JPH09305707A - Image extracting system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly precisely extract a character or the like from a gradation image constituted by containing a straight line and the character, etc., at least. SOLUTION: An interference picture element likelihood calculating means 904 regards picture elements, which exist near the part of contact with the character among picture elements (ruled line picture elements) belonging to the straight line identified by a ruled line removing means 902, as the cluster of picture element strings of width 1 arranged in linear direction, the average density value calculated by an average ruled line density calculating means 903 is compared with the density of respective ruled line picture elements belonging to that picture element string for each picture element, and the possibility of a picture element (interference picture element) overlapping the character and the straight line is found for each picture element. A character stroke likelihood calculating means 905 verifies the continuity of picture elements having the high possibility of the interference picture element and expresses this continuity as the exponent of character stroke likelihood. An interference picture element restoring means 906 extracts the picture element having the high exponent of character stroke likelihood as the interference picture element.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image extraction system,
In particular, the present invention relates to an image extraction method for extracting an image for extracting characters, figures, symbols, etc. that interfere with straight lines such as character frames and ruled lines in an optical character reading device.

[0002]

2. Description of the Related Art A technique disclosed in Japanese Patent Application Laid-Open No. 7-334619 has been proposed as a method for reading a form image including straight lines such as ruled lines and characters using a multivalued image as an input image. In the invention disclosed in this publication, it is assumed that characters are darker than ruled lines, and the threshold for character recognition and the threshold for straight line detection are determined from the state of the input image. If a pixel having a pixel value larger than the threshold value for character recognition is selected, only the character can be uniquely extracted.

Further, as a method for extracting a character interfering with a straight line from a binary image including a straight line such as a frame line and a ruled line and a character, Japanese Patent Publication No. 63-251874 and Japanese Patent Application Laid-Open No. 6-309498. The technology disclosed in US Pat. JP-B-63-251874 and JP-A-6-3
In the method described in Japanese Patent Publication No. 09498, a form image is captured as a binary image, and characters and the like are extracted based on features such as the line width of the character frame and the continuity of character strokes.

[0004]

However, the method described in Japanese Patent Application Laid-Open No. 7-334619 is very effective when the density of ruled lines and characters is clear, but it is not always necessary in actual forms. And the density of letters are not always clearly separated. For example, even a ruled line printed with the same ink has a different shade value between a thick ruled line and a thin ruled line, and even a single ruled line has greatly different densities in the center and peripheral portions of the ruled line.
If the ruled line has a slight inclination, the density value changes subtly along the direction of the ruled line. With regard to the written character, the density is completely different between the part where the writing pressure is applied and the part where it is not.

On the other hand, Japanese Patent Publication No. 63-251874,
A method based on a binary image described in Japanese Patent Laid-Open No. 6-309498 may not be able to uniquely separate a character from a ruled line. This is because the features such as the character frame width and the character stroke continuity on the binary image are not always the features that can uniquely separate the character and the ruled line.

To solve this problem, Japanese Patent Publication No. 63-25
In the invention disclosed in Japanese Patent No. 1874, a hypothesis verification method is adopted in which a plurality of separation states are generated as hypotheses and the hypotheses are verified by character recognition. However, in a form composed of block character frames and ruled lines, compared to a form composed of one character frame, straight lines such as frame lines and ruled lines (hereinafter, the two are simply referred to as ruled lines) and characters are more complicated. There is a problem that the number of hypotheses increases exponentially due to the increased number of cases that interfere with. Also, in order to be able to correctly extract characters even if the horizontal line at the bottom of the number “2” or the horizontal line at the center of the English letter “A” overlaps the ruled line, a large number of separation states are hypothesized. There is also a problem that it needs to be generated as.

In view of the above-mentioned conventional circumstances, the present invention has been made to solve the above-mentioned problems inherent in the prior art. Therefore, the object of the present invention is to provide straight lines such as ruled lines and the density of characters. An object of the present invention is to provide a novel image extraction method that enables highly accurate and unique separation of characters and the like even from form images in which the difference is not so clear.

