JPH09282461A - Method and system for dividing and sorting important constituting element of color image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To clearly express an important part by dividing a color image into a prescribed number, selectively deciding the importance of each divided part and sorting them to image-process in a state like visualizing. SOLUTION: The color picture is divided into prescribed number, median- filtered for removing noise from an original image and averaging a texture as preprocessing and reduced in addition. After then area extraction is executed by clustering which roughly extract an area from color information and positional information is added to merge-process fine areas. Next the deciding processing of the importance of an object in each area is executed by fuzzy inference based on feature values in each area to sort according to the importance of each area after then. The feature values are a prominent thing easily noticed by people, a thing with special meaning or interest for an individual, a size (area), a position in the image, the state of collecting, an adjacent degree to the tip of the picture, the color of an area, etc.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and system for dividing and selecting an important constituent element of a color image for the purpose of dividing a color image into a predetermined number, selectively determining the degree of importance of the color image, and selecting. It is a thing.

[0002]

2. Description of the Related Art Conventionally, selective compression, storage, and transmission of color images have been performed equally in each area. On the other hand, since the human eye more strongly recognizes a more prominent object in a color image, the degree of recognition by region in the same color image is usually different from the represented real image.

[0003]

According to the above-mentioned prior art, since the important area and the unimportant area are selected in the same manner, there is a problem in that the important portion and the unimportant portion are treated equally in the reproduction. .

Therefore, each part of the transmitted color image is uniformly restored and an image different from the visual state is represented.

[0005]

SUMMARY OF THE INVENTION However, the present invention is characterized in that a color image is divided into a predetermined number, and the degree of importance of each divided portion is selectively determined and selected. It is a method of dividing and selecting components, and in order to determine the degree of importance, the original image is reduced or filtered, it is excessively divided using color information, and regions are integrated from various adjacent relationships on the original image. It is characterized in that the region is divided into regions, the feature amount of the region is calculated from the image, and the importance of each region is determined by fuzzy inference based on this. In addition, the system invention performs preprocessing for excessively dividing a color image, then performs area extraction processing, and then uses fuzzy reasoning based on the calculation of the feature amount of each area to determine the importance of the object in each area. It is a division and selection system of important components of a color image, which is characterized by making a decision. Next, after performing preprocessing by applying a median filter for noise removal and reduction from color images, area extraction processing by clustering that roughly extracts areas from color information, and minute areas by taking positional information into account Is merged. Then, based on the calculation of the feature amount in each area, after performing the process of determining the importance of the object in each area by fuzzy inference, the color image characterized by selecting according to the importance of each area This is a division and selection system for important components, and in order to perform region extraction processing by clustering, the original image is excessively divided using color information, and regions are integrated from various adjacent relationships on the original image. As a result of the region division, the feature amount of the region, which is an element that determines an important object in the region, is calculated from the original image, and the feature amount is calculated.

The color image division methods described above can be summarized as follows: noise removal or reduction is performed from the original image, preprocessing is performed to facilitate processing, and then area extraction is performed from color information to determine the position. The merging process of the minute areas is performed in consideration of the information. A method of calculating visual information of a human being, that is, a feature amount for the human attention to the extracted region, and determining the importance of the object by fuzzy reasoning is adopted.

By following these steps, not only the required calculation can be reduced, but also the features of each area are averaged, so that it is easy to capture the overall features, and the importance of an object whose importance is difficult to determine is important. It is easy to determine the degree.

Further, it is difficult to determine the importance of an object because it depends on the subjectivity of an individual, but in order to easily determine the importance, fuzzy rules are obtained by learning from a sample in an experiment and fuzzy reasoning is performed. This eliminates the complexity of processing due to individual differences among humans.

In the present invention, the object is regarded as "a set of several areas". This is because, in order to capture an object as a whole, it is necessary to have a semantic knowledge about the object as is done in the field of recognition in the current state of the art. In the present invention, only knowledge that is objectively felt by a person who does not rely on such knowledge is handled.

In the present invention, in order to define an important object and automatically determine the degree of regularity of the area, it is first necessary to divide the original image into parts having some meaning or characteristics.

