JPS58163075A - Picture quality deciding method of binary linear graphic picture - Google Patents

Abstract

PURPOSE:To facilitate the decision of partial picture quality and to reduce an effect due to the number of lines, by deciding the picture quality after observing a legal pattern of unit meshes emerging in a subject binary picture of a linear pattern. CONSTITUTION:A binary linear pattern is observed through the unit meshes (e.g., a square window of 3X3 picture element size) which are thinner than the line of the binary linear pattern, and the emerging frequencies are calculated for white/ black patterns (legal patterns) which can emerge in the unit meshes. The 22 types of parameters are calculated in terms of the correlation between the good quality and inferior quality including contamination, blur, etc. and the emerging frequencies of 66 legal patterns (or 7 groups). Then the quality of pictures is decided from the result of the parameter calculation. For a picture to be decided for its quality, the ''good quality'' of the picture is first decided from 9 types of parameters. For the picture decided as ''inferior quality'', the ''contamination'' or ''blur'' is decided from other parameters.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for determining the image quality of a binary image of a line figure such as a document.

Binary images obtained by optically scanning and reading line figures such as documents and binarizing them can be broadly classified into "faint" images, "contamination", "double images", and "good quality" images based on image quality. can.

``'Faint'' drawing method is when the lines of characters etc. become thinner, °
``Contamination and contamination.'' The image is not of good quality, including blurring of lines such as characters, noise due to the background of the manuscript, and other stains.
There is almost no "fading" or "contamination". Of course, the above image quality is greatly influenced by the image quality of the original itself, but even in the case of a "good quality" original, if the threshold for binarization processing is inappropriate, the binary image will be If the image is misaligned, the image will become ``contaminated''.

When storing, transmitting, or recognizing binary images, it goes without saying that it is desirable for the image quality to be of "good quality." In order to meet this requirement, it is necessary to The premise is that the image quality of the image is determined in advance. If the image quality can be determined in advance, processing according to the image quality can be applied, such as ``additional black-on-black processing'' for ``faint'' images, and ``black removal'' processing for contaminated images, to bring it closer to a ``good quality'' image. becomes possible.

Therefore, an object of the present invention is to provide a method for determining the image quality of binary line graphic images.

When determining image quality, it must be taken into account that image quality is not necessarily uniform throughout the image. This is because the image quality of the image frequently changes depending on the location due to aging of the image reading mechanism or the like. Also,
The proportion of black pixels cannot be simply used as a parameter for image quality determination. If the number is the same, there are images such as literary calendars that are crowded with lines, and there are images that have extremely few lines such as graphs, and it is normal for the density of lines to vary greatly depending on the location.

Therefore, another object of the present invention is to provide an image quality determination method that can easily determine the partial image quality of a binary line graphic image and is less affected by the number of lines.

Therefore, the image quality determination method of the present invention is based on the M
OL D Theory (Mesh Oriented Line
Drawings Theory), the image quality is determined by observing a legal pattern of unit meshes appearing in a binary image of a target line figure.

Unlike natural images, binary line figures are artificially generated images and have strong constraints. The MOLD theory described above expresses this constraint and east condition in a unit mesh.

Lines forming a binary line graphic image input by facsimile or the like do not form digital arcs as defined in digital geometry. These lines are generally wide, realistic lines. In MOLD theory, the target line is called a digital line. A digital line has two or more arbitrary black pixels included in it in the sense of 4 adjacent pixels,
It is defined as having black pixels adjacent to each other in the sense of 8 adjacent pixels, and white pixels in the complementary set thereof also satisfying the same condition. A line that actually exists will always become a digital line if the sampling interval is made fine enough. A necessary and sufficient condition for this is to sample at a fineness that is less than half the sampling interval determined by the sampling theorem.

When a digital line is observed with a smaller unit mesh (generally a square window of m x m pixel size),
The black and white patterns that can occur within it are limited. That is, patterns can be divided into patterns that match the constraint that the target is a digital line (this is called a legal pattern) and other patterns (this is called an illegal pattern).

As an example, FIG. 1 shows a legal pattern when the mesh is a window with a size of 3×3 pixels. The number above each legal pattern is the symbol number of that legal pattern, which is numbered in descending order of frequency of appearance in "good" images. For a binary image such as an alphabet screen, if we take statistics on the unit mesh pattern of 3 x 3 strokes, if the assumption that the thickness of the line is larger than the unit mesh is satisfied, these 66 legal patterns The cumulative rate of only is 99th coefficient or higher of the whole. Even excluding the most frequent all-white legal pattern (symbol number 1), the cumulative rate is 95%
exceed.

