JP3187894B2 - Document image tilt detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention detects the inclination of a document image.
The present invention relates to a method of detecting a document image inclination.

[0002]

2. Description of the Related Art Conventionally, as a method of detecting the inclination of an image, a method of detecting the inclination of an image by performing a two-dimensional Fourier transform process on a document image (22nd National Convention of Information Processing Society, pp. 677-678). ), Or a method of dividing an image into a plurality of band-like regions, finding peripheral distributions in each region, and detecting an image inclination from a phase shift between the peripheral distributions (Transactions of the Institute of Electronics, Information and Communication Engineers, 1983 V).
ol. J66-D No. 1 pp. 111-11
8).

[0003]

However, each of the above-mentioned methods has a drawback that the amount of data to be processed increases because the inclination of the image is obtained by performing processing on the entire image. An object of the present invention, by obtaining the inclination of the character string region or <br/> et document image is to provide a document image tilt detection method reduces the amount of processing.

[0004]

To achieve the above object, according to the first aspect of the present invention, a binarized document image is provided.
(Hereinafter referred to as “original image”)
Reducing the original image and determining whether the reduced image
A character string region from the original image corresponding to the character string region.
About the circumscribed rectangle of the black pixel connected component in the character string area on the image
Create a histogram of the height of each circumscribed rectangle,
The height of the circumscribed rectangle that takes the maximum value is used as the reference value, and
To convert a circumscribed rectangle larger than a predetermined threshold into a reference value,
The position of the circumscribed rectangle converted to the reference value and the reference value
Of the document image based on the position of the circumscribed rectangle not to be converted to
Is detected .

[0005]

A document is captured as a binary image by an image input unit such as a scanner, and the input image is reduced by an image reduction unit in order to extract a character string area from the input image. The block generation and classification unit extracts a block by performing a blurring process in a character string direction of the reduced image, classifies the block into a figure, a photograph, and a character string region based on the size of the block. Then, an area whose horizontal length is three times or more the vertical length is selected from the areas extracted as the character string areas . The inclination detection unit calculates a straight line having the smallest deviation from the center point of each circumscribed rectangle in the selected area by using the least squares approximation method, and the inclination of the calculated straight line becomes the inclination of the character string.

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be specifically described below with reference to the drawings. FIG. 1 is a block configuration diagram of an embodiment of the present invention. Reference numeral 101 denotes an image input unit that reads and inputs a document image such as a scanner; 102, an image reduction unit that reduces an input image; A block generation unit that extracts blocks from the reduced image and classifies the extracted blocks into a photograph and a character string area ; a classification unit; 104, an image inclination detection unit; 105, an input image, a reduced image, block data, and other data Is a control unit that controls each unit.

FIG. 2 shows a processing flowchart of the present invention. In step 201, a document is captured as a binary image by an image input unit 101 such as a scanner. Next, in order to extract a character string region from the input image, the image reduction unit 102 reduces the input image. This image reduction is to reduce the image vertically and horizontally so that the black pixels of the upper, lower, left and right character strings do not touch (step 202). The block generation / classification unit 103 extracts a block by performing a blurring process in a character string direction of the reduced image, and extracts a diagram, a photograph, and a sentence from the size of the block.
Classify into character string area .

[0008] Here, the blur processing means that the run-length data of an image is separated by a long white run .
Black orchid and a short white orchid ties one of the range (run of the collector
), And a collection of the blurred runs
A block that is connected vertically is called a block.

The block generation / classification unit 103 further outputs a sentence
From the regions extracted as character string regions , a region having a horizontal length that is at least three times the vertical length is selected (step 204).
By selecting an area including a plurality of characters as described above, many black pixel connected components are to be calculated, and therefore, the accuracy of tilt detection can be improved.

FIG. 3 shows that the horizontal length is three times or more the vertical length.
A character string area 301 is shown, and 302 to 308 are circumscribed rectangles of black pixel connected components. Note that this black pixel is not a pixel on the reduced image, but the input original image ( binary document
Image ). That is, the circumscribed rectangles 302 to 30
8 is a sentence on the original image corresponding to the extracted character string area
The circumscribed rectangle of the black pixel connected component in the character string area 301.

The inclination detecting unit 104 calculates the center point of each circumscribed rectangle (the center point of the circumscribed rectangle 302 is 310).
Is calculated using the least squares approximation method, and the slope of the calculated straight line is the slope of the character string to be obtained. When calculating the straight line, each center point is weighted in proportion to the area of the circumscribed rectangle.

