JP3046652B2 - How to correct the inclination of text documents - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for correcting the inclination of a character document, and more particularly to a method for correcting the inclination of a character document, such as a copying machine, a facsimile, an OCR,
In a device that reads a set document as an image,
The present invention relates to a correction method for eliminating an unexpected inclination of a document.

[0002]

2. Description of the Related Art Due to the inclination of a character document, the data compression rate decreases in facsimile, and the recognition rate decreases in OCR. Further, in a copying machine, in order to stabilize the copy quality, it is necessary to automate the inclination correction of the original.

[0003] Conventionally, as the inclination detecting method, expansion, contraction,
Known methods using thinning (Nori Nakamura, Makoto Ujiie,
Noriyoshi Okamura and Minatoshi Minami, "Character Region Extraction Algorithm for Mixed Mode Communication", IEICE Transactions VO
L. J67-D, No. 11, pp1277-1284,
November 1984).

According to this method, a character document is binarized,
The letters are connected by expansion. Further, the character string is displayed as a relatively thin band by shrinking and filling, and this is thinned. Further, irregular portions at both ends of the thinned image are deleted by reduction. In such a thinned image, after labeling, the inclination of a group having a length equal to or more than a predetermined value is obtained,
This inclination is regarded as the inclination of the document. Here, the number of times of the expansion was determined based on an empirical rule, and an optimum value could not be obtained unless the taking-in size, quality, etc. of the character document were visually judged.

[0005]

SUMMARY OF THE INVENTION The present invention has been made in order to solve such a conventional problem. The present invention is capable of executing an expansion process optimal for detecting tilt, and has high detection accuracy. It is intended to provide a correction method.

[0006]

According to a method for correcting the inclination of a character document according to the present invention, an image of a character document is sequentially expanded, and an optimum number of times of expansion is obtained by detecting a decrease in expansion of a figure area before and after expansion. An image is generated by expanding the original image by a predetermined number of times based on the optimal number of expansions, and in the generated image, the vertical and horizontal components of each part of the figure contour are calculated to determine which component in the vertical and horizontal directions is larger. Only in the direction of, the original image is expanded by the optimal number of times of expansion, the image obtained by this expansion is thinned, the vertical and horizontal components of each part of the image after the thinning are calculated, and based on the statistics of more components, The inclination is calculated and the correction is performed. According to the character document inclination correction method according to the present invention, each line of the character document can be processed as a substantially rectangular band, and by thinning the line, the inclination can be detected and corrected with high accuracy.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, an embodiment of a method for correcting the inclination of a character document according to the present invention will be described with reference to the drawings. FIG. 1 is a flowchart showing the embodiment. First, a document is read from input means such as a scanner (step 1-1), and the size of an image is adjusted so as to conform to the subsequent processing (step 1).
-2) For example, if the output image of the scanner is 4096 × 40
96 pixels, and the processing target pixel of the image processing unit is 1024 pixels.
If it is × 1024, a reduction of １ ／ is required. In this reduction process, if the process of setting the average value of the predetermined convolution to the density of the specific pixel is performed, the maximum density information of the original image can be preserved. That is, in the 1/4 reduction, the average value of the 4 × 4 convolution is set as the density of the specific pixel (for example, the upper left pixel in the convolution). Assuming that the pixel value at the coordinates (x, y) of the digital image is f (x, y) and the image after the convolution processing is g (x, y), the “predetermined convolution” is (x, y). ), Which includes an m × n pixel area, for example, g (x, y) is f (x, y), f (x +
1, y),. . . , F (x + m-1, y) f (x, y + 1), f (x + 1, y + 1),. . . , F
(X + m-1, y + 1). . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . f (x, y + n-1), f (x + 1,
y + n-1),. . . , F (x + m-1, y + n-1)
Is a function of Here, assuming that 0 <= x <= M−1 and 0 <= y <= N−1 for f (x, y), 0 <= x <= M− for g (x, y) m, 0 <= y <= N−n. Then, g (x, y) becomes f (x, y).
When there is a linear relationship with y),

(Equation 1) G (x, y) is calculated by the following calculation. Where k
(I, j) is a coefficient of f (x + i, y + j). Note that the “4 × 4 convolution” means a process of m = n = 4, and when m = n = 3, is referred to as “3 × 3 convolution”.