[0008]

In order to achieve the above object, the invention according to claim 1 (first invention) includes at least a straight line such as a frame line or a ruled line, and characters, figures or symbols. In an image extraction method for extracting characters, figures, symbols, etc. from a grayscale image composed of, a binary image is obtained by separating lines such as ruled lines, characters, figures, symbols, etc. from the grayscale image and the background 2 A binarizing unit and a straight line such as a ruled line are removed from the binary image, and pixels belonging to the removed ruled line (hereinafter abbreviated as ruled line pixels) and pixels included in the binary image after the ruled line removal are also included. Ruled line removing means for storing the position (hereinafter abbreviated as character pixel), and seed pixel selection for selecting the pixel with the highest density among the ruled line pixels adjacent to the character pixel as a seed pixel at each adjacent position Means and the seed pixel On the other hand, an average ruled line density calculating means for calculating the average density of ruled line pixels existing in a straight line direction with reference to the grayscale image, and a pixel in the grayscale image in the straight line direction starting from the position of the seed pixel. When the pixel value is scanned, and the pixel value is larger than the value obtained by adding a constant to the average ruled line density, it is determined to be a pixel in which a character and a straight line overlap (hereinafter referred to as an interference pixel), and the position coordinates are stored, and one pixel value Has an interference pixel restoration means for ending the pixel scanning in that direction when is smaller than a value obtained by adding a constant to the average ruled line density value, and when an interference pixel is extracted by the interference pixel restoration means, The extracted interference pixel is treated in the same manner as the character pixel, and the character pixel and the like are extracted by repeatedly applying the processing after the seed pixel selection means to the location where the ruled pixel is adjacent to the interference pixel.

The invention according to claim 3 (second invention)
Is an image extraction method for extracting characters, figures, symbols, etc. from a grayscale image including at least straight lines such as frame lines and ruled lines, and characters, figures, symbols, etc. Binarization means for extracting a binary image by extracting characters / figures / symbols and a background, and removing straight lines such as ruled lines from the binary image, and removing the ruled line pixels and the binary image after the ruled lines are removed. Ruled line removing means for storing the position of the included character pixel, and for each position where the character pixel and the ruled line pixel are adjacent to each other, the ruled line pixel near the adjacent position is divided into a pixel row of width 1 arranged in the linear direction, and each pixel is divided. Seed pixel selection means for extracting a pixel having the maximum density value from a column as a seed pixel, and an average ruled line density for calculating an average density of ruled line pixels existing in a pixel column to which the seed pixel belongs by referring to the grayscale image Calculating means, and For a ruled line pixel near the contact point, an interference pixel likelihood calculating means for calculating the probability that the pixel is an interference pixel from the difference between the gray value in the gray image and the average density of the pixel row to which the pixel belongs. And a character stroke likelihood calculating means for calculating how consecutively the pixels having a high interference pixel likelihood exist as the character stroke likelihood of the pixel, and the direction of the ruled line to which the seed pixel belongs from the seed pixel as a starting point. Scan the pixels to
If the character stroke likeness is larger than the threshold value, it is determined as an interference pixel and the position coordinates thereof are stored, while if the character stroke likeness is smaller than the threshold value, the pixel scanning in that direction is ended. And an interfering pixel restoring means for performing the same.

[0010]

[Function] Even in a form in which the straight lines such as ruled lines and the densities of characters and the like are not clearly separated, the interference pixels in which the ruled lines and the characters overlap have a higher density than the ruled line pixels of the gold rod existing on the same pixel line in the ruled line direction. Tends to be high.

In the invention of claim 1 (first invention), among the ruled line pixels identified (set) from geometrical characteristics and the like, there is a high possibility that interference pixels exist, that is, ruled line pixels and character pixels are By checking the average density of each pixel row for the ruled line pixels existing around adjacent areas, and determining the pixel with a density higher than this average density as an interfering pixel, the density values of the ruled lines and characters can be clarified. Interference pixels can be extracted with high accuracy and uniquely even from images that are not separated. A character can be extracted by combining the character pixel extracted by the ruled line removing means and the interference pixel. Also,
Interference pixels are restored in order from the periphery of the seed pixel, and if a pixel that is not an interference pixel is found during restoration, the pixel search in that direction is stopped, so that the input image contains some noise components. Even in an image, it is possible to prevent pixels that are not originally interference pixels from being mistakenly restored.