A certain correlation is recognized between the frequency of appearance of individual legal patterns and groups of legal patterns and the image quality of the target binary line graphic image. Figure 2 shows the results of investigating the appearance frequency of the legal pattern shown in Figure 1 for binary images of three types of image quality (faint, good quality, and contamination) of CCITT's facsimile test manuscript No. 1 (English letter). . The image size is 1728 x 2352 pixels, the unit mesh size is 3 x 3 pixels, and the unit mesh is moved for each cumulative stroke, so the total number of samples is 4.056.100.

The following can be seen from Figure 2.

The appearance frequency of legal patterns of symbol numbers 3 to 10 is “
Although it is stable in 'good quality' images, it is very uneven in 'contaminated' and blurred images. This unevenness means that noise is superimposed on the digital line of the target image. That is, since the unit mesh is moved pixel by pixel, if the image is of "good quality", for example, if a legal pattern with symbol number 6 exists, there will always be a legal pattern with symbol number 50 next to it. For this reason, °°If both images are of good quality, symbol numbers 3 to 10
The frequency distribution of the 0-legal pattern becomes flat (
However, there is an exception if there is a bias in the direction of the line).

The appearance frequency distribution of legal patterns for symbol numbers 11 to 18 is also higher in ``contaminated'' images and ``good quality'' images than in ``good quality'' images.
Faintness'' Both images have severe unevenness.If the digital line is broken or there are many isolated points, black pixels may form one legal pattern (
It is easy to understand that the appearance frequency of 11°12, 13.15) increases. A legal pattern with one white pixel (1
4゜16, 17.18) increases overall in the "'good" image in the "'good"' image (this is because these patterns also appear in the noise);
It is significantly reduced in "gray" images. symbol number 19
A similar trend can be seen for the ~26 legal patterns.

In the present invention, the legal pattern of the unit mesh of the target binary line figure image is observed, and the parameters related to the correlation between the image quality and the appearance frequency of the legal pattern (or group) as described above are calculated, The image quality is determined from the results.

This is the main feature of the image quality determination method according to the present invention.

”-Before explaining the parameters used for image quality judgment,
Add a supplementary explanation about the characteristics of legal patterns.

The 66 legal patterns of 3×3 stroke size shown in FIG. 1 can be classified into 7 groups A to G shown in FIG. 3 based on their shapes. The numbers (in parentheses) below each legal pattern in each group indicate the total number of that pattern and its rotations.

For example, the legal pattern on the left side of group B makes it 9
A total of four patterns are represented, including three patterns rotated by 0°.

If classified according to directionality, they can be divided into seven groups (■ to ■) in Figure 4. For example, the group of legal patterns that appear in the degree mesh located at the bottom right of the diagonal digital line is ■
It's a group.

Next, the results of examining parameters for image quality determination will be described. In the following description, the size of the unit mesh is 3×3 pixels, and the legal patterns are the 66 patterns shown in FIG. 1.

The parameters considered were 29 types of parameters shown in FIG.

Parameter 1 is the ratio of legal patterns. This parameter 1 is an all-white legal pattern (symbol number 1,
By excluding the W pattern (abbreviated as W pattern) from the statistics, it is possible to obtain a parameter that does not depend on the amount of black pixels in the target image.

Parameter 2 is the ratio of the left legal pattern (11, 12, 13, 15) to the right legal pattern (14, 16, 17, 18) of group B in FIG. Parameter 3 is the legal pattern (19
~26) and the legal pattern on the right (27, 28° 30, 3
1.35.36.39.40).

Regarding these two parameters, in a blurred image, the legal pattern on the denominator side decreases, and the legal pattern on the numerator side increases.

Parameter 4 is the statistical ranking of the all-black legal pattern (2). Although the boundary portion of the digital line is divided into 64 legal patterns (3 to 66), an all-black legal pattern (2) always appears inside the line. Therefore, if the line is not broken or blurred, this legal pattern (2) will definitely come in second place.

Parameters 5 and 6 are parameters indicating the connectivity of the W pattern. Parameter 5 has 3 W patterns in the scanning direction.
This is the ratio of 0 consecutive occurrences.

Parameter 6 is the ratio at which all of the i5 x 155 pixel square area has a W pattern. These two parameters depend on the amount of characters and lines in the target image.

