JPH08263647A - Area division device - Google Patents

Abstract

PURPOSE: To provide an area division device which can appropriately divide an area in accordance with the strength of the texture of a picture. CONSTITUTION: The area division device 1 changes the condition of noise removal in the case of dividing the area in accordance with the value of the strength of a texture element in quantized picture in full colors by a noise removal strength decision means 51. A deformation strength decision means 52 changes the magnitude of the deformation of the area in the case of dividing the area in accordance with the value of the strength of the texture element. A contribution rate change means 53 changes the contribution rate of prescribed color space and xy-coordinates, which appear in the condition expression of division in the case of dividing the area, in accordance with the value of the strength of the texture element.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an area dividing device for performing area division according to a pattern of quantized full-color image data, mainly as image preprocessing.

[0002]

2. Description of the Related Art The process of dividing a read image into appropriate areas is an important technique mainly used as preprocessing in various fields of image processing. For example, in image editing, it is necessary to cut out an area to be edited, and at that time, area division of the image is used. Also, in the recognition processing, area division is used as a pre-processing for finding an object to be recognized. Since the performance of this area division directly affects the performance of the entire processing, it is necessary to perform area division with high performance.

Further, in recent years, when extracting and organizing highly common image features used in various image processes, a process is attempted in which the regions of the image are grouped together so that they can be easily used in the next image process. However, even in that case, it is necessary to appropriately divide the image area.

As a method of area division, there is a K-means algorithm which has been known for a long time.
A method of performing a K-means algorithm in five dimensions by adding the xy coordinates of the image to such color components (A study of a segmentation method using both color information and position information, see Proceedings of the 1991 IEICE Spring National Convention) Have been proposed. Further, as another approach, one using MAP (maximum posterior probability) estimation has been proposed.

[0005]

However, in dividing a color image by these area dividing techniques, it is a difficult problem to determine an appropriate number of areas. Also, RG
When area division by the K-means algorithm is performed in five dimensions by adding xy coordinates to a color space such as B or L ^* a ^* b ^* , each area is gathered at a short distance on the image, but it is inevitably excessive. It is easy to be divided, and even parts that are not necessarily divided are divided.

Further, considering that artistic value is added to an image by utilizing such a characteristic of being over-divided, it is difficult to determine a proper area division standard.
When the same condition is used over the entire image, inadequate division is often made partially. Therefore, it is an object of the present invention to provide an area dividing device capable of performing appropriate area division according to the strength of image texture.

[0007]

SUMMARY OF THE INVENTION The present invention is an area dividing device made to achieve the above object. That is, the present invention is an area dividing device that uses a predetermined color space, xy coordinates of pixels, and a texture element of pixels when dividing the quantized full-color image data into areas according to the pattern. Depending on the value of the strength of the texture element of, the noise removal condition at the time of area division is changed, or the size of the deformation of the area at the time of the area division is changed according to the value of the strength of the texture element of the pixel. It is for changing or changing the contribution rate of a predetermined color space and xy coordinates appearing in the conditional expression of the division at the time of the area division according to the value of the strength of the texture element of the pixel.

[0008]

In the present invention, the noise removal condition at the time of region division is changed according to the strength value of the texture element of the pixel. Therefore, the more the texture is complicated, the more the noise removal amount can be increased. it can. Further, since the size of the deformation of the area at the time of the area division is changed according to the value of the strength of the texture element of the pixel of the texture element of the pixel, the deformation amount of the area increases as the texture becomes more complicated. it can. Furthermore, since the contribution ratios of the predetermined color space and the xy coordinates appearing in the conditional expression of the division at the time of the area division are changed according to the value of the strength of the texture element of the pixel, xy becomes more complicated as the texture becomes more complicated. The contribution of the position by the coordinates can be increased, and the contribution of the color space can be increased as the texture is simpler.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT An embodiment of the area dividing device of the present invention will be described below with reference to the drawings. 1A and 1B are views for explaining an area dividing device according to the present invention. FIG. 1A is an overall configuration diagram, and FIG.
FIG. 4 is an internal configuration diagram of the area dividing means. As shown in FIG. 1A, the area dividing device 1 includes an image input means 2, a direction component adding means 3, an image buffer 4, an area dividing means 5, a divided area buffer 6, and a division result output. The image dividing processing in the area dividing means 5 is particularly characterized.