The above-mentioned “process for setting the average value of a predetermined convolution to the density of a specific pixel” is expressed by the following equation (1).
(I, j) = 1 / (m × n) means “reduction processing”. When the pixel value of the processed image is h (x, y), h (x, y) ) = G (m × x, n × y). Here, the range of (x, y) of h (x, y) is as follows: 0 <= x <= M / (m−1), 0 <= y <= N / (n−
1) Further, g (m × x, n × y) is a pixel obtained as a result of the “processing of setting the average value of the predetermined convolution to the density of the specific pixel”.

Next, a density histogram of the image is obtained, binarized by an optimal threshold (step 1-3), and noise is further removed (step 1-4). As a method of calculating the optimum threshold, a method using the median of the density as a threshold, a mode method,
There are a discriminant analysis method, a P-tile method, and the like. As a noise removing method, an isolated point removal, a median filter, or the like is effective. In detecting the inclination, the image of the character document is expanded in the main component direction, that is, the horizontal writing document is expanded in the horizontal direction, and the vertical writing document is expanded in the vertical direction by a predetermined number of times based on the optimum number of expansions, and each line is converted into a striped figure. By thinning this figure, inclination detection becomes easy. Here, a method of calculating the optimum number of times of expansion Ndir will be described (step 1-5).
１−1-13).

In step 1-5, the expansion is repeated thereafter.
Loop count for counting the number of inflation
The parameter i is initialized (i = 1). Then, step 1-6
The number of non-background pixels (the number of white pixels) Nw1 after noise removal
(I = 1) is counted, and the initial value of the number of non-background pixels is obtained.
You. Next, in step 1-7, expansion is performed once, and
In step 1-8, the loop counter i is incremented (i
= 2), in step 1-9, the number of non-background pixels Nw2 (i =
2) is counted. Next, in step 1-10, the non-
The third derivative (Δ ³ Nwi) of the number of background pixels is obtained . The third derivative is calculated by the following equation.

The third derivative Δ ³ Nwi = △ ² Nw (i + 1)
− △ ² Nwi However, △ ² Nwi = △ Nw (i + 1) − △ Nwi ΔNwi = Nw (i + 1) −Nwi Therefore, △ ³ Nwi = Nw (i + 3) −3 · Nw
(I + 2) + 3 · Nw (i + 1) −Nwi

When the value of the ^third derivative Δ ³ NWi becomes negative, it means that the inter-line fusion has already occurred, and that the increase in the figure area due to the expansion has been detected (step 1-10).
Therefore, at this stage, the process exits the expansion loop, and the number of expansions up to the previous time, Ndil = i-1, is adopted as the optimum number of expansions (step 1-13). Then, the previous dilated image is used for the following main component detection (step 1-14). In this way, if the area change of the figure is traced for each expansion of one layer, the global start of fusion can be reliably detected, and the optimum number of expansions can be obtained.

The start of fusion may be unclear due to the influence of noise or the like that cannot be completely removed by the noise removal. Therefore, the upper limit I of the number of times of expansion is set in advance, and even before the third derivative reaches a negative value,
The expansion is terminated (step 1-11). In this case, the optimal number of times of expansion is (I-1). If the third derivative is positive,
First, the content of the loop counter i is less than the upper limit I of the number of times of expansion.
, The process returns to step 1-6 again, and
Repeat Step 1-7, Step 1-8, Step 1-9
And again, at step 1-10, the sign of the third derivative is determined.
Refuse. In the loop of steps 1-6 to 1-11,
Table 1 below.

[Table 1] Note that in order to determine whether or not ＜ ³ Nwi <0 is satisfied,
Requires that the current i be 4 or greater, i = 2 and
In the case of 3, “△ ³ Nwi <0” in step 1-10
Cannot be determined, so the same branch as when not satisfied (to N
Is performed). Next, the main component of the document, that is, whether the document is a horizontally written document or a vertically written document is determined. This determination is made by synthesizing the pattern inclination components of each part of the image (step 1-14).

