JPH01232894A - Noise elimination method for color picture - Google Patents

Abstract

PURPOSE:To eliminate noise in a versatile picture in an excellent way by applying probability processing to vagueness of noise. CONSTITUTION:A color picture being a processing object is read by an external scanner and its quantization data is stored in a picture memory 101. A CPU 102 applies the processing for noise elimination and its pre-processing according to a processing program 105 stored in a program memory 103. The CPU 102 in the processing for noise rejection measures a value depending on the color difference and distance of two adjacent areas as the noise-like degree and incorporates the areas in response to the value. A data memory 104 is used to store the data on the way of processing and the processing result data.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field] The present invention relates to a method for removing noise in color images in devices that handle image information.

[Prior art]

Traditionally, image filtering is used to perform similar processing1.
Noise in the image is removed.

However, there is a problem in that parameters need to be adjusted (tuned) depending on the attributes of the image. Furthermore, it is difficult to achieve both smoothing processing and sharpening processing, and when processing for removing granular noise (smoothing processing) is performed, there are problems such as blurring of the outline of the image.

Publicly known materials regarding this type of noise removal include ``Enhancement and Smoothing Process'', Nakamura February Monthly Magazine ``O Plug E J, &75. 1986, 2
Month (published by Shintechi Co., Ltd. Communication), No. 83
There are pages 97 to 97.

〔the purpose〕

An object of the present invention is to provide a noise removal method that can effectively remove noise (especially noise caused by color density unevenness) from various color images without adjusting parameters for each image.

〔composition〕

The present invention attempts to achieve the object by performing probabilistic processing on the ambiguity of noise.

In other words, according to the present invention, instead of checking the degree of noise-likeness of a region in a color image and uniformly integrating the regions when the degree exceeds a certain limit, the probability is determined according to the degree of noise-likeness. Integrate adjacent areas with . As the degree of noise-likeness, a value E=f (color difference, distance) determined by the color difference and distance between two adjacent areas is measured. The function f is a function in which the value E increases as the color difference increases and the distance decreases, and by integrating the smaller area of two adjacent areas into the other area with probability according to this value E, Removes noise caused by uneven color density. Such area integration is preferably performed with priority given to areas with large areas.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows an example of an apparatus configuration for implementing the present invention. 101 is an image memory, 102 is a central processing unit (
CPU), 103 is a program memory, and 104 is a data memory.

A color image to be processed is read by an external scanner, and its quantized data is stored in image memory 101. Regarding this color image, processing for noise removal and its preprocessing are executed by the central processing unit 102.
The processing program 105 is stored in the program memory 103. The rules for noise removal are as described below.

There are three types, (1) and (2), which are incorporated into the processing program 105, but may also be made independent as a rule base.The data memory 104 is used to store data in the middle of processing and processing result data. .

Next, the overall flow of the process will be explained with reference to the flowchart shown in FIG.

First, a label is attached to each uniformly colored region of the color image in the image memory 101 (step 201). Next, the color image after labeling is recursively divided into four to describe it in a color image quadtree structure, and the description data is stored in the image description data section 10 in the data memory 104.
6 (step 202), the identification number of each area is provided as attached information to the leaf of the tree structure.

Note that the process up to the knee is not new, so a detailed explanation will be omitted. Further, instead of the quadtree structure, a tree structure such as a binary tree may be used.

When an image is described in a tree structure in this way, the amount of data can be reduced, and a topological spatial search for a region in the image can be executed at high speed in subsequent processing.

Next, by referring to the image description data, table 1 shown in FIG. 3 and table 2 shown in FIG. .

Table 1 stores, for each labeled area, area identification number, color identification number, area (number of pixels), and area circumscribing rectangle information.

FIG. 5 is an explanatory diagram of this circumscribed rectangle information, where 501 is one labeled uniform color area, and the coordinates (minx+m1ny) of the diagonal vertices of the circumscribed rectangle (broken line),
(maxx, maxy) is stored in table 1 as circumscribed rectangle information.

When performing a topological spatial search of an image region, by limiting the range of the region to be searched using this circumscribed rectangle information, efficient search becomes possible.

Table 2 stores data for the three primary colors of R (red), G (green), and B (blue) corresponding to each color identification number, and based on these three primary color data, , converts the color corresponding to the color identification number into a uniform color space, ie, a Lea umbrella space, and stores its Lm value, a middle value, and b umbrella value. In this way, by using the L-skew a*b-skew color system, which is a uniform color space, it becomes possible to perform noise determination that matches human visual characteristics.

Steps 204 to 206 refer to the image description data and Tables 1 and 2 to determine Rules I and II.

This is the processing part that performs noise removal according to (2).

At step 204, a region of interest a is set, and at step 205
Then, rules I, n, and m are sequentially applied to remove noise.

Then, in step 206, it is determined whether all areas have been set as attention areas, and if there are any unset areas remaining, the process returns to step 204 to set the next attention area and perform processing.

Therefore, in this embodiment, "background ring" and "edge degree" are used as the degree of noise-likeness. The background ring is determined by the density and area of the area of interest, and the degree of edge is determined by the distance and color difference between the two areas of interest.
Each is defined as follows.

