JPH0954828A - Similarity calculation device and method, and position detection device using the same - Google Patents

Abstract

(57) Abstract: Provided is a similarity calculation device capable of accurately producing a result even when calculating a similarity between a model image and an input image in which a part of the background has changed. A difference between a density gradient direction of each pixel of an input image stored in an image memory 3 and a density gradient direction of each corresponding pixel of a model previously stored in a memory 9 is obtained by a subtraction unit 20, and a direction difference evaluation unit is obtained. 21. The evaluation value is weighted by the weight obtained from the model, the product-sum calculation unit 22 adds the weighted values, and the division unit 2
In 3 the added value is the weighted sum calculated by the weighted sum calculation unit 19,
The division is performed to calculate the similarity R.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a similarity calculating apparatus and method for calculating the similarity between an input image and a model, and a position detecting apparatus using the same.

[0002]

2. Description of the Related Art Conventionally, a normalized cross-correlation has been known as a method for calculating the similarity between grayscale images that have little effect on illumination variations (Image Analysis Handbook, supervised by Mikio Takagi and Yohisa Shimoda, University of Tokyo. Publishing Association). FIG. 23 shows the configuration of a conventional image processing apparatus using this normalized cross correlation.
This image processing apparatus includes a camera 1 for capturing an object, an A / D converter 2 for converting a captured image into a digital signal, and an image memory 3 for storing a digitized image.
A D / A converter 4 for converting a digital image into an analog signal for display, a CRT display 5, an address / data bus 6, a timing controller 7, an input image capture, display, similarity calculation, etc. A CPU 8 that executes various processes and controls, a memory 9 that stores image data for similarity calculation, a similarity calculation unit 10, and a threshold determination unit 11 are provided.

As shown in FIG. 24, the similarity calculating section 10 includes a covariance calculating section 12, a standard deviation calculating section 13, and
The integrating unit 14 and the dividing unit 15 are provided. In this image processing apparatus, the analog video signal output from the camera 1 is A / D converted by the A / D converter 2 in synchronization with the timing signal from the timing control unit 7, and then the image memory 3 as a grayscale image. Memorized in. Image memory 3
The input image stored in is converted into an analog signal through the D / A converter 4, and then displayed on the CRT display 5. On the other hand, the similarity calculation unit 10 calculates the similarity between the input image stored in the image memory 3 and the model image stored in the memory 9 in advance, and stores it in the memory 9. The threshold determination unit 11 compares the similarity stored in the memory 9 with a preset threshold value, and determines OK / NG. The result is memory 9
Is stored. Passing data between modules is
This is done through the address / data bus 6. Further, the CPU 8 issues a start command for each module.

In the similarity calculator 10, the sizes of the model image and the input image are set to (mx, my), the density value of the input image is I (x, y), and the density of the model image is M (x, y).
Then, the similarity (CC) is calculated by the following formula.

[0005]

[Equation 1]

The density value I (x, y) of the input image stored in the image memory 3 and the density value M (x, y) of the model image stored in the memory 9 are similar through the address / data bus 6. It is taken into the degree calculation unit 10. In the covariance calculation unit 12, the covariance of I (x, y) and M (x, y)

[0007]

[Equation 2]

Is calculated. Also, the standard deviation calculating unit 13
At, the standard deviation of I (x, y)

[0009]

(Equation 3)

Is calculated. In the integration unit 14, I
Product of standard deviation of (x, y) and standard deviation of M (x, y)

[0011]

(Equation 4)

Is calculated. The division unit 15 calculates the normalized cross-correlation value (CC) and stores it in the memory 9 as the similarity between the model image and the input image.

[0013]

The above-mentioned conventional method for obtaining the normalized cross-correlation is a matching method that is invariable under the movement or expansion / contraction of the density gradation. However, when a linear conversion relationship does not hold between the model image and the input image such as shading or background change, the ratio of the product of each standard deviation to the covariance of the model image and the input image becomes large, There is a problem that the degree of similarity becomes small and stable pattern recognition cannot be performed. For example, in FIG.
As described above, when the similarity between the input image and the model image in which a part of the background of the model image is changed is calculated, the similarity becomes about 0.61, which is compared with the correlation value of 1.0 between the model images. In that case, it is considerably lowered and stable recognition cannot be performed.

The present invention has been made in view of the above-mentioned problems, and an apparatus and a device capable of accurately producing a result even when the similarity between an input image and a model image in which a part of the background of the model image is changed is calculated. It is intended to provide a way.

[0015]

A similarity calculation device according to claim 1 of the present application obtains a density gradient direction Iθ (x, y) in each pixel of an input image. A value f for evaluating the difference in the density gradient direction between the means and the density gradient direction Iθ (x, y) of the input image and the density gradient direction Mθ (x, y) of the corresponding pixel of the predetermined model. Concentration gradient direction evaluation value calculating means for obtaining {Iθ (x, y) -Mθ (x, y)}, and weighting for weighting the evaluation value with a weight M _W (x, y) obtained from the predetermined model. Method and weighted evaluation value

[0016]

(Equation 5)

And means for calculating Further, the similarity calculating device according to claim 2 is a density gradient direction calculating means for obtaining a density gradient direction Iθ (x, y) in each pixel of the input image, and a density gradient direction Iθ (x,
y) and the density gradient direction M of the corresponding pixel of the predetermined model
A density gradient direction evaluation value calculating means for obtaining a value f {Iθ (x, y) -Mθ (x, y)} for evaluating the difference in the density gradient direction from θ (x, y), and this evaluation value. A weighting means for weighting with a weight M _W (x, y) obtained from the predetermined model, an adding means for adding weighted evaluation values, and a division of the weighted addition value by a total weight. Degree R

[0018]

(Equation 6)

And means for calculating Further, the similarity calculation apparatus according to claim 3 is the similarity calculation apparatus according to claim 1 or 2, wherein the weight of the weighting means is
The density gradient strength of the model image is used. The similarity calculation apparatus according to claim 4 is the one according to claim 1, claim 2, or claim 3, wherein the evaluated value is the cosine of the difference in the density gradient direction.

The similarity calculation device according to claim 5 is
The value evaluated in claim 1, claim 2, or claim 3 is 1 if the difference in the concentration gradient direction is within a predetermined range including 0 °, and within a predetermined range other than that. If it is -1, if it is a value within a predetermined range other than that, it is 0
And The similarity calculation device according to claim 6 is
According to the first or second aspect of the present invention, the similarity calculation is performed only on pixels whose weight value weighted by the weighting means is a predetermined value or more.