According to the third aspect of the invention (the second aspect of the invention), the character stroke likeness is calculated from the characteristic of how continuously the ruled line pixels darker than the surrounding ruled line pixels exist, and this is calculated. Since it is used as a reference for extracting interference pixels, it is possible to extract interference pixels with a smooth shape if there are some ruled line pixels in which shade differences are clear even if there are several ruled line pixels in which shade differences cannot be extracted. become.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will now be described in detail with reference to the drawings for each of its preferred embodiments.

FIG. 1 is a block diagram showing an embodiment of a character extraction method from a form image constructed according to the invention (first invention) described in claim 1. FIG. 2 shows an example of the input image of this embodiment.

Referring to FIGS. 1 and 2, the binarizing means 10
0 binarizes the input image and separates the character and ruled line from the background. As a binarization method, a method of binarizing the entire image with a uniform threshold value, a method of changing the threshold value for each local region, etc. are generally known. FIG. 3 is a diagram showing a part of an image obtained by binarizing the input image shown in FIG. 2 with a uniform threshold value. The ruled line removing means 101 is the binarizing means 1
The ruled line printed on the form is extracted from the binary image output by 00 and removed. Further, the ruled line removing means 101 stores the positions of the removed ruled line pixels. As a specific storage method, an image area having the same size as the binary image may be secured to store the ruled line position, or the ruled line pixel in the binary image may be stored in the third position.
The third value may be written to form a ternary image. In this embodiment, the ruled line position is stored by the latter method.

FIG. 4 is a diagram showing a part of the ternary image generated by the ruled line removing means 101. As can be seen from FIG. 4, when the operation of the ruled line removing means 101 is completed, the background pixel 400, the character pixel 401, and the ruled line pixel 402 (however, the ruled line pixel includes a pixel where the character and the ruled line overlap = interference pixel). ) Is identified (set).

Next, the operation principle of the seed pixel selection means 102 will be described with reference to FIGS. 4 and 5.

In FIG. 4, there are two places where the character pixel 401 and the ruled line pixel 402 are adjacent to each other. Since the interference pixel to be extracted is a part of the character stroke, there is a high probability that it will be present at a position adjacent to the pixel 401 already known as the character pixel. Further, on the assumption that the interference pixel has a higher density than the surrounding ruled line density, the pixel having the highest probability of being an interference pixel at each adjacent portion is selected as the seed pixel. For example, for the image shown in FIG. 4, two seed pixels 500 and 501 shown in FIG. 5 are selected.

The average ruled line density calculating means 103 calculates the average ruled line density around the various pixels selected by the seed pixel selecting means 102. In the case of horizontal ruled lines is calculated an average density of w _h pieces of ruled lines pixels existing on the left and right directions around the seed pixel. In the case of a vertical ruled line, the average density of w _v ruled line pixels existing in the vertical direction centering on the seed pixel is calculated. Generally, towards the overlapping length between the lateral borders shorter than the overlap length between the vertical ruled lines, so-w _h <w _v.
Even if the number of ruled line pixels for calculating the average ruled line density does not reach the above number, the average value is calculated only from the ruled line pixels existing there.

Further, the reason why the average value is calculated only from the pixels on the same column as the seed pixel will be described with reference to FIG.

FIG. 6 is a diagram showing the density value of a portion where a horizontal ruled line and a character stroke in the vertical direction intersect, which is observed in an actual form image. From this figure, the values of the interfering pixel and the other ruled line pixels are clearly separated in the pixel row 600 arranged in the ruled line direction, but the density values are different near the center and around the ruled line. I understand. For example, in the ruled line pixel 601 in the center of the ruled line and the interference pixel 602 in the ruled line outline, the ruled line pixel 601 has a higher density value. This is because one CCD pixel included in the image input device has a diameter of a region in which light is sensed that is considerably larger than the ruled line width. Therefore, the ruled line pixel is separated into an interference pixel and other pixels. In order to do this, it is effective to check the change in density for each pixel column.