Parameters 7 to 26 are parameters related to rules between legal patterns (legal pattern rules 1 to 5) for diagonal components. These legal pattern rules 1 to 5 are rules such that the relationship shown in FIG. , 30) is equal to the number of appearances of legal pattern θ4).The same applies to other legal pattern rules 2 to 5.

However, FIG. 6 illustrates only the group ■ in FIG. 4, and each legal pattern rule also holds true for the groups ■, ■, and ■ in FIG. 4. That is, legal pattern rules 1 to 5 also hold true between legal patterns obtained by rotating each pattern on both sides by 90 degrees. Therefore, for legal pattern rule 1, the formula 4m holds true, for legal pattern rule 2, 4 sets of formulas hold true, and similarly for legal pattern rule 3.4.
Four sets of equations hold true for each.

That is, parameter 7 is legal pattern rule 1 in FIG.
The legal pattern (41,46゜3(
+) is the ratio of the total number of occurrences of legal pattern 0→ on the right side. 90 each of these legal patterns
Similar ratios between legal patterns rotated by degrees are parameters 8 and 9.10, respectively.

Similarly, parameters 11-14 are legal pattern rules 20
This is the ratio of the total number of legal patterns appearing on the left side of the equation (full) to the number of legal patterns appearing on the right side, and a total of four sets of parameters are determined, including the case where each legal pattern is rotated by 90 degrees.

Parameters 15 to 18 are the appearance ratio of legal patterns on the left and right sides of legal pattern rule 3, and parameters 19 to 22
are the appearance ratios of legal two-turns on the left and right sides of legal pattern rule 4, and parameters 23 to 26 are the appearance ratios of legal two-turns on the left and right sides of legal no-no-turn rule 5.

Parameters 27 and 28 are parameters related to the ratio between the straight legal pattern and the diagonal legal pattern.

Parameter 27 is the appearance ratio of legal patterns in group B and legal patterns in group D in FIG. The parameter path is the appearance ratio of the legal patterns of the C group and the D group. Since deterioration in image quality may be associated with a decrease in linear patterns and an increase in diagonal line patterns, these parameters should be taken into consideration.

Parameter 29 is the appearance ratio of the legal pattern of group 2 in group D and the legal pattern of group C in group D.

In order to evaluate whether each parameter in Fig. 5 is effective as a parameter for image quality judgment (-, the ratio of the variance within an image quality class to the variance between image quality classes (F ratio) can be used. Calculate the F ratio using the following formula.

-Σ (Xr(k)-ma(k))2mr! :However, Xrj(k) is the value of the parameter of the sample of image quality class rOJth, Xr(k) is the average value of parameter 1 for all image quality classes within the image quality class, and n is the number of samples in each image quality class. , m is the number of image quality classes.

CCITT facsimile test manuscript No. 1 (English letter), No. 2 (handwritten line drawing), No. 4, (
For the four types (French type), No. 5 (Illustrated French text), we created binary images of three image quality classes: "Faint", "Good quality", and "Contamination", and 255 x
Forty 255-pixel image samples were cut out and the values of each parameter were calculated. In addition, the size of the image sample marked with "-" (255 x 255 pixels) is A4 size (
When inputting in facsimile high quality mode, 1728
x 2352 pixels).

Then, according to the cutting allowance, F is applied to each parameter for the entire 3 image quality classes and 3 combinations of 2 image quality classes.
The results of determining the ratio are shown in FIG.

The following can be seen from Figure 7.

The parameters that are effective for distinguishing between "good quality" and others ("contaminated" and faded) are parameter 1, lO ~ 1709 parameters.For more detailed evaluation (2), When we looked at the average values of parameters 1, 10.14 within the same image quality class, we found that the graph in Figure 8 became 2 (each class is further divided into -3 subclasses). In this graph, Parameter 1, 1
For 0.1.4, the average value is maximum near ``good quality'', and it is judged whether the images are ``good quality'' or not (``good quality'').
It can be seen that it is effective for separating ′).

From Fig. 7, it is determined whether it is ``contamination'' or ``fading'' (separating ``contamination'' and ``fading'') -1
- Valid parameters are parameters 4 to 10.19 to
It can be seen that there are 11 parameters of 22. That is,
These parameters have smaller values for °°pollution. Parameters 4 to 7.18 within the three image quality classes
Figure 9 shows the average value of .

Note that parameters 2 and 3 are not very effective, but they show large values when the target image becomes blurred, so they can be used to detect extremely "faint" images.Parameters 23 to 29 are also used to determine the image quality. It can hardly be used as a parameter for determination.

