JPH03218590A - Optimal binarization method - Google Patents

Abstract

PURPOSE:To deal with an original printed badly by normalizing the density of an input picture by expressing the number of black picture elements at every threshold by rate on the basis of the black picture element obtained by considering the picture element excepting the picture element of the lightest level as the black picture element, and obtaining the optimal threshold from the rate of the number of the black picture elements at each threshold. CONSTITUTION:The number of the black picture elements at every threshold is counted by counting the number of the picture elements of the multi-level-quantized picture for every density, and accumulating successively the number of picture elements as changing the threshold from a dark level to a light level. In this case, when the density of the input picture is normalized by expressing the number of the black picture elements at every threshold by the rate on the basis of the number of the black picture elements obtained by considering the picture element other than the picture element of the lightest level as the black picture element, an obtained result comes to express the worn-down state of a character at that threshold to the most worn-down character. Thus, by determining the optimal threshold by consulting the normalized rate value of the character of the best state after such normalization, the original of a bad printed state can be dealt with.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an optimal binarization method in a pattern recognition device such as character recognition.

[Conventional technology]

In general, images processed in pattern recognition devices such as character recognition are obtained by converting values such as the output of a CCD image sensor of a scanner into black and white binarized values using a threshold value (threshold level). At this time, in order to enable optimal binarization even for documents with poor printing conditions, it is necessary to generate respective optimal binarization thresholds corresponding to differences in the density of the documents.

Various methods have been proposed regarding such binarization methods. For example, [Hideyuki Tamura, Souken Publishing, 1985
There are the mode method, the differential histogram method, and the p-tile method, which are shown in the document "Introduction to Computer Image Processing J, p. 67". The mode method calculates a histogram of the density values of a given image, and when the distribution has two peaks, a threshold is determined at the valley between the two peaks. In addition, in the differential His1-Durham method, the boundary between the object and the background in the image is . Since this is considered to be located in a portion where the 1 degree value changes suddenly, the threshold value is determined using a differential value (rate of change in density) rather than directly using the density value of the image. The p-tile method performs processing based on the area of the entire image.

In addition, [1972 National Conference of the Institute of Electronics and Communication Engineers, Information Section,
Nobuyuki Otsu, “Threshold Determination Method from Concentration Distribution”, 145”
There is a method for determining the threshold value from the concentration distribution, which is shown in the literature.

This method uses only the 0th and 1st moments of the concentration distribution and determines the optimal threshold based on integration.

Furthermore, there is an f-optimal binarization method disclosed in Japanese Patent Publication No. 60-37952, which stores a multilevel video signal in a video buffer and converts the video signal read from the video buffer into a variable slice level. For each of the plurality of binary video signals created by slicing the multilevel video information at different slice levels, The line width amplification factor (number of points)/(number of circumferences) is determined, and the slice level of the slice circuit is set based on the plurality of line width amplification factors and the standard line width amplification factor.

[Problem to be solved by the invention]

However, the mode method cannot be applied to documents with poor printing conditions because no clear valleys are produced in the hiss 1 to gram range. The differential histogram method does not work effectively when the density value near the boundary between the object and the background changes in a complex manner. In the p-tile method, since the area of the entire image is used as a standard, an optimal threshold value may not be obtained depending on the number of characters in the document, the size, complexity, and number of strokes of each character.

Further, the method of determining a threshold value from the density distribution is not an effective method for processing blurred or blurred "lines" as images used in pattern recognition such as character recognition.

Further, as a result of experiments, it has been found that the optimal binarization method disclosed in the above-mentioned publication is unstable in determining an appropriate threshold value depending on the density of the original.

In addition, in any case, the density of the actual original image often changes partially (the density of the image may vary due to uneven printing within a single original, shading of the input device, etc.). If the conventional method is used, it is difficult to generate an optimal binary image that corresponds to local density changes such as the orientation of the document.

[Means to solve the problem]

In the binarization method for converting a multi-level quantized image into a black and white binary image, the invention described in claim (1) provides that when the threshold value is changed from a dark level to a light level, The density of the input image is normalized by counting the number of black pixels, and expressing the number of black pixels at each threshold as a percentage based on the black pixels excluding pixels at the lightest level, and calculating from the ratio of the number of black pixels at each threshold. Find the optimal threshold.

In the invention described in claim (2), the optimal threshold value is determined using the level at which the darkest pixel begins to appear, instead of the ratio of the number of black pixels at each threshold value.

In the invention described in claim (3), the optimum threshold value is determined by determining the original density from the rate of change in the ratio of the number of black pixels at each threshold value.

In the invention described in claim (4), a multivalued image for a fixed small area is held in a memory, and the number of black pixels is counted when the threshold value is changed from a dark level to a light level within the small area. , the density of the input image is normalized by expressing the number of black pixels at each threshold as a percentage, with all pixels other than those at the lightest level as black pixels, and the optimal threshold for the small area is determined from the ratio of the number of black pixels at each threshold. is determined, an optimal threshold for the entire image is determined by changing the small region, and a binary image is output for each small region, which is binarized using the optimal threshold for that small region.

In the invention set forth in claim (5), instead of the ratio of the number of black pixels at each threshold value in the invention set forth in claim (4), the optimum threshold value is determined using a level at which the darkest pixel begins to appear.

In the invention described in claim (6), instead of the ratio of the number of black pixels at each threshold value in the invention described in claim (4), the document density is determined from the rate of change in the ratio of the number of black pixels at each threshold value. Find the optimal threshold.

In the invention described in claim (7), a part of the image is divided into a plurality of small areas, it is determined whether each small area is a character area, and the number of small areas determined to be a character area is a certain number. an optimal threshold for binarizing the entire image by the method according to claim (1), (2), or (3) for the integrated area of the small area determined to be a character area in the image portion when the threshold is exceeded; seek.