The image input means 2 is for reading an image on a document, for generating an image by an image scanner or computer processing, or for taking in an image generated outside the area dividing device 1 inside. In this embodiment, the case where a full-color image represented in the L ^* a ^* b ^* color space recommended by CIE (International Commission on Illumination) is obtained from the image input means 2 will be described as an example.

The directional component adding means 3 is a part for calculating a directional component for indicating a texture element of an image and adding it to the captured image data. Also, the image buffer 4
Stores the direction component d calculated by the direction component adding unit 3 together with the image data (for example, L ^* a ^* b ^* ) captured by the image input unit 2. Area dividing means 5
Is the L ^* a ^* b ^* color space of the captured image, xy of pixels
This is a portion for performing area division processing using coordinates and a direction component d which is a texture element of pixels. The divided area buffer 6 is used as a working buffer when the area dividing means 5 divides an image into areas. Further, the division result output means 7 is composed of a display, a printer, etc. for outputting the division result.

Next, the outline of the processing in the area dividing means 5, which is a feature of the present invention, will be described with reference to FIG. That is, the area dividing unit 5 obtains the texture component added image in which the texture component is added to L ^* a ^* b ^* of the full color image from the image buffer 4 (see FIG. 1A). Here, the direction component calculated by the direction component adding means 3 shown in FIG. 1A is used as the texture component.

The area dividing means 5 includes a noise removal strength determining means 51, a deformation strength determining means 52, a contribution rate changing means 53, and an area determining means 54 and a division end determining means 55. The texture component added image input to the area dividing unit 5 is input to all or one or two of the noise removal strength determining unit 51, the deformation strength determining unit 52, and the contribution rate changing unit 53, and the predetermined condition is satisfied. The area determining means 54 determines the area to which the pixel of interest belongs. Then, when the division end determination unit 55 determines that all the pixels belong to an appropriate division area, the division result is output, and when it is not appropriate, the area division processing is repeated under a new condition. become.

Next, the image area dividing process, which is a feature of the present invention, will be described in detail with reference to the drawings. First, FIG. 1 (a)
The image read by the image input means 2 shown in FIG. The image data is also input to the direction component adding means 3 to extract the direction element. The direction component d extracted here is written in an empty space of the image buffer 4. Direction element extraction is
For example, the method described on page 80 of Report No. 835 of Electrotechnical Laboratory is applied.

That is, for example, only L ^* of the input image G shown in FIG. 2 is used as an input image, and the directional element is extracted by the directional element extraction mask shown in FIG. In addition,
Although only L ^* is used here, direction elements may be extracted using all L ^* a ^* b ^* . To extract the directional element, the L ^* value of a 3 × 3 pixel centered on the pixel of interest is multiplied by the corresponding mask value and the sum is taken. FIG.
3A shows a mask for the horizontal direction, and FIG. 3B shows a mask for the vertical direction. For example, an image G made up of 3 × 3 pixels and a mask M are combined, and the pixels where the mask M overlaps are G (1,1), G (1,2), ..., G (3, respectively.
3), the corresponding elements of the image G and the mask M are multiplied and added. This operation is (G * M).

CC = (G * M1), LL = (G
* M2) is calculated as shown in Equation 1.