At this stage, first, the original image is reduced, noise is removed by filtering, and excessive region extraction is performed by clustering in the color space, and then the color and adjacency relation of the region on the original image are considered. Then, a process of collecting the excessively divided regions is performed.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION According to the present invention, a color image is divided into a predetermined number, and the degree of importance of each divided portion is selectively determined.
This is a method of dividing and selecting important constituent elements of a color image.

Further, the color image is pre-processed, clustering and merging are performed on minute areas, calculation is performed according to features, and then processing for determining the importance of the object in each area by fuzzy inference is performed. It is a division and selection system of important constituent elements of a color image that is selected according to importance.

[0014]

[Embodiment 1] Next, the following preprocessing is performed in order to extract the area and determine the degree of importance. A median filter is applied to remove noise from the original image and average the texture, and further reduce the size. Here, the feature amounts of the original image are red (Red), green (Green), and blue (Blue).
From the additive color mixture (hereinafter referred to as the RGB color system) based on the three stimulus values of e), Adams (Adm) belonging to the global uniform color system
as) Convert to an improved type of color system (hereinafter referred to as CIELab color system). Future processing will be performed using this CIELab color system.
In addition, the final result is color still image compression technology (Joint
Since the compression of Photigraphic Experts Group (hereinafter referred to as JPEG) is used, it is necessary to keep the reduction in a range where the outline of the object in the image can be seen, and excessive reduction must be avoided.

The reduction processing is performed as follows. 1. In order to match the pixel after reduction with the JPEG quantization unit (block of 8 × 8 pixels), 2 × 2 pixels to 4 × 4 pixels are used as the reduction unit, and this unit corresponds to the pixel after reduction. . When the size is reduced to 1 / T of the whole size, it is as follows. 2. For pixels in each reduction unit, the average value of the values of the elements L, a, and b is output.

[0016]

[Equation 1]

The filtering process is performed in the following procedure.
When the size is halved vertically and horizontally and the median filter is applied three times, the result is as shown in FIG.

One set of 1.3 × 3 pixels is set as one set.
Process as a unit. 2. For the original image, for 9 pixels in 1 unit,
The median (median value) of the values of the L, a, and b elements is taken as the output P (i, j).

[0019]

[Equation 2]

By combining the above-described median filter and reduction processing, noise removal, texture smoothing, and calculation amount reduction can be performed at the same time.

Next, after performing the above-described preprocessing, area extraction is performed to divide the image into rough color areas. As the feature quantity used here, using the CIEL, a, b color system,
Clustering is performed in the coordinate space of this color system. At this stage, since processing is performed on the reduced image,
The extracted area does not exactly match the original image.

The area extraction here is performed by a method applying a peak-climbing clustering algorithm. This is because, even if only peak-climbing is performed, the cluster state obtained depends on the size of the population (cell) for which the histogram is taken, so what size should be used to extract the optimum region? This is because the cluster evaluation problem is added. Also, if the number of extracted regions is too large, the subsequent processing becomes complicated, and conversely, if the number of regions is too small, the state of the object cannot be seen. It becomes a problem and these things need to be considered.

[0023]

Second Embodiment Next, the processing procedure of clustering will be described below.

1. Region extraction processing is performed using the color information (CIEL, a, b) color system. Here, the cluster number c is
The division is done with about 20 as a guide. (A) Since the histogram is taken on the CIEL, a, and b spaces,
CIEL, a, b space is divided into cubes (cells) with one side D,
Substitute an appropriate value for D and initialize. (B) For the cell size D, peak-climbing is performed in the (CIEL, a, b) space. (C) The value of the cell size D is reduced by a constant value until the number of clusters exceeds c, and 1b is repeated. (D) When the number of clusters exceeds c, the process is terminated and m clusters region _i (i = 1, 2, ...,
m) is obtained and the region is extracted.

2. The region extracted by clustering in the CIEL, a, and b spaces is inverse-mapped on the original image and smoothed by applying a threshold value filter, and the number of pixels in the region is less than half that before smoothing processing. The area that has become a non-meaningful area is removed and included in the adjacent area. The minimum value filter is a process of replacing the pixel P (i, j) with the value of the pixel having the highest appearance frequency with respect to 9 pixels including the surrounding 8 pixels.