When detecting the inclination component, in the 3 × 3 convolution, the patterns of FIGS. 2A to 2L are set as the horizontal component, and the patterns of FIGS. 3A to 3L are set as the vertical component. The main component of the entire image is obtained depending on whether the number is large. These patterns are the outlines of the figure. The inclination of the outline is defined for each two-pixel row, and those with the total inclination of the convolution near horizontal are considered as horizontal patterns, and those near vertical are considered as vertical patterns. It is a pattern of directions. Note that a pattern including one background pixel on the periphery of the convolution ((K) in FIGS. 2 and 3,
(L)) is a vertical pattern when the background is on the left and right peripheral edges, and a horizontal pattern when the background is on the upper and lower peripheral edges, and does not include any oblique 45-degree pattern.

Assuming that the number of vertical patterns is Nv and the number of horizontal patterns is Nh, if Nv <= Nh, the main component is determined to be horizontal, otherwise it is determined to be vertical (step 1-). 15-1-17). Here, the index of the main component is set to horv.
"v" is substituted for orv, and "h" is substituted for horv in the horizontal direction (steps 1-16 and 1-17).

Here, the original image before expansion is expanded a predetermined number of times in the direction of the main component, and the characters on each line are connected. The number of times of expansion should be the minimum number of times that characters are surely connected, and empirically, a value such as 9 · Ndir / 4 [times] is used based on the Ndir. In an image generated by a predetermined number of dilations based on the optimum number of dilations, each row has a substantially rectangular shape. Next, the substantially rectangular figure is thinned. Thinning refers to making a graphic a line graphic having a pixel width of 1 by sequentially removing pixels on the periphery of the graphic without changing the topology of the graphic. As an image processing method of thinning, for example, HILDIT
The method of CH and the method of JP-A-64-13680 are effective. By this thinning, a substantially center line of the substantially rectangular figure is obtained, and an effective element for calculating the inclination of the figure can be obtained. However, when thinning is performed, the shape of the character at the end of the figure or the noise causes
The end of the thin figure as shown in FIG. 1C may be branched or curved. Therefore, in order to remove these irregularities that hinder the calculation of the inclination with high accuracy, the thinning is performed and then the end points are reduced (FIG. 4D). This allows
The irregularities at both ends are removed, and the accuracy of the subsequent inclination calculation increases. Here, the "degeneration process" is a process of removing both ends of a line figure, and the determination as to whether or not to remove the line figure is determined by neighboring pixels. For example, for a specific pixel, 3 × 3 in FIG.
Referring to the pixel pattern, when the neighboring pixels match these patterns, it is determined that the central pixel should be removed, and the pixel is converted to a background pixel. Such processing is performed on all the pixels of the processing target image, and the processing ends.

It should be noted that the number of times of degeneration should be the minimum number of times that the irregular part can be completely erased. According to experience, a value such as 15 · Ndir / 4 + 1 [times] is used based on the Ndir. Adopted. This empirical formula is approximately one letter to 1.5
This means deleting a line of a length corresponding to a character.

FIGS. 6A, 6B, 6C, 6D, and 6D show the calculation of the inclination of the image mainly composed of the horizontal component.
The number of each of the patterns (E), (F), (G), (H), and (I) is counted, and the counted values are multiplied by a coefficient and totaled. Here, the shape of each pattern and the numerical value to be multiplied by the number of each pattern are as follows.

Horizontal pattern shape and coefficient (A) Horizontal pattern coefficient of 3 pixels = 0 (B) (downward right pattern of 2 pixels) + (horizontal pattern of 2 pixels) Coefficient = 0.5 (C) (2 pixels) (Horizontal pattern) + (two pixel right down pattern) coefficient = 0.5 (D) (two pixel horizontal pattern) + (two pixel right upper pattern) coefficient = −0.5 (E) (two pixel right (Up pattern) + (horizontal pattern of 2 pixels) coefficient = −0.5 (F) right down pattern of 3 pixels coefficient = 1 (G) upper right pattern of 3 pixels coefficient = −1 (H) (right of 2 pixels) Downward pattern) + (upper right pattern of 2 pixels) coefficient = 0 (I) (upper right pattern of 2 pixels) + (downward right pattern of 2 pixels) coefficient = 0

These coefficients are error components with respect to the horizontal component of each pattern, and give an absolute value "1" to a diagonal pattern of three pixels. These patterns (A) ~
The total number of the entire image of (I) is Na to Ni.
Then, the inclination angle θ is given by the following equation.

