition of the step is satisfied, the two phonemes constituting "two" are deterministically integrated into one phoneme.

If L2 ≦ L1 in the step,
Character width w obtained by integrating the next character element 2 and the next character element 3
Perform a comparison judgment between 1.2.3 and standard character width (step), w
If 1.2.3 <standard character width, compare the length L2 of character element 2 with the length L3 of character element 3 (step), and if L3> L2, deterministic integration of character element 1.2.3 I do. For example, the third
Assuming that the uppermost line segment of “3” shown in the figure is the target character unit 1, the condition of the step is satisfied.
Are deterministically integrated.

As is apparent from the above description, the character element integration result shown in FIG. 4 is obtained for a character string that is difficult to extract in the normal processing mode shown in FIG. 3, so that each character can be correctly extracted. .

Next, the contents of the normal processing mode will be described with reference to FIGS. However, for convenience, it is assumed that the line is a horizontal writing line.

1) If two adjacent primal circumscribed rectangles a and b satisfy the conditional expressions (1) and (2) at the same time, they are integrated into one primal circumscribed rectangle (see FIG. 5).

 Wab ≦ [standard character width] Conditional expression (1) where Wab is the character width of the combined rectangle of a and b.

Xsb−Xea <[minimum character interval] Conditional expression (2) where Xsb is the starting point x coordinate of b, and Xea is the ending point x coordinate of a.

Conditional expression (1) indicates that the width does not exceed the width Aab [standard character width] of the combined circumscribed rectangle, and conditional expression (2) indicates a, b
Is smaller than [minimum character spacing].

However, if the condition (3) is met, the characters are integrated even if they exceed the standard characters.

Xeb <Xeb Condition (3) where Xeb is the end point x coordinate of b. The positional relationship between a and b is such that a is on the left (toward the beginning of the line) and b is on the right.

That is, as shown in FIG. 6, two completely overlapping circumscribed rectangles a and b satisfying Dab <0 are integrated even if Wab ≧ [standard character width].

2) In the case where the conditional expression (1) is satisfied in the above 1) but the conditional expression (2) is not satisfied, the right-hand circumscribed rectangle (b)
Satisfies all of the following conditional expressions (4) to (7), it is regarded as a character circumscribed rectangle of a voiced or semi-voiced dot, and exceptionally integrated (see FIG. 7).

Wb ≦ fth Conditional expression (4) Hb ≦ fth Conditional expression (5) (Hb is the height of b) Yeb−Tsa ≦ fth Conditional expression (6) (Yeb is the y coordinate of the end point of b, Yea is the y coordinate of the start point of a) Xsb−Xea <[minimum character spacing] + α Conditional expression (7) where fth (threshold) = [line height] / β. α and β
Is an empirical value, for example, α = 2 and β = 3.

3) In the above 1), if conditional expressions (1) and (2) are satisfied, but the right-hand circumscribed rectangle (b) satisfies the following conditional expressions (8) to (10), punctuation marks are used. Assuming that there is, there is no exceptional integration (see FIG. 8).

Wb ≦ fth Conditional Expression (8) Hb ≦ fth Conditional Expression (9) base−Yeb ≦ fth / 7 Conditional Expression (10) where base is the y coordinate of the baseline, and 7 is an empirical value, for example, 4 is selected.

One line of character element data in which integration processing in the normal processing mode as described above is performed (excluding character elements determined in the exception processing mode)
Each time is completed once, the parameter setting unit 8 resets the parameter based on the corrected character element data, and rewrites the parameter memory 9. Using the rewritten parameters, the character element integration unit 10 executes the processing in the normal processing mode again.

Thus, the character element integration process is repeated while dynamically modifying the parameters. This repetition ends, for example, when the parameter stops changing or when the amount of change in the parameter becomes equal to or less than a certain threshold.

FIG. 9 shows an example of character element integration in the normal processing mode for horizontal writing lines. For the line image shown in FIG.
Extracts character elements as shown in FIG. These character elements are integrated as shown in (c) by the first integration processing, and integrated as shown in (d) by the second integration processing.

At the stage where the integration processing is completed as shown in (d), most of the character circumscribed rectangles correspond to the circumscribed rectangle (character frame) of one character. That is, circumscribed rectangles of most characters are generated. However, a character element to be integrated such as “small” in this example may remain separated.

Since there is a possibility that such unfinished character elements may remain, the character image cutout processing unit 2 in the character pattern synthesizing unit 11 simply sets each character element circumscribed rectangle after the integration processing to a character frame from the original image. Rather than cutting out an image, continuous rectangles of circumscribed characters are combined so as not to exceed [maximum character width] (stored in the parameter memory), and the combined rectangles (including one circumscribed rectangle of character) Are used as character frames to extract character image candidates from the original image and output them to the character image memory 3. However, for a character element determined in the exception processing mode of the character element integration process, only the character image cutout is performed alone.

The character recognition processing unit 4 performs character recognition on the character image candidates stored in the character image memory 3. However, for character image candidates with duplicated phonemes, the respective recognition results are evaluated, and the most probable character is selected as a foreign character image candidate (that is, character extraction is determined from the character recognition result), and the recognition is performed. Only the result is output to the recognition result memory 5.

The method of recognizing such a combination of character elements and determining the character cutout by evaluating the recognition result is described in detail in the specification and the drawings of the above-mentioned 1988 Patent No. 133424.

[Results of the Invention]

As described above in detail, according to the present invention, when a specific character consisting of only a vertical line segment in the row direction (for example, a kanji numeral "123" in a vertical writing line) appears continuously. In addition, accurate character image cutting can be performed.

[Brief description of the drawings]

FIG. 1 is a block diagram of a character recognition system according to an embodiment of the present invention, FIG. 2 is a flowchart for explaining an exception processing mode of character element integration, and FIG. 3 is a character string to be subjected to the exception processing mode. FIG. 4 is a diagram showing a result of the integration processing of the character strings shown in FIG. 3, FIGS. 5 to 8 are explanatory diagrams of integration conditions for a normal processing mode of character element integration, and FIG. It is a figure which shows the example of figure letter element integration. 1 ... Original image memory, 2 ... Character image cutout processing unit, 3 ... Character image memory, 4 ... Character recognition processing unit, 5 ... Recognition result memory, 6 ... Line cutout unit, 7 ... Character element Data memory 8, Parameter setting section 9, Parameter memory 10, Character element integration section 11, Character pattern synthesis section

Claims (1)

(57) [Claims] In a character image extracting method for extracting a character image from a line image by obtaining a character frame by integrating a character circumscribed rectangle extracted from a line image, attention is paid to a character circumscribed rectangle having a width smaller than a predetermined value. Then, condition determination is performed on the length of the focused vowel circumscribed rectangle and the succeeding vowel circumscribed rectangle and the width at the time of integration, and according to the result of the condition determination, the focused vowel circumscribed rectangle and the subsequent vowel circumscribed are determined. A character image clipping method characterized by obtaining a character frame of a specific character consisting of only line segments perpendicular to the line direction by controlling the integration of rectangles.