JPH0765164A - Method for generating template - Google Patents

Abstract

PURPOSE:To generate the template which can exert larger and larger negative influence on a matching rate as a pixel value is larger and larger by generating values at respective points on the template with specific functions. CONSTITUTION:A mean template generating means 12 calculates the mean value of plural template data to generate mean template data. A histogram generating means 14 generates a density histogram of the mean template data. A function generating means 18 generates a function which gives values corresponding to respective sections. Then a template generating means 20 generates the respective points on the template in a section from the minimum value of the mean template data to a specific reference value by a function which gives negative values for a decrease from the reference value to the maximum value side and in a section from the reference value to the maximum value with a function which gives positive values for an increase from the reference value to the maximum value side.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention, in a pattern matching process for extracting a pattern similar to a recognition target pattern from an image using a template, is similar to an image recognition target input under the same condition as the image to be processed. By cutting out an arbitrary number of patterns of the specified size,
The present invention relates to a template generation method capable of extracting the recognition target pattern.

[0002]

2. Description of the Related Art Conventionally, for example, in the matching by normalized cross-correlation described in Makoto Nagao, "Digital Image Processing", Modern Science, p.306-312, the image pattern to be detected is used as it is as a template. Then, the degree of matching (matching rate) is calculated by the following calculation formula (1).

[0003]

[Equation 1]

Here, g is a template, f is an image to be matched, (u, v) are coordinates on the template, and (i
, J) are coordinates on the matching target image, and Xt and Yt are template sizes in the u direction and v direction, respectively.

The numerator f (i + u, j + v) of the above formula (1)
The sum of g (u, v) with respect to u and v is a convolution operation that is frequently used in image processing for the purpose of edge extraction, noise removal, etc., and requires a large amount of calculation, but it is executed at high speed. For example, Japanese Patent Laid-Open No. 3-13196
A large number of them have been created, including "5. Two-dimensional contraction array for neural network and its method".

Further, in Japanese Patent Laid-Open No. 4-283881, each pixel of an image is ranked according to a brightness value, run-length coding is performed for each ranked image, and a template is created. Calculate the correlation value,
Seeking a degree of matching.

[0007]

However, the above (1) shown in the above-mentioned "digital image processing" is used.
In the denominator of the equation, the sum of g (u, v) ² in the latter half with respect to u and v is a constant, but f (i + u, j + v) ^{2 in the} first half
The sum of u and v of requires the same amount of calculation as convolution for each (i, j), so this execution requires a large amount of calculation time,
Or, there is a problem that a device is required.

Further, in Japanese Patent Laid-Open No. 4-283881, when pixels of an image are ranked by luminance value and run length coding is performed for each ranked image to form a template, the template is classified by the difference in luminance rank. However, the difference between the pixel value in the template and the pixel value of the matching target pixel can have the same negative influence regardless of the size of the pixel value of, and it is correctly reflected in the matching rate. There is a problem that a complicated procedure is required instead of a simple multiply-add operation.

The present invention has been made to solve the above-mentioned conventional problems. In the recognition target pattern, when there is a large pixel value at a position where a small pixel value should be taken, matching is performed as the pixel value becomes larger. An object of the present invention is to provide a template generation method capable of generating a template that can have a large negative influence on the rate.

[0010]

SUMMARY OF THE INVENTION The present invention relates to a template generation method in pattern matching for extracting a pattern similar to a target recognition target pattern from an input image, and a plurality of recognition target patterns included in the input image. A plurality of masks having an arbitrary size including only one of them are set so that the positions of the recognition target patterns with respect to the masks are all the same in each of the plurality of masks. The plurality of parts enclosed by each mask in the inside are cut out as template data, and the arithmetic mean or geometric mean of the values at each coordinate similarly defined in each template data is calculated, and is taken as the value at each coordinate. Average template data is generated, and a histogram of the density values of all pixels in the average template data is generated. A suitable threshold value that clearly separates the average template data or the pattern to be extracted when binarizing the entire surface of the image or the background by calculating the maximum value and the minimum value from the histogram, or the average value. The arithmetic mean or geometric mean of the template data or the entire surface of the image is calculated, and this is used as a reference value, and the value on each coordinate of the average template data is as described above for the section from the minimum value to the reference value. A negative value is given so that it becomes smaller as it goes from the reference value to the minimum value side, and when it is equal to the reference value, it is 0, and further, for the section from the reference value to the maximum value,
The object is achieved by generating the value of each point on the template by a function that gives a positive value so that the value becomes larger from the reference value toward the maximum value side.