3. In the image after the threshold value filtering process, the color region extracted from the region is labeled not by the color space but by the position information on the image.

The labeling procedure is as follows. (A) For each pixel, its L, a, b value is the region region _m
If it belongs to, assign 1.

[0028]

(Equation 3)

(B) Labeling processing is performed on the binary (here, (i, j) = (0, 0).) Binary (i, j) = (0, 0) (background) Pixel),
No label. Label (i, j) = 0 • If binary (i, j) = 1 (graphic pixel), attach a label. -Label (i -1, j) = 0 and Label (i, j -1)
) = 0, a new label new is attached. Label (i, j) = new-Label (i-1, j) = Label (i, j-1) (≠ 0
), Attach the same label as both pixels. Label (i, j) = label (i-1, j) -label (i-1, j) (≠ 0) and label (i, j)
If −1) = 0, give the same label as the pixel above. Label (i, j) = Label (i-1, j) -Label (i, j-1) (≠ 0) and Label (i-1)
, J) = 0, give the same label as the left pixel. Label (i, j) = Label (i, j-1) -Label (i-1, j)> Label (i, j-1) (≠ 0
), Then give the same label as the pixel above. Label (i, j) = label (i −1, j). Further, all label labels (i, j −1) are labeled (i
−1, j). ◎ ・ Shift (i, j) by one and return.

4. From the labeled image, a minimum value filter (FIG. 3) is applied to remove smaller areas. Here, the processing is performed on the label value.

5. Area adjacency and color values L, a, b
Alternatively, the regions are merged by using fuzzy inference, considering the color difference between the regions.

(A) L, a, of each pixel in the area label _m
A representative color of the label area label _m is obtained by taking an average value of colors from b and the pixel number area.

[0033]

(Equation 4)

(B) Considering the problem that when the lightness component is included, the region is divided into excessive regions due to the influence of shadows and shading, and the hue (metric-ch) of the a and b chromaticities of the entire region is taken into consideration.
roma) value is averaged.

(C) For a certain region region _m , pixels region _n (m ≠ n at the boundary between adjacent regions other than that region _m )
) Representative color in the CIEL, a, b space (average value of colors)
And obtain the color difference.

[0036]

(Equation 5)

(D) Simplified fuzzy inference is performed for each area label _m from the adjacency relationship between areas to calculate the degree of conformity with other areas and merge the areas with high degree of conformity.
The processing procedure is as follows. Find all areas adjacent to the area of interest. The three color differences are calculated for each adjacent area. Fuzzy inference is performed using that value as input, and the fitness of merging with each adjacent region is obtained. Search for the one with the highest merging suitability, and if that value is greater than a certain value (0.3 to 0.5), merge that region into the region of interest. When the processing is completed for the next area in ascending order, the order of the size of the area is updated and steps 1 to 5 are executed. This is repeated until there is no set of regions whose degree of merging is greater than or equal to the threshold.

Next, a process of determining the degree of importance of a region corresponding to a visually important object (hereinafter referred to as an important region) from the regions divided by the region extraction is performed.
At this stage, the image has just been labeled, and the region must have some "meaning" for use in determining the importance of the region.
It is necessary to examine the features of the areas in the image.

It is not easy to make a decision as to whether or not it is visually important on a computer because it is greatly affected by human subjectivity and individual differences, and it is difficult. Therefore, the degree of importance is determined by fuzzy reasoning, which can express vague knowledge of humans and can perform judgment processing similar to humans.

In order to determine the importance, the following steps are performed. 1. The feature amount of the area in the image is calculated. 2. Teacher data is obtained from the experiment based on the obtained feature amount of the region. 3. A fuzzy rule is obtained by learning the fuzzy rule based on the teacher data. 4. Fuzzy reasoning is performed according to the acquired fuzzy rules to determine the importance.

[0041]

[Third Embodiment] The processing in each stage will be described below.

