JPH08265585A - Area division method for color picture - Google Patents

Abstract

PURPOSE: To narrow down the range of a small number of cluster representative colors even in the case of a small difference between object colors by clustering the cluster representative colors, which are obtained by clustering, again to divide them into groups. CONSTITUTION: A color space conversion part 10 converts an input picture signal to an L*U*v* signal. A feature space conversion part 20 converts the L*U*v* signal to points L, μ, and (v) on the feature space adapted to area division. An area division part 30 performs clustering on the space and clusters the cluster representative colors obtained as the result again to divide the cluster representative colors into groups. By this method, the range of narrowed down to a sufficiently small number of cluster representative colors corresponding to the number of areas in the picture even if the variance of each object color is large and the difference between object colors is small. It is preferable that cluster representative colors are clustered again after conversion of cluster representative colors to hue hi and saturation si.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color image area dividing method, and more particularly to extracting a specific area (so-called clipping) or area corresponding to each object in an image when processing image data of a printing plate. The present invention relates to a color image area dividing method suitable for selective color tone correction or for structurally describing an image and for use in, for example, structure extraction coding used for data compression or the like.

[0002]

2. Description of the Related Art A clustering method on a color space is one of typical methods for dividing a region of a digital color image.

As shown in FIG. 1, this is done by mapping pixels in an image plane into a color space (for example, RGB color space) and using a cluster boundary set according to the distance from the cluster centroid in that space. This is a method in which each pixel is classified and labeled as 1, 2, 3, etc. by clustering, and an area-divided image is obtained by inverse mapping on an image plane to perform area division.

This conventional method can properly divide an area when pixels in the same area are always localized in a specific place on the color space, but in many cases, it is not so in an actual image.
Therefore, even though they are the same area, they are divided into a plurality of areas, different areas are regarded as one area, and there are many problems in using them independently.

In order to solve such a problem, the applicant has already proposed JP-A-6-111016.

In this method, when the color image is divided into regions, the characteristic of the distribution in the color space estimated from the reflection model of the object is expressed as a curve from “black” to “white” through “object color”. By modeling each pixel and mapping each pixel on the parameter space that distinguishes the curve, a distribution in which pixels in the same region are localized at a specific location is obtained, and clustering is performed by performing clustering on that space. The color is obtained, and when the variation of each object color is relatively small, the same region in the image can be represented by a small number of cluster representative colors.

[0007]

However, for example, when the variation in each object color is large and the difference between the object colors is small, a large number of cluster representative points appear, which corresponds to the number of regions in the image. There is a problem that it may not be possible to narrow down to a sufficiently small number of cluster representative colors.

The present invention has been made to solve the above-mentioned conventional problems, and corresponds to the number of regions in an image even if, for example, each object color has a large variation and the difference between the object colors is small. An object of the present invention is to provide a clustering method capable of narrowing down to a sufficiently small number of cluster representative colors.

[0009]

According to the present invention, a characteristic of a distribution in a color space estimated from a reflection model of an object is defined as "black".
From the “object color” to the “white”, and by mapping each pixel on the parameter space that distinguishes the curve, pixels in the same region are localized at a specific place. The object is achieved by obtaining a distribution, performing clustering on the space, clustering the cluster representative colors obtained as a result again, and grouping the cluster representative colors into groups.

The cluster representative color is re-clustered after the cluster representative color is converted into hue and saturation.

Further, the re-clustering is performed according to ISODA.
The TA method is used for two-dimensional clustering based on hue and saturation.

Alternatively, the re-clustering is performed by IS
The ODATA method is used to perform one-dimensional clustering by hue.

[0013]

In general, an image is photographed by illuminating an object having a three-dimensional spread with illumination light, passing the reflected light through an appropriate lens system, and mapping the reflected light on a two-dimensional plane.

One object has the same spectral reflectance, that is, the same color, and various gradations occur depending on the relationship between the angle of the illumination light and the surface of the object and the position of the viewpoint. However, this gradation changes along a certain curve in the color space. This is because the shadowed area where the illumination light does not reach becomes black, and the exit angle determined by the incident angle when the illumination light is incident on the surface of the object and the area of the object surface where the line of sight to the object is close, In many cases, it is a specular reflection state, which is a so-called catch light place. In this part, the illumination light reflects light of all wavelengths with a certain reflectance, so it becomes the same "white" as the illumination light. Since it has a distribution, pixels in the same area (same object) are
In the color space (for example, L * u * v * space), it is considered that there is a tendency to concentrate on the curve of black → object color → white as shown in FIG. You can confirm that you are doing. Here, L * represents lightness, the distance from the L * axis in the u * v * plane represents saturation, and the rotation angle represents hue.

Therefore, when viewed in the entire image, there are a plurality of banana-shaped distributions, such as a bent ellipsoid as shown in FIG. 2, corresponding to the number of regions (the number of objects), as shown in FIG. Become.