Next, the operation principle of the interference pixel restoration means 104 will be described with reference to FIG. If the horizontal direction is the x-axis and the vertical direction is the y-axis, the interference pixel restoration unit 104 searches for pixels in the horizontal direction for horizontal ruled lines and in the vertical direction for vertical ruled lines, and the grayscale value g ( (x, y) and the average density value calculated by the average ruled line density calculation means 103 for the seed pixel, g (x _s + x, y)> average value + α (1) Pixel P (x _s + x, y) satisfying the above condition is restored as an interference pixel, and the fourth value corresponding to the interference pixel is written in the label image. Here, x _s indicates the x coordinate of the seed pixel, and α indicates a constant. FIG. 7 shows a case of a horizontal ruled line. First, the seed pixel 70
The density of 0 is compared with the average value, and if the expression (1) is satisfied, it is determined to be an interference pixel, and the density of the pixel 701 on the left is checked. If the pixel 701 also satisfies the expression (1), the pixel in the left direction is further searched, and if a pixel that does not satisfy the expression (1) appears, the pixel on the right side of the seed pixel is similarly searched. Go. When the search for the pixel on the right side is completed, the restoration of the interference pixel for this kind of pixel is completed.

Then, the extracted interference pixel is treated in the same manner as the character pixel, and the processing from the seed pixel selection means 102 and the subsequent processing is repeated for the adjacent portion of this pixel and the ruled line pixel. Then, the interference pixel is further restored, and a label image as shown in FIG. 8 is obtained for the image as shown in FIG. Here, by extracting only the pixels having the character pixel label and the pixels having the interference pixel label, the character can be correctly extracted. In FIG. 8, reference numeral 803 is an interference pixel.

Next, an embodiment of a character extraction method from a form image constructed according to the invention (second invention) described in claim 3 will be described with reference to FIG.

Referring to FIG. 9, the binarizing means 900 and the ruled line removing means 901 have completely the same functions as those shown in the above-mentioned first invention.

First, the function of the seed pixel selection means 902 will be described with reference to FIG. Seed pixel selection means 90
In the image 2 output by the ruled line removing unit 901, the pixels 1004 to 1007 having the maximum density value are used as seed pixels in each pixel column 1000 to 1003 near all interference points where the ruled line label pixels and the character label pixels are adjacent to each other. elect. The length of each pixel column, that is, the search range of the seed pixel may be determined as follows. First, the ruled line pixel column 1000 adjacent to the character pixel column 1008 and the ruled line pixel column 1002 adjacent to the character pixel column 1009 are composed only of ruled line pixels that are in four-connection with the character pixel. And character pixel row 1
As the distance from 008, 1009 increases, the length of the pixel row is set to be slightly longer. By doing so, even when a character stroke intersects a ruled line at an acute angle, it is possible to increase the probability that a pixel that should originally be an interference pixel is selected as a seed pixel.

The average ruled line density calculation means 903 calculates the average ruled line density of the ruled line pixels existing on the same column for the various pixels selected by the seed pixel selection means 902.

The interference pixel likeness calculating means 904 compares the gray value of the ruled line pixel existing on the same column as the seed pixel with the average ruled line density calculated for the seed pixel, and the certainty factor that each ruled line pixel is an interference pixel. (Interference pixel likeness) is calculated.
The likelihood of interfering pixels is calculated only for w _h ruled lines that are on the left and right of the seed pixel in the case of a horizontal ruled line and w _v ruled lines that are above and below the seed pixel for a vertical ruled line, and is farther than the seed pixel. It is not necessary to find the ruled line pixels that are far apart. The interference pixel likelihood O (x, y) is defined by O (x, y) = f (g (x, y) -average value) (2). Here, the function f represents a monotonically increasing function shaped like a sigmoid function. The upper limit of the range of this function is Omax and the lower limit is Omin.