Further, the similarity calculation device according to claim 7 is
Density gradient direction Iθ (x, y) in each region of the input image
Between the density gradient direction Iθ (x, y) of this input image and the density gradient direction Mθ (x, y) of the corresponding region of the predetermined model. Value f {Iθ (x, y) −Mθ (x,
y)} for the density gradient direction evaluation value calculation means, and a weight M _W (x,
y), a weighting means for weighting, an adding means for adding weighted evaluation values, and a similarity R by dividing the weighted addition value by the total weight.

[0022]

(Equation 7)

And means for calculating In the similarity calculation method according to claim 8, the density gradient direction Iθ (x, y) is obtained for each pixel of the input image, and the model density gradient direction Mθ (x, y) is calculated between corresponding pixels of a predetermined model.
The value obtained by evaluating the difference in y) is weighted with a weight M _W (x, y) obtained from a predetermined model, added, and divided by the total weight to calculate the similarity between two grayscale images.

[0024]

(Equation 8)

Calculate In the similarity calculation method according to the ninth aspect, the weight obtained from the model is the density gradient strength of the model image, and the similarity between the two grayscale images is calculated. The similarity calculation method according to claim 10 is the method according to claim 8, wherein the evaluated value is the cosine of the difference in the density gradient direction, and the similarity between the two grayscale images is calculated.

The similarity calculation method according to an eleventh aspect is the method according to the eighth aspect, in which the evaluated value is 1 if the difference in the density gradient direction is a value within a predetermined range including 0 °. And if the value is within the other predetermined range, −
1 and 0 if it is a value within a predetermined range other than that,
The degree of similarity between the two grayscale images is calculated. Claim 12
A position detecting device according to the above, a model image storage means for storing a model image, an input image storage means for storing an input image, a scanning means for moving the model image within the input image, and a model image in the scanning process. The similarity calculation device according to claim 1 or 2 for calculating the similarity between the position and the input image, and the position of the input image having the highest calculated similarity as the position of the model in the input image. And a position determining means for performing the position determination.

A similarity calculating apparatus according to a thirteenth aspect is the one according to the first aspect, the second aspect, or the seventh aspect, in which the density gradient direction storage for storing the density gradient direction in each pixel of a plurality of model images is stored. Means, density gradient strength storage means for storing density gradient strength in each pixel of the plurality of model images, representative density gradient direction calculation means for calculating a representative density gradient direction from the density gradient direction, the density gradient direction, and A representative concentration gradient strength calculating means for calculating the representative concentration gradient strength from the concentration gradient strength is provided.

Further, a similarity calculation method according to a fourteenth aspect is the method according to the ninth, tenth, or eleventh aspect, wherein the density gradient direction in each pixel of a plurality of model images is stored and the plurality of model images are stored. The density gradient strength in each pixel of the model image is stored, the representative density gradient direction is calculated from the density gradient direction, and the representative density gradient strength is calculated from the density gradient direction and the density gradient strength.

A similarity calculating apparatus according to a fifteenth aspect of the present invention is the apparatus according to the thirteenth aspect, which obtains an average of a plurality of images in the density gradient direction as the representative density gradient direction. A similarity detection apparatus according to a sixteenth aspect of the present invention is the apparatus according to the thirteenth aspect, which obtains, as the representative density gradient strength, means for adding density gradient strengths of a plurality of images.

The similarity calculation apparatus according to a seventeenth aspect is the one according to the thirteenth aspect, wherein the representative density gradient strength is the average of the density gradient directions of the plurality of images and the density gradient strength of the plurality of images. The arithmetic mean of the representative concentration gradient direction components of the concentration gradient strength is obtained from the representative concentration gradient direction obtained from. Further, the similarity detection device according to claim 18 obtains the length of a vector whose components are the average of sine and the average of cosine in the density gradient direction from the density gradient directions of a plurality of images, and the calculated value and the The product of the representative concentration gradient strength is obtained as a new representative concentration gradient strength.

A similarity calculating apparatus according to a nineteenth aspect of the present invention is a density gradient direction calculating means for obtaining a density gradient direction in each pixel of a plurality of images in advance, and a density for calculating a density gradient intensity in each pixel of the plurality of images. A model using a gradient strength calculating means, a representative concentration gradient direction calculating means for calculating a representative concentration gradient direction from the concentration gradient direction, and a representative concentration gradient strength calculating means for calculating a representative concentration gradient strength from the concentration gradient strength A storage unit for creating and storing data, a density gradient direction of an input image, and a density gradient strength are represented between a representative density gradient direction and a representative density gradient strength of each corresponding pixel of the model data. A representative concentration gradient evaluation value calculation means for obtaining a value for evaluating the difference between the gradient direction and the representative concentration gradient strength, and a means for calculating the similarity by adding the evaluation values are provided. And it was painting.

The similarity calculating apparatus according to a twentieth aspect of the invention is a density gradient direction calculating means for obtaining a density gradient direction in each pixel of a plurality of images in advance, and a density calculating a density gradient intensity in each pixel of the plurality of images. Model using gradient strength calculation means, representative concentration gradient direction calculation means for calculating a representative concentration gradient direction from this concentration gradient direction, and representative concentration gradient strength calculation means for calculating a representative concentration gradient strength from the concentration gradient strength A storage unit for creating and storing data, a density gradient direction of an input image, and a density gradient strength are represented between a representative density gradient direction and a representative density gradient strength of each corresponding pixel of the model data. Representative density gradient evaluation value calculation means for obtaining a value for evaluating the difference between the gradient direction and the representative density gradient strength, addition means for adding the evaluation values, and division by the number of pixels And a dividing means for calculating a degree of similarity by.

Further, claim 21, claim 22, claim 2
The similarity calculation device according to claim 3 or claim 24 has the same characteristics as claim 15, claim 16, claim 17, or claim 18 in claim 19 or claim 20, respectively. is there.

In the similarity calculating device according to the first to sixth aspects, the similarity between the two grayscale images is calculated by adding the values evaluated for the difference in the density gradient direction. The density gradient direction is characterized in that the values are the same even when the contrast of the background portion and the target portion changes. In addition, since the density gradient direction is calculated using a mask calculation, even when the background portion changes continuously,
When viewed locally, the background portion has the same density. Therefore, there is a feature that the direction of the concentration gradient is hard to change. Also, even when the target is drawn on a complicated background,
The density gradient direction becomes strange in the discontinuous portion of the background change, but since the density gradient direction is calculated locally, it does not propagate to the entire image. Therefore, it is possible to recognize marks and characters in a pattern such as shading in which the gray value of the background part changes continuously, or in a pattern such as a complex background that has a discontinuous part in the background part but has a small proportion. Will be possible.

Claim 1, Claim 2, Claim 7, and Claim 8
In the similarity calculation device according to the first aspect, since the division is performed by the number of pixels or the total weight value, the meaning of the similarity does not change even when the number of pixels or the total weight value changes, and the threshold value determination becomes easy. In the similarity calculation device and method according to claims 3 and 9, since the correlation value weighted by the density gradient intensity is used, even in a pattern such as a complicated background in which the background differs depending on the image, the mark or the character It becomes possible to recognize.

