JPH10162143A - Similarity calculating device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To calculate the similarity of images at a high speed by adding an evaluated value, which has been weighted at the time of similarity calculation, to a corresponding pixel in the decreasing order of the density gradient of a model image, and then discontinuing the adding process in its early stage. SOLUTION: A similarity calculation part 7 calculates the similarity by using the density gradient intensity and density gradient direction of the model image, which are previously calculated by a density gradient intensity calculation part 6 and density gradient direction calculation part 5 and stored in a memory 4 and the density gradient direction of an input image. In this case, a value that evaluates a difference in density gradient direction is found, the evaluation value is weighted by the density gradient intensity value of a pixel obtained from the model image, and the weighted evaluation values is added other corresponding pixel in the decreasing order of the density gradient intensity of the model image to calculate halfway similarity. According to the result obtained by comparing the halfway similarity in the middle of the addition with a previously set threshold value, the addition is discontinued and on the basis of the current halfway similarity, similarity is calculated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a similarity calculating apparatus for calculating the similarity of an input image to a model image.

[0002]

2. Description of the Related Art Conventionally, a normalized cross-correlation has been known as a method of calculating the similarity between gray-scale images having little effect on illumination fluctuations. . The technique for finding this normalized cross-correlation is
This is a matching method that does not change under the movement or expansion and contraction of the density gradation. However, when a linear transformation relationship is not established between the model image and the input image due to shading, background change, or the like, the similarity is increased by increasing the ratio of the product of each standard deviation to the covariance of the model image and the input image. However, there is a problem that the degree of reduction is small and stable pattern recognition cannot be performed.

In order to solve such a problem, the present applicant has previously invented a similarity calculating apparatus (Japanese Patent Application No. Hei 8-1).
No. 20278). As shown in FIG. 12, the similarity calculating device includes an address / data bus 1, a CPU 2, an image memory 3, a memory 4, a density gradient direction calculator 5, a density gradient intensity calculator 6, a similarity calculator 7, an addition It includes a termination threshold calculation unit 8.

[0004] The similarity R in this device is determined by the density gradient direction Mθ (x, y) at each pixel of the model image, the density gradient intensity M _W (x, y) at each pixel of the model image, and each pixel of the input image. Gradient direction Iθ (x, y) at
Is mx × my (see FIG. 13), it is calculated by the following equation 1.

[0005]

(Equation 1)

The following expressions 2 and 3 are provided as functions for evaluating the difference in the direction of the density gradient between the model image and the input image.

[0007]

(Equation 2)

[0008]

(Equation 3)

In calculating the actual similarity, the above equation (1) is modified as shown in equation (4).

[0010]

(Equation 4)

Here, i = mx × between x, y and i.
There is a relationship of y + x. This indicates that, as shown in FIG. 13, the equation 5 is added by mx × my times the size of the model image one line at a time from the upper left to the lower right of the model image.

[0012]

(Equation 5)

By the way, this kind of similarity calculating device includes:
It is required that the similarity be calculated at high speed so that pattern recognition of marks, characters, and the like can be performed in real time. To this end, it is necessary to reduce the number of times of addition of Expression 5.
For example, if it is certain that the similarity exceeds a certain threshold, or if it is certain that the similarity does not exceed a certain threshold, it is not necessary to calculate an accurate value. Once it has been assured that exceeds the threshold theta ₁ set in advance, discontinue the addition of equation 5. (B) at the time that the similarity R does not exceed the threshold theta ₂ that has been set in advance is guaranteed, discontinue the addition of Equation 5. There is a method that can be considered. The above-described device includes the addition cutoff threshold value calculation unit 8, and the addition cutoff threshold value calculation unit 8 adopts such a method. That is, the similarity r at the time when the addition of Expression 5 is performed k times.
(K) is represented by Equation 6.

[0014]

(Equation 6)

The intermediate similarity r (k) is expressed by
₁ (k), the similarity R [= r (mx × my
+1)] exceeds the threshold value θ ₁ [the method (A)]. Similarly, if the intermediate similarity r (k) is smaller than Θ ₂ (k) in Equation 8, the similarity R [= r
(Mx × my + 1)] is guaranteed not to exceed the threshold value θ ₂ [the method (B)].

[0016]

(Equation 7)

[0017]

(Equation 8)

Therefore, it can be understood that if the addition of Expression 5 is interrupted in the middle, the number of additions will be within the size mx × my times of the model image.

[0019]

In the above apparatus, the number of times of addition can be reduced by the addition cutoff threshold value calculating section 8. However, the size of the model image mx,
When my becomes large, there is a problem that similarity cannot be calculated at high speed. Accordingly, the present invention has been made in view of such a conventional problem, and an object of the present invention is to provide a similarity calculating apparatus capable of calculating the similarity at high speed.

[0020]

According to another aspect of the present invention, there is provided a similarity calculating apparatus according to the present invention, comprising: a density gradient direction calculating means for determining a density gradient direction in each pixel of a model image and an input image; Density gradient intensity calculating means for calculating the density gradient intensity at each pixel of the model image; density gradient intensity extracting means for extracting the density gradient intensity of the model image in descending order; the density gradient direction of the model image and the density gradient direction of the input image A density gradient direction evaluation value calculating means for calculating a value for evaluating a difference in a density gradient direction between the image processing apparatus; a weighting means for weighting the evaluation value with a density gradient intensity value of the pixel obtained from the model image; The evaluation value
An adding means for calculating an intermediate similarity by adding in accordance with the corresponding pixels in descending order of the density gradient intensity of the model image, and terminating the addition based on a result of comparing the intermediate similarity during the addition with a preset threshold value It is characterized by comprising an addition discontinuation determining means and a means for calculating a similarity based on an intermediate similarity at the time when the addition is discontinued.