First, the characteristics of the image area are examined. However, the feature here must be a feature amount that can be quantitatively calculated in order to be used in the process of determining the importance.

Elements that characterize a visually important object are: It is "prominent" that is easily noticed by humans. 2. The object is personally of special interest and interest. That seems to be a big factor.

Here, in the latter case, it is premised that the content of the object in the image is understood. But,
Semantic comprehension is in the field of cognition, and various studies in such fields have been conducted, and the results of the research in that field are adopted.

In the present invention, the concept of approaching to the unconscious perceptual processing that humans naturally perform, that is, when they look at an object, suddenly pay attention to a certain object, without using semantic knowledge about the content, is adopted. Based on this, the importance of the area is determined.

Among the characteristic amounts of the regions in the image, there are those that each region has uniquely and those that depend on the features of the surrounding regions whether or not the region is important.

Regarding image processing, some characteristic amounts are listed, but the characteristic amounts used in the present invention are as follows.

Size (area) The size belongs to the area with respect to the size of the image,
This is because the number of pixels cannot be ignored.

[0049]

(Equation 6)

Position in image The area located near the center of the image is easy to be noticed. Therefore, the distance between the center of gravity of the area and the center of the image is examined.

The center of gravity (iregion _r , jregion _r ) of the area is as follows.

[0052]

(Equation 7)

From the coordinates of the center of gravity, the position can be obtained next.

[0054]

(Equation 8)

Here, (icenter, jcenter) is the coordinates of the center of the image.

Cohesiveness The shape of the cohesiveness is defined as follows.

[0057]

[Equation 9]

Here, the boundary represents the perimeter, and corresponds to the number of pixels on the boundary of the area. Also, an area represents the number of pixels in the area.

This value is a value indicating whether or not the shape is simple, and becomes 1 in the case of the simplest circle, and the value becomes smaller as the shape (complexity) deviates from the circle.

Areas that have a meaningful appearance as "shapes" in the image are areas with good cohesion.

Degree of adjacency to the edge of the image

(Equation 10)

When the position of the center of gravity is near the center, it is easily perceived by humans. On the contrary, when the degree of adjacency to the edge of the image is large, it tends to be the background region.

Area Colors L, a, b For each element L, a, b, the "representative color" of each area is represented by the arithmetic mean value of all the pixels in the area.

Conspicuousness of Color When the difference in color between the regions is large as compared with the characteristics of the overall color of other regions, the hue is noticeable (pops out). Also, it can be said that due to the proximity factor, the color difference in the proximity region is more noticeable than in the distant region.

[0065]

[Equation 11]

Here, distance _ik is

(Equation 12)

Is calculated as

From the above feature amounts, for example, a person is vaguely located in the vicinity of the center, small, collected, red, and a portion (area) that pops out with respect to the whole green. With the knowledge we have, we will be able to capture the characteristics of regions in the image.

In the present invention, the degree of importance of a region is determined by using fuzzy reasoning to obtain robustness.
There arises a problem of how to correlate the obtained feature amount with the importance. In other words, it is necessary to have a basis for what people pay attention to and what they consider important when they see things.

The fuzzy inference is represented by an input / output relationship in which the feature amount of each area is input and the importance of the area is output.

Here, the result of inference depends on the basis for performing fuzzy inference (fuzzy inference rule). Inference rules for this kind of purpose are often determined by human experience or subjectivity, and it is difficult to objectively determine them by human hands.

Therefore, learning is performed by using the teacher data actually obtained by human experimentation, the fuzzy inference rule is automatically determined, and the difficulty in setting the rule is eliminated. By such a method, a value closer to the desired target can be obtained.

The teacher data used in the present invention is obtained by an experiment by a human. In an actual experiment, two images, an original image and a region-extracted image, are shown to the subject, and each region in the region-extracted image is shown. On the other hand, whether or not the area is important was selected, and the values were normalized by a value of 0 to 1 by averaging by the number of test subjects, and the importance of the area serving as teacher data was determined.

An image for teacher data was prepared, and further, an experiment was performed on an image for evaluation for checking whether or not the inference result is actually correct without using it for learning.