According to the present invention, after the sections are set to arbitrary numbers v and w of 0 or more and 100 or less respectively, the high density side and the low density side of the histogram are removed by v% and w%, respectively. The maximum value of the concentration value of the upper limit value, the minimum value as the lower limit value, the interval from the lower limit value to the reference value, and the interval from the reference value to the upper limit value, to achieve the same object It was done.

[0012]

The above-mentioned conventional example is described in "Digital Image Processing" p.
As indicated by 306 to 312, the numerical value of the matching rate has the property of correctly reflecting the geometric distance between the template and the matching target image. this is,
Since the value of each pixel in the template is different only by superimposing, when the value of the target pixel in the matching target image is different from the template, the effect on the matching rate will be mixed, and the term of the denominator corrects it. This is because it has an effect.

The reason why the value reflecting the similarity between the template and the matching target image cannot be obtained only by superimposing without the correction term is as follows. That is, for a pixel having a large value in the template, when the value of the corresponding pixel in the matching target image is large, that is, similar, the matching rate is large, and when the value of the corresponding pixel is small, that is, not similar. Sometimes the matching rate becomes smaller. On the other hand, with respect to the pixel having a small value in the template, it is similar to the fact that the value of the corresponding pixel in the matching image is small, but this has an adverse effect on the matching rate.

Further, since the absolute value is small, the degree of influence is small, and the degree of influence on the matching rate varies for each pixel in the template.

Therefore, according to the present invention, the image pattern to be detected is not used as it is in the template, but in correspondence with this, for a pixel having a certain fixed value, for example, an average value, a negative value is obtained in the template. By giving the value of, the matching rate becomes small when the matching target image has a large value in the pixel which should have a small value in the image pattern to be detected.

In addition, the absolute value of the pixel in the template, each pixel in the image pattern to be detected, and the absolute value of the difference between the constant values are given a monotonically increasing relationship to affect the matching rate. It was decided to adjust the degree of.

Further, in this adjustment, not all the densities in the image pattern to be detected are set, but the upper limit value and the lower limit value are set, so that the brightness or the darkness is saturated and an error occurs. This variation can be eliminated in the concentration range where only variation appears. This makes it possible to obtain a matching rate that reflects the degree of similarity between the detection target pattern and the matching target image, as in the case of normalized cross-correlation.

According to the present invention, any number of patterns similar to the detected pattern are cut out from the image as template data, and the average of the pixel values on the same coordinates of each template data is calculated to obtain the average template data.

Here, an appropriate threshold value that clearly separates the detection pattern and the background when binarizing the average template data or the entire image, or a value obtained by calculating the average value of the average template data or the entire image is used as a reference value. Ask.

From the reference value and the average template data, the value on each coordinate of the average template data becomes smaller in the section from the minimum value of the average template data to the reference value as going from the reference value to the minimum value side. As described above, depending on the function that gives a negative value, and when it is equal to the reference value, it is set to 0, and for the section of the maximum value of the average template data from the reference value, it becomes larger as going from the reference value to the maximum value side. By generating the value of each point on the template by the function that gives a positive value,
When there is a large pixel value at a position where a small pixel value should be taken in the recognition target pattern, it is possible to generate a template that can have a large negative influence on the matching rate as the pixel value increases.

[0021]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the flow of the basic processing of the automatic template weight generation method according to the present invention.

In FIG. 1, 10 is a template data cutting means by a mask, 12 is an average template data generating means, 14 is an average template data histogram generating means, 16 is a reference value setting means, 18 is a function generating means, and 20 is a function. It is a template generation means.

In the template data cutting means (10) using a mask, as shown in FIG. 2, n × p is used as a mask of an arbitrary size surrounding a target pattern portion.
(N and p are arbitrary positive integers) are set, and then the target pattern is cut out from the image with a mask having the size described above to obtain template data. Here, the size is n × n. Here, it is assumed that m pieces of the template data can be cut out. In addition, as shown in FIG. 2, the upper left of each template data is (1, 1),
Set coordinates such that the lower right corner is (n, n).

Average template data generating means (12)
In the above, from the m template data, as shown in FIG. 3, from each value on the same coordinate, the arithmetic mean is obtained from the following equation (2), or as shown in FIG. Then, an average value such as a geometric mean is calculated to generate average template data.

[0026]

[Equation 2]

[0027] ^{_{m √ (N1 jk · N2 jk}} ···· Nn jk) ... (3)

Next, the average template data histogram generating means (14) generates a density value histogram of the average template data as shown in FIG.