When there are a plurality of interference points between ruled lines and characters in close proximity to each other, it is possible to define a plurality of interference pixel-like pixels for one ruled line pixel. In such a case, the maximum value is set as the interference pixel of that pixel. Let's be like However, the minimum value is adopted for the ruled line pixel at the intersection of the vertical ruled line and the horizontal ruled line.

The character stroke likelihood calculating means 905
With respect to the pixels for which the likelihood of interference pixels has been calculated, how consecutively the pixels with high likelihood of interference pixels are arranged is represented by a quantitative value called "character stroke likelihood". First, the directivity in a minute area is defined by the following equation (3).

[0031]

However, in this calculation, the likelihood of interfering pixels with respect to the character label pixel and the background label pixel is O, respectively.
Let (x, y) = Omax and O (x, y) = 0 min. To facilitate understanding of this calculation formula, as an example, D
_The principle of calculating ₀ (x, y) will be described with reference to FIG. D
_O (x, y) is two filters 1100 that are long in the lateral direction,
And 1101 is the same as using the larger one of the product-sum operation results, and thus controls the extraction of lateral strokes having a width of about 2 pixels or more. D ₄₅ (x, y), D
₉₀ (x, y) and D ₁₃₅ (x, y) can also be calculated with a filter that is inclined by 45 degrees.

The character stroke likelihood of each pixel is L (x, y) = MAX (L _O (x, y), L ₄₅ (x, y), L ₉₀ (x, y), L ₁₃₅ (x, y) )) ・ ・ ・ ・ ・ ・ (4)

[0034]

It is defined by Value L _O (x, y), L ₄₅
(X, y), L ₉₀ (x, y), and L ₁₃₅ (x, y) are rows of interference pixels near the pixel (x, y) in the horizontal direction, the right diagonal direction, the vertical direction, and the left diagonal direction. Indicates the likelihood of existence.

To facilitate understanding of this calculation formula, FIG.
The calculation method of L _O (x, y) will be described using 2. The fact that there is an interfering pixel row (= character stroke) in the horizontal direction at the pixel 1200 means that there should be a character stroke connected to it as well. Therefore, pixel position 1
In each of 201 to 1203 and 1204 to 1206, the one having the maximum value of D in the illustrated direction is selected, and from this and the value of D _{0 at} the pixel position 1200, the horizontal character stroke at the pixel position 1200 can exist. Calculate the sex. L ₄₅ (x, y), L ₉₀ (x,
y) and L ₁₃₅ (x, y) are also the same as when they are inclined by 45 degrees.

The interference pixel restoration means 906 restores a pixel having a large character stroke likeness as an interference pixel. As in the first aspect of the invention, the restoration order starts from the ruled line pixel row in contact with the character label pixel and proceeds in the direction opposite to the ruled line. In one ruled line pixel row,
A pixel is searched in the direction of the ruled line using the seed pixel as a starting point, and if the character stroke likeness is larger than the threshold value, it is restored and the pixel next to it is focused, and if it is smaller than the threshold value, the search is made in that direction. Discontinue.

Form image 5 of a blue ruled line form using the present invention
An experiment was conducted to extract characters from 0 sheets. As a result, an extraction success rate of 94.1% was obtained for a portion where character extraction is relatively difficult, where ruled lines and character strokes overlap in a long section.

In this embodiment, an example in which the present invention is applied to a form image has been described, but it is effective when a character, a symbol, or a figure is extracted from an image in which a straight line segment touches the character, the symbol, or the figure other than the form image. Is. An example is underline removal.

Next, an embodiment of a character code extraction method from a form image constructed according to the invention of claim 2 or claim 3 will be described with reference to the block diagram shown in FIG. In this embodiment, the interference pixel restoration means 130 shown in FIG.
4 (104 in FIG. 1), the character cutting means 13
05 and character recognition means 1306 are added to the system. When a grayscale image as shown in FIG. 14 is input to the system shown in this block diagram, the final output image of the invention of claim 1, that is, the interference pixel restoration means 1304 is an image containing only the characters shown in FIG. .

The character cutting means 1305 is means for generating an image for each character from the main image. As a concrete implementation method of this means, a method of using valley points of a black pixel histogram obtained when black pixels are projected in the vertical direction, a method of using a labeling result of connected black pixels, etc. are generally known. There is.