As an approach for automatically generating fuzzy inference rules, simplified fuzzy inference by delta rules, which is an extension of sequential fuzzy modeling, is used, and the membership function is adjusted by learning assuming that the knowledge obtained from human experiments is correct. Tuning), we have automatically derived objective inference rules.

Central value of membership function A _{ij in} the antecedent part
Set the initial values of a _ij, width b _ij , and real value c _ij of the consequent part.

As for the real value of the consequent part which is the evaluation result, no matter what value is taken in the initial state when learning is started, there is no problem if the value is within the range. In the above, c _i = 0, i = 0, ...

For each antecedent part variable, it is necessary to determine the number of fuzzy sets and the initial value of the width of each set. However, note that the initial state of the membership function in the antecedent part has a large effect on subsequent learning.
In the fuzzy reasoning here, the antecedent variable x ₀ , ...,
x ₇ is the seven feature values. This feature value is represented by a fuzzy set indicating each degree. First, the number of fuzzy sets for each input variable was set to "small", "normal", and "large" except for the degree of adjacency and color to the edge.

As for the degree of adjacency to the edge, two conditions, that is, the background area and the absence thereof, are sufficient in view of whether or not they are adjacent to each other, that is, the edge portion is likely to become the background.

The color values a and b are (a and b)
The range of values actually taken on the plane (actually −100 ~
(It takes between 100)) and expressed it with 5 fuzzy sets.

The initial state of the width of each fuzzy set is the feature quantity in each area, and the output is the importance of each area obtained from the experimental results.

Fuzzy inference is performed on each input data by using fuzzy inference, and the goodness of fit u of each inference rule u
_i and the inference result c _i are obtained.

When written by the fuzzy IF-THEN rule,
It looks like this:

Rule i: IF large is A _a position is B _b cohesiveness is C _c adjacency is D _d L ^* is E _e a ^* is F _f b ^* is G _g color conspicuity is H _h THEN importance _i = c _i a, b, c, e, h = small, medium, large d = small, large f, g = verysmall, small, medium, large, verylarge i = 1, 2, ..., l (= 3 ⁵ × 2 × 5 ² = 12150)

By the image segmentation focusing on the importance of the object proposed in the present invention, the importance of the region representing the object in the image is determined.

Based on this importance, it is considered that the compression rate is suppressed in the important area to maintain the quality, and the compression rate is increased in the other areas as much as possible to perform the compression.

The image compression method targeted by the present invention basically uses the standard JPEG basic method. For basic processing, JPEG is used as it is, and a part of the portion related to the importance is expanded. The compression is performed in the direction of permitting.

By adopting such a compression method, J
Since compatibility with PEG is maintained, there is no need to modify the basic JPEG processing part, and expansion is easy. Further, the contents can be confirmed by an image display tool compatible with JPEG which is widely available in the world.

In this way, the advantages of JPEG "standardization" can be utilized to the fullest extent, and in addition to human vision,
The image size can also be efficiently compressed.

Various researches have been made in the field of image compression, the world is entering the world of intelligent image compression and structural compression, and some methods focusing on local features of images have been proposed. There is. Also, it becomes extremely difficult because it is related to the internal structure and knowledge of the image.

Focusing on the degree of importance of an object according to the present invention, J
The PEG image compression is similar to the area division encoding called the second generation encoding, but the area division code is encoded by utilizing the fact that the color characteristics of one area are similar. It is different from the conventional method in that it corresponds to the perception that a human sees an image, and that the importance of an object and the compression rate are linked.

The overall flow of compression performed by the present invention is shown in FIG.
As shown in FIG.

First, when performing compression, a distribution of importance is used based on the importance of the area obtained by image division. It is necessary to describe the boundary information of the area. In the present invention, the distribution of the importance determined by the image division is passed to the compression side in the form of a data structure called a block map, and the object information is described based on this block map.

The block map has a minimum data unit (MDU) of JPEG of 8 × 8.
Since it is a block of pixels or 16 × 16 pixels, a two-dimensional array of MDUs for the original image ^width MDU ^{× height} M
Given in DU. All pixels that belong to the same MDU
Treated as the same quality when compressed. That is,
^It can be regarded as an original image of ^width MDU ^{× height} MDU pixels.