As mentioned above, parameters 7 to 10 (legal pattern rule 1
) and parameters 19 to 22 (legal pattern rule 4) are °
Effective in determining 'contamination' and 'fading', parameters 11 to 14 (legal pattern rule 2) and parameters 15 to
The reason why 18 (legal pattern rule 3) is effective in distinguishing between high-quality patterns and others can be considered as follows.

As can be seen from Figure 2, the legal pattern θ→ included in legal pattern rule 1 and its rotation pattern, and the legal pattern (27) of legal pattern rule 40 and its rotation pattern become extremely small as the image becomes blurred. Become. on the other hand,
The frequency of appearance of these legal patterns does not change much between ``contaminated'' and ``good quality.''

Therefore, parameters 7-10 and 19-22 are useful for distinguishing between "contamination" and "fading".

On the other hand, legal pattern rule 2 legal ; turn [η
Figure 2 shows that when the image quality of a legal pattern and its rotation pattern deteriorates and becomes dirty or blurred, the number of legal patterns increases dramatically compared to other legal patterns. Kararekaru. Therefore, parameters 11 to 18 are "good quality"', "contamination"' and °'
It becomes effective in ``distinguishing from ``fading.'' 2.

Next, the order of determining image quality will be described. Most of the parameters mentioned above have extreme values near "good quality". From this, "contamination", "fading", "good quality"
It is thought that it is more advantageous to perform the image quality determination in two stages than to classify and determine the three image quality classes of '' at once without prioritizing them. That is, for an image whose image quality should be judged, it is first judged whether the image is of ``good quality'' (first judgment), and then the image judged not to be of ``good quality'' is checked for ``contamination'' or ``fading''. This is a two-stage determination, in which it is determined which one is (second determination).

Next, in order to clarify the effectiveness of image quality determination according to the present invention, the results of an image quality determination experiment will be described.

The system for the experiment had a cycle time of 650 ns,
It was created using FOR and TRAN on a Panafacom computer system with a main memory capacity of 64 bytes. The procedure for image quality judgment in this system is as follows: First, for a sample of a binary line figure image to be judged, parameter 1 (legal pattern ratio), parameters 11 to 14 (legal pattern ratio 2),
), parameters 15 to 18 (legal pattern rule 3) are used to make a first determination as to whether or not the pattern is of good quality.
If it is judged as "good quality", the judgment ends. ``Good quality''
For image samples that are determined to be not, parameter 2 (proportion within B group), parameter 3 (proportion within C group), parameter 4 (all black pattern frequency ranking), parameters 5 and 6 (W pattern continuity), Parameters 7 to 10 (legal pattern rule 1), parameter 1
A second determination is made using rules 9 to 22 (legal pattern rule 4) to determine whether it is "contamination" or "fading".

The first and second judgments are made using a method in which the image sample (weighted average of the evaluation values for a pair of individual parameters is compared with a threshold value (the weight is determined based on the F ratio).Evaluation of the above individual parameters In order to carry out
facsimile test 1 Imperial manuscript No. 0.1 ~ No. 5
The three classes of Nitsuite, t+kasushi''', "a a", and pa pollution are each further divided into three subclasses,
A total of nine types of images are prepared in advance, and 40 samples from each image (one sample is 255 x 255
Pixels) were extracted, statistics were taken for parameters 1 to 29, and the average value of the parameters for each class was taken to form a standard pattern. At the same time, the value of the variance within each class was also calculated for each ζ parameter.

and unknown (the system tries to determine the image quality)
The method for evaluating each parameter for an image sample is to evaluate the possibility of belonging to a certain class as a fuzzy value based on the average distance of Euclidean distance and Mahalanabis distance to a standard pattern.

In addition, in the experimental system, the second determination is "'contamination".
When the distinction between ``good quality'' and ``contamination'' is unclear, it is considered ``rejection''.
Although a judgment error between "Paper contamination" and "Painting" does not cause much damage, the later image quality improvement process is completely opposite for "Paper contamination" and "Fading", so This is because we believe that a judgment error between the two is unacceptable.

A total of 165 samples, 145 samples from the image used to create the standard pattern described in -1- and 20 samples from other images (each sample is 255 x 255 pixels), were extracted and input into the experimental system to judge the image quality. The results are shown in Figure 10. The correct answer was determined by visual judgment. The time required to judge the image quality of one image sample was approximately 80 seconds. In Figure 10, the experimental system is
Of the 54 samples that were determined to be of good quality, one sample resulted in a judgment error (visually "faint"). All 43 samples that were determined to be "contaminated" by the experimental system were also determined to be "contaminated" by visual inspection. Of the 36 samples that the experimental system judged as "faint", 4 samples had a judgment error (visual inspection: "Pacific contamination").As described above, although some judgment errors occurred, the overall evaluation was sufficient. Good judgment performance was obtained.