In the similarity calculating apparatus and method according to the fifth and eleventh aspects, the value obtained by evaluating the difference in the density gradient direction is 1,
Since the three values of 0 and -1 are used, there is an effect that the amount of calculation is reduced. In the similarity calculation device and method according to claims 5 and 11, the maximum value of the differences evaluated in the density gradient direction is set to 1, and the minimum value is set to -1, and the similarity calculation of claims 2 and 8 is performed. Since the apparatus and the method divide by the number of pixels or the total weight value, even when the number of pixels or the total weight value changes, the maximum similarity value becomes 1 and the minimum value becomes -1, and the same input image as the model image. Is 1
Becomes Therefore, since a plurality of models can be recognized with one threshold value, the threshold value setting becomes easy.

In the similarity calculating apparatus and method according to the fourth and tenth aspects, since the value obtained by evaluating the difference in the density gradient direction is used as the cosine of the difference in the density gradient direction, the model image and the black-and-white inverted input image are compared. The degree of similarity is -1. In the similarity detection devices of claim 15, claim 16, claim 17, claim 18, claim 19, claim 20, claim 21, claim 22, claim 23, and claim 24, the representative concentration gradient direction Since the representative density gradient strength is calculated to obtain the degree of similarity, it is possible to describe even when it is not possible to prepare a work in which marks, characters, etc. are clearly drawn on a plain surface.

[0038]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, the present invention will be described in more detail with reference to embodiments. 1 is a block diagram showing a configuration of an image processing apparatus according to a first embodiment of the present invention.
The device of this embodiment is the same as the conventional image processing device shown in FIG. 23: camera 1, A / D converter 2, image memory 3, D / A converter 4, CRT display 5, address / data bus 6, timing. Control unit 7, CPU
8, a memory 9, and a threshold value determination unit 11 are provided, and instead of the similarity degree calculation unit 10, a concentration gradient direction calculation unit 16, a concentration gradient strength calculation unit 17, and a similarity degree calculation unit 18 are provided.

As shown in FIG. 2, the similarity calculating section 18 includes a weight sum calculating section 19, a subtracting section 20, a direction difference evaluating section 21, a product sum calculating section 22, and a dividing section 23. ing. In the image processing apparatus according to this embodiment, the analog video signal output from the camera 1 is A / D converter 2 and is synchronized with the timing signal from the timing control unit 7 to perform A / D conversion.
After D conversion, it is stored in the image memory 3. The image captured in the image memory [(a) in FIG. 4] is converted into an analog signal through the D / A converter 4, and then the CRT is displayed.
It is displayed on the display 5. On the other hand, the image captured in the image memory 3 is converted into the input density gradient direction [(b) in FIG. 4] by the density gradient direction calculation unit 16 and the memory 9
Is stored. In the similarity calculation unit 18, the model concentration gradient strength [(b) in FIG. 3] calculated in advance by the concentration gradient strength calculation unit 17 and the concentration gradient direction calculation unit 16 and stored in the memory 9 and the model concentration gradient direction. The degree of similarity is calculated using [(c) of FIG. 3 and the input concentration gradient direction [FIG. 4 (b)]. The calculated similarity is stored in the memory 9. The threshold determination unit 11 compares the similarity stored in the memory 9 with a preset threshold to determine OK / NG. The result is stored in the memory 9. Data is passed between the modules via the address / data bus 6. Also,
The CPU 8 issues a start command for each module.

In the similarity calculating section, the size of the model concentration gradient direction, the model concentration gradient strength, and the input concentration gradient direction are set as (mx, my), and the model concentration gradient direction is Mθ (x,
y), the model concentration gradient strength is M _W (x, y), and the input concentration gradient direction is Iθ (x, y), the similarity (R) is calculated by the following formula.

[0041]

[Equation 9]

Here, the input concentration gradient direction Iθ (x, y)
And the model concentration gradient direction Mθ (x, y), the evaluation value of the difference value is the cosine of the difference in the concentration gradient direction. (Claim 6) f (ω) = cosω As an example and as an evaluated value, If the difference in the concentration gradient direction is within a range of ± 45 ° including 0 °, it is set to 1,
± (45 ° to 135 °) is 0, ± (135 °
˜180 °), it is set to −1 (claim 7).

[0043]

(Equation 10)

Is as follows. The input density gradient direction stored in the image memory 3 is Iθ (x, y), and the model density gradient strength stored in the memory 9 is M _W (x, y) and the model density gradient direction Mθ (x, y). Are taken into the similarity calculation unit 18 via the address / data bus 6. In the weight sum calculation unit 19, the total value of the model concentration gradient strength

[0045]

[Equation 11]

Is calculated. Further, in the subtraction unit 20, the difference between the input density gradient direction Iθ (x, y) and the model density gradient direction Mθ (x, y) of each pixel.

[0047]

(Equation 12)

Is calculated. In the direction difference evaluation unit 21,

[0049]

(Equation 13)

Is calculated, and in the product-sum calculation unit 22,

[0051]

[Equation 14]

Is calculated, and the division unit 23 calculates the degree of similarity according to the present embodiment. The model is created in advance by the following procedure and stored in the memory 9. A model image is taken from the camera 1 into the image memory 3 through the A / D converter 2 [(a) in FIG. 3]. The model image is processed by the density gradient direction calculation unit 16 in the model density gradient direction [FIG.
(C)], and is converted into model concentration gradient strength [(b) in FIG. 3] in the concentration gradient strength calculation unit 17.
The model concentration gradient direction and the model concentration gradient strength are stored in the memory 9
Is stored. In the concentration gradient strength calculation unit 17, the concentration gradient strength is calculated by the following formula (Sobel operator).

[0053]

(Equation 15)

An operator such as a gradient may be used as the method of calculating the concentration gradient strength. Concentration gradient direction calculation unit 16
In, the concentration gradient direction [0 °, 360 °] is calculated by the following formula (Sobel operator). Mθ (x, y) = atan2 (Dx, Dy) Here, atan2 refers to an arctangent function of Dx-Dy coordinates represented by Dx coordinates and Dy coordinates (FIG. 5). As the density gradient direction calculation method, another operator such as a Prewitt operator may be used.