Θ = tan ⁻¹ [｛0.5 (Nb + Nc)
−0.5 (Nd + Ne) + Nf−Ng｝ / (Na + Nb)
+ Nc + Nd + Ne + Nf + Ng + Nh + Ni)]

The same applies to the processing for a vertically written character document. FIGS. 5A to 5D show the processing contents of a horizontally written document for FIGS. 4A to 4D. Then, FIG.
FIGS. 7A to 7I show vertical patterns corresponding to (I). The shapes and coefficients of these patterns are as follows.

(A) 3 pixel vertical pattern coefficient = 0 (B) (2 pixel left-down pattern) + (2 pixel vertical pattern) coefficient = 0.5 (C) (2 pixel vertical pattern) + ( Coefficient = 0.5 (D) (vertical pattern of two pixels) + (right-down pattern of two pixels) Coefficient = -0.5 (E) (right-down pattern of two pixels) + ( (Vertical pattern of two pixels) Coefficient = −0.5 (F) Downward pattern of 3 pixels Coefficient = −1 (G) Upper right pattern of 3 pixels Coefficient = 1 (H) (2 pixels left down pattern) + ( (2 pixels right down pattern) coefficient = 0 (I) (2 pixels right down pattern) + (2 pixel left down pattern) coefficient = 0

Assuming that the total number of these patterns (A) to (I) for the entire image is from Na to Ni, the inclination angle θ is given by the following equation. θ = tan ⁻¹ [｛0.5 (Nb + Nc) −0.5 (N
d + Ne) -Nf + Ng｝ / (Na + Nb + Nc + Nd)
+ Ne + Nf + Ng + Nh + Ni)]

As described above, an expanded figure having the optimum number of expansions in the main component direction is generated, the expanded image is thinned, and the vertical and horizontal components of the thinned image are calculated. By calculating the inclination based on the statistics for the component, the inclination can be easily and accurately calculated. Next, in order to correct the inclination, after calculating the inclination angle, the image is subjected to affine transformation so as to eliminate the inclination. The affine transformation can be executed by a matrix operation by software or an address conversion process by a dedicated IC.

[0026]

As described above, the method of correcting the inclination of a character document according to the present invention sequentially expands the image of the character document, and detects the slowdown of the expansion amount of the graphic area before and after expansion, thereby obtaining the optimum number of expansions. An image is generated by expanding the original image by a predetermined number of times based on the optimal number of expansions, and in the generated image, the vertical and horizontal components of each part of the figure contour are calculated to determine which component in the vertical and horizontal directions is larger. Only in the direction of, the original image is expanded by the optimal number of times of expansion, the image obtained by this expansion is thinned, the vertical and horizontal components of each part of the image after the thinning are calculated, and based on the statistics of more components, Since the inclination is calculated and the correction is performed, there is an excellent effect that the expansion processing optimal for the inclination detection can be executed, and the inclination of the character document can be corrected with high detection accuracy.

[Brief description of the drawings]

FIG. 1 shows a method of correcting the inclination of a character document according to the present invention.
It is a flowchart of Example 1-1 to Example 1-13 which shows an Example.

FIG. 2 is a conceptual diagram for explaining a horizontal component pattern. (A) is a conceptual diagram showing a horizontal component pattern including an error component of 0.5. (B) is a conceptual diagram showing another horizontal component pattern including a 0.5 error component. (C) is a conceptual diagram showing another horizontal component pattern including an error component of 0.5. (D) is a conceptual diagram showing another horizontal component pattern including an error component of 0.5. (E) is a conceptual diagram showing another horizontal component pattern including an error component of 0.5. (F) is a conceptual diagram showing another horizontal component pattern including an error component of 0.5. (G) is a conceptual diagram showing another horizontal component pattern including an error component of 0.5. (H) is a conceptual diagram showing another horizontal component pattern including an error component of 0.5. (I) is a conceptual diagram showing a horizontal horizontal component pattern. (J) is a conceptual diagram showing another horizontal horizontal component pattern. (K) A conceptual diagram showing a horizontal component pattern including one background pixel. (L) It is a key map showing other horizontal component patterns containing one background pixel.