However, in the conventional method, since these matters are not taken into consideration at all, as shown by the broken line in FIG. 4, even though they are the same area (object), they are divided into a plurality of areas, or in FIG. As shown by the broken line, different regions (objects) may be regarded as one region, and the conventional clustering method is incomplete as a region division method.

In the present invention, based on the above-described distribution model, cluster representative colors are obtained by performing clustering so as to separate a plurality of curved ellipsoidal distributions corresponding to respective regions, and the cluster representative colors calculated as a result are obtained. The problem of the conventional method is solved by re-clustering and grouping.

The clustering method of the present invention will be described in detail below.

Now, as shown in FIG. 6, a color space (for example, L * u
Consider a group of curves passing through a point A representing "black" and a point B representing "white" on the * v * space). Then any pixel becomes
Although represented as one point in the color space, it always belongs to one curve in the curve group. That is, one curve corresponds to one pixel.

Since two parameters for distinguishing each curve are sufficient, each pixel can be represented by mapping to one point on the two-dimensional parameter space. As an example of this curve group, consider a set of semi-ellipses passing through points A and B as shown in FIG. 7, and move the position of the origin on the lightness axis according to L = L * -Lc ... (1) The range of L * is -Lc≤L * ≤
Normalize to 100-Lc. Here, Lc is a median value of the lightness range in the medium, and for example, if the lightness range is 0 to 100, Lc = 50.

Here, since the plane where one semi-ellipse intersects the L = 0 plane is uniquely determined, the semi-elliptic curve group has a parameter depending on the (u *, v *) coordinate value (μ, ν) of this intersection. Can be converted.

Therefore, each pixel can be mapped on the L = 0 plane by a kind of mapping operation along the semi-ellipse.

When the distribution of the actual image in the color space is as shown in FIG. 8, for example, such mapping results in the distribution shown in FIG. 9 on the (μ, ν) plane. It is possible to obtain a localized distribution for each region. In this way, the cluster representative color can be efficiently obtained by using the normal clustering method after mapping in the feature space.

Further, according to the present invention, the cluster representative colors can be clustered again to narrow down the groups, so that a single group can represent the region in the image.

[0025]

Embodiments of the present invention will be described below in detail with reference to the drawings.

An apparatus for carrying out this embodiment is shown in FIG.
As shown in FIG. 3, the color space conversion unit 10, the feature space conversion unit 20, and the region division unit 30 are included.

The color space conversion unit 10 receives the input image signal,
For example, an RGB signal is converted into an L * u * v * signal. RG
B to XYZ coordinate system tristimulus values XYZ, XYZ to L
Conversion to * u * v * has been established and RGB to XY
The conversion to Z is defined by the following equation (see Makoto Miyahara “Systematic Image Coding” IPC Co., Ltd., page 134).

[0028]

[Equation 1]

The conversion from XYZ to L * u * v * is
It is defined by the following formula (see "Image Analysis Handbook" edited by Mikio Takagi and Yoshihisa Shimoda, pages 107-108 of the University of Tokyo Press).

[0030]

[Equation 2]

Here, (X0, Y0, Z0) is (X, Y, Z) for a perfect diffused white surface under a standard light source.
Of the standard light source C (color temperature 6774K), (X0, Y0, Z0) = (98.072, 100.0)
0, 118.225).

In this embodiment, L * u * v * is used as the color space.
Although the color system is used, the equivalent effect can be obtained by considering an appropriate curve group that passes through “black → object color → white” in other color spaces such as the L * a * b * color system. Can be expected.

The conversion unit 20 into the feature space is (L *,
u *, v *) is a point (L,
μ, ν). From FIG. 11, the conversion formula is obtained by algebraic geometry, and is as follows.

[0034]

(Equation 3)

The area dividing unit 30 samples all images, creates training data from the values of (μ, ν) for each pixel, and ISODATA (Ite
rative Self Organizing Data Analysis Techniq
ues) method is used for two-dimensional clustering (see “Image Analysis Handbook” edited by Mikio Takagi and Yohisa Shimoda, pages 648-651 of the University of Tokyo Press).

In the ISODATA method, an appropriate cluster is given as an initial state, and its members are recombined to gradually obtain a "better cluster", which is a representative method of non-hierarchical clustering (relocation method). , Is a method that has been known since ancient times.

The procedure of this ISODATA method is shown in FIG.

In this ISODATA method, first in step 100, parameters such as relocation convergence conditions, minute clusters, isolated data determination conditions, and split / fusion branch conditions are set.

Then, in step 102, the center of gravity of the initial cluster is determined. For example, the total average can be used with the number of initial clusters being 1.

Next, in steps 104 to 108, the rearrangement method is performed. Specifically, first, step 104
Then, the distance between each piece of data and the center of gravity of the cluster is calculated, and each piece of data is arranged in the cluster having the smallest distance.
Then, in step 106, the center of gravity of each cluster is recalculated and corrected in the rearranged clusters. Then
In step 108, if the average of the distance between each data and the center of gravity of the cluster does not change as a result of the rearrangement / correction of the center of gravity, it is considered to have converged, and otherwise returns to step 104 to repeat.

After the rearrangement method is completed, the process proceeds to step 110,
Small clusters with a significantly small amount of data and isolated data that are significantly separated from other clusters are excluded from subsequent clustering.

Next, in step 112, when the minimum value of the distance between the cluster centroids is greater than or equal to a certain threshold value and the maximum value of the variance of the clusters is less than or equal to a certain threshold value, it is determined that the clustering has converged, and the clustering ends. Otherwise, go to steps 114 and / or 116.

If the minimum distance between the cluster centroids is less than or equal to a threshold, the process proceeds to step 114, the cluster pairs are integrated, and a new cluster centroid is calculated.

On the other hand, when the maximum value of the variance of the cluster is equal to or more than a certain threshold value, the process proceeds to step 116, the cluster is divided until the maximum value of the cluster variance is equal to or less than the threshold value, and a new cluster centroid is calculated. .

Integration by step 114 or step 1
After the division by 16 is completed, the process returns to step 104 again to repeat the processing.

In this way, cluster division (cluster boundaries are indicated by broken lines) and cluster representative points 1 to 8 (indicated by circles) as shown in FIG. 13 are obtained.

As described in the section of the operation of the invention, since each of the representative points 1 to 8 represents a single region or a plurality of regions having the same hue in the image, each of the representative points after the division is performed. It is determined which group the pixel belongs to, labeling is performed, and the region division is completed.

That is, although the individual distributions represented by the closed curve and the slanted line in FIG. 13 correspond to one area in the image, a small number of cluster representative points (representative colors) may be included. Therefore, all the cluster representative points are newly regarded as samples, the samples are converted into hue and saturation, and clustering is performed again.

When the set of cluster representative color samples belonging to the "cluster representative color" calculated by this calculation is called a group, one group represents a single region or a plurality of regions of the same hue in the image.

Here, the hue / saturation conversion of the sample is performed as follows.

The cluster representative color sample is (μi, νi)
Then, the hue hi and the saturation si are obtained by the following equations.

Hi = tan ^-1 (νi / μi) 0 ≦ hi ≦ 2π (9) si = √ (μi ² + νi ² ) (10)

The re-clustering for grouping uses the ISODATA method and 2 by (hi, si).
Dimensional clustering may be used, or one-dimensional clustering with hi may provide sufficient results.

[0054]

As described above, according to the present invention, even when the variation in each object color is large and the difference between the object colors is small, a sufficiently small number corresponding to the number of regions in the image can be obtained. It is possible to narrow down to the cluster representative color.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a conventional clustering method on an RGB color space.

FIG. 2 shows L * u * v * for explaining the principle of the present invention.
Diagram showing an example of distribution of object color of one object in space

FIG. 3 is a diagram similarly showing an example of distribution of object colors of a plurality of objects in the entire image.

FIG. 4 is a diagram showing an example of a division result by a conventional clustering method in an L * u * v * space.

FIG. 5 is a diagram showing another example in the L * u * v * space.

FIG. 6 is a diagram showing points representing “black” and “white” on the L * u * v * space for explaining the clustering method of the present invention.
Diagram showing a group of curves passing through points

FIG. 7 is a diagram showing a state in which the range of L * can be normalized.

FIG. 8 is a diagram showing an example of distribution of an actual image in a color space.

FIG. 9 is also obtained by mapping (μ,
ν) Diagram showing the distribution on the plane

FIG. 10 is a block diagram showing an overall configuration of an apparatus for carrying out an embodiment of the present invention.

FIG. 11 is a diagram for explaining a conversion formula used in a conversion unit for converting into a feature space in the embodiment.

FIG. 12 is a block diagram of 2 used in the area dividing unit of the embodiment.
Flow chart for explaining an example of the dimension clustering method

FIG. 13 is a diagram showing cluster division before re-clustering and cluster representative points, which are obtained by the embodiment.

[Explanation of symbols]

 10 ... Color space conversion unit 20 ... Feature space conversion unit 30 ... Region division unit

Claims (4)

[Claims] 1. A distribution characteristic in a color space estimated from a reflection model of an object has a characteristic of "white" passing from "black" to "object color".
By modeling each curve as a curve leading to, and mapping each pixel on the parameter space that distinguishes the curve, a distribution in which pixels in the same region are localized at a specific place is obtained, and clustering is performed on that space. A method for dividing an area of a color image, which comprises performing clustering on the cluster representative colors obtained as a result, and grouping the cluster representative colors. 2. The area dividing method of a color image according to claim 1, wherein the cluster representative color is re-clustered after converting the cluster representative color into hue and saturation. 3. The color image area dividing method according to claim 2, wherein the re-clustering is performed by two-dimensional clustering based on hue and saturation using the ISODATA method. 4. The color image area dividing method according to claim 2, wherein the re-clustering is performed by one-dimensional clustering according to hue using the ISODATA method.