The method according to the present invention is applied to the pattern of FIG. 25, which is poor in the normalized cross-correlation. However,
Since the 3 × 3 mask calculation is performed for the calculation of the density gradient strength and the density gradient direction, as shown in FIGS. 6A and 6B, there is a blank space of 1 pixel in both the model image and the input image. I shall. When the model concentration gradient direction and the input concentration gradient direction are calculated using the above method, the results are as shown in (c) and (d) of FIG. 6. Further, when the model concentration gradient strength is calculated using the above method, it becomes as shown in (e) of FIG. Here, since the label ND in the model density gradient direction is 0 for both Dx and Dy, it indicates a pixel whose direction is undefined. The evaluation of the direction difference of this pixel is f (ω) = 0. The difference between the input concentration gradient direction and the model concentration gradient direction is shown in (f) of FIG. f
When the direction difference is evaluated using (ω) = cosω and the above numerical example, the results are as shown in (g) and (h) of FIG. 7, and the background (excluding the part where the direction is undefined) (ND) is excluded. The value becomes small in the discontinuous portion of the density value. However, in other cases, it can be seen that the evaluation value is close to 1 without being affected by the contrast between the background and the character portion. The similarity when using f (ω) according to claim 6 is about 0.92, and the similarity when using f (ω) according to claim 7 is about 0.94.
And the similarity using normalized cross-correlation (about 0.61)
It can be seen that the evaluation value does not decrease even when compared with, and that the similarity calculation method according to the present invention is strong against background changes.

In addition, in the pattern having shading in the background portion (FIG. 8), the similarity due to the normalized cross-correlation is about 0.91, whereas in the similarity calculation method according to the present invention, When f (ω) according to claim 6 is used, it is about 0.98, and when f (ω) according to claim 7 is used, it is 1.0, and this method is also effective for shading. Was shown.

FIG. 9 is a block diagram showing a position detecting device according to the second embodiment of the present invention. The position detecting apparatus according to this embodiment includes the same-numbered components of the image processing apparatus shown in FIG. 1, and further includes a similarity calculating section 18 and a threshold determining section 11.
Instead of the above, the collation position detection unit 24, the weighted sum calculation unit 25, the similarity calculation unit 26, and the position detection unit 27 are provided.

As shown in FIG. 10, the similarity calculation section 26 includes a subtraction section 20, a direction difference evaluation section 21, and a sum of products calculation section 2.
2 and a division unit 23. It should be noted that the weighting sum calculated by the weighting sum calculation unit 25 is stored in the division unit 23 at the address /
It is designed to be added via the data bus 6.

In the position detecting apparatus of this embodiment, the camera 1
The analog video signal output from the A / D converter 2 is A / D converted in synchronization with the timing signal from the timing control unit 7, and then stored in the image memory 3.
The image [FIG. 12 (a)] captured in the image memory 3 is
After being converted into an analog signal through the D / A converter 4, it is displayed on the CRT display 5. On the other hand, the image captured in the image memory 3 is processed by the density gradient direction calculation unit 16
, The input density gradient direction [(b) in FIG. 12] is converted and stored in the memory 9. The matching position calculation unit 24 calculates the matching position (i, j) between the model and the input image (the upper left corner coordinate reference of the model). Similarity calculation unit 26
11, the model concentration gradient strength [(b) of FIG. 11] calculated in advance by the concentration gradient strength calculation unit 17 and the concentration gradient direction calculation unit 16 and stored in the memory 9 and the model concentration gradient direction [(of FIG. 11). c)] and the sum of the model concentration gradient intensities calculated by the weight sum calculation unit 25 and stored in the memory 9, the similarity [R (i,
j)] is calculated. The calculated similarity and its matching position are stored in the memory 9. The collation position calculation unit 24 sequentially
At (i, j), the value is changed to calculate the similarity at all possible matching positions. The matching position calculation unit 2
When the message indicating the end of search is returned in 4, the processing is transferred to the position detecting unit 27. The maximum value of the similarity stored in the memory 9 is compared with a preset threshold value by the position detection unit 27, and the OK / NG determination is made. The result is stored in the memory 9. In the case of OK, the position of the maximum value, that is, the collation position (imax, jmax) between the model and the input image is stored in the memory 9. Data transfer between the modules is performed through the address / data bus 6. Further, the CPU 8 issues a start command for each module.

FIG. 14 shows the flow of processing by the collation position calculation unit. The predetermined search range (FIG. 13) is set to (sx, s
y)-(ex, ey). As an initial value (i, j)
= (Sx-1, sy) is set. The model is scanned from the upper left to the lower right of the search range so that the model fits in the search range, and the matching position between the model and the input image is obtained. When the matching is completed up to the lower right corner, the position detection unit returns a search end determination.

In the similarity calculating section, the collation position (i,
The similarity R (i, j) in j) is calculated by the following formula.

[0062]

(Equation 16)

The calculation procedure is the same as in the first embodiment. Also in the present embodiment, the first embodiment
Similar to the above, the effect that the background change and shading can be recognized stably can be obtained. FIG. 15 is a block diagram showing the configuration of the stamp character reading device according to the third embodiment of the present invention. The stamped character reading device of this embodiment is basically the same in structure as the position detecting device of FIG.
In addition to the components having the same numbers as those of the position detecting device, the position detecting unit 27 is replaced by a character reading unit 28.

In this embodiment, the marking character reading device,
The analog video signal output from the camera 1 is A / D converted by the A / D converter 2 in synchronization with the timing signal from the timing control 7, and then stored in the image memory 3. The image [(a) in FIG. 17] captured in the image memory 3 is converted into an analog signal by the D / A converter 4 and then displayed on the CRT display 5.
On the other hand, the image captured in the image memory 3 is converted into the input density gradient direction [(b) of FIG. 17] by the density gradient direction calculation unit 16 and stored in the memory 9. The matching position calculation unit 24 calculates the matching position (i, j) between the model and the input image (the upper left corner coordinate reference of the model).

For each character [(a) in FIG. 16], the model density gradient strength [(b) in FIG. 16] and the model density gradient direction [(c) in FIG. 16] are sent to the density gradient strength calculation unit 17. Also, it is calculated in advance by the concentration gradient direction calculation unit 16 and stored in the memory 9. In addition, the total value of the model concentration gradient strength is calculated by the weight sum calculation unit 25 for each character and stored in the memory 9. The similarity calculation unit 26 uses the total value of the model concentration gradient strength [(b) of FIG. 16], the model concentration gradient direction [(c) of FIG. 16] and the model concentration gradient strength to calculate the similarity [R (i , J. Mode
l)] is calculated. The calculated similarity [R (i, j.
model)], its matching position (i, j) and model name (model) are stored in the memory 9. The values of (i, j) are sequentially changed, and the similarities at all possible matching positions are calculated. Do this for every character you read.

Using the similarity stored in the memory 9,
The character reading unit 28 reads characters. Among the similarities to all models at a certain matching position (i, j), the model showing the maximum is model-max. The similarity R (i, j, model-max)
Is greater than or equal to the threshold value set in advance and stored in the memory 9, the model (model-
It is determined that there is a character corresponding to (max), and the character and position are stored in the memory 9. Do this for all possible (i, j).