Here, the level of importance is described for each MDU in the block map. Based on this level of importance, the quality of the block in the image is
It is determined by the scale table described later.

According to the present invention, compression (encoding) is performed based on the importance of an object in an image or a region as an element thereof. However, all features of the object are determined by a block map. Processing is performed by the block map.

The model to be processed in this manner is not a block having a fixed size anymore, but objects having various shapes. In order to realize such processing, it is necessary to investigate and describe not only the content of the object, but also the characteristics such as the structure and shape (boundary).

According to the present invention, the shape of an important object is described based on the degree of importance of the object, the shape of the important object is described, and the information of the pixels inside the area is also required. By filling and filling with the same importance, JP
It takes the form of writing to EG.

According to the present invention, in order to efficiently reduce the amount of data, all connected blocks of the same importance level are regarded as one object and are grouped together, and the shape is described and processed.

As a method of describing a shape, a Freeman chain code by contour line tracing of a binary image is generally well known and is used for character recognition and the like. In the present invention, the shape is described using this Freeman chain code.

The Freeman chain code is shown in FIG.
, A numerical value from 0 to 7 is assigned to eight directions around a certain pixel, and contour tracing is performed from that point,
Arrange the numbers in the direction of travel. Here, an actual example is shown in FIG.

However, even if the boundary information is described by this method, there may be a region (hole) inside the region. B in FIG. 6 represents a hole.

If there is such a hole, the inside area can be examined and the inside can be regarded as another area. However, regarding that portion, it is necessary to describe the areas in descending order of importance. To do.

Even if a shape corresponding to the importance of the block map is described for each MDU by the Freeman chain code described in the above description, it must be effectively used when actually applied to JPEG. Is the meaning.

Therefore, it is necessary to provide the compressed image with information as to which part of the region has what shape and with what degree of importance.

When applied to standard JPEG, JPEG
Has several data areas reserved for future expansion and other uses. Among them, the application data segment (Application Data Segment
There is an area called t) that can be used independently by the application. In the present invention, the description of the boundary information is stored in this application data segment (FIGS. 7A and 7B).

Here, AP _n , LP, and BND _i indicate the following, respectively.

AP _n application data marker 2 bytes LP information of application area 2 bytes BND _i Information of i-th area The contents of each are as follows. BND Start point that describes boundary information Level Level of quality of block X ₀ X coordinate of start point of area Y ₀ Y coordinate of start point of area Lb Number of boundary blocks B _{1 to} B _n Freeman chain code

Here, in the present invention, since the shape is described by making a binary comparison with the level of background importance as a reference, the level of background importance is taken into consideration. This is because, depending on the image, all of the images are judged to be considerable in importance in some cases, and the importance of the background area portion is not 0, and a reasonably high importance level may be required.

[0110]

Fourth Embodiment In the present invention, a block map and
A scale table, which is a correspondence table of importance and quality level, is used to obtain quality from the level of importance obtained from image segmentation.

The quality level value ranges from 0 to 100, with 0 being the worst and 100 being the best quality. Each element of the base quantization table (8x8)
And perform a calculation including integration to set a new quantization table.

By determining the quality corresponding to the importance of the block map, the quantization table of each block is set by the above-described arithmetic method. In this way, quantization is performed using a different quantization table for each block.

When compression is actually performed by JPEG, the quality after compression, the compression size, etc. change depending on how the quantization table is determined. However, since there is no standard in this quantization table and the decision is left to the discretion of the user, how to determine this quantization table becomes a problem.

In the present invention, the quantization table prepared as the sample is used. By using this quantization table, the parameters can be aligned with those of standard JPEG, and there is an advantage that it is easy to compare the compression results of JPEG and this method.

The actual compression procedure of the JPEG image compression focusing on the importance of the object proposed by the present invention is as follows.