## EQU1 ## d1 = 16 (tan ^-1 (LL / CC) + π / 2) / π d2 = Log (LL × LL + CC × CC + 1)

This d1 is the direction of the direction component of the image 0 to 1.
5 is represented by 16 stages, and d2 represents the magnitude of the direction component of the image by 16 stages of 0-15. That is, d
Since 1 represents the direction of the contour line when the density of the image is displayed in contour lines, the direction elements of 180 degrees opposite directions are the same, that is, as shown in FIG. It is sufficient to be able to describe in 16 steps. Therefore, 8 bits are required for both d1 and d2. This 8-bit data is added to the input image as a direction component of L ^* a ^* b ^* , and a region division based on this is performed as a texture component added image.

Next, the area dividing means 5 shown in FIG. 1 divides the image into areas. As the area division method, a 6-dimensional K-means algorithm using L ^* a ^* b ^* stored in the image buffer 4, the direction component d (d1 and d2) calculated previously, and the xy coordinates that are the pixel position information is used. Apply.

The K-means algorithm is one of the clustering methods that can be applied to segment a color image. That is, K initial region centers are set for the input image, and the region division is repeated over several orders until the divided regions including the initial region division converge to a certain state,
By calculating the area center of each divided area after each subsequent area division after the initial area division, it is determined whether or not the divided areas converge to a certain state.

For example, as shown in FIG. 5A, K = 1
As 00, make vertical and horizontal 10 × 10 cells, and each cell
The pixel located at the center of gravity is set to the region center Ci, j (i is the
The number of calculations and the initial value are represented by i = 0. Ie the initial area
The center is C0, j. Further, j is a divided area number and j =
1, 2, ..., 100) will be described. Further
, The position of the center of the initial region is represented as (x0, j, y0, j),
In addition, L of the pixel^*a ^*b^*Representative color of the divided area
(L^*0, j, a^*0, j, b^*0, j). Furthermore,
Regarding the directional component, if d2 <1, it has no direction.
Substituting the non-directional symbol “16” into d1
Keep it.

Then, as shown in FIG. 6A, the value of d2 is not recorded in the feature buffer in the divided area table, but only the value of d1 is recorded. Hereinafter, these are collectively summarized as the region center (xi, j, yi, j, L ^* i, j, a ^* i, j, b
^* i, j, d1i, j).

Next, the area dividing means 5 (see FIG. 1) displays the image.
The constituent pixels are displayed in the image buffer 4 (see FIG. 1A).
From each pixel in turn, and the pixel value (xn, yn
 , L ^*n, a^*n, b^*n, d1n, d2n) (n is
Pixel number) and the center of all areas (xi, j, yi, j,
L^*i, j, a^*i, j, b^*i, j, d1i, j) distance L
i, j, n is calculated by Equation 2 (the center of the initial region is i =
0). Here, first, the difference of each component is obtained.

## EQU2 ## dxi, j, n = (xi, j-xn) dyi, j, n = (yi, j-yn) dli, j, n = (L ^* i, j-L ^* n) dai, j , n = (a ^* i, j-a ^* n) dbi, j, n = (b ^* i, j-b ^* n)

Further, since the distance of the directional component is an angular component of polar coordinates, when d1i, j = 16 and d1n <1, there is no directional component, so the difference is small.
It becomes like number 3.

## EQU3 ## ddi, j, n = 0

On the other hand, when d1i, j <16 and d1n ≥1, the following equation 4 is obtained.

Ddi, j, n = | d1n-d1i, j | However, in the case of ddi, j, n> 8, ddi, j, n = 16- | d1n-d1i, j |

(D1i, j = 16 and d1n ≧ 1)
Alternatively, in the case of (d1i, j ≠ 16 and d1n <1), it is considered that the distance between the direction components whose angles are different from each other by π / 2 is shorter and the distance between the direction components facing the same direction is greater. Think about it, and take a value between 0 and 7 as in Equation 5.

## EQU5 ## ddi, j, n = 3.5

The area dividing means 5 (see FIG.
Is the distance between each pixel value and the center of all regions,
The area to which the pixel belongs is updated. At this time, an isolated point generated at a place where the two areas are complicatedly intermingled becomes an annoyance like noise, and therefore noise is deleted by the noise removal strength determination means 51 shown in FIG. 1B.