Data is transferred between the modules (components) through the address / data bus 6. Further, the CPU 8 issues a start command for each module. The process of the matching position calculation unit is similar to that of the second embodiment. In the similarity calculation unit, a model (mo
del) at the matching position (i, j) (i, j.m)
odel) is calculated by the following formula.

[0068]

[Equation 17]

The calculation procedure is the same as in the first embodiment. In the case of engraved characters [(a) in FIG. 18],
When the density gradient direction is viewed, there is a possibility that the direction may be shifted by 180 ° even at the same position depending on the direction of illumination [(f) and (g) in FIG. 19]. This is because the position of the shadow varies depending on how the light hits [(b), (c) of FIG. 18], and the appearance of the image changes [(d) of FIG. 18,
(E)]. To deal with this, the function f of the similarity equation is
If (ω) is f (ω) = cos (2ω), the direction is 1
Even when they are deviated by 80 °, f (ω) = 1 and the directions can be handled as the same. That is, f (θ)
= F (θ + 180).

For example, when the degree of similarity with an image having a different illumination direction [(d) of FIG. 21] is obtained using the model image of FIG. 20 (a), the model density gradient direction [(c of FIG. 21)] is obtained. )] And the input density gradient direction [(e) of FIG. 21] have an angle difference of 180 ° depending on the direction of illumination [(f) of FIG. 22]. Therefore, f (ω) in the expression of similarity is f (ω) =
By setting cos (2ω), the value of f (ω) becomes 1.
It becomes 0 [(h) in FIG. 22]. Therefore, when the similarity is calculated, it becomes 1.0, and there is no decrease in the similarity. On the other hand, when the similarity between the model image [(a) in FIG. 20] and the input image [(d) in FIG. 21] is calculated using the conventional method, it is −0.8.
4, the absolute value was 0.84, indicating the superiority of the method according to the present invention.

In each device such as the similarity detection device of the embodiment shown in FIGS. 1 to 25, a work in which an object such as a mark or a character is clearly drawn on a plain image is picked up and imaged at each pixel position. The concentration gradient direction and concentration gradient intensity of are registered as models. However, here, as an example, let us consider a work in which marks printed on a transparent sheet can overlap various positions on a background pattern as shown in FIG. Images of these works are images 0 to 4 in FIG. Image 0
The mark is overlaid on the plain fabric. The density gradient direction and density gradient strength of each image are as shown in FIGS. 28 and 29, for example. In the density gradient direction of FIGS. 28 and 29, the blackened portions mean that the direction is undefined.

The equation for calculating the similarity is

[0073]

(Equation 18)

It is assumed that Here, the size in the model concentration gradient direction, the model concentration gradient strength, and the input concentration gradient direction is mx
Xmy, the model concentration gradient direction was Mθ (x, y), the model concentration gradient strength was M _w (x, y), and the input concentration gradient direction was Iθ (x, y). The function f that evaluates the direction difference is

[0075]

[Equation 19]

It is assumed that When a model is created from a work in which marks are overlapped on a plain surface as in image 0, the correlation value R is about 0.93 even in the case of a work in which marks are overlapped on the background pattern as in image 1. This is quite close to the correlation value of 1.0 between model images, and stable recognition can be performed.

However, when it is not possible to prepare a work in which the target is drawn on a solid color, if a model created from the work in which the target is drawn on the background pattern is used to perform matching, the similarity will be the same as that of the conventional model. However, there is a problem in that the recognition accuracy becomes poorer than that of the above. For example, if a model is created from image 4, the similarity with image 1 is 0.7.
It becomes as small as 4 and stable recognition cannot be performed.

Due to the above-mentioned problems, it is understood that sufficient performance cannot be obtained when a work on which an object is drawn cannot be prepared. In the following, by having a means for creating one model from images obtained by capturing a plurality of workpieces, an embodiment similarity degree with which sufficient performance can be obtained even if a workpiece on which an object is drawn cannot be prepared. The calculation device will be described.

FIG. 30 is a block diagram of a similarity calculation device according to the fourth embodiment of the present invention. Further, FIG. 31 shows a block diagram of the position detection device of the fifth embodiment. 30 shows the similarity calculation device of FIG. 1, FIG. 31 shows the position detection device of FIG. 9, and the representative concentration gradient direction calculation unit 29 and the representative concentration gradient strength calculation unit 3 are shown.
0 is added. The process until the representative concentration gradient direction calculating unit 29 and the representative concentration gradient strength calculating unit 30 obtain information necessary for calculating the representative concentration gradient direction and the representative concentration gradient strength will be described below.

Images of a work in which objects are drawn on various background patterns are picked up by the camera 1. The output analog video signal is A / D converted by the A / D converter 2 in synchronization with the timing signal from the timing control section 7, and then stored in the image memory 3. By the above procedure, N model images of size mx × my are stored in the image memory 3. Image i stored in the image memory 3
Position (x, y) of (1 ≦ i ≦ N) (0 ≦ x ≦ (mx−
1), the concentration gradient direction M in 0 ≦ y ≦ (my-1))
θ (i, x, y) is calculated by the concentration gradient direction calculation unit 16 as follows.
The concentration gradient strength M _w (i, x, y) is calculated by the concentration gradient strength calculator 17. Concentration gradient direction Mθ (i, x,
y) and the concentration gradient strength M _w (i, x, y) are stored in the memory 9. Representative concentration gradient direction Mθ at position (x, y)
_p (x, y) is calculated by the representative concentration gradient direction calculation unit 29 based on Mθ (i, x, y). The representative concentration gradient strength M _wp (x, y) at the position (x, y) is M
It is calculated in the representative concentration gradient strength calculation unit 30 based on θ (i, x, y) and M _w (i, x, y).

FIG. 32 shows the calculation of the representative density gradient direction and the representative density gradient strength at all positions (x, y) in the model image.
It is a flowchart which shows what is performed in (0 <= x <(my-1), 0 <= y <(my-1). First, variables x and y
Is set to 0 (ST11), and then the processing for obtaining the representative concentration gradient direction and the representative concentration gradient strength at each position (Flow chart 3
3 to 36) (ST12), the variable x is incremented by 1 (ST13), and the processes of ST12 and ST13 are repeated until this variable reaches mx. x = m
When it becomes x, the determination in ST14 becomes NO, this time y is incremented by 1, x is set to 0 (ST16), and ST is set.
Returning to step 12, the processing of ST12 and ST13 is performed again with x = mx
Repeat until When x = mx, the variable y is incremented by 1 again. Then, the processes of ST11 to ST16 are repeated until y = my-1.