In FIG. 16, the character cutting means 1305 is
It is a figure which shows the example of the character image extracted from the uppermost character string of the image shown in FIG. 15 by these methods. The character recognition means 1306 recognizes each character image shown in FIG.
Output each character code. Interference pixel restoration means 1
As shown in FIG. 15, since the image output by 304 is restored to a state in which the character shape in the vicinity of the portion where the character interferes with the straight line component such as ruled line is approximated to the actual shape, the conventional individual character recognition is performed. High read rates can be obtained using the method. However, in the character reading method using the invention disclosed in Japanese Patent Publication No. 63-251874, the character shape near the point where the line component such as a ruled line interferes with the character has a stair shape. There is a problem in that an inappropriate value is extracted as a feature amount for making the character recognition difficult.

[0043]

As described above, according to the present invention,
Characters, figures, and symbols can be extracted with high accuracy from grayscale images that include straight lines such as frame lines and ruled lines, and characters, figures, and symbols.

Further, since the character extracted by the present invention faithfully restores the original character shape, an image of excellent quality as an input image for the character recognition processing which is likely to be executed in the latter stage of the present invention. Can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment according to the first invention.

FIG. 2 is a diagram showing an example of an input grayscale image in the first invention.

FIG. 3 is a diagram showing a part of a binary image of the form shown in FIG.

FIG. 4 is a diagram showing an example of a label image output by a ruled line removing unit according to the first invention.

FIG. 5 is a diagram showing an example of positions of seed pixels set by a seed pixel selection unit according to the first invention.

FIG. 6 is a diagram showing an example of a gray value at an intersection of a ruled line and a character.

FIG. 7 is a diagram showing an order in which an interference pixel restoration unit according to the first invention searches for a pixel.

FIG. 8 is a diagram showing an example of a label image after restoration of interference pixels.

FIG. 9 is a block diagram showing an embodiment according to the second invention.

FIG. 10 is a diagram showing an example of positions of seed pixels set by a seed pixel selection unit in the second invention.

FIG. 11 is a diagram showing two types of filters for calculating continuity (horizontal direction) of interference pixels in a minute area.

FIG. 12 is a diagram illustrating a method of calculating character stroke likelihood (horizontal direction).

FIG. 13 is a block diagram showing an embodiment of the invention described in claim 2.

FIG. 14 is a diagram showing an example of an input grayscale image.

FIG. 15 is a diagram showing an output image of the interference pixel restoration unit when FIG. 14 is an input image.

FIG. 16 is a diagram showing an example of a character image output by a character cutout unit.

[Explanation of symbols]

Reference numeral 100 ... Binarization means 101 ... Ruled line removal means 102 ... Seed pixel selection means 103 ... Average ruled line density calculation means 104 ... Interference pixel restoration means 400 ... Background pixels 401 ... Character pixels 402 ... Ruled line pixels 500, 501 ... Seed pixels 600 ... Pixel column arranged in ruled line direction 601 ... Ruled line pixel in the center of horizontal ruled line 602 ... Interference pixel near horizontal ruled line outline 700 ... Seed pixel 701 ... Left pixel of seed pixel 800 ... Background pixel 801 ... Character pixel 802 ... Ruled line pixel 803 ... Interference pixel 900 ... Binarization means 901 ... Ruled line removal means 902 ... Seed pixel selection means 903 ... Average ruled line density calculation means 904 ... Interference pixel likeness calculation means 905 ... Character stroke likeness calculation means 906 ... Interference pixel restoration means 1000, 1001, 1002, 1003 ... Search range of seed pixels 1004, 1005, 1006, 007 ... Seed pixel search range 1000, 1001, 1002, 1003, selected seed pixel 1008, 1009 ... Character pixel adjacent to ruled line pixel 1100, 1101 ... Directionality of interference pixel in minute area (horizontal direction) Filters for calculating the values 1200, 1201, 1202, 1203, 1204,
1205, 1206 ... Pixels 1300 ... Binarization means 1301 ... Ruled line removal means 1302 ... Seed pixel selection means 1303 ... Average ruled line density calculation means 1304 ... Interference pixel restoration means 1305 ... Character cutout means 1306 ... Character recognition means