To remove noise, first, FIG. 7 is shown.
Of the pixels adjacent to the pixel of interest n, it differs from the pixel of interest n.
The number of pixels Nneighbor that belong to the area
hbor ²Ask for. For example, the image shown by hatching in FIG.
The element belongs to a region different from the region to which the pixel of interest n belongs.
The number of pixels is four in this example. Therefore Nne
ighbor = 4 and Nneighbor²= 16.

Then, this Nneighbor ² term is added to the previously calculated distance Li, j, n. As a result, the pixel of interest n
When the pixel belongs to the divided area j and is close to the isolated point, many pixels belonging to the divided area different from the pixel of interest n out of the pixels in contact with the pixel of interest n increase the value of Nneighbor ² and increase the distance. Li, j, n also becomes large. Therefore, it becomes difficult for the divided area j to become an isolated point, and noise elimination can be realized.

Further, in the deformation strength determining means 52 of the area dividing means 5 shown in FIG. 1B, the functions Fx () and Fy () for controlling the deformation of the area when the area division is performed are as follows. Define it. First, | d1n -16π / (tan
^{In the case of -1} (-dyi, j, n / dxi, j, n) + π / 2) | <8, it is defined as in Equation 6.

Fx (d1n, dyi, j, n, dxi, j, n) = | d1n −16π / (tan ⁻¹ (−dyi, j, n / dxi, j, n) + π / 2) | / 3.5-1 Fy (d1n, dyi, j, n, dxi, j, n) = | d1n-16 [pi] / (tan- ¹ (-dyi, j, n / dxi, j, n) + [pi] / 2) | /3.5-1

Also, | d1n-16π / (tan ^-1 (-
In the case of dyi, j, n / dxi, j, n) + π / 2) | ≧ 8, it is defined as in Expression 7.

## EQU00007 ## Fx (d1n, dyi, j, n, dxi, j, n) = 16- | d1n-16.pi ./ (tan ^-1 (-Dyi, j, n /dxi,j,n)+π/2)|/3.5-1 Fy (d1n, dyi, j, n, dxi, j, n) = 16- | d1n-16π / ( tan ^-1 (-Dyi, j, n / dxi, j, n) + π / 2) | /3.5-1

As shown in equations (6) and (7), the functions Fx () and Fy () for controlling the deformation of the area when the area division is performed are defined by the deformation strength determining means 52 (see FIG. 1 (b)). This allows the ratio of area deformation to be changed according to the direction component, that is, the value of the strength of the texture element.

Further, H (d2i, j) is defined by the contribution rate changing means 53 shown in FIG.

## EQU00008 ## H (d2i, j) = 1 / (1.1 + exp (11. (0.8-N))) where N = N2i, j / Ni, j where Ni, j is the divided region j The number of pixels included in N
2i, j is the number of pixels included in the divided area j and satisfying d2i, j> 1.

The area dividing means 5 (see FIG. 1) performs the calculation shown in the equation 9 by combining the above terms.

Li, j, n = kx (1 + kh1.H (d2i, j)). (1 + kh2.H (d2i, j) .Fx (d1n, dyi, j, n, dxi, j, n)). dxi, j, n ² + ky (1 + kh1.H (d2i, j)). (1 + kh2.H (d2i, j) .Fy (d1n, dyi, j, n, dxi, j, n)). dyi, j, n ² + kl · dli, j, n ² + ka · dai, j, n ² + kb · dbi, j, n ² + kh3 · H (d2i, j) Nneighbor ² + kd · ddi, j, n ²

In Equation 9, i is the number of times of repeated calculation, n is a pixel number, j is a divided area number, kx, k.
y, kh1, kh2, kh3, kl, ka, kb, kd
Is an arbitrary coefficient, and here kh2 = 0.9 and all other coefficients are "1".