In the invention described in claim (8), claims (1), (2),
In the method described in (3) or (7), the determined optimal threshold is set in the scanner, and the binary image is directly input from the scanner.

In the invention described in claim (9), in the method described in claims In (1) to (8), the number of black pixels counted using a certain level as a threshold and the number of black pixels counted using the next level as a threshold are Determine the reference level based on the percentage,
The number of black pixels using this reference level as a threshold is used as a reference for normalizing the density of the input image instead of "black pixels other than pixels at the lightest level".

[For production]

According to the invention described in claim (1), the number of pixels of a multi-level quantized image is counted for each density level, the threshold value is changed from the dark level side to the light level side, and the number of pixels is accumulated. By doing this, the number of black pixels for each threshold value is counted. In this case, the number of black pixels, with pixels other than those at the lightest level being black, changes depending on the number of characters in the document, etc. In this regard, if we normalize the density of the input image by expressing the number of black pixels at each threshold as a percentage based on the number of black pixels with pixels other than those at the lightest level as black pixels, the obtained result is It represents the extent to which characters are blurred (or faded) at that threshold value. Therefore, after such normalization, by determining the optimal threshold value with reference to the normalized ratio value of the character in the best condition, it is possible to deal with documents with poor printing conditions.

Also, the level where the value of the normalized ratio takes a very small value,
In other words, if we focus on the level at which the image is at its faintest and is on the verge of disappearing, as in the invention set forth in claim (2), it is possible to determine the optimal threshold even when considering the level at which the darkest pixels begin to appear for the original. Seek.

Furthermore, when examining the relationship between the normalized ratio value and the concentration level, linearity is observed at a certain level. Therefore, by determining the document density from the rate of change in the ratio of the number of black pixels at each threshold value, as in the invention described in claim (3), instead of the ratio itself of the number of black pixels at each threshold value,
Furthermore, the optimal threshold value is determined.

According to the invention described in claim (4), (5), or (6), the above-mentioned processing is not performed on the entire document image at once, but is divided into fixed small areas and processed on each small area. Since the binarization process is performed by finding the optimum threshold value for each image, an optimum binary image corresponding to local density changes can be obtained.

According to the invention described in claim (7), by reducing the number of processing steps for threshold determination, it is possible to speed up the processing, and also to integrate small areas determined to be character areas, Since the processing for determining the threshold value is performed for each region, variations in threshold value determination for each region are reduced, and a good binary image can be obtained.

According to the invention described in claim (8), since the binarized image using the optimal threshold value is directly input from the scanner, the amount of data processing is smaller than the method of obtaining a binary image by processing multi-valued image data. Binary images can be obtained at high speed even with a relatively slow processing system.

Further, according to the invention described in claim (9), the density of the input image can be normalized using the lightest density level excluding the background in the document image as the reference level, so that the density of the input image can be normalized using the lightest density level excluding the background in the document image. Also, binarization is performed using an optimal threshold value, and a good recognition rate can be obtained.

〔Example〕

Example 1 ↓ An example of the invention described in claim (1) will be described based on FIGS. 1 and 2.

FIG. 2 shows a block diagram for implementing this embodiment.

This relates to the processing from the multivalued image reading section 1 to the binary image outputting section 2. Briefly, first, the multi-value image reading unit 1 reads a multi-value image (in this case, 1 image) from the scanner 3.
A six-value quantized image) is read and stored in the multi-value image memory 4. Next, concentrated hiss 1 and duram count section 5
Then, a 16-gradation multi-value image ('a degree level 0 or 615) is read from this multi-value image memory 4, and the number of pixels at each density level is counted. Then, based on the density histogram obtained in this way, the threshold calculation unit 6 calculates the characteristic value indicating the density of the image to calculate the optimal threshold, and the binarization unit 7 calculates the optimal threshold based on the optimal threshold. 2 multivalued images
Value. The binarized image is stored in a binary image memory 8.
On the other hand, it is stored in the character recognition unit 9 via the binary image output unit 2.
etc., and is used for processing such as character recognition.

Here, there is a particular feature in the processing by the threshold calculation unit 6,
This threshold calculation unit 6 includes a cumulative histogram memory 10,
Normalized histogram memory 11, optimal percentage constant value 12
and threshold table 13 are connected to each other.

In such a configuration, the processing according to this embodiment is performed in the first
An example of handling a 16-value quantized image will be explained with reference to the flowchart shown in the figure. First, 16 input from scanner 3
Value quantization Counts the number of pixels at each density level of the image. In Figure 1, "con"' indicates the concentration level, and '"Qv"
′ indicates the obtained density histogram (number of pixels).

Here, in general, the density distribution of something like a character image written on a white background is dominated by pixels that are perceived to be the lightest (these are called 0-level pixels). There are many cases. What is sensed as pixels other than level 0 are characters and their surroundings (or noise).

At this time, as before, specify the image area to be read,
According to the method of determining the threshold value by considering all pixels in the area to be the same, the threshold value for this level 0 pixel will differ depending on the format of the document, that is, whether there is a lot of white background or not. Even if the document is written in high density, the threshold value will change depending on the amount of white background.