It should be noted that the four samples that were erroneously determined to be blurred and blurred contained extremely small proportions of original characters and lines, and were images that were difficult to determine even by visual inspection.

The present invention has been described in detail above, and provides an effective means for determining the image quality of binary line graphic images, and its effects are extremely large.

[Brief explanation of the drawing]

Figure 1 is a diagram showing an example of legal patterns, Figure 2 is a histogram showing the appearance frequency distribution of legal patterns at different image qualities, Figures 3 and 4 are diagrams explaining grouping of legal patterns, Figure 5 is a diagram showing an example of parameters for image quality determination, the sixth tooth is a diagram showing legal pattern rules, and Figure 7 is a diagram showing the F ratio values of each parameter determined to evaluate the image quality determination parameters. , FIG. 8 and FIG. 9 are graphs showing the relationship between specific parameters and image quality, and FIG. 10 is a diagram showing the results of image quality determination by the experimental system. Agent Patent Attorney Makoto Suzuki 1898-163075 (4th Sen Figure 6 Procedural Amendment (voluntary) Revenue stamp amount 0 yen July 29, 1981 1, Table of cases 1982- 45298 No. 2, Invention
Name Image Quality Judgment Method for Binary Line Graphic Images 3, Relation to the Amendment Case Applicant 4, Agent 8151 Address 2-38-12 Yoyogi, Shibuya-ku, Tokyo
Line Department Building 201 Name (7376) Patent Attorney Makoto Suzuki Telephone 03 (374) 9671
No. 5 Number of inventions increased by amendment None 6, Specification and drawings subject to amendment 7 Contents of amendment (1) r from page 6, line 16 to line 17 of the specification 172
8 The description of X23521 is l'-1,728X 2,3
Corrected to 52 J. (2) Add "10 e.g." after "W pattern ga" on page 10, line 15 of the specification. (3) "Baramake 7" on page 11, line 18 of the specification is 1
Correct it to "Parameter 10". (4) “Parameter 8.9.1” on page 12, line 3 of the specification
0" is corrected to "Parameter 7.8.9". (5) Add 1 "Fig. 4" before "■Group" on the 4th line of page 13 of the specification and before "■Group" on the 5th line of the same page. (6) In the fourth line from the bottom of page 13 of the specification, change "-<n3% in the head of household image quality class" to 1. (k) is corrected to "within image quality class r." (Replace “1o~17” on page 14, line 18 of the specification with “1”
1 to 18”. (8) Parameter 1.10.14 on page 14, line 19 and page 15, line 3 of the specification
．． 11.15]. (9) “determined by” in line 8 of page 15 of the specification is “determined by”
Correct to. (10) Delete the statement [that is, the values of these parameters are smaller in "contamination"] from lines 10 to 12 on page 15 of the specification. (11) Page 13 of the specification Change "18" in the line to "I9"
Correct to. (12) After “bian 5” on page 18, line 19 of the specification, “(
Add the description ``Excluding 3 <)''. (13) Correct "smaller by 1" from line 16 to line 17 on page 20 of the specification, "difficult to judge visually" to "smaller." (14) Correct each of FIG. 1, FIG. 2, FIG. 4, FIG. 5, FIG. 8, and FIG. 9 as shown in the attached sheet. 8. Inventory correction drawings for accompanying documents 1. Japanese Patent Application Publication No. 1983
-1 to 3075 (13)

Claims (1)

[Scope of Claims] (1) A binary line figure image characterized in that the image quality of the binary line figure image is determined by observing a legal pattern of unit meshes on the binary line figure image. Image quality judgment method. (2) The image quality of the binary line graphic image is determined by observing a legal pattern of unit meshes on the binary line graphic image and calculating parameters related to the frequency of appearance of a specific legal pattern. An image quality determination method for binary line graphic images. (31) The method for determining the image quality of a binary line graphic image according to claim 2, wherein the parameters include a ratio of appearance frequencies between a specific legal pattern or a group of legal patterns. (4) The image quality of the binary line graphic image according to claim 2, wherein the parameters include a parameter related to the frequency at which a predetermined number of specific legal patterns appear consecutively. Judgment method. A method for determining the image quality of a binary line graphic image according to claim 1.