FIG. 33 shows a flow chart of the representative concentration gradient direction calculation unit 29 according to claim 16. Here, first, the variable α is set to 0, the variable β is set to 0, the variable i is set to 1, and initialization is performed (S
T21), then Mθ (i, x, y) ≠ direction indeterminate? Is determined (ST22). If not indefinite, α = α + cos
Mθ (i, x, y), β = β + sinMθ (i, x,
y) is calculated (ST23). Next, the variable i is incremented by 1 (ST24), and the processes of ST22 to ST24 are repeated until the variable i exceeds N (ST25).
If the determination in ST25 is YES, α = α / N, β = β
/ N is calculated (ST26). In step ST26, the value of α is the average of the cosine of the concentration gradient direction Mθ (i, x, y), and the value of β is the average of the sine of the concentration gradient direction Mθ (i, x, y). That is,

[0083]

(Equation 20)

The value of the concentration gradient direction Mθ (i, x, y) is
When it is fixed, the determination is made in ST22, and ST
23 is skipped and the addition in ST23 is not performed.
(Α²+ Β²) Is the square root of the concentration gradient direction Mθ (i, x,
The larger the variation between y), the smaller the value.
Then, the density gradient direction Mθ in the target pixel (x, y)
It is an index of the stability of (i, x, y). Based on the values of α and β
Then, the representative concentration gradient direction Mθ _p(X, y) is

[0085]

(Equation 21)

[0086] Here, δ is a preset threshold value. atan2 (X, Y) is the X coordinate,
The arctangent function of the XY coordinate represented by the Y coordinate is shown (FIG. 5). FIG. 34 shows a flow chart of the representative concentration gradient strength calculating unit 30 according to claim 17. Here, in order to calculate the representative concentration gradient strength, first, the variable M _{1 is set} to 0 and the variable i is set to 1
(ST31), M ₁ = M ₁ + M _W (i, x, y) is calculated (ST32), the variable i is further incremented by 1 (ST33), and ST32, S until the variable i becomes N.
The process of T33 is repeated. Then, when i = N in ST34, the representative concentration gradient strength M _WP1 (x, y): (M
Calculate _WP1 (x, y) = M ₁ ÷ N). The representative concentration gradient strength M _wp1 (x, y) is the concentration gradient strength M _w (i, x, y)
The arithmetic mean of,

[0087]

(Equation 22)

Is as follows. A flow chart of the representative concentration gradient strength calculating unit 30 according to claim 18 is shown in FIG. Step ST
When NO is determined in 46, the value of M ₂ is a vector whose magnitude is 1, the direction is the representative concentration gradient direction Mθ _p (x, y), and the magnitude is the concentration gradient strength M _w (i, x, y), the sum of inner products with a vector whose direction is the concentration gradient direction Mθ (i, x, y),

[0089]

(Equation 23)

## EQU10 ## Concentration gradient direction Mθ (i, x, y)
When the value of is indefinite, the determination is made in step ST42, the processes of steps ST43 and ST44 are skipped without being performed, and the process proceeds to ST45. The representative concentration gradient strength M _wp2 (x, y) is the quotient of M ₂ and N,

[0091]

(Equation 24)

(ST47) Representative concentration gradient strength M
_wp2 (x, y) is the concentration gradient direction Mθ (i, x, y)
Is farther from the representative concentration gradient direction Mθ _p (x, y),
There is a property that the contribution of the concentration gradient strength M _w (i, x, y) becomes small. FIG. 36 shows a flowchart of the representative concentration gradient strength calculating unit 30 according to claim 19. Step ST58
In, the value of α is the average of the cosine of the concentration gradient direction Mθ (i, x, y), and the value of β is the concentration gradient direction Mθ (i, x, y).
It is the average of the sine of. That is,

[0093]

(Equation 25)

It becomes The value of M ₃ is a vector whose magnitude is 1 and whose direction is the representative concentration gradient direction Mθ _p (x, y),
The magnitude is the concentration gradient strength M _w (i, x, y) and the direction is the concentration gradient strength Mθ (i, x, y), and the average of the sum of inner products with the vector,

[0095]

(Equation 26)

## EQU10 ## Concentration gradient direction Mθ (i, x, y)
When the value of is indefinite, the determination is made in step ST52, and the processes of ST53, ST54, and ST55 are not performed. The representative concentration gradient strength M _wp3 (x, y) is M
_3, the root of the product of the density gradient direction Mθ (i, x, y) becomes the index of stability of (α ² + β ^2).

[0097]

[Equation 27]

It becomes Representative concentration gradient strength M _wp3 (x,
y) has a property that the value decreases as the concentration gradient direction Mθ (i, x, y) varies. The representative concentration gradient direction and the representative concentration gradient strength are shown in FIG.
7 is calculated from images 1 to 4 and the degree of similarity to image 1 is shown.
The threshold δ for calculating the representative concentration gradient direction Mθ _p (x, y) and the representative concentration gradient strength M _wp3 (x, y) is
It was set to 0.

From images 1 to 4, representative density gradient direction Mθ
37 and 38 _{show p} (x, y) and M _wp1 (x, y) for the representative concentration gradient strength. The similarity between this model and image 1 is calculated to be about 0.89. Image 1
FIG. 39 shows the representative concentration gradient direction Mθ _p (x, y) and the representative concentration gradient strength M _wp2 (x, y) obtained from FIG. When the similarity between this model and image 1 is calculated, it is about 0.
It becomes 89.

From images 1 to 4, representative density gradient direction Mθ
_p (x, y), comprising a representative concentration gradient intensity as M _wp3 (x, y) Figure 40 when seeking. The similarity between this model and image 1 is calculated to be about 0.97. In either case,
Since the similarity between image 0 and image 1 (approximately 0.93) is close,
Stable recognition is possible.

As described above, with the model creating method according to the present invention, a model can be created even when a work having marks or characters drawn on a plain surface cannot be prepared.

[0102]

According to the inventions of claim 1, claim 2, claim 7 and claim 8, the similarity between two images is calculated by adding the values evaluated for the difference in the density gradient direction. The density gradient direction has the same value even when the contrast of the background portion and the target portion changes. Since the calculation is performed using the mask calculation, even when the background portion changes continuously, the background portion has the same density locally, and the density gradient direction hardly changes. Even when the target is drawn on a complicated background,
The density gradient direction becomes strange at the discontinuous portion of the background change, but since the density gradient direction is calculated locally, it does not propagate to the entire image. Since there are features such as shading, in a pattern where the gray value of the background part changes continuously, such as shading, or in the background part where there is a discontinuous part such as a complicated background, in a pattern with a small proportion, This has the effect of enabling recognition of marks and characters.