In this regard, in this embodiment, when calculating the threshold value,
These level 0 pixels are excluded and processed as follows. That is, the threshold calculation unit 6 uses the density histogram as input data, and accumulates the number of pixels starting from the darkest pixels of level 1 or higher. In Figure 1, "s Q v
``'' indicates this cumulative value. This cumulative value is calculated for each level and stored in the cumulative histogram memory 10. This cumulative value corresponds to the number of black pixels when that level is taken as a threshold. Such pixel accumulation processing is performed from a dark pixel level to a light level of level 1. The cumulative value of level 1 is the number of black pixels of the character and its surrounding area (or noise) with respect to the above-mentioned level O. However, the cumulative value of level l is also the same as the number of characters in the manuscript or 1
It changes depending on the size, complexity, number of strokes, etc. of the characters.

Therefore, successively, the cumulative value at each level is normalized using the cumulative value of level 1 as a reference. Specifically, level 1
The cumulative value at each level is expressed as a percentage of level 1, assuming that the cumulative value is 100.

In Figure 1, 11, QvI1 indicates the percentage of the cumulative value, and the percentage "r・Rv[jl" at a certain level j is s
Calculated based on Qv[j/sQv[l].

This percentage will be referred to as the "standardized cumulative value." Then, the normalized cumulative value at each level represents the degree of blurring (or blurring) of the characters at that level relative to the most blurred character. Therefore, the optimal threshold value is determined by referring to the normalized cumulative value of the character in the best condition.

The first specific method of this determination method will be explained.

The above-mentioned normalized cumulative value, which indicates the best degree of blurring, is obtained by actually recognizing several manuscripts in advance to find the optimal threshold, and then calculating backwards from there. That is, the normalized cumulative value at each level and the optimal normalized cumulative value determined in advance (optimal percentage constant value 12 in Figure 1).
This method compares the two, finds the level that minimizes the difference, and uses that level as the optimal threshold.

If this optimal normalized cumulative value is 70%, then the first
In the example of the table, the value of density level 5 is closest to 70%, so the optimal threshold value is 5.

Table 1 Example of normalized cumulative value Also, a second specific method of the above determination method will be explained.

Some scanners 3 can output both multivalued images and binary images;
If you want to convert into a value, you can use the first specific method, but if you decide on the binarization threshold and then read the binary image again, the first specific method will not yield the optimal threshold. There may be cases where this is not possible.

Therefore, in the second specific method, first find the level at which the normalized cumulative value closest to the optimal regularized cumulative value is obtained, and then normalize between this level and the next closest level. The density level at which the cumulative value is closest to the optimal normalized cumulative value is calculated and approximated to the first decimal place. The first above
According to the example in the table, the level closest to 70% is 72.86.
% level 5, and the next closest level is 63.30
Since it is level 6 of %, the value obtained by dividing the difference by 10 (
72.86-63.30)/io=0.956, the following calculation is performed.

63.30+0.956 =62.256 ・
・Level 5.96 3.3 0+0.9 5 6 x
2=6 2.212 ・Level 5.863.3
0+0.956X7=69.992 ...Level 5.363.30+0.956X8=70.948
...Level 5,263.30+0.956x9=7
1.904...Level 5.1 Then, from the levels approximated to the first decimal place, select the level that is closest to the optimal normalized cumulative value. In this example, when determining the level closest to the optimal normalized cumulative value of 70%, level 5.
It becomes 3.

The optimum threshold value is determined by referring to a table (threshold value table 13 in FIG. 2) for determining the optimum threshold value from this level.

Note that the flowchart in FIG. 1 shows this second specific method.

According to this embodiment, even in the case of a document with poor print quality, such as a document printed by a wire dot printer, the optimal binarization threshold can be automatically set, and the most A good recognition rate can be obtained.

Embodiment 2 An embodiment of the invention described in claim (2) will be described with reference to FIGS. 3 and 4. Components that are the same as those described in FIGS. 1 and 2 are designated by the same reference numerals, and their description will be omitted. In this embodiment, instead of the optimum percentage constant value 12, the threshold calculation unit 6A is provided with a maximum level percentage 14.

In such a configuration, as in the case of the first embodiment, the histogram is accumulated from the dark level to the level 1, and the histogram cumulative value is normalized.
The cumulative value at each level is expressed as a percentage of level 1, assuming that the cumulative value is 100. In this case as well, the normalized cumulative value at each level represents the degree of blurring (or blurring) of the characters at that level relative to the most blurred character.

In this embodiment, the threshold value is determined by focusing on the level at which the normalized cumulative value takes a very small value, that is, the level at which the image is the faintest and on the verge of disappearing. This is the level at which the darkest pixels begin to appear for the original.

The contents of this process will be explained using a first specific example in this embodiment. Now, let us assume that this normalized cumulative value (the darkest level 100%*14 in FIG. 3) is set to 5%. The level at which the difference between the normalized cumulative value and this normalized cumulative value of 5% is the smallest is determined.

According to the example in Table 1 described above, density level 11 with a normalized cumulative value of 5.58 corresponds. If the relationship between this density level and the optimum threshold value of the document is determined in advance, the optimum threshold value can be determined from the density level determined in this way.

Also, a second specific method of the determination method according to this embodiment will be explained (FIG. 4 shows this method). Depending on the scanner 3, the span at each level may differ, so in order to pursue higher accuracy, find the level whose normalized cumulative value is closest to 5%, and then set the next closest level. The density level is obtained by calculating the difference from the cumulative value divided by 10 and approximating it to the first decimal place. According to the example in Table 1 above, the closest level is level 11 with 5.58%, and the next closest level is level 12 with 1.51%, so the difference was divided by 10. value (5
．． Using 58-1.51)/io=0.407, levels 11 to 12 are calculated as follows.

1.51+0.407=l. 197 Level 11.91.51+0.407x2=2.324
Level 11.81.51+0.407x8=4.766
Level 11.21.51+0.407X9=5.
173 L/Bell 11.1 Then, from the levels approximated to the first decimal place, the level closest to the normalized cumulative value of 5% is found. In this example, the level closest to the optimal normalized cumulative value of 5% is found to be level 11.1.