In the reference value setting means (16),
Set the reference value. The reference value is, for example, a threshold value obtained by a discriminant analysis method, a p-tile method, a mode method when binarizing the average template data or the entire surface of the image, or the average template data or the phase of the entire surface of the image. In addition, the geometric mean value is used. As an option, v and w% (v, w ≧ 0) are set on the high density value side and the low density value side, respectively, from the density value histogram, and they are removed. High density side v of the histogram
%, The maximum density value after removing w% on the low density value side is the upper limit value, and the minimum value is the lower limit value.

From this, for example, the interval between the lower limit value (V _d ) and the reference value (V _s ) and the interval between the reference value and the upper limit value (V _u ) are divided into p sections (in FIG. 5). Is divided into 6), and the function generating means (18) considers a function that gives a value corresponding to each section.

The width on the low density side of the reference value is H _d ,
If the width on the high density side is H _u , then H _d = (V _s −V _d ) / p (4) _Hu = (V _u −V _s ) / p (5), and the average template data If the value of V _j is V _j , the template value W _j is i) when V _j <V _d , Wj = −120 (6) ii) when V _d ≦ V _j <V _s , P _i is P _i = V _d + H _d × i (7) (where i = 0, 1, ..., P), if P _i <V _j <P _{i + 1} , then W _j −120 + i × 20. (8) iii) When V _j = V _s , W _j = 0 (9) iv) When V _s <V _j ≤V _u , Q _j is Q _j = V _s + H _u × i (10) (However, i = 0,1, ···, p ) when the <if _{Vj ≦ Qj + 1, W j} = i × 20 ... (11) v) V j> Qj time of V _u, W _j = 120 (12) and the template can be generated.

The template generation as described above is performed by the template generation means (20), and the template as shown in FIG. 6 is generated.

[0033]

As described above, according to the present invention, when there is a large pixel value at a position where a small pixel value should be taken in the pattern to be recognized, the larger the pixel value, the greater the negative influence on the matching rate. This has the effect of generating a template that can be used.

[Brief description of drawings]

FIG. 1 is a block diagram showing a template generation procedure in the present embodiment.

FIG. 2 is an explanatory diagram showing cutout of template data according to the present embodiment.

FIG. 3 is an explanatory diagram showing generation of average template data by arithmetic averaging in the present embodiment.

FIG. 4 is an explanatory diagram showing generation of average template data by geometric mean in the present embodiment.

FIG. 5 is a diagram showing a density value histogram of average template data in the present embodiment.

FIG. 6 is a diagram showing an example of template generation in this embodiment.

[Explanation of symbols]

 10 ... Template data cutting means by mask 12 ... Average template data generating means 14 ... Average template data histogram generating means 16 ... Reference value setting means 18 ... Function generating means 20 ... Template generating means

Claims (2)

[Claims] 1. A method for generating a template in pattern matching for extracting a pattern similar to a target recognition target pattern from an input image, wherein the template is generated for each of several recognition target patterns included in the input image. A plurality of masks having an arbitrary size including only one are set so that the positions of the recognition target pattern with respect to the masks are all the same in each of the plurality of masks, and the masks are surrounded by each mask in the image. Cut out a plurality of parts as template data, calculate the arithmetic mean or geometric mean of the values at each coordinate defined similarly in each template data, and generate the average template data having the value at each coordinate, Generate a histogram of the density values of all pixels in the average template data, The maximum and minimum values are calculated from the program, and the average template data or the entire surface of the image is calculated as 2
When binarizing, an appropriate threshold value that clearly separates the pattern to be extracted from the background, or the average template data or the arithmetic mean or geometric mean of the entire surface of the image is calculated, and this is used as a reference value, and the average template Regarding the section from the minimum value to the reference value, the value on each coordinate of the data is given a negative value so as to become smaller as going from the reference value to the minimum value side, and the reference value Is equal to 0, and for the section from the reference value to the maximum value, a function that gives a positive value that becomes larger as going from the reference value to the maximum value side A method of generating a template, which is characterized by generating a value at each point. 2. The method according to claim 1, wherein an arbitrary number v, w of 0 or more and 100 or less is set for each of the sections, and v% is removed on the high density side and w% is removed on the low density side of the histogram. The maximum value of the density value after the upper limit value, the minimum value as the lower limit value, the section from the lower limit value to the reference value, and the section from the reference value to the upper limit value of the template characterized by Generation method.