[0116] 1. Create a block map by dividing the importance distribution of the image segmentation results into blocks. 2. The compression side receives the block map. 3. Examine the level of importance of each block from the received block map. 4. From the level of importance, the scale table gives the corresponding quality. 5. Freeman chain code is used to describe the shape of the area according to each importance. 6. The importance level, the background importance, the chain code start point (X ₀ , Y ₀ ), and the chain code are written in the JPEG application data segment. 7. JPEG compression is performed with the obtained quality (FIG. 8).

According to the procedure described above, the quality (compression rate) can be increased according to the degree of importance of the object which can be easily expanded by directly using JPEG without directly modifying the internal internal parts. It can be changed freely by compression.

The actual decoding procedure is the opposite of the encoding,
It looks like this: This operation is basically a standard JP
Same as EG decoding, but the contents of the application data segment cannot be read as it is, so JPEG
Before decrypting, you need to make some changes to read that part.

[0119] 1. Import the image file to be decrypted. 2. Read the application data segment from the imported JPEG file. 3. Read the importance level of the background area from the application data segment. 4. The application data segment is read from the starting point (x, y) and data of each area, and the boundary information of each area is obtained by the Freeman chain code. 5. Create a block map. 6. Read the scale table and determine the quality from the block map. 7. JPEG decoding is performed based on the quality of each block (FIG. 9).

First, in order to roughly extract an important area, as a preprocessing thereof, the image of FIG. 3 is subjected to reduction processing and median filter processing.

FIG. 4 shows an image which has been subjected to the reduction processing once and the median filter three times. Here, in FIG.
It is enlarged to the same size as the original image so that the details of the processing result can be clearly discriminated.

A value called an inference error was considered in order to show the progress of learning of fuzzy rules. The inference error is the difference between the degree of importance judged by humans obtained by experiments and the degree of importance judged by the system.

Therefore, the graph shows the relationship between the transition of the inference error with respect to the learning image and the transition of the inference error with respect to the unknown image. Here, the inference error of the learning image is the average of the inference errors of all the 20 images used for learning, and the inference error of the unknown image is the average of the inference errors of the 10 images not used for learning. Using.

The inference error with respect to the learned image naturally decreases as the learning progresses. However,
The inference error for an unknown image converges to some extent, but increases at a certain point. This is because as the learning progresses, the rules are adjusted to fuzzy rules that calculate the importance of the region, but if it goes too far, it will be specialized in the characteristics of the learning image itself and will be compatible with other images. It shows that it will disappear. Therefore, we decided to use the fuzzy rule, which has the smallest inference error for unknown images, because it is the most versatile.

[0125]

According to the present invention, since the color image is divided and the degree of importance is set for each divided portion and the extraction processing is performed, the image processing is performed in a visual state, and the important portion is clearly identified. There is an effect expressed in.

[Brief description of drawings]

FIG. 1 is a conceptual diagram of reduction processing according to the present invention.

FIG. 2 is a conceptual diagram of labeling as well.

FIG. 3 is a conceptual diagram of a threshold filter.

FIG. 4 is a block diagram of a compression process in the same embodiment.

FIG. 5A is a diagram similarly showing a Freeman chain code. (B) The figure which similarly shows another chain code.

FIG. 6 is a diagram when there is a hole inside the same region.

FIG. 7A is a block diagram of an application data segment. (B) Similarly, a block diagram of a format describing a shape.

FIG. 8A is also a quantization table of lightness. (B) Similarly, a quantization table for hue.

FIG. 9 is a block diagram of the same compression process.

FIG. 10 is a block diagram of the decoding process.

[Procedure amendment]

[Submission date] April 26, 1996

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claim 2

[Correction method] Change

[Correction contents]

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] Claim 3

[Correction method] Change

[Correction contents]

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] Claim 5

[Correction method] Change

[Correction contents]

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0005

[Correction method] Change

[Correction contents]

[0005]