The optimum threshold value is determined by referring to a table (threshold value table 13) for determining the optimum threshold value from this level.

According to this embodiment, the threshold value for optimal binarization can be automatically set even in the case of a non-K document with poor printing condition, such as a document printed by a wired printer. The best imperial skewers are t (pickled).

Embodiment 3 An embodiment of the invention recited in claim (3) will be described with reference to FIGS. 5 to 8. Components similar to those described in FIGS. 1 and 2 are denoted by the same reference numerals, and description thereof will be omitted. In this embodiment, a regression line memory 16 with a linear level 15 is provided in place of the optimum percentage constant value 11 in the IJ value calculation unit 6B.

In such a configuration, as in the case of the above-mentioned embodiment, through the accumulation processing of histograms from the deep level to level 1 and the normalization processing of the histogram accumulation value, the accumulation value of level 1 is set to 100, The cumulative value at each level is expressed as a percentage of level 1. In this case, the percentage value at each level represents the degree of blurring (or blurring) of the characters at that level relative to the most blurred character.

Now, when we examine the relationship between this percentage and the degree a degree, we get
There is some linearity at some level. When only this portion is extracted and a regression line is determined by the least squares method, the slope of the regression line, that is, the rate of change, differs depending on the density of the original, as shown in FIG. 7. That is, it can be seen that the lighter the document, the steeper the slope, and the darker the document density, the gentler the slope. By utilizing this fact, it is possible to represent the density characteristics of the original.

Therefore, in this embodiment, the relationship between the optimal threshold value of the original and the slope of the regression line as shown in FIG. This table is provided as the threshold table 13B. Then, a regression line is calculated using the linearity found between the normalized cumulative value and the concentration level (the regression line memory 16 with linear level 15 is provided for this calculation). The optimum threshold value is obtained by referring to the threshold value table 13B using the calculated slope of the regression line.

Embodiment 4 An embodiment of the invention recited in claim (4) will be described with reference to FIGS. 9 to 11. FIG. 9 shows a block diagram for implementing this embodiment.

This embodiment relates to processing up to the multi-value image reading section 21 or the 62-value image output section 22. Briefly, first, the multi-value image reading section 21 reads a predetermined fixed number of lines of a multi-value quantized image from the scanner 23 and stores it in the multi-value image memory 24 . The multivalued image for the fixed number of lines read in this way is divided into fixed small regions by the small region dividing section 25. The density hiss 1 to gram counter 26 counts the number of pixels at each density level for the divided small area image. Then, the cumulative histogram calculation unit 2
7, the number of pixels is accumulated starting from the darker pixels, the normalized cumulative value calculating section 28 calculates the normalized ratio of the density of the input image, and the threshold calculating section 29 calculates the normalized cumulative value and the normalized cumulative value. An optimal threshold value is calculated from the reference standardization cumulative value 36, and the multivalued image is binarized by the binarization unit 30 and stored in the binary image memory 11. Thereafter, processing is transferred to the adjacent small area and similar binarization is performed. After the binarization of all areas of the read multivalued image is completed, the next fixed number of lines are read and the same processing is performed. This is done for the entire κ draft to generate an optimal binary image. This binary image is then sent to a character recognition unit 32 or the like for processing such as character recognition.

Here, a density histogram memory 33 is connected to the density histogram counting section 26 and the cumulative histogram calculating section 27, and a cumulative histogram memory 34 is connected to the cumulative histogram calculating section 27 and the normalized cumulative value calculating section 28.
A normalized cumulative value memory 35 is connected to the normalized cumulative value calculation section 28 and the threshold value calculation section 29, and further. A reference normalized cumulative value memory 36 is connected to the threshold calculation unit 29 .

In such a configuration, the processing according to this embodiment is performed in the first
An example of handling a 16-value quantized image will be explained with reference to the flowchart in FIG. First, a fixed number of lines of multivalued image are read from the scanner 23, and the multivalued image memory 24 is loaded with a fixed number of lines.
to be held in The multivalued image memory 24 is divided into small areas by the small area dividing unit 2.
5, the area is divided into fixed small areas. For the divided small areas, the density histogram counting unit 26 calculates the number of pixels (first
Qv) in Figure 0 is counted and stored in the density histogram memory 33. Next, the cumulative histogram calculation unit 2 reads the number of pixels of each level from the density histogram memory 33.
7, accumulate the number of pixels starting from the dark level pixels, calculate the cumulative value (sQv in Figure 10) for each level,
The cumulative histogram memory 34 holds the cumulative histogram. This cumulative value for each level corresponds to the number of black pixels when that level (1 in FIG. 10) is taken as a threshold. Such pixel accumulation processing is performed from a dark level to a light level (level 1). The cumulative number of level 1 is the black pixels of characters and their surroundings (or noise) for level O, but this varies depending on the number of characters in the manuscript or the size, complexity, number of strokes, etc. of each character. It turns out. For this reason, next, the cumulative value at each level is normalized using the cumulative value of level 1 as a reference. That is, the cumulative value is read from the cumulative histogram memory 34, and the normalized cumulative value calculating section 8 calculates the ratio of the cumulative value at each level to the cumulative value up to level 1. The reason for limiting the number to level 1 is to consider excluding level 0 pixels, which correspond to the white background area and are overwhelmingly numerous in the distribution of a document image. Specifically, the cumulative value of level 1 is set to 100, and the cumulative value of each level is expressed as a percentage of level 1. Figure 10 11J. “r fl v ” indicates the percentage of the normalized cumulative value, and the percentage at a certain level j “rQv[j]
''' is calculated based on SQv [jl /s Qv [1 piece.