That is, N is closer to "0" when the strength value of the texture element in the area is smaller, and is closer to "1" as the strength value of the texture element is larger. Therefore, if the value of the strength of the texture element is small, H (d2i, j) ≈0. Therefore, the contribution of the xy coordinates to the distance Li, j, n, the ratio of area deformation, and the strength of noise removal are minimized. . On the contrary, when the value of the strength of the texture element increases, the contribution of the xy coordinates to the distance Li, j, n, the area deformation ratio, and the noise removal strength increase.

When such processing is executed for all pixels, the divided areas are updated and the shape changes (see FIG. 5B). Also, the center of gravity (xi +
1, j, yi + 1, j), representative color (L ^* i + 1, j, a ^* i +)
1, j, b ^* i + 1, j), representative image direction component d1i + 1, j
Is calculated again, and the value is divided into the divided area buffer 6 (see FIG.
(See (a)) to update the value (see FIG. 6 (b)).

In general, the representative colors (L ^* i + 1, j, a ^* i +
1, j, b ^* i + 1, j) is a simple arithmetic average of pixels in the newly divided divided area. Further, the center of gravity is a value obtained by summing the xy coordinate values of all pixels included in the divided areas and dividing both by the number of pixels included in the area. By such calculation, a virtual pixel having the above six values is obtained and set as a new area center.

When this process is repeatedly executed, the region gradually converges to a constant shape and is not deformed so much. FIG.
The division end determination means 55 shown in (b) calculates the deformation amount of this area and determines to terminate the division processing based on this value.

That is, whether or not the center of the region has converged is determined by the new center of the region (xi, j, yi, j, L ^* i, j, a).
^* i, j, b ^* i, j, d1i, j) and the previous region center (xi-1,
j, yi-1, j, L ^* i-1, j, a ^* i-1, j, b ^* i-1, j,
d1i-1, j) and the deviation amount Di, j (i is the number of times of repeated calculation, j is the divided area number) are calculated, and if the deviation distance becomes negligibly small, it is determined that the area has converged. The division process is ended. This shift amount Di, j
In the case of calculating, the difference between the components is first obtained by the equation 10.

## EQU10 ## Dxi, j = (xi, j-xi-1, j) Dyi, j = (yi, j-yi-1, j) Dli, j = (L ^* i, j-L ^* i-1) , j) Dai, j = (a ^* i, j-a ^* i-1, j) Dbi, j = (b ^* i, j-b ^* i-1, j)

The direction component is as follows. First, when d1i, j = 16 and d1i-1, j = 16, that is, when there is no directional component, the equation 11 is used.

[Equation 11] Ddi, j = 0

Further, D1i, j <16 and D1i-1, j <1
In the case of 6, the formula 12 is used.

## EQU12 ## Ddi, j = | d1i, j -d1i-1, j | When Ddi, j> 8, Ddi, j = 16- | d1i, j -d1i-1, j |

Further, (d1i, j = 16 and d1i-1, j
<16) or (d1i, j <16 and d1i-1, j = 1
In the case of 6), it is considered that they are closer than the distance between the direction components whose angles are different from each other by π / 2 and are farther than the distance between the direction components facing in the same direction. As in Equation 13.

[Equation 13] Ddi, j = 3.5

Then, based on these, the shift amount Di, j between the center of the new area and the center of the previous area is calculated by the equation (14).

Equation 14] Di, j = kx (1 + kh1 · H (d2i, j)) · Dxi, j 2 + ky (1 + kh1 · H (d2i, j)) · Dyi, j 2 + kl · Dli, j 2 + ka · Dai, j ² + kb ・ Dbi, j ² + kd ・ Ddi, j ²

In the equation 14, i is the number of times of repeated calculation, j is the divided area number, kx, ky, kh1, k
l, ka, kb, and kd are arbitrary coefficients, and are all set to "1" here.