Claims (4)

[Claims] 1. An image extraction method for extracting characters, figures, symbols, etc. from a grayscale image including at least straight lines such as frame lines and ruled lines and characters, figures, symbols, etc., wherein a ruled line is created from the grayscale image. Binarizing means for separating a straight line, characters, figures, symbols, etc. from a background to obtain a binary image, and removing straight lines such as ruled lines from the binary image, and pixels belonging to the removed ruled lines ( A ruled line removal unit that stores the positions of pixels (hereinafter abbreviated as ruled line pixels) and the positions of pixels (hereinafter abbreviated as character pixels) belonging to a character or the like included in a binary image after ruled line removal, and the ruled line adjacent to the character pixel A seed pixel selection unit that selects the pixel with the highest density among the ruled line pixels as a seed pixel at each adjacent position, and the average density of the seed pixel and the ruled line pixels existing in a straight line direction with respect to the seed pixel, with reference to the grayscale image. Calculated And an average ruled line density calculation means for scanning the pixels in the grayscale image in the direction of a straight line starting from the position of the seed pixel, and when the pixel value is larger than a value obtained by adding a constant to the average ruled line density. A pixel in which a character and a straight line overlap each other (hereinafter referred to as an interference pixel) is determined and its position coordinates are stored. On the other hand, when the pixel value is smaller than the value obtained by adding a constant to the average ruled line density value, the direction to that direction is set. An interfering pixel restoring means for ending the pixel scanning, and when the interfering pixel is extracted by the interfering pixel restoring means, the extracted interfering pixel is treated in the same manner as a character pixel, and a portion where the ruled line pixel and the ruled line pixel are adjacent to each other. An image extraction method characterized in that characters and the like are extracted by repeatedly applying the processing after the seed pixel selection means to the. 2. A character cutout unit that determines a region where each character or the like exists from an image including the character or the like generated by the interference pixel restoration unit and outputs a character image that includes only one character, and the character image A character recognition means for recognizing and outputting a character code is further provided.
Image extraction method described in. 3. A straight line such as a frame line or ruled line and a character
From a grayscale image that includes figures or symbols,
An image extraction method for extracting characters, figures, symbols, etc., comprising: binarizing means for extracting straight lines such as ruled lines, characters, figures, symbols, and background from the grayscale image to obtain a binary image;
Ruled line removing means for removing straight lines such as ruled lines from the value image and storing the positions of the removed ruled line pixels and the character pixels contained in the binary image after the ruled line removal, and a position where the character pixel and the ruled line pixel are adjacent to each other. A seed pixel selecting unit that divides ruled line pixels near adjacent portions into pixel columns of a width 1 arranged in a linear direction for each pixel, and extracts a pixel having the maximum density value as a seed pixel from each pixel column, and the seed pixel. And an average ruled line density calculating means for calculating an average density of ruled line pixels existing in a pixel row to which the pixel belongs, and a gray value and its pixel in the gray image for ruled line pixels near the adjacent portion. From the difference in the average density of the pixel row to which the pixel belongs, an interference pixel likelihood calculating means for calculating the probability that the pixel is an interference pixel, and how consecutively the pixels with high interference pixel likelihood are present. Character stroke likelihood calculating means for calculating as the character stroke likelihood of the pixel of, and scanning the pixel in the direction of the ruled line to which the seed pixel belongs, starting from the seed pixel, when the character stroke likelihood is larger than a threshold value. Includes an interfering pixel restoring unit that stores the position coordinates of the interfering pixel and stores the position coordinate, and ends the pixel scanning in that direction when the character stroke likelihood is smaller than the threshold value. Image extraction method. 4. A character cutout unit that determines a region where each character or the like exists from an image including the character or the like generated by the interference pixel restoration unit, and outputs a character image including only one character, and the character image A character recognizing means for recognizing and outputting a character code is further provided.
Image extraction method described in.