This percentage will be referred to as a "normalized cumulative value." Then, the normalized cumulative value at each level represents the degree of blurring (or blurring) of the character at that level relative to the most blurred character.

Next, in the threshold calculation unit 29, the normalized cumulative value memory 35
The normalized cumulative value is read from , and the normalized cumulative value for the character in the best state determined in advance (the content of the reference normalized cumulative value memory 36 is set to the optimal percentage constant value 12 in the first embodiment). The level that takes the value closest to (equivalent to) is determined as the optimal threshold.

Next, the binarization unit 30 binarizes the multivalued image of the small area using the optimal threshold value obtained as described above, and stores it in the binary image memory 31. Regarding the adjacent small area,
Perform similar processing. Then, when all the read multi-value images have been binarized, the multi-value iiji for the next fixed number of lines is
The i-image is read into the multivalued image memory 24. Similar processing is performed on this multivalued image as well. Do this for the entire manuscript.

According to this embodiment, even if the document is printed in poor quality, such as a document printed by a wire dot printer,
The optimal binarization threshold can be automatically set,
The best recognition rate will be obtained. In particular, this kind of processing is not performed on the entire original image at once, but rather on dividing it into fixed small areas and binarizing it by finding the optimal threshold for each small area, so local density changes can be avoided. An optimal binary image corresponding to the above can be obtained.

FIG. 11 shows a conceptual diagram of the local optimal binarization processing of this embodiment described above. As shown in the figure, the input image is divided vertically and horizontally into a large number of fixed small areas, and binarization is performed using an optimal threshold value for each small area.

Embodiment 5 An embodiment of the invention set forth in claim (5) will be described with reference to FIGS. 12 and 13. Figure 9 and 1
The same reference numerals are used for the same parts as those explained in FIG. In this embodiment, instead of the reference normalized cumulative value memory 36 in the threshold value calculation unit 29A, the remotest level calculation unit 17
and the threshold table 18 are connected to each other.

In this embodiment, the processing up to finding the normalized cumulative value for each level for each small region of the input image is the same as in the fourth embodiment, but the reference normalization is Instead of determining the optimal threshold value with reference to the normalized cumulative value, as in Example 2, the normalized cumulative value is determined at a level where the normalized cumulative value takes a very small value, that is, the level at which the image is the faintest and about to disappear (the darkest). This method attempts to determine the optimal threshold value by paying attention to the pixel appearance density level. this is,
This is the level at which the darkest pixels begin to appear for the original.

The contents of this threshold value determination process will be explained using a specific example.

Now, let us assume that the normalized cumulative value at this level is set to 5%. The darkest level calculation unit 37 determines the level at which the difference between the regularized cumulative value at each level and the regularized cumulative value of 5% is the smallest. Here, since the span at each level differs depending on the scanner 23, in order to pursue higher accuracy, the normalization cumulative value should be set at 5%.
After determining the level closest to the level, the density level is calculated by approximating the difference from the normalized cumulative value of the next closest level by 10 to the first decimal place. The first mentioned above
According to the example in the table, the closest level is level 11 with 5.58%, and the next closest level is level 12 with 1.51%, so the difference is divided by 10 (5. 58
-1.51)/10=0.407, level 11
to level 12 are determined by the following calculation.

1.51+0.407=l. 197 Level 11.91.51+0.407X2=2.324
Level 11.81.51+0.407X8=4. ．． 76
6 Level 11.21.51+0.407x9=5
．． 173 Level 11.1 Then, among the levels approximated to the first decimal place, find the level that is closest to the normalized cumulative value of 5%. In this example, when finding the level closest to the optimal normalized cumulative value of 5%, the level is 11.1.
becomes. Such processing is performed by the darkest level calculation section 37. Then, in the threshold value calculation unit 29A, a table (IJ value table 38) for calculating the optimal threshold value from this level.
, and determine the optimal threshold value for the relevant small area. This threshold value table: 38 was determined in advance through experiments as shown in Table 2.

In this example, the optimal threshold (scanner reading level) is 7.

Table 2: Threshold Table According to this embodiment, it is possible to automatically set the optimal threshold for binarization even in the case of a document with poor printing quality, such as a document printed by a wire dot printer. , the best recognition rate is obtained.

An embodiment of the invention recited in claim (6) of the Patent Claims will be described with reference to FIGS. 14 and 15. This embodiment differs from the fourth embodiment in that a regression line calculation section 39 and a threshold table 40 are provided in the threshold calculation section 9B instead of the reference normalized cumulative value memory 36. The only difference in processing content from the above-mentioned embodiment is the processing of determining the optimal threshold value from the normalized cumulative value obtained for each small region.

The contents of this threshold value determination process are basically described in Example 3 of I1i.
However, the regression line calculation unit 39 calculates the slope of the regression line, and the threshold calculation unit 29B refers to the threshold table 40 using the calculated slope of the regression line, and calculates the optimum value for the small area. Determine the threshold.

Note that the threshold table 40 corresponds to the threshold table 13B in the third embodiment, and an example thereof is shown in Table 3.

Table 3 Threshold Table Example 7 An example of the invention described in claims (7) and (8) will be explained with reference to FIGS. 16 to 18.

First, the outline of the process will be explained with reference to FIG.

In this embodiment, a multivalued image is read for a predetermined fixed number of lines, and these partial images (a, b, c in FIG. 18, etc.)
Divide into small regions. Then, it is determined whether each small area is a character area or not, and it is checked whether the number of small areas determined to be character areas is less than or exceeds a certain value. In a partial image such as a shown in FIG. 18, the number of character areas is less than a certain value, so the next fixed number of lines of the multivalued image are read and the same processing is performed. If the number of character areas in the read partial image exceeds a certain value, as shown in FIG. decide. This determined optimal threshold is for the entire image.