FIG. 3 is a conceptual diagram for explaining a vertical component pattern. (A) is a conceptual diagram showing a vertical component pattern including an error component of 0.5. (B) is a conceptual diagram showing another vertical component pattern including an error component of 0.5. (C) It is a key map showing other longitudinal component patterns containing an error component of 0.5. (D) is a conceptual diagram showing another vertical component pattern including an error component of 0.5. (E) is a conceptual diagram showing another vertical component pattern including an error component of 0.5. (F) is a conceptual diagram showing another vertical component pattern including an error component of 0.5. (G) is a conceptual diagram showing another vertical component pattern including an error component of 0.5. (H) is a conceptual diagram showing another vertical component pattern including an error component of 0.5. (I) A conceptual diagram showing a vertical component pattern. (J) is a conceptual diagram showing another vertical component pattern. (K) A conceptual diagram showing a vertical component pattern including one background pixel. (L) It is a key map showing other longitudinal component patterns containing one background pixel.

FIG. 4 is a conceptual diagram for explaining a horizontally written original. (A) is a conceptual diagram showing an example of a horizontally written original. FIG. 3B is a conceptual diagram illustrating a horizontally expanded figure of a horizontally written document. (C) is a conceptual diagram showing a thinned image of a horizontally expanded figure of a horizontally written document. FIG. 4D is a conceptual diagram showing a state in which a thinned image of a horizontally expanded figure of a horizontally written original is degenerated.

FIG. 5 is a conceptual diagram for explaining a vertically written original. (A) is a conceptual diagram showing an example of a vertically written original. (B) is a conceptual diagram showing a vertically expanded figure of a vertically written document. (C) is a conceptual diagram showing a thinned image of a vertically expanded figure of a vertically written original. (D) is a conceptual diagram showing a state in which a thinned image of a vertically expanded figure of a vertically written original is degenerated.

FIG. 6 is a conceptual diagram for explaining a horizontal pattern. (A) is a conceptual diagram showing an example of a horizontal pattern. (B) is a conceptual diagram showing a pattern including an error component of 0.5 with respect to a horizontal pattern. (C) is a conceptual diagram showing another pattern including an error component of 0.5 with respect to the horizontal pattern. (D) is a conceptual diagram showing still another pattern that includes an error component of 0.5 with respect to the horizontal pattern. (E) is a conceptual diagram showing still another pattern that includes an error component of 0.5 with respect to the horizontal pattern. (F) It is a key map showing the pattern of a right descent. (G) is a conceptual diagram showing an upward-sloping pattern. (H) It is a key map showing the pattern which changes from right descent to right descent. (I) is a conceptual diagram showing a pattern that changes from rising right to falling right.

FIG. 7 is a conceptual diagram for explaining a vertical pattern. (A) is a conceptual diagram showing an example of a vertical pattern. (B) is a conceptual diagram showing a pattern including an error component of 0.5 with respect to a vertical pattern. (C) is a conceptual diagram showing another pattern including an error component of 0.5 with respect to the vertical pattern. (D) is a conceptual diagram showing still another pattern that includes an error component of 0.5 with respect to the vertical pattern. (E) is a conceptual diagram showing still another pattern that includes an error component of 0.5 with respect to the vertical pattern. (F) It is a key map showing the pattern of a right descent. (G) is a conceptual diagram showing an upward-sloping pattern. (H) It is a key map showing the pattern which changes from left descent to right descent. (I) It is a conceptual diagram showing the pattern which changes from right descent to left descent.

FIG. 8 is a flowchart of steps 1-14 to 1-23.

FIG. 9 is a conceptual diagram showing a pattern for degeneration processing.

Claims (2)

(57) [Claims] 1. An original image of a character document is sequentially expanded, an optimum expansion number is obtained by detecting a decrease in expansion amount of a graphic area before and after expansion, and an image is generated by expanding the original image by the optimum expansion number. Then, in the generated image, the vertical and horizontal components of each part of the figure contour are calculated to determine which of the vertical and horizontal components is larger, and only in the direction of the larger component, the original image is expanded a predetermined number of times based on the optimal expansion number. The image obtained by this expansion is thinned, the vertical and horizontal components of each part of the image after the thinning are calculated, and the inclination is calculated based on the statistic of a larger number of components. A method for correcting the inclination of a character document, characterized by rotating a document. 2. The dulling of the amount of expansion is performed by using the original image and each stage.
The number of non-background pixels in the expanded image is determined by the
Detected when this third derivative became negative,
The number of inflation immediately before that is the optimal number of inflation.
2. The method of correcting inclination of a character document according to claim 1, wherein: