JP2004104318A - Image processing method, image processing unit, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To find a binary threshold for obtaining a satisfactory binary image from a grey level image. <P>SOLUTION: A likelihood reference value and total dispersion are found from a gray level histogram obtained from the density value of each pixel in the grey level image and a gradation conversion density value where a real density value in which the density value of each pixel in the grey level image is converted is converted for processing further. A normalization separation degree normalized to 0 to 1 is found from the likelihood reference value and the total dispersion. A normalization binary evaluation value is found from the normalization separation degree and the total dispersion, where the normalization binary evaluation value is a binary evaluation value for judging an appropriate binary threshold. It is judged that the normalization binary evaluation value found to each value of a gradation conversion parameter is appropriate. The binary threshold value as the normalization binary evaluation value in the optimum value of the gradation conversion parameter is set to be an optimum binary threshold. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention provides an image processing method and apparatus for selecting a binarization threshold for binarizing a multi-valued gray-scale image as a pre-processing when performing character recognition, or as an image compression processing, and It relates to an image processing program.
[0002]
[Prior art]
When character recognition is performed after binarizing a character observed as a grayscale image, or when binarization is performed to compress a large number of grayscale images, a high-speed binarization algorithm is required. As a method of performing the binarization at high speed, a method using a density histogram is effective. That is, for all pixels of the image, a density histogram is obtained by counting the number of pixels for each density value, and a binarization evaluation value when binarizing each density value is calculated on the basis of the density histogram, and the optimal histogram is calculated. The density value giving the binarization evaluation value is used as the binarization threshold. Accordingly, the binarized evaluation value calculation processing is proportional to the number of density levels, and therefore requires a calculation amount significantly smaller than the number of pixels of the image.
[0003]
Among methods using a density histogram, a method called a maximum likelihood method using a likelihood criterion as a binary evaluation value is known to be effective. The maximum likelihood method has been used for a long time and is a mathematical method for finding the optimum value of the likelihood criterion. This method can be applied to various situations, depending on the target. Here, the maximum likelihood method is applied to the problem of selecting a binary threshold. The likelihood criterion is also a mathematical function that has been used in statistics for a long time, and a plurality of likelihood criterion can be formulated by different assumptions regarding the histogram.
[0004]
As an example of the likelihood criterion for selecting a binarization threshold, it is assumed that two normal distributions are superimposed on a density histogram of a grayscale image. Furthermore, there are an occurrence probability and a variance as parameters expressing each normal distribution, and the following four types of assumptions can be divided from the relationship between the occurrence probability and the variance of two normal distributions. (1) Assuming that occurrence probabilities are different and variances are equal. (2) When the probability of occurrence and the variance are assumed to be equal. (This is called Otsu's method because he devised it.) (3) When it is assumed that the occurrence probability and variance are different. (4) When it is assumed that occurrence probabilities are equal and variances are different. (This is called Kurita's method because he devised it.)
[0005]
As a conventional technique, a binarization threshold is selected based on the likelihood criterion formulated by Kurita's method assuming that the relationship between the occurrence probability and the variance of two normal distributions is equal and the variances are different. There is one method (see Non-Patent Documents 1 and 2).
This method will be described with reference to the block diagram of FIG. 2 while defining symbols to be used in FIGS. 5, 6, and 7.
First, the image 6 is input and the gray image data 102 is obtained by the image input processing 101 of FIG.
[0006]
In FIG. 5, the number of density levels of the grayscale image data is a number Lg = imax−imin + 1 (Equation (501) in FIG. 5) from the minimum integer density value imin (FIG. 5) to the maximum integer density value imax (FIG. 5). A density value smaller than imin and a density value larger than imax by 1 are added to this, and the number of processing density levels is set to L = Lg + 2 in equation (502) in FIG. The real number density value u (j) is obtained by subjecting the density value to affine conversion so that the integer density values imin-1 to imax + 1 become 0 to 1. The reason why the density value is converted to a real number from 0 to 1 is to facilitate the gradation conversion performed later. Even if the density value is subjected to the affine conversion, the relative position of the optimum binarization threshold value before and after the affine conversion does not change. It uses the property that there is no. The density histogram p (j) is obtained by dividing the frequency h (i) of a pixel having an integer density value i by the total number of pixels N, as shown in Expression (505) in FIG. ), The sum of p (j) is 1. In FIG. 2, a density histogram 104 and a real number density value 105 are obtained from the grayscale image data 102 by a density histogram creation process 103. This corresponds to the density histogram p (j) (Equation (505)) and the real number density value u (j) (Equation (503)) in FIG.
[0007]
Next, assuming that the gradation conversion parameter is g, the real number density value u (j) is converted to a gradation conversion density value v (g; j) by equation (508) in FIG. This corresponds to obtaining a gradation conversion density value 107 from the real number density value 105 in FIG. Equation (508) in FIG. 5 shows that when j = 0, v (g; 0) = u (0) = 0, and when j = L−1, v (g; L−1) = u (L−1). = 1, and the density values of j = 1 to L-2 on the way change depending on the parameter g. g is a positive real number and can take a value from around zero (0) to infinity (∞), but in practice, it is sufficient to take a range of about 0.45 to 2.23 in 0.01 steps. It is. When g = 1, no gradation conversion is performed, and u and v have the same value for all j. The gamma conversion of Expression (509) is used as a normal tone conversion function. However, when Expression (509) is used, two types of images are obtained: an image in which the density value of the grayscale image data is inverted, and an image that is not inverted. The quantification result will be different. Since it is unnatural that the binarization result is different even though it is simply inverted, the function for performing the gradation conversion is such that even if the pixel value is inverted, the binarization result is the same. A function that maintains symmetry is preferred. Any function may be used as long as the function maintains the symmetry. Here, equation (508) shown in FIG. 5 is used as an example.
[0008]
Next, a statistic when the provisional threshold is set to the kth is defined. Here, binarization using the k-th density value v (g; k) as a threshold value means that a density value equal to or less than v (g; k) falls into class 0 and exceeds v (g; k). This means that the density value is set to class 1. When the provisional threshold value is the kth and the density values are divided into two classes, the occurrence probability, the density average value, the total density average value, the variance, and the intra-class variance, FIG. 6 defines the statistics of the inter-variance and the total variance. Although each statistic in FIG. 6 changes with the parameters g and k, only the total density average value and the total variance change only with the parameter g, and do not change with k.
[0009]
Next, using the statistics of FIG. 6, a likelihood criterion based on the binarization evaluation value is defined in FIG. Kurita's likelihood criterion formulated assuming that two normal distributions, which are one of the likelihood criterion, have the same probability of occurrence and different variances, is given by equation (701). Equation (702) of FIG. 7 is obtained by adding the total variance of Equation (611) in FIG. 6 so that it becomes an invariant function of the density affine transformation, and this is called Kurita's degree of separation. The upper limit of this equation (702) is 1, but the lower limit is unknown, and can be 0 or less. For comparison, Otsu's likelihood criterion formulated assuming that both the occurrence probability and variance of two normal distributions are equal is shown in equation (703), and Otsu's degree of separation is shown in equation (704). Otsu's degree of separation has an upper limit of 1 and a lower limit of 0, and is a scale normalized to 0-1. The likelihood criterion, the degree of separation, and the binarization evaluation value described later are all defined such that the larger the evaluation value, the higher the evaluation as the binarization threshold. Depending on the definition method, the smaller the binary evaluation value, the higher the evaluation. The above process corresponds to obtaining a likelihood standard value 109 and a total variance 110 from the density histogram 104 and the gradation conversion density value 107 in FIG.
[0010]
The degree of separation defined above has the property that it becomes larger and approaches 1 as the gradation conversion parameter g approaches either 0 or ∞ or both depending on the form of the density histogram of the grayscale image data. There is. This results in that it is optimal to set the parameter g to 0 or an extreme value of ∞. That is, no matter what the grayscale image data is, an extreme and abnormal state of the parameter g is adopted, and it is not only meaningless to introduce the parameter g but also a good binarization threshold. Means that it cannot be obtained.
[0011]
Therefore, the degree of separation is corrected using the total variance (Equation (611) in FIG. 6). However, for the Otsu separation degree for comparison (Equation (704) in FIG. 7), the gradation conversion parameter g is fixed at g = 1, and the gradation conversion is not performed and the separation degree is not corrected. . This is because Otsu's degree of separation is formulated assuming that the occurrence probability and variance are also equal, so even if gradation conversion or correction of the degree of separation is performed, the performance as a binary evaluation value does not improve. It is.
[0012]
The number L of processing density levels is not the number Lg of density levels from the minimum to the maximum existing in the grayscale image data, but is a number obtained by adding an extra density value to each of the maximum and the minimum. By doing so, the frequency of the density value at both ends for performing the grayscale conversion is always 0, and as the grayscale conversion parameter g approaches 0 or ∞, the portion where the frequency of the density histogram is 1 or more is pushed to one side. Go. By using the calculation formula in FIG. 5 (508), the density histogram is pushed to the density value 1 side as g becomes smaller than 1, and the density histogram is pushed to the density value 0 side as g becomes larger than 1. Also, the total variance (Equation (611) in FIG. 6) approaches 0.
[0013]
For example, the number of processing density levels is L = 11, and the real number density value (Equation (503) in FIG. 5) is 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6. , 0.7, 0.8, 0.9, and 1.0, and the frequency (Equation (504) in FIG. 5) of the pixel having the density value is 0, 2, 5, 1, 2, 4, 5, 4, Consider the case of 3, 1, 0. FIG. 3 shows the shape of the density histogram when the gradation conversion parameter is g = 1. On the other hand, the shape of the density histogram when g = 0.5 is as shown in FIG. 4. By making g a value smaller than 1, the whole is shifted rightward (in the direction of the density value 1) as compared with FIG. Being pushed away.
By making use of this property that the total variance approaches 0, a function that combines the degree of separation and the total variance is created, so that the binary conversion evaluation value can be obtained before the gradation conversion parameter g becomes 0 or an extreme value of 値. Can be corrected so as to find the maximum value of.
[0014]
The function of the total variance is shown in FIG. 7 (705), and Kurita's corrected separation is obtained by combining Kurita's separation (see FIG. 7 (702)) and FIG. 7 (705). I do. This is the binarized evaluation value. The function of the total variance (Equation (705)) may be any function as long as it holds the magnitude relation of the total variance. Here, the total variance itself (Equation (706) in FIG. 7) is used as the function of the total variance. , The square root of the total variance (Equation (707) in FIG. 7). Kurita's corrected separation degree may be any formula as long as it holds Kurita's separation degree and the magnitude relationship of the total variance. Here, the multiplication of Kurita's degree of separation and Equation (706) or Equation (707), which is a function of the total variance, is defined as Kurita's corrected degree of separation. The Kurita's corrected separation (Equation (709), Equation (710)) is a binarized evaluation value according to the related art.
[0015]
The density value of an image, which is originally an analog value, is quantized by passing through an image input process, is taken in as an integer value, and is quantized, thereby causing a quantization error. When the number L of processing density levels is large, the effect of the quantization error is small, but when the image data is text, the number L of processing density levels may be small. At this time, the influence of the quantization error appears greatly. Would. Therefore, it is necessary to consider this quantization error. Considering the quantization error, an increase in the variance shown in equation (712) of FIG. 7 occurs.
[0016]
The Kurita's likelihood criterion when using the increase in variance and considering the quantization error is shown in equation (713) in FIG. 7, the Kurita's degree of separation is shown in equation (714), and the Kurita's corrected degree of separation is shown in equation (714). (715). Kurita's corrected separation factor taking into account this quantization error (Equation (715)), in fact, Expressions (716) and (717) are also one of the prior arts. The Otsu likelihood criterion (Equation (703)) and the Otsu separation degree (Equation (704)) are obtained by simply adding the increase in variance (Equation (712) in FIG. 7) as a constant. It does not affect the evaluation, whether or not it is taken into account.
[0017]
From the above, Otsu's degree of separation (Equation (704) in FIG. 7) for comparison and five of Eq. (709), Eq. (710), Eq. (716) and Eq. It is a valuation evaluation value (Equation (718) in FIG. 7). In FIG. 2, a binarized evaluation value 122 is obtained from a likelihood reference value 109 and a total variance 110 by a binarized evaluation value calculation process 121 using one of the binarized evaluation values. Corresponding.
[0018]
For the provisional threshold number k, a binary evaluation value 122 for k = 0 to L-2 is calculated for each g value, and this binary evaluation value (equation (718) in FIG. 7) Find the maximum value of and k at that time. For the parameter g, starting from 1, the binarization evaluation value (equation (718)) is calculated alternately for g larger than 1 and g smaller than 1 in increments of 0.01. When the maximum value is found, the process ends at that point. The k-th density value at the time when the local maximum value is taken is the optimum binary threshold. This corresponds to obtaining the optimal binarization threshold value 114 from the binarization evaluation value 122 by the local maximum value determination process 113 in FIG.
[0019]
Here, the method of finding the maximum value instead of the maximum value is shown, but the method of finding the maximum value is because the processing speed is faster than the method of finding the maximum value because the processing is terminated midway. A similar result is obtained with the method of finding. In the method of finding the maximum value, g is changed from 0.45 to 2.23 in steps of 0.01 to find the maximum value of the binarization evaluation value, and the k-th density value at that time is the optimal binary value. Threshold.
[0020]
In the above, the binarization threshold value selection method of the related art has been described using the definitions of FIGS. 5, 6, and 7 and along the block diagram of FIG.
Next, the performance of this method will be investigated. A binary value indicating how good each density value is as a threshold for a character image data of a character database ETL1 in a web page (http://www.etl.go.jp/@etlcdb/gtbin/). There is data for evaluating the performance of the optimization algorithm. In this method, each density value of an image is classified according to the goodness as a binary threshold, and all 141 and 217 samples of the character database ETL1 are classified into excellent, good, acceptable, limited, and unacceptable. Data is given. Here, the performance of the conventional binarization threshold value selection method will be evaluated using the data of this web page.
[0021]
The threshold values obtained by the conventional binarization threshold selection method are classified into any of excellent, good, acceptable, limit, and unacceptable according to the data of this web page. it can. FIG. 8 shows the numbers obtained in the experiment. A weight of 1.00 (excellent), 1.00 (good), 0.90 (acceptable), 0.50 (limit), 0.00 (impossible) is given to each classification, and a weighted average of the number is calculated. Thus, the performance of the binarization threshold value selection method can be displayed as a numerical value. The evaluation value of FIG. 8 is obtained by calculating the weighted average for each of the conventional techniques. The larger the value, the better the binarization performance.
[0022]
In FIG. 8, among the conventional techniques, the evaluation value regarding the Otsu separation (formula (704)) investigated for comparison is 0.991714, which is the result that the performance as the binarization algorithm is the best. Kurita's four methods (Equation (709), Equation (710), Equation (716), and Equation (717)) all have lower evaluation values than the Otsu method.
[0023]
[Non-patent document 1]
Taiichi Saito: "Binarization threshold selection method based on likelihood criterion with density gradation correction", IEICE Transactions on Electronics (D-II), Vol. J80-D-II, No. 6, pp. 1343-1351 (1997-06).
[Non-patent document 2]
Taiichi Saito and Hirozo Yamada: "Creation of threshold information for binarization and performance comparison of likelihood criterion using it", "Research Report of Electronic Technology Research Institute", Vol. 62, No. 4, pp. 187-199 (1998-04).
[0024]
[Problems to be solved by the invention]
As described above, as a conventional technique, a method of calculating an optimum binarization threshold value by performing gradation conversion of image density and finding a maximum value of a binarization evaluation value obtained by combining a function of separation and total variance. However, it was shown that the binarization performance was lower than the method using Otsu's degree of separation for comparison. As a cause of this low performance, it can be considered that while the separation degree of Otsu is normalized to be a value from 0 to 1, the separation degree of Kurita is not normalized. In the method of performing density gradation conversion by changing the value of the parameter g, calculating the likelihood criterion, the degree of separation, and the binarization evaluation value using the converted density value, and finding the optimum value, The valuation value must be strictly defined. The fact that Kurita's degree of separation is not normalized from 0 to 1 means that the degree of rigor is lacking, which is considered to lower the binarization performance.
[0025]
Therefore, the present invention provides a binarization algorithm with high performance to improve the quality of a binarized image so that the binarization image recognition processing or compression processing is not adversely affected. The purpose is.
[0026]
[Means for Solving the Problems]
In order to achieve the above object, it is necessary to normalize Kurita's degree of separation to a function that changes from 0 to 1. A theoretical study of Kurita's degree of separation was performed, and as a result, a method of normalizing from 0 to 1 with a simple addition was discovered.
In FIG. 9, a normalized separation degree obtained by normalizing the separation degree of Kurita from 0 to 1 is shown in Expression (901). Kurita's corrected separability (Equation (708)) is changed to normalized separability. The corrected normalized separability is expressed by Equation (902), and Kurita's corrected separability (Equation (704) Equation (906) shows the corrected normalized separation in consideration of the quantization error in which the separation of ()) is changed to the normalized separation. Thus, the equations (903), (904), (907), and (908) in FIG. 9 are new binarized evaluation values.
[0027]
The performance evaluation of the new binarized evaluation value was evaluated in the same manner as that performed for the conventional technology using the data of the web page (http://www.etl.go.jp/@etlcdb/gtbin/). Try. The result is FIG. 10, which also shows the result of Otsu's degree of separation (formula (704)) for comparison (the same value as in FIG. 8). Among the new binarized evaluation values, the evaluation value of the corrected normalized separability (Equation (908)) taking the quantization error into consideration, taking the function of the total variance as the square root of the total variance, is 0.991839. The evaluation value of the degree of separation exceeds 0.991714, indicating that the binarization performance is improved by using the normalized degree of separation in which Kurita's degree of separation is normalized to 0 to 1. .
[0028]
According to the image processing method of the present invention, a density histogram is obtained from the density values of each pixel of the grayscale image, the density value of each pixel of the grayscale image is converted into a real number density value obtained by converting an integer to a real number, and The image data is converted into a gradation conversion density value by a gradation conversion function that changes according to the tone conversion parameter. From the density histogram and the gradation conversion density value, a likelihood standard value and a total variance formulated when it is assumed that the occurrence probabilities of the two normal distributions are equal and the variances are different are obtained. A binarization evaluation value for determining a normalized separation degree normalized to 0 to 1 from the likelihood standard value and the total variance, and determining an optimal binarization threshold from the normalized separation degree and the total variance. Is obtained. It is determined whether the normalized binarized evaluation value obtained for each value of the gradation conversion parameter is optimum, and the normalized binarized evaluation value at the optimized gradation conversion parameter value is determined. Is taken as the optimal binarization threshold.
[0029]
The image processing apparatus according to the present invention includes a density histogram obtained from density values of respective pixels of the grayscale image, in order to select a binarization threshold for binarizing the grayscale image having a multi-valued pixel value, The likelihood standard value and the total variance are obtained from a tone conversion density value obtained by further converting a real density value obtained by converting the density value of each pixel of the grayscale image from an integer to a real number with a tone conversion function that changes with a tone conversion parameter. Is calculated from the likelihood standard value and the total variance, and the binarized evaluation value is determined to obtain an optimal binarized threshold. Means for obtaining a normalized separation degree normalized to 0 to 1 from the likelihood standard value and the total variance; and a means for determining an optimal binarization threshold value from the normalized separation degree and the total variance. Means for calculating a normalized binarized evaluation value, which is a binarized evaluation value, and an optimum value determination for judging where the normalized binarized evaluation value obtained for each value of the gradation conversion parameter is optimal Means. A binarization threshold value that takes the normalized binarization evaluation value at the value of the gradation conversion parameter determined to be optimal by the optimal value determination unit is set as an optimal binarization threshold value.
[0030]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, the processing according to the present invention will be described with reference to the block diagram of FIG. The image 6 is input by the image input processing 101 to obtain the gray image data 102. A density histogram 104 (formula (505) in FIG. 5) and a real density value 105 (formula (508) in FIG. 5) are obtained from the grayscale image data 102 by the density histogram creation processing 103. By the density gradation conversion processing 106, a gradation conversion density value 107 (formula (509) in FIG. 5) is obtained from the real number density value 105. 6 is calculated by the likelihood criterion calculation processing 108 in FIG. 1, and the likelihood criterion value 109 and the total variance 110 are calculated. The normalized binarized evaluation value calculation processing 111 shown in FIG. 1 calculates the normalized binarized evaluation value 112 using a part of the equations in FIG. 6 and the equations in FIG. The normalized binarized evaluation value is a normalized value obtained by normalizing the degree of separation of Kurita formulated assuming that the occurrence probabilities of two normal distributions are equal and the variances are different so as to have a value of 0 to 1. The degree of chemical separation was used. From the obtained normalized binarized evaluation value 112, an optimal binarized threshold 114 is selected by a local maximum value judging process 113 in FIG.
[0031]
FIG. 1 is substantially the same as FIG. 2 except that FIG. 1 differs from FIG. 2 in that the binarized evaluation value calculation processing in FIG. It is in the process of calculating the chemical conversion evaluation value.
[0032]
【Example】
FIG. 11 shows a configuration of an apparatus for performing a new binary threshold selection process. Each unit of the device is connected to a bus 1, and the overall control is controlled by a control unit 2. The program describing the processing method is stored in the memory 40 from the hard disk 3, and the control unit 2 controls the processing according to the program in the memory 40. The processing of the program includes addition, subtraction, multiplication, division, function calculation, and the like.
[0033]
First, in the first process, the image 6 is input from the input device 7 and the grayscale image data 102 (FIG. 1) is stored in the memory 41. The grayscale image data already on the hard disk 3 may be stored in the memory 41 in some cases. A density histogram (Equation (505) in FIG. 5) and a real number density value (Equation (503) in FIG. 5) are created from the grayscale image data in the memory 41. Then, the tone conversion density value (Equation (508) in FIG. 5) is calculated from the formula and stored in the memory 42. A likelihood criterion, a total variance, and a normalized binarized evaluation value (Equation (909) in FIG. 9) are calculated from the density histogram and the gradation conversion density value in the memory 42 and stored in the memory 43. From the normalized binarized evaluation value in the memory 43, a binarized threshold is obtained by a local maximum value detection process. The binary threshold is stored in the memory 44 or the hard disk 3 or output to the output device 8. Alternatively, using the binarization threshold, the grayscale image data in the memory 41 is binarized, and the binary image data is stored in the memory 45 or the hard disk 3 or output to the output device 8.
[0034]
【The invention's effect】
The binarization threshold selection method according to the present invention evaluates the performance of a binarization algorithm using data of a web page (http://www.etl.go.jp/@etlcdb/gtbin/), and performs a conventional technique. It was shown to be more efficient. If the performance of the binarization algorithm is low, the binarized image may be of low quality, which adversely affects the recognition processing of the binary image or the compression processing. The higher the binarization algorithm used, the less this effect can be reduced.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a method of selecting a binarization threshold using a normalized binarization evaluation value in which a degree of separation normalized to 0 to 1 is introduced according to the present invention.
FIG. 2 is a block diagram showing a conventional technique of selecting a binarization threshold value using a binarization evaluation value based on a non-normalized degree of separation.
FIG. 3 shows real number density values of 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0. Is a diagram showing the shape of a density histogram when the frequency of a pixel having that density value is 0, 2, 5, 1, 2, 4, 5, 4, 3, 1, 0 (the gradation conversion parameter is g = 1 as in the case of gradation conversion).
FIG. 4 is a diagram showing a shape of a density histogram when a gradation conversion parameter is set to g = 0.5 with respect to FIG. 3;
FIG. 5 is a definition formula of information relating to gray-scale image data to be binarized.
FIG. 6 is a definition expression showing various statistics when binarizing with a provisional threshold.
FIG. 7 is a definition formula of a binarization evaluation value according to the related art, which is obtained from the statistics in FIG. 6;
FIG. 8 shows, for performance evaluation, a large number of grayscale image data, the number of binarization thresholds obtained by the conventional binarization threshold value selection method classified according to the binarization goodness, and the number of bins. A table displaying valuation evaluation values.
FIG. 9 is a definition formula of a normalized binarized evaluation value according to the present invention, which is obtained from the statistics in FIG.
FIG. 10 shows the number of binarization thresholds obtained by the binarization threshold value selection method of the present invention classified into good binarization values and the number of binarization values for a large amount of grayscale image data for performance evaluation. A table displaying valuation evaluation values.
FIG. 11 is a block diagram showing a configuration of an apparatus for selecting a binarization threshold value using a normalized binarization evaluation value according to the present invention.
[Explanation of symbols]
1: Bus
2: Control unit
3: Hard disk
4, 40, 41, 42, 43, 44, 45: Memory
5: Operation unit
6: Image
7: Input device
8: Output device
101: Image input processing
102: Grayscale image data
103: Density histogram creation processing
104: density histogram
105: Real number density value
106: density gradation conversion processing
107: tone conversion density value
108: Likelihood criterion calculation processing
109: likelihood standard value
110: Total dispersion
111: Normalized binary conversion evaluation value calculation processing
112: Normalized binarized evaluation value
113: Local maximum value judgment processing
114: Optimal binary threshold
121: Binary evaluation value calculation processing
122: Binary evaluation value

Claims (6)

Obtaining a density histogram from the density values of each pixel of the gray image, converting the density value of each pixel of the gray image from an integer to a real number density value, and converting the real number density value by a tone conversion parameter. A step of converting to a gradation conversion density value by a changing gradation conversion function; and formulating a case where it is assumed that the probability of occurrence of two normal distributions is equal and the variance is different from the density histogram and the gradation conversion density value. Obtaining the normalized likelihood standard value and the total variance; calculating the normalized separation degree normalized to 0 to 1 from the likelihood standard value and the total variance; Obtaining a binarized binarization evaluation value that is a binarization evaluation value for determining a binarization threshold; An optimal value judging step of judging a position where the normalized binarized evaluation value is optimal, and taking the normalized binarized evaluation value at the value of the gradation conversion parameter optimized in the optimal value judging step. An image processing method using a binarization threshold as an optimal binarization threshold. The tone conversion function is an optimal binarization obtained by the normalized binarization evaluation value, even for an inverted gray image in which each density value of the gray image is inverted, and for a gray image that is not inverted. The image processing method according to claim 1, wherein the image processing method is a gradation conversion function that maintains symmetry such that threshold values are the same. In order to select a binarization threshold for binarizing a multi-valued gray-scale image, a density histogram obtained from the density values of each pixel of the gray-scale image, and a density value of each pixel of the gray-scale image Is calculated from an integer to a real number, and a tone conversion density value further converted by a tone conversion function that changes with a tone conversion parameter, a likelihood standard value and a total variance are obtained. And the binarization evaluation value is obtained from the total variance, and in the image processing apparatus for determining the binarization evaluation value and determining the optimal binarization threshold value by performing the determination processing,
Means for obtaining a normalized separation degree normalized to 0 to 1 from the likelihood reference value and the total variance,
Means for determining a normalized binarization evaluation value that is a binarization evaluation value for determining an optimal binarization threshold from the normalized separation degree and the total variance,
Optimal value determining means for determining where the normalized binarized evaluation value obtained for each value of the gradation conversion parameter is optimal,
An image processing apparatus in which a binarization threshold value that takes the normalized binarization evaluation value at the value of the gradation conversion parameter optimized by the optimal value determination unit is an optimal binarization threshold value. The tone conversion function is an optimal binarization obtained by the normalized binarization evaluation value, even for an inverted gray image in which each density value of the gray image is inverted, and for a gray image that is not inverted. The image processing apparatus according to claim 3, wherein the image processing apparatus is a gradation conversion function that maintains symmetry such that threshold values are the same. In order to select a binarization threshold for binarizing a multi-valued gray-scale image, a density histogram obtained from the density values of each pixel of the gray-scale image, and a density value of each pixel of the gray-scale image The tone conversion density value obtained by further converting the real number density value obtained by converting the integer into a real number, obtains the likelihood standard value and the total variance, and obtains the binarized evaluation value from the likelihood standard value and the total variance. In the image processing program for determining the binarization evaluation value to determine the optimal binarization threshold,
From the likelihood standard value and the total variance, determine a normalized separation degree normalized to 0 to 1,
Obtain a normalized binarization evaluation value that is a binarization evaluation value for determining an optimal binarization threshold from the normalized separation degree and the total variance,
Determine where the normalized binarized evaluation value obtained for each value of the tone conversion parameter is optimal,
An image processing program for executing each procedure of setting a binarization threshold value that takes the normalized binarization evaluation value at the optimized gradation conversion parameter value as an optimal binarization threshold value. The tone conversion function is an optimal binarization obtained by the normalized binarization evaluation value, even for an inverted gray image in which each density value of the gray image is inverted, and for a gray image that is not inverted. The image processing program according to claim 5, wherein the image processing program is a gradation conversion function that maintains symmetry such that threshold values are the same.

Images