In this way, when the optimal threshold value is determined at the time of reading a certain partial image, the threshold value determination process ends, and the determined threshold value is set as the reading level (binarization slice level) in the scanner for image input. , by having the scanner read the image again. A binarized image is directly input from a scanner using an optimal threshold value.

Next, this embodiment will be explained in detail. FIG. 16 is a block diagram showing the construction of this embodiment. The same reference numerals as those in FIG. 9, which is a block diagram of the fourth embodiment, represent the same parts, so a description thereof will be omitted. Said Example 4
The difference in configuration is that a character area determination unit 41 for determining whether or not the area is a character area, and a binary image reading unit 42 for reading a binary image from the scanner 23 are added. , the operation of the cumulative histogram calculation unit 27C is partially changed.

The details of the processing will be explained below. A flowchart of the process is shown in FIG.

As in the fourth embodiment, the multi-value image information reading section 21 reads the multi-value image for a fixed number of lines from the scanner 23, stores it in the multi-value image memory 24, and divides the multi-value image into the multi-value image by the small area dividing section 25. The memory 24 is divided into fixed small areas, and the density histogram counting section 26 creates a density histogram lv[con] of one small area and stores it in the density histogram memory 33.

Every time the density histogram of one small area is obtained, the character area determination unit 41 reads the number of pixels Qv[O] of the level O of the small area from the density histogram memory 33 and compares it with the determination threshold CHR, and calculates Qv[ 0]≧CHR, the small area is determined to be a character area, and the number of character areas char
Increment num. Pixels at level 0 correspond to pixels on a white background, and while there are many white background pixels in text areas, there are almost no white backgrounds in photo areas, so this type of text area determination is possible. . Cumulative histograms from 1 to 1 are applied only to small areas determined to be text areas.
1 calculation unit 27C creates a cumulative histogram sQv[i] similar to that in the fourth embodiment and stores it in the cumulative histogram memory 34.

Similar processing is repeated for all small areas of the partial image for a fixed number of lines in the multi-valued image memory 24. When the processing is completed up to the last small area, the character area determination unit 41 determines the number of character areas (char num) determination threshold value C H R
Determine TH. If the determination result is No, the multivalued image for the next fixed number of lines is read, and the same process is executed for this partial image.

If the result of this determination is YES, the cumulative histogram calculation unit 27C reads the cumulative histograms of all the small areas determined to be character areas and accumulates the cumulative values at the same level. A cumulative histogram at each level for the integrated area of the area is created and stored anew in the cumulative histogram memory 34. The cumulative value at each level means the number of black pixels in the integrated area of the character area when that level is taken as a threshold.

In this way, the integration of cumulative His 1 and Durham (sQ in Figure 17)
When the process of (integration) is completed, the normalized cumulative value calculation unit 28 performs the same normalization process as in the fourth embodiment on the obtained integrated cumulative histogram, and the obtained normalized cumulative value is normalized. stored in the cumulative value memory 35. Then, the optimum threshold value is determined by the threshold value calculation unit 29 using the same method as in the fourth embodiment.

When the optimal threshold value is determined, the binary image reading unit 42 sets this optimal threshold value as the reading level and reads the scanner 23.
to read the fJK manuscript again, and set 2 using the optimal threshold.
A value image is directly read from a scanner 23 and stored in a binary image memory 31.

Note that even if the number of readings by the scanner 23 per fixed number of lines exceeds the count value, if the optimum threshold value cannot be determined, the binary image is read by the scanner 23 at a fixed reading level.

According to this embodiment, the number of processing parts for multivalued images is reduced compared to the fourth embodiment, etc., so that processing speed can be increased. Furthermore, since the optimal threshold value is determined based on an area that is a combination of small areas as character areas, there is less variation in threshold value determination for each area, and a high-quality binary image can be obtained.

Another embodiment of the invention described in claims (7) and (8) of the Shishi-Sei-Sametan patent will be described with reference to FIGS. 19 and 20.

As is clear from a comparison between FIG. 19 and FIG. 16, the configuration of this embodiment is almost the same as that of the seventh embodiment, and the difference is that the portion related to threshold determination is the same as that of the fifth embodiment. That has changed.

Regarding the processing contents, as is clear from the flowchart in FIG.
The processing for determining the threshold value is the same as that in the fifth embodiment.

Implementation Example 9 Another example of the invention described in claims (7) and (8) will be described with reference to FIGS. 21 and 22.

As is clear from a comparison between FIG. 21 and FIG. 16, the configuration of this embodiment is almost the same as that of the seventh embodiment, and the difference is that the portion related to threshold determination is the same as that of the sixth embodiment. That has changed.

Regarding the processing contents, as is clear from the flowchart in FIG.
The processing for determining the threshold value is the same as that in the sixth embodiment.

Implementation jj! jlO- An embodiment described in claim (9) of the claims is
This will be explained with reference to the drawings and FIG. 24.

As is clear from the comparison between FIG. 23 and FIG. 9, the configuration of this embodiment is almost the same as that of the fourth embodiment. To,
A norm level calculation unit 43 for adaptively determining a reference level for normalizing cumulative values is added.

Regarding the processing contents, as is clear from the flowchart 1 in Fig. 24, histogram generation etc. are performed without performing region segmentation, and the base C for normalizing the cumulative value is used.
! A part for determining the level has been added, and the cumulative value is normalized using the reference level determined here.

First, the process of determining the reference level will be explained.

The basic q level is defined as the lightest density level in a document image excluding the background. This is because in a document image, the number of pixels corresponding to the background is overwhelmingly large among the number of pixels for each level, so it is impossible to judge areas where the number of pixels changes drastically as the background. be.

Therefore, in the reference level calculation section 43, each level j
The value (cumulative value) of the cumulative histogram at sQv[j] and the cumulative value sQv[i+l] of the level (j+1) above it
are sequentially read from the cumulative histogram memory 34 from the lowest level, sAv[j+l/s Qv[jl
≧pth determination processing is performed. Then, the level j that first satisfies this determination condition is set as a reference level and set in the normalized cumulative value calculation unit 28.

The threshold value pth for this determination is normally set at 0.75.
In the case of Pth=0.75 and the cumulative histogram shown in Table 4, the determination condition is first met at j=2, so the reference level is level 2.

As is clear from the flowchart of FIG. 24, the normalized cumulative value calculation unit 28 normalizes the cumulative value at each level using the cumulative value at the reference level set by the reference level calculation unit 43 as a standard. I do.

Therefore, in the case of cumulative hysteresis of 11 grams shown in Table 4, if the reference level=2, the normalized cumulative value is obtained as shown in Table 4.

As described above, in this embodiment, the reference level for normalization is not fixedly set to level 1, but is determined adaptively according to the background density level of the original image. It is possible to obtain a binary image using an optimal threshold value even for a document, and a high recognition rate can be achieved.

Table 4 Examples of Normalized Cumulative Values Embodiments ↓−↓ Another embodiment of the invention set forth in claim (9) will be described with reference to FIGS. 25 and 26.

As is clear from a comparison between FIG. 25 and FIG. 23, the configuration of this embodiment is almost the same as that of the tenth embodiment, and the difference is that the threshold value determination part has been changed to the same structure as the fifth embodiment. It is.

Regarding the processing contents, as is clear from the flowchart of FIG. 26, the only difference from the tenth embodiment is that the threshold value is determined in the same manner as in the fifth embodiment.

Embodiment 12 Another embodiment described in claim (9) will be described with reference to FIGS. 27 and 28.

As is clear from a comparison between FIG. 27 and FIG. 23, the configuration of this embodiment is almost the same as that of the tenth embodiment, and the difference is that the threshold determination part has been changed to the same structure as the sixth embodiment. It is.

Regarding the processing contents, as is clear from the flowchart of FIG. 28, the only difference from the tenth embodiment is that the threshold value is determined in the same manner as in the sixth embodiment.

Note that a configuration in which the above embodiments are combined is also possible. For example, in the embodiments 1, 2, 3, to, 11, or 12, the binary image may be input directly from the scanner as in embodiments 7, 8, or 9. Moreover, the above-mentioned Example 1
to 9, the reference level for normalization may be determined in the same manner as in Example 10.11 or 12. [Effects of the Invention] As explained above, according to the invention of claim (1) or (4), the number of black pixels is counted when the threshold value is changed from the dark level to the light level, and the pixel at the lightest level is counted. The density of the human image is normalized by expressing the number of black pixels at each threshold as a percentage, with all other pixels as black pixels. The optimal threshold value is determined from the ratio of the number of black pixels at each threshold value, and according to the invention described in claim (2) or (5), instead of the ratio of the number of black pixels at each threshold value, the level at which the darkest pixel begins to appear is determined. According to the invention described in claim (3) or (6), the document vagueness is determined from the rate of change in the ratio of the number of black pixels at each threshold value to determine the optimal threshold value.

Therefore, it is possible to automatically set the optimum binarization threshold value even for a document whose printing condition is not good, such as a document printed by a wire dot printer, and to obtain the best recognition rate.

According to the invention as set forth in claim (4), (5) or (6), the original image is divided into small regions, the optimal threshold value is determined for each small region, and the respective binarization processing is performed. It is possible to obtain the optimal binary image that corresponds to local density changes, such as when there is uneven printing in the middle of the paper, or when the image density value changes due to shading of the input device, etc. The recognition rate will be obtained.

According to the invention set forth in claim (7), the optimal threshold value for the entire image is determined by processing the integrated region of the small regions determined to be character regions within the image portion, so that the amount of processing for determining the threshold value is reduced. By doing so, it is possible to speed up the processing, and since the small areas determined to be character areas are handled in an integrated manner, there is less variation in threshold value determination for each area, and a good binary image can be obtained.

According to the invention set forth in claim (8), the determined optimal threshold value is set in the scanner, and the binary image is directly input from the scanner, so even a relatively slow processing system can produce a good binary image at high speed. can get.

According to the invention described in claim (9), the background in a document image is removed based on the ratio between the number of black pixels counted using a certain level as a threshold and the number of black pixels counted using the next level as a threshold. Since the density of the input image can be normalized using the lightest density level as the reference level, the optimal threshold value can be determined and binarized even for original images with background noise, resulting in good recognition. rate will be obtained.

[Brief explanation of drawings]

1 and 2 are a flowchart and a block diagram showing an embodiment of the invention as claimed in claim (1), respectively, and FIGS. 3 and 4 show an embodiment of the invention as claimed in claim (2), respectively. 5 and 6 are block diagrams and flowcharts showing one embodiment of the invention as claimed in claim (3), respectively. FIG. FIG. 8 is a characteristic diagram showing the correlation between the slope of the circular line and the optimum level, and FIGS. 9 and 10 are block diagrams each showing an embodiment of the invention as claimed in claim (4). and a flowchart, FIG. 11 is a conceptual diagram of the local optimal binarization method, and FIGS. 16 and 17 are block diagrams and flowcharts showing an embodiment of the invention as claimed in claims (7) and (8), respectively, FIG. 18 is a conceptual diagram of the optimal binarization method, and FIGS. 19 and 20 are The figures are a block diagram and a flowchart showing a second embodiment of the invention as claimed in claims (7) and (8), respectively, and FIGS. 21 and 22 show a second embodiment of the invention as claimed in claims (7) and (8), respectively. A block diagram and a flowchart showing the third embodiment, FIGS. 23 and 24 are respectively a block diagram and a flowchart showing an embodiment according to claim (9), and FIGS.
The figures are a block diagram and a flowchart showing a second embodiment of the invention as claimed in claim (9), FIG. 27, and a second embodiment, respectively.
FIG. 8 is a block diagram and a flowchart, respectively, showing a third embodiment of the invention as claimed in claim (9). 1. Multivalued image reading section, 2... Binary image outputting section,
3...Scanner, 4...Multi-value image memory, 5...
- Resignation histogram counting unit, 6... Threshold calculation unit, 7... Binarization unit, 8...2
Value image memory, 9...Character recognition section, 10...Cumulative histogram memo IJ, 11...Normalized histogram memory, 12...Optimum 6
Constant fraction value, 13...Threshold value table. *7 Zukaku U1 8th Figure 9Q Kabuto (aZ+t [To Shiito L/9r-1 I0 Shin) Kari 0
-l Figure 11 [Ge/T1I7 Agony 2J 恍259β Attractive Ward j IS Figure C4 Ko≦%21! rc Saijif;! ! L4”ij 14th
Figure mouth: Go23 Kakakuhari 1' Mouth Sago32 Figure 16 Figure 17 Figure 19 1st figure $23 figure r23 One animal draw = Ox Royo? = Figure 27 Entrance = 132 Figure 3 σ2 = D C = i Procedural Amendment (Method) December 18, 1989

Claims (9)

[Claims] (1) In a binarization method that converts a multilevel quantized image into a black and white binary image, the number of black pixels is counted when the threshold value is changed from a dark level to a light level, and the pixels at the lightest level are Optimal 2 is characterized in that the density of the input image is normalized by expressing the number of black pixels at each threshold value as a percentage, with all other pixels as black pixels, and the optimal threshold value is determined from the ratio of the number of black pixels at each threshold value. Value method. (2) In a binarization method that converts a multilevel quantized image into a black and white binary image, the number of black pixels is counted when the threshold is changed from a dark level to a light level, and the pixels at the lightest level are Optimization is characterized in that the density of the input image is normalized by expressing the number of black pixels at each threshold as a percentage, with all other pixels as black pixels, and the optimal threshold is determined using the level at which the darkest pixels begin to appear. Binarization method. (3) In a binarization method that converts a multilevel quantized image into a black and white binary image, the number of black pixels is counted when the threshold value is changed from a dark level to a light level, and the pixels at the lightest level are The density of the input image is normalized by expressing the number of black pixels at each threshold as a percentage, with all other pixels as black pixels, and the document density is determined from the rate of change in the number of black pixels at each threshold to determine the optimal threshold. An optimal binarization method characterized by determining. (4) In a binarization method that converts a multi-value quantized image into a black and white binary image, the multi-value image for a fixed small area is held in memory, and the threshold value is set to a dark level within the small area. Count the number of black pixels when changing from to light level, and normalize the density of the input image by expressing the number of black pixels at each threshold as a percentage, with pixels other than the lightest level as black pixels, The optimum threshold value for the small area is determined from the ratio of the number of black pixels at each threshold value, the optimum threshold value for the whole image is obtained by changing the small area, and each small area is binarized using the optimum threshold value for that small area. An optimal binarization method characterized by outputting a value image. (5) In a binarization method that converts a multi-value quantized image into a black and white binary image, the multi-value image for a fixed small area is held in memory, and the threshold value is set to a dark level within the small area. Count the number of black pixels when changing from to light level, and normalize the density of the input image by expressing the number of black pixels at each threshold as a percentage, with pixels other than the lightest level as black pixels, Using the level at which the darkest pixel begins to appear, the optimum threshold value for the small area is determined, the small area is changed to find the optimum threshold value for the entire image, and each small area is binarized using the optimum threshold value for that small area. An optimal binarization method characterized by outputting a binary image. (6) In a binarization method that converts a multi-value quantized image into a black and white binary image, the multi-value image for a fixed small area is held in memory, and the threshold value is set to a dark level within the small area. The density of the input image is normalized by counting the number of black pixels when changing from to light level, and expressing the black pixels at each threshold as a percentage based on the black pixels other than the lightest level pixel. The document density is determined from the rate of change in the proportion of the number of black pixels in the threshold value, and the optimal threshold value for the small area is determined.The optimal threshold value for the entire image is determined by changing the small area. An optimal binarization method characterized by outputting a binary image binarized using an optimal threshold value. (7) Divide the image into multiple small areas, determine whether each small area is a text area, and when the number of small areas determined to be text areas exceeds a certain number, A claim (
An optimal binarization method characterized by determining an optimal threshold for binarizing an entire image by the method described in 1), (2), or (3). (8) Claim characterized in that the determined optimal threshold value is set in the scanner, and the binary image is directly input from the scanner.
The optimal binarization method described in 1), (2) or (7). (9) In the optimal binarization method according to claims (1) to (8), based on the ratio of the number of black pixels counted using a certain level as a threshold to the number of black pixels counted using the next level as a threshold. The method is characterized in that a reference level is determined, and the number of black pixels using this reference level as a threshold is used as a reference for normalizing the density of an input image, instead of using pixels other than those at the lightest level as black pixels. Optimal binarization method.