Skin ring = COMB (wl ・D, w2 ・S)
(1) Edge degree = COMB (w3 ・L, w4
・E) (2) With this N, C=COMB (a,
b) is defined as follows.

■ If a==1 or b=1, then C=1 ■ a > O
And if b>o, then C=a+b-ab■ ab≦0 and a
# If ±1 and b ~ ±1 then C = a + b ■ If a < O and b < o then C = a + b +
a b ■ If a = -1 or b = -1, then C = -1 In equations (1) and (2), wi is the weighting coefficient, D is the concentration level, S is the area, L is the distance, and E is It's the color difference.

D, S, L, and E are expressed as values from -1 to +1 depending on their degree. In other words, both the background ring and edge degree are values from -1 to +1, and the density level is low (close to white).
, the larger the area, the higher the background ring, and the smaller the distance and larger the color difference, the higher the edge degree. Note that the width value of the smaller of the widths of the two regions is substituted for L, and the Lm value is substituted for D.

Furthermore, E is the color difference in the L umbrella a*b color system.

Next, the process of step 205 will be explained for each applied rule.

First, processing according to rule (2) is performed. Rule (2) is an integration rule between adjacent areas, and this process mainly removes noise caused by color density unevenness. FIG. 6 is a flowchart.

The tree structure is searched for the region set B = (bilbi is the region adjacent to a) adjacent to the region of interest a (step 6
01). The degree of edge between the region of interest a and one adjacent region b1 is measured (step 602).

A comparison is made between the area of the attention area a and the area of the adjacent area bi (step 603).

When the area of the attention area a is larger than the area of the adjacent area bi, the adjacent area bi is integrated into the attention area a; otherwise, the attention area a is integrated into the adjacent area bi (step 6
04,605).

This integration is not performed uniformly, but with a probability that depends on the degree of edgeness.As shown in Figure 7, this integration probability is 0% when the degree of edgeness is above a certain value, and when the degree of edgeness is less than that value. The smaller the value, the larger it becomes, and the edge degree = -1 is 100.
%. In other words, when the integration probability according to the degree of edge is 100%, integration is always performed, and when the integration probability is 0.
When the integration probability is 50%, no integration is performed, and when the integration probability is 50%, integration is performed on average once every two times.Probabilistic integration processing is performed according to the degree of edge.

In this way, in order to stochastically integrate ambiguous noise regions, unlike the method of uniformly integrating regions when the edge degree exceeds a fixed threshold, the noise removal processing parameters are adjusted for each image. Good noise removal is possible for a variety of images without any adjustment (tuning).

After the above processing for the region of interest a and one adjacent region bi is performed, it is determined whether the processing has ended for all regions bi in the region set B (step 606), and if there is any remaining region bi, the processing starts from step 602. resume.

One image description data and table 1 are updated for the combined areas through such processing.

Note that if the region integration processing according to this rule (2) is executed without considering the size of the area of the region of interest, the integration may proceed mainly in regions with small areas. In general, regions with small areas contain uncertain information, so it is possible to achieve more accurate integration by starting with the regions with large areas and proceeding with the integration. Therefore, it is preferable to sort all regions into areas, set regions of interest in descending order of area, and proceed with the integration process. However, there is a problem in that the sorting process requires a very long processing time.

FIG. 8 shows an example of a processing method for obtaining the same effect as when sorting is performed without performing this sorting. Step 801 corresponds to step 201 to step 20 in FIG.
3, and steps 804 and 807 are the same processing steps as steps 204 and 206 in FIG. Step 806 is a processing step similar to steps 601 to 606 in FIG.

Step 802 is a step of initializing the counter n, and step 808 is a step of comparing and determining the value of the counter n with the maximum value.

Step 805 is a comparison judgment between the area of the attention area a set in step 804 and 81 and S2, and only the attention area that satisfies this determination condition becomes the processing target of step 806.

Here, SL and S2 are values that change depending on the value of n. For example, when n=1, 5L=200. S2=infinity When n=2, 5L=100, 52=200 When n=3, 51=1
0. When 52=100n=4, 51=O, 52=
It is 10.

After the noise removal process according to rule I as described above, the noise removal process according to rule (2) is executed. This rule (2) is an integration rule for a small area included in a large area, and this process performs noise removal assuming white noise, halftone dots, and color toner adhesion noise in electrophotography. FIG. 9 is a flowchart of this process. First, it is checked whether the area of the attention area a is equal to or larger than the threshold value Th1 (step 901), and if the area is equal to or greater than Th1, the area set B included in the attention area a= (bilbi is a
) is searched for on the tree structure (step 902).

This region set B is clustered into regions Ci with small color differences (step 903). The areas Ci are a subset of B, each consisting of one or more areas bi in a tree structure, and do not overlap with each other.

That is, B=CIUC2U...CkU...

Each region Ci is defined as the color difference with respect to the region of interest a, the area and "
The degree of aggregation is classified into categories 1.2.3 (step 904).
It is the value obtained by dividing the number i) by the area of region a.

Category 1 is assumed to be white noise,
Areas where the color difference is fairly small and the area is not large are classified into this category. Category 2 assumes halftone dots,
Regions with medium color difference, small area, and medium aggregation degree are classified into this category. Category 3 is assumed to be color toner adhesion noise, and areas with a fairly large color difference, a small area, and a fairly small aggregation degree are classified into this category.

Regarding category 3, a color condition may be added, such as, for example, that the ink color is Y (yellow), M (magenta), C (cyan), or BK (black).

Areas Ci that are not classified into any of the above categories are not noise (they are valid information) and are therefore excluded from area integration targets.

If at least one C1 in B is classified into one of categories 1, 2, and 3, the degree of texture of the attention area a is measured (step 906), and according to the degree of the degree of background,
Region bi within region Ci classified as category 1, 2, or 3 is probabilistically integrated into region of interest a (step 9
07).

As shown in FIG. 10, the integrated probability in this region is 0% when the background texture is below a certain value, but increases as the background texture increases above that value.

Such area integration can mainly remove the above-mentioned granular noise. Furthermore, since area integration is not performed in a regular manner, but stochastic integration is performed according to the degree of texture, it is possible to remove noise from a variety of images without having to tune parameters for each image.

Next, processing is performed according to rule 3. Rule 3 is a rule for integrating ambiguous areas (noise) in the outline of the image,
This processing makes it possible to sharpen the outline. 1st
Figure 1 is a flowchart of this process.

First, check whether the width of the attention area a is smaller than the threshold Th2 (
In step 1001), if the width is equal to or greater than Th2, the region of interest a is not an outline region and is therefore excluded from processing.

For the region of interest a whose width is smaller than Th2, a nearby region bi (not necessarily adjacent to the region of interest a) within a certain range is searched on the tree structure (step 1002).

In this proximal area, select two areas bl and b2 that are in a positional relationship sandwiching the attention area a from both sides (does not need to be in contact with a) and have a width of Th3 or more (Step 1
003). If such areas bl and b2 cannot be selected, the process ends.

Next, it is determined whether the color of the attention area a has a mixed color relationship with the colors of the areas bl and b2 (step 1004). If there is no color mixing relationship, the process ends.

If it is a color mixture relationship, the attention area a is selected from the areas bl and b2.
Let the area with the smaller color difference be bl, the other b2,
The color differences ΔEl and ΔE2 with respect to each region of interest a are determined (step 1005).

A comparative judgment is made between ΔE1 and Δ2 (step 1006),
If ΔE1≧ΔE2, the process ends, but ΔEl<ΔE
If it is 2, the degree of edge between the region of interest a and the region b1 is measured (step 1007).

Then, the attention area a is determined with probability according to the degree of edge.
The color is changed to the color of area b1 (step 1008).
, if the two areas are in contact, this color change is the integration of the areas themselves, but if they are apart, it is the area integration of only the color planes.

As shown in FIG. 12, the probability of this color change (area integration) increases as the edge degree increases.

By stochastically changing the area in this way, it is possible to remove ambiguous areas (noise) in the image outline, and instead of randomly changing the color (integration), we can change the probability according to the degree of edgeness. Since the colors are changed using , good results can be obtained for a variety of images without having to tune parameters for each image.

Through the above processing, the quadtree-structured image description data and table 1 are updated, and these are obtained as processing results in the respective storage units 106 and 107 of the data memory 104.

Note that the rules and algorithms for noise removal, the noise-likeness index, the configuration of the processing device, etc. can be changed as necessary.

〔effect〕

As is clear from the above description, according to the present invention, parameters are tuned for each image by stochastic region integration (fuzzy processing) according to the degree of noise-likeness determined by the color difference and distance between two adjacent regions. Noise such as color density unevenness can be effectively removed from various color images without causing any problems.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an example of a device configuration for implementing the present invention, FIG. 2 is a flow chart of the entire process, FIG. 3 is an explanatory diagram of table l, FIG. 4 is an explanatory diagram of table 2, Figure 5 is an explanatory diagram of circumscribed rectangle information, Figure 6 is a flowchart of processing according to rule 1, Figure 7 is a diagram of the relationship between edge degree and integrated probability, and Figure 8 is a diagram of the relationship between edge degree and integrated probability.
Figure 9 is a flowchart showing a modification of processing related to Rule 1, Figure 9 is a flowchart of processing according to Rule 2, Figure 10 is a relationship diagram between texture and integrated probability, Figure 11 is a flowchart of processing according to Rule 3, FIG. 12 is a diagram showing the relationship between edge degree and color change probability. 101... Image memory, 102... Central processing unit, 103... Program memory, 104... Data memory. Figure 2 Figure 0 (Tape 1) Figure 4 Figure 6 Figure 10 Figure 12 Enson (

Claims (1)

[Claims] (1) Measure the value E = f (color difference, distance) determined by the color difference and distance between two adjacent areas in a color image, where the function f is such that the larger the color difference and the smaller the distance, the larger the value E becomes. A color image noise removal method characterized by integrating the smaller area of the two regions into the other region with probability according to a value E.