The division end judging means 55 shown in FIG. 1 (b).
Repeats the calculation of the deviation amount Di, j, ends the division processing when the deviation amount becomes small enough to be ignored, and outputs the result (division result). In this way, the image when the center of the region converges is in a state in which the region division is made according to the human sense.

For example, in the image of the human hair portion, a flowing texture appears strongly, and the value of the strength of the texture element is large. When this area is divided into areas by the area dividing device 1 of the present invention, the amount of noise removal is increased, the size of the area deformation is large, the contribution ratio of the xy coordinates in the division conditional expression is large, and the pattern of the hair portion is followed. The area is divided into long and narrow shapes (which match the appearance).

On the other hand, since the image of the human face portion has a small amount of texture, the strength value of the texture element becomes small. When this area is divided into areas by the area dividing device 1 of the present invention, the amount of noise removal is reduced, the magnitude of area deformation is small, and the contribution ratio of the xy coordinates in the division conditional expression is small (the contribution ratio of the color space is large. ), A substantially circular area division is performed so that facial features often appear.

In this embodiment, the noise removal condition is changed according to the value of the strength of the texture element of the pixel, the size of the region deformation at the time of region division is changed, and the conditional expression of the division is further obtained. The explanation has been centered on an example in which all the contributions of the color space and the xy coordinates in the above are changed.
The present invention is not limited to this, and the conditions may be changed individually. Further, the operator may select which condition to apply. Further, although L ^* a ^* b ^* is used as the color space in the present embodiment, the same is true when other color spaces such as the L ^* u ^* v ^* color space other than this are used.

[0049]

As described above, the area dividing device of the present invention has the following effects. That is, when the area is divided, the noise removal condition, the area deformation ratio, the color space and the xy coordinate contribution rate appearing in the conditional expression of the division are adaptively converted according to the value of the strength of the texture of the image itself. Therefore, it is possible to perform the region segmentation that suits the human sense. This makes it possible to perform preprocessing suitable for image processing such as image editing.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an area dividing device of the present invention,
(A) is an overall block diagram, (b) is an internal block diagram of the area dividing means.

FIG. 2 is a diagram showing extraction of image data.

3A and 3B are diagrams showing directional element extraction masks, in which FIG. 3A is for a horizontal direction and FIG. 3B is for a vertical direction.

FIG. 4 is a diagram showing a display example of a directional component.

FIG. 5 is a diagram showing a state of area division, in which (a) is an initial state and (b) is a state after updating.

FIG. 6 is a diagram showing an example of a divided area table, in which (a) is an initial value and (b) is after i times of updating.

FIG. 7 is a diagram illustrating noise removal.

[Explanation of symbols]

 1 area dividing device 2 image input means 3 direction component adding means 4 image buffer 5 area dividing means 6 divided area buffer 7 division result output means 51 noise removal strength determining means 52 deformation strength determining means 53 contribution rate changing means 54 area determining Means 55 Division end judging means

Claims (3)

[Claims] 1. A region dividing device that uses a predetermined color space, xy coordinates of pixels, and a texture element of the pixel when dividing the quantized full-color image data into regions according to the picture pattern. A region dividing device, characterized in that the noise removal condition at the time of the region division is changed according to the value of the strength of the texture element. 2. A region dividing device that uses a predetermined color space, xy coordinates of pixels, and a texture element of pixels when dividing the quantized full-color image data into regions according to the picture pattern. The area dividing device, wherein the size of deformation of the area at the time of area division is changed according to the strength value of the texture element. 3. A region dividing device that uses a predetermined color space, xy coordinates of pixels, and a texture element of pixels when dividing the quantized full-color image data into regions according to the pattern thereof. The area dividing device, wherein the contribution ratios of the color space and the xy coordinates appearing in the conditional expression of the division at the time of the area division are changed according to the strength value of the texture element.