That is, when y (x) = a + bx, a and b are determined so as to minimize J at J = {mi {y (xi) -fi} * 2. Here, each center point is represented by coordinates (xi, fi), and the area of each circumscribed rectangle is represented by mi. * 2 represents a square. 309 in FIG.
Is a straight line calculated in this way. Then, the above processing is performed for some character string regions , and the inclination detection unit 104 obtains the inclination of the document image by taking the average of the inclinations (step 205).

<Embodiment 2> In a character string centered on the alphabet such as an English document, characters extending upward such as "d" and "l", characters extending downward such as "g" and "y", and There is a capital letter, and when the inclination of the image is obtained from the arrangement of the center points of the circumscribed rectangle as in the above-described embodiment, the error increases. Therefore, the second embodiment relates to a method for correctly detecting the inclination of a document image even in an English character string having a variation in the size and position of a circumscribed rectangle.

FIG. 4 is a diagram showing a specific example of an English character string in which the size and position of the circumscribed rectangle vary. The inclination detection unit 104 of the present embodiment creates a histogram of the height of such a circumscribed rectangle (FIG. 5), and sets the circumscribed rectangle having the highest frequency as the reference size. Then, the upper part of the circumscribed rectangle larger than the reference value by a threshold value or more (that is, higher) is cut off. That is, in FIG. 6, reference numeral 601 denotes a circumscribed rectangle of a black pixel connected component of the original image. When the reference value of the height of the circumscribed rectangle obtained from the histogram of FIG. Is converted to a reference value to become a circumscribed rectangle 602.

In the example of FIG. 4, circumscribed rectangles 402, 403,
Assuming that 405 is a circumscribed rectangle larger than the reference value, the height is converted as indicated by 602 in FIG. Then, for the converted circumscribed rectangle, a straight line with the smallest deviation from the center point of each circumscribed rectangle is calculated using the least squares approximation method in the same manner as in the above-described embodiment, thereby obtaining the inclination of the character string. Ask for.

As another embodiment, the circumscribed rectangles 404 and 411 that are much smaller than the reference value of the height of the circumscribed rectangle obtained from the histogram of FIG. 5 are often periods, commas, or the like. Since there is a slight deviation from the arrangement of the character rectangles, errors can be reduced by excluding these rectangles from the processing for tilt detection.

Further, among the circumscribed rectangles much smaller than the reference value, the embodiment can be modified so that only the small circumscribed rectangle on the rightmost wing of the block is excluded from the processing. That is, in FIG. 4, the circumscribed rectangle 404 is a circumscribed rectangle much smaller than the reference value, but is not a period or a comma, but a part of a character. Such a circumscribed rectangle appears in characters “i”, “j”, etc., but when the circumscribed rectangle 404 is excluded, an error occurs because the circumscribed rectangle 405 is below. Therefore, only the small circumscribed rectangle of the rightmost wing of a block having a high probability of being a period or a comma is excluded from the inclination detection processing.

Further, the height of uppercase and lowercase letters may not be as large as the threshold depending on the font and character size. Even in such a case, since the size deviation appears as an error, the circumscribed rectangle of the leftmost wing corresponding to the first character of the block is likely to be an uppercase letter,
The embodiment can be further modified so that the conversion processing is always performed to the reference height.

[0019]

As described above, according to the present invention, the inclination of the document image is obtained from the character string area .
Processing amount as compared with conventional methods Ru is greatly reduced. Also,
Performs processing to cope with variations in the size and position of the circumscribed rectangle.
Therefore, the inclination of the image can be correctly detected even in an English document.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 shows a processing flowchart of the present invention.

FIG. 3 is a diagram illustrating a character string region whose horizontal length is three times or more the vertical length.

FIG. 4 is a diagram showing an example of an English character string in which the size and position of a circumscribed rectangle vary.

FIG. 5 is a diagram illustrating a height histogram of a circumscribed rectangle;

FIG. 6 is a diagram for converting the size of a circumscribed rectangle of an original image into a reference value.

[Explanation of symbols]

 Reference Signs List 101 Image input unit 102 Image reduction unit 103 Block generation and classification unit 104 Tilt detection unit 105 Data storage unit 106 Control unit

Claims (1)

(57) [Claims] 1. A binarized document image (hereinafter referred to as an original image)
A document image inclination detecting method for detecting the inclination of the original image, wherein the original image is reduced, a character string area is extracted from the reduced image, and a character string area on the original image corresponding to the character string area is extracted. For the circumscribed rectangle of the black pixel connected component, a histogram of the height of each circumscribed rectangle is created, and the height of the circumscribed rectangle having the highest value is set as a reference value. Is converted to a reference value, and the inclination of the document image is detected based on the position of the circumscribed rectangle converted to the reference value and the position of the circumscribed rectangle not to be converted to the reference value. Detection method.