Further, according to the inventions of claim 2, claim 7, and claim 8, since the division is performed by the number of pixels or the total weight value, even if the number of pixels or the total weight value changes, the similarity can be calculated. There is an effect that the meaning does not change and the threshold value is easily determined. Further, according to the inventions according to claims 3 and 9, since the correlation value weighted by the density gradient strength is used, even in a pattern such as a complicated background in which the background differs depending on the image, recognition of the mark or the character is performed. There is an effect that it becomes possible.

According to the inventions of claim 2, claim 7, and claim 8, a plurality of models can be recognized with one threshold value, so that there is an effect that threshold value setting becomes simple.
Further, according to the inventions of claims 4 and 10, the value obtained by evaluating the difference in the density gradient direction is used as the cosine of the difference in the density gradient direction. The similarity is -1.

According to the twelfth aspect of the invention, the position of the model in the input image can be accurately detected. Further, according to the inventions of claims 13 to 24,
Even if it is not possible to prepare a work in which marks, letters, etc. are clearly drawn on a plain surface, it is possible to accurately recognize the work.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an image processing apparatus according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a similarity calculation unit of the image processing apparatus according to the embodiment.

FIG. 3 is a diagram illustrating a model image in the image processing apparatus of the same embodiment.

FIG. 4 is a diagram illustrating an input image in the image processing apparatus of the same embodiment.

FIG. 5 is a diagram for explaining calculation of a density gradient direction of a model image of the image processing apparatus of the same embodiment.

FIG. 6 is a diagram illustrating an example of similarity calculation of the image processing apparatus according to the embodiment.

FIG. 7 is a diagram illustrating an example of similarity calculation of the image processing apparatus of the embodiment, together with FIG. 6.

FIG. 8 is a diagram illustrating an example of similarity calculation of the image processing apparatus according to the embodiment.

FIG. 9 is a block diagram showing a configuration of a position detection device according to another embodiment of the present invention.

FIG. 10 is a block diagram showing a configuration of a similarity calculation unit of the position detection device of the embodiment.

FIG. 11 is a diagram illustrating a model image in the position detection device of the embodiment.

FIG. 12 is a diagram illustrating an input image in the position detection device of the same embodiment.

FIG. 13 is a diagram illustrating position search in the position detection device of the embodiment.

FIG. 14 is a flowchart illustrating a position search process in the position detection device of the embodiment.

FIG. 15 is a block diagram showing a configuration of a stamp character reading device according to another embodiment of the present invention.

FIG. 16 is a diagram illustrating a model image in the marking character reading device according to the embodiment.

FIG. 17 is a diagram illustrating an input image in the marking character reading device according to the embodiment.

FIG. 18 is a diagram illustrating a difference in a read image due to a difference in illumination direction in the stamped character reading device according to the embodiment.

FIG. 19 is a view for explaining the difference in the read image due to the difference in the illumination direction in the stamped character reading device according to the same with FIG. 18.

FIG. 20 is a diagram showing an example of a stamp character collation in the stamp character reading apparatus according to the embodiment.

FIG. 21 is a diagram showing an example of stamped character collation in the stamped character reading device according to the same embodiment as FIG. 20.

FIG. 22 and FIG. 21 are views showing an example of a stamp character collation in the stamp character reading apparatus according to the embodiment.

FIG. 23 is a block diagram showing a configuration of a conventional image processing device.

FIG. 24 is a block diagram showing a configuration of a similarity calculation unit of the conventional image processing apparatus.

FIG. 25 is a diagram showing an example of a model image and an input image of the conventional image processing apparatus.

FIG. 26 is a diagram showing a case where a work covered with a transparent sheet is installed on a work having a background pattern.

FIG. 27 is a diagram showing an example of an image obtained by capturing an image of the work shown in FIG. 26.

28 is a diagram showing a density gradient direction and a density gradient strength in the case of the image 0 shown in FIG.

29 is a diagram showing the density gradient direction and density gradient strength in the case of Image 1, Image 2, Image 3, and Image 4 shown in FIG.

FIG. 30 is a block diagram showing the configuration of a similarity calculation device according to another embodiment of the present invention.

FIG. 31 is a block diagram showing the configuration of a position detecting device according to still another embodiment of the present invention.

32 is a flow chart in the case where the representative density gradient direction and the representative density gradient strength are calculated using all pixels in the model in the apparatus of the embodiment shown in FIGS. 30 and 31. FIG.

FIG. 33 is a flowchart showing a method of calculating the representative concentration gradient direction in the apparatus shown in FIGS. 30 and 31.

FIG. 34 is a flowchart showing a method of calculating the representative concentration gradient strength in the apparatus shown in FIGS. 30 and 31.

FIG. 35 is a flowchart showing another method of calculating the representative concentration gradient strength in the apparatus shown in FIGS. 30 and 31.

FIG. 36 is a flowchart showing still another calculation method of the representative concentration gradient strength in the apparatus shown in FIGS. 30 and 31.

FIG. 37 is a diagram showing representative density gradient directions from images 1 to 4 obtained by the calculation flowcharts shown in FIGS. 32 and 33.

38 is a diagram showing representative concentration gradient intensities from images 1 to 4 obtained by the calculation flowcharts shown in FIGS. 32 and 34. FIG.

FIG. 39 is a diagram showing representative density gradient intensities from images 1 to 4 obtained by the calculation flow charts shown in FIGS. 32 and 35.

40 is a diagram showing representative density gradient intensities from images 1 to 4 obtained by the calculation flowcharts shown in FIGS. 32 and 36. FIG.

[Explanation of symbols]

 3 image memory 9 memory 19 weighted sum calculation unit 20 subtraction unit 21 direction difference evaluation unit 22 product sum operation unit 23 division unit

Claims (24)