SUMMARY OF THE INVENTION However, the present invention is characterized in that a color image is divided into a predetermined number, and the degree of importance of each divided portion is selectively determined and selected. It is a method of dividing and selecting components, and in order to determine the degree of importance, the original image is reduced or filtered, the image is divided using the color information, and regions are integrated from various adjacent relationships on the original image. It is characterized in that the region is divided into regions, the feature amount of the region is calculated from the image, and the importance of each region is determined by fuzzy inference based on this. The invention of the system performs pre-processing for split a color image, and then after the area extraction processing, the importance of an object in each region by fuzzy inference based on the calculation of the feature amount of each region It is a division and selection system of important components of a color image, which is characterized by making a decision. Next, after performing preprocessing by applying a median filter for noise removal and reduction from color images, area extraction processing by clustering that roughly extracts areas from color information, and minute areas by taking positional information into account Is merged. Then, based on the calculation of the feature amount in each area, after performing the process of determining the importance of the object in each area by fuzzy inference, the color image characterized by selecting according to the importance of each area is an important component divided sorting system to do area extracting process by the clustering, and split the original image by using the color information in the original image on integrates regions from various adjacency regions As a result of the division, the feature amount of the area, which is an element that determines an important object in the area, is calculated from the original image, and the calculation is performed.

[Procedure Amendment 5]

[Document name to be amended] Statement

[Correction target item name] 0011

[Correction method] Change

[Correction contents]

At this stage, first, the image is reduced from the original image, noise is removed by filtering , the region of the image is extracted by clustering in the color space, and then the color of the region on the original image and the adjacency relation are taken into consideration. , performs a process of summarizing the previous Symbol divider area.

[Procedure correction 6]

[Document name to be amended] Statement

[Correction target item name] 0034

[Correction method] Change

[Correction contents]

(B) Considering the problem that when the lightness component is included, the region of the image is divided due to the influence of shadow and shading, the hue (metri) of the a and b chromaticity of the entire region is taken into consideration.
c-chroma) value is averaged.

Front page continuation (72) Inventor Kenichi Okada 4-25-12 Hongo, Bunkyo-ku, Tokyo (72) Inventor Rina Hayasaka 3-14-1, Hiyoshi, Kohoku-ku, Yokohama-shi, Kanagawa Keio University Faculty of Science and Engineering, Department of Measurement Engineering Matsushita Laboratory (72) Inventor Zhao Ji-ei, 14-1, Hiyoshi 3-chome, Kohoku-ku, Yokohama-shi, Kanagawa Matsushita Laboratory, Keio University Faculty of Science and Technology (72) Yoshihisa Shimazu 3-chome, Hiyoshi, Kohoku-ku, Yokohama-shi, Kanagawa 14-1 Keio University Faculty of Science and Engineering Department of Measurement Engineering Matsushita Laboratory (72) Inventor Koji Ota 3-14-1 Hiyoshi, Kohoku Ward, Yokohama City, Kanagawa Prefecture Keio University Faculty of Science and Engineering Department of Measurement Engineering Matsushita Laboratory

Claims (5)

[Claims] 1. A method of dividing and selecting important components of a color image, which is characterized by dividing a color image into a predetermined number and selectively determining the degree of importance for each division portion and performing selection. 2. In order to determine the importance level, the original image is reduced or filtered, excessively divided using color information, and regions are integrated on the original image from various adjacency relations and divided into regions. 2. The method for dividing and selecting important components of a color image according to claim 1, wherein the feature amount of the region is calculated from the image, and the degree of importance of each region is determined by fuzzy inference based on the feature amount. . 3. Pre-processing for excessive division of a color image, then area extraction processing, and then fuzzy inference based on the calculation of the feature amount of each area to determine the importance of the object in each area. A division and selection system for important components of color images. 4. A region extraction process by clustering that roughly extracts a region from color information after preprocessing by applying a median filter for noise removal and reduction from a color image, and position information. The merging process of minute areas is performed. Then, based on the calculation of the feature amount in each area, after performing the process of determining the importance of the object in each area by fuzzy inference, the color image characterized by selecting according to the importance of each area Divided sorting system for important components. 5. In order to perform a region extraction process by clustering, an original image is excessively divided using color information, regions are integrated from various adjacency relations on the original image to obtain a region division result, and 5. The division / selection system for important constituent elements of a color image according to claim 4, wherein the characteristic amount of the area, which is an element that determines an important object in the area, is calculated from the original image.