[Claims] 1. A density gradient direction I at each pixel of an input image.
Between the density gradient direction calculating means for obtaining θ (x, y), the density gradient direction Iθ (x, y) of the input image, and the density gradient direction Mθ (x, y) of the corresponding pixel of the predetermined model. Is a value f {Iθ (x, y) for evaluating the difference in the concentration gradient direction.
Concentration gradient direction evaluation value calculating means for obtaining −Mθ (x, y)}, weighting means for weighting this evaluation value with the weight M _W (x, y) obtained from the predetermined model, and weighted evaluation value And a means for calculating the degree of similarity by adding. 2. A density gradient direction I in each pixel of an input image
Between the density gradient direction calculating means for obtaining θ (x, y), the density gradient direction Iθ (x, y) of the input image, and the density gradient direction Mθ (x, y) of the corresponding pixel of the predetermined model. Is a value f {Iθ (x, y) for evaluating the difference in the concentration gradient direction.
Concentration gradient direction evaluation value calculating means for obtaining −Mθ (x, y)}, weighting means for weighting this evaluation value with the weight M _W (x, y) obtained from the predetermined model, and weighted evaluation value And a means for calculating the degree of similarity R by dividing the weighted addition value by the total weight. 3. The similarity calculation apparatus according to claim 1, wherein the weighting means uses the density gradient strength of the model image as the weight. 4. The similarity calculation device according to claim 1, wherein the evaluated value of the difference value is the cosine of the difference in the density gradient direction. 5. The evaluated value of the difference value is 1 if the difference in the density gradient direction is a value within a predetermined range including 0 °, and is -1 if the value is within another predetermined range. The similarity calculation device according to claim 1, claim 2 or claim 3, wherein a value within a predetermined range other than each is set to 0. 6. The similarity calculation apparatus according to claim 1 or 2, wherein the similarity calculation is performed only on pixels whose weight value weighted by the weighting means is a predetermined value or more. 7. A density gradient direction I in each region of an input image
Between the density gradient direction calculation means for obtaining θ (x, y), the density gradient direction Iθ (x, y) of this input image, and the density gradient direction Mθ (x, y) of the corresponding region of the predetermined model. Is a value f {Iθ (x, y) for evaluating the difference in the concentration gradient direction.
Concentration gradient direction evaluation value calculating means for obtaining −Mθ (x, y)}, weighting means for weighting this evaluation value with the weight M _W (x, y) obtained from the predetermined model, and weighted evaluation value And a means for calculating the degree of similarity by dividing the weighted addition value by the total weight. 8. A density gradient direction I for each pixel of an input image
theta (x, y) obtains the corresponding model between pixel density gradient direction Mθ of predetermined model (x, y) values obtained by evaluating the difference between the weight M _W (x obtained from a predetermined model, y ) By weighting, adding and dividing by the total weight,
A similarity calculation method comprising calculating the similarity between two grayscale images. 9. The similarity calculation method according to claim 8, wherein the weight obtained from the model is the density gradient strength of the model image, and the similarity between the two grayscale images is calculated. 10. The similarity calculation method according to claim 8, wherein the evaluated value is the cosine of the difference in the density gradient direction, and the similarity between the two grayscale images is calculated. 11. The evaluated value is set to 1 if the difference in the concentration gradient direction is within a predetermined range including 0 °, and is set to -1 if the value is within a predetermined range other than that. 9. The similarity calculation method according to claim 8, wherein a value within a predetermined range other than is set to 0 and the similarity between two grayscale images is calculated. 12. A model image storage means for storing a model image, an input image storage means for storing an input image, a scanning means for moving the model image within the input image, a model image in this scanning process, and its The similarity calculation device according to claim 1 or 2, which calculates the similarity of the position to the input image, and the position of the input image having the highest calculated similarity,
A position detecting device comprising a position determining means for determining the position of the model in the input image. 13. A density gradient direction storage means for storing a density gradient direction in each pixel of a plurality of model images, a density gradient strength storage means for storing a density gradient strength in each pixel of the plurality of model images, and the density gradient. The representative concentration gradient direction calculating means for calculating the representative concentration gradient direction from the direction, and the representative concentration gradient strength calculating means for calculating the representative concentration gradient strength from the concentration gradient direction and the concentration gradient strength. Alternatively, the similarity calculation device according to claim 7. 14. A density gradient direction in each pixel of a plurality of model images is stored, a density gradient strength in each pixel of the plurality of model images is stored, a representative density gradient direction is calculated from the density gradient direction, and the density is calculated. The similarity calculation method according to claim 9, 10, or 11, wherein the representative concentration gradient strength is calculated from the gradient direction and the concentration gradient strength. 15. The similarity calculation apparatus according to claim 13, wherein an average of the plurality of images in the density gradient direction is obtained as the representative density gradient direction. 16. The similarity calculation apparatus according to claim 13, wherein an arithmetic mean of density gradient intensities of a plurality of images is obtained as the representative density gradient intensity. 17. The representative density gradient strength is calculated from a density gradient direction of a plurality of images, a density gradient strength obtained from an average of the density gradient directions of a plurality of images, and a phase of a component of the representative density gradient direction of the density gradient strength. 14. The similarity calculation device according to claim 13, wherein the arithmetic mean is calculated. 18. The length of a vector whose components are the average of sine and the average of cosine in the density gradient direction is calculated from the density gradient directions of a plurality of images, and the product of the value and the representative density gradient strength is newly represented. 18. The similarity calculation device according to claim 17, which is obtained as a density gradient strength. 19. A density gradient direction calculation means for obtaining a density gradient direction in each pixel of a plurality of images in advance, a density gradient strength calculation means for calculating a density gradient strength in each pixel of the plurality of images, and the density gradient direction. A storage unit for creating and storing model data using a representative concentration gradient direction calculating means for calculating a representative concentration gradient direction from the above and a representative concentration gradient strength calculating means for calculating a representative concentration gradient strength from the concentration gradient strength. , The density gradient direction of the input image, and the density gradient strength between the representative density gradient direction of each corresponding pixel of the model data and the representative density gradient strength A similarity calculation apparatus comprising: a representative density gradient evaluation value calculation means for obtaining a value to be evaluated; and a means for calculating similarity by adding the evaluation values. 20. A density gradient direction calculating means for obtaining a density gradient direction in each pixel of a plurality of images in advance, a density gradient strength calculating means for calculating a density gradient strength in each pixel of the plurality of images, and from this density gradient direction A storage unit that creates and stores model data using a representative concentration gradient direction calculation unit that calculates a representative concentration gradient direction and a representative concentration gradient strength calculation unit that calculates a representative concentration gradient strength from the concentration gradient intensity. , The density gradient direction and the density gradient strength of the input image, the difference between the representative density gradient direction and the representative density gradient strength between the representative density gradient direction and the representative density gradient strength of each corresponding pixel of the model data is evaluated. A representative density gradient evaluation value calculation means for obtaining a value, an addition means for adding the evaluation values, and a division means for calculating the similarity by dividing the evaluation value by the number of pixels. For example was the similarity calculation device. 21. The similarity calculation device according to claim 19, wherein an average of the density gradient directions of a plurality of images is obtained as the representative density gradient direction. 22. The similarity calculation device according to claim 19 or 20, wherein an arithmetic mean of density gradient intensities of a plurality of images is obtained as the representative density gradient intensity. 23. As the representative density gradient strength, the phase of the representative density gradient direction component of the density gradient strength is calculated from the density gradient direction and the density gradient strength of a plurality of images and the representative density gradient direction obtained from the average of the density gradient directions of the plurality of images. 21. The similarity calculation device according to claim 19 or 20, which calculates an arithmetic mean. 24. A length of a vector having an average of sine and an average of cosine in the density gradient direction is obtained from the density gradient directions of a plurality of images, and a product of the value and the representative density gradient strength is newly represented. 21. The similarity calculation device according to claim 19 or 20, which is obtained as the density gradient strength.