In this apparatus, when calculating the similarity, the weighted evaluation values are added in accordance with the corresponding pixels in descending order of the density gradient strength of the model image, so that the necessary similarity range (predetermined threshold value) is set in advance. If it is set, the addition process is terminated earlier than in the conventional device, and the similarity can be calculated at a higher speed.

[0022]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below based on embodiments. FIG. 1 is a block diagram showing a configuration of a similarity calculation device according to an embodiment. This device is obtained by adding a concentration gradient intensity aligning section (concentration gradient intensity extraction means) 15 to the conventional device shown in FIG. That is, the apparatus comprises an address / data bus 1, a CPU 2, an image memory 3, a memory 4, a density gradient direction calculator 5, a density gradient intensity calculator 6,
It includes a similarity calculation unit 7, an addition cutoff threshold calculation unit 8, and a density gradient strength alignment unit 15, which are not shown in FIG.
Further, it is provided with an imaging camera, an A / D converter, a D / A converter, a CRT display, and the like. The density gradient strength alignment unit 15 extracts the density gradient strengths of the model images in descending order. As shown in FIG. 2, the similarity calculation unit 7 includes a weight sum calculation unit 9, a subtraction unit 10, a direction difference evaluation unit 11,
Product-sum operation unit 12, addition discontinuation determination unit 13, division unit 14
Is provided. The transfer of data between the modules is performed via the address / data bus 1, and the activation command issuance of each module is performed by the CPU 2.

In this similarity calculation device, an analog video signal output from a camera is synchronized with a timing signal to generate an analog video signal.
After being converted into a digital signal by the / D converter, it is stored in the image memory 3. The image fetched into the image memory 3 (see input image 1 in FIG. 5) is converted into an analog signal through a D / A converter, and then displayed on a CRT display. On the other hand, the image taken in the image memory 3 is converted in the density gradient direction by the density gradient direction calculation unit 5 and stored in the memory 4. In the similarity calculator 7, the density gradient intensity (FIG. 6) and the density gradient direction (FIG. 7) of the model image previously calculated by the density gradient intensity calculator 6 and the density gradient direction calculator 5 and stored in the memory 4 are stored. And the density gradient direction of the input image (FIG. 8) to calculate the similarity. The calculation of the similarity will be described later.

The similarity calculation unit 7 sets the density gradient direction and density gradient intensity of the model image and the magnitude of the density gradient direction of the input image to (mx, my) as shown in FIG. Mθ (x, y), the intensity of the concentration gradient is M _W
(X, y), the direction of the density gradient of the input image is Iθ (x, y)
As described above, the similarity R is calculated using Expression 1 and Expression 4 as a modified expression thereof.

Here, the density gradient direction Iθ of the input image
(X, y) and the density gradient direction Mθ (x,
Assuming that the evaluation value of the difference value of y) is the cosine of the difference in the density gradient direction, Expression 2 is satisfied. The evaluation value is set to 1 if the difference in the concentration gradient direction is within a range of ± 45 ° including 0 °, 0 if ± (45 ° to 135 °), and ± (13
In the case of (5 ° to 180 °), if −1 is set, Expression 3 is established.

The density gradient direction Iθ (x, y) of the input image stored in the image memory 3, the density gradient intensity M _W (x, y) and the density gradient direction Mθ of the model image stored in the memory 4. (X, y) is the address / data bus 1
Is input to the similarity calculation unit 7 through In the weight sum calculation unit 9, the density gradient intensity M _W (x, y) of the model image
(The denominator of Expressions 1 and 4) is calculated. or,
In the subtraction unit 10, the density gradient direction Iθ of the input image
(X, y) and the density gradient direction Mθ (x, y) of the model image
(Iθ (x, y) −Mθ (x, y)) is calculated. In the direction difference evaluation unit 11, f ｛Iθ (x, y) −
Mθ (x, y)} is calculated, and the product-sum operation unit 12 calculates the molecular formulas of Expressions 1 and 4. Then, the addition discontinuation determination unit 13 determines whether or not to terminate the addition, and the division unit 14 calculates the similarity R.

Next, the process of calculating the similarity will be described with reference to the flowchart of FIG. This flow chart adopts the methods (A) and (B) described above,
T3 and ST4 correspond to the processing. Here, as an example, a case will be described in which when the similarity R is larger than 0.8 or smaller than 0.6, an accurate value is not required. The function for evaluating the difference in the direction of the density gradient between the model image and the input image is represented by Expression 3, where the threshold θ _{1 in} the case of (A) is 0.8 and the threshold in the case of (B) is 0.8. θ
₂ becomes 0.6.

First, let image 1 in FIG. 4 be a model image and image 2 in FIG. In this example, the input image 1 is an image having a large similarity to the model image. These images 1 and 2 are stored in the image memory 3. 6 and 7 show the density gradient intensity and the density gradient direction of the image 1 (model image), respectively. ND in the density gradient direction indicates that there is no direction at that location (pixel). The Sobel operator was used to calculate the density gradient strength and the density gradient direction. FIG.
Indicates the density gradient direction of the image 2 (input image 1). The calculation of the density gradient direction and the density gradient intensity is performed by the density gradient direction calculation unit 5 and the density gradient intensity calculation unit 6, respectively.

Further, the density gradient intensity aligning section 15 extracts the density gradient intensities (FIG. 6) of the model image calculated by the density gradient intensity calculating section 6 in descending order.
..., 63, and so on, from the upper left to the lower right. The result is shown in FIG. According to the corresponding pixels in this order, the addition according to Expression 5 is performed. Based on the calculated value of the density gradient intensity of the model image, the addition cutoff threshold calculator 8 calculates two types of thresholds for the addition cutoff of (A) and (B), and the memory 4 Is stored in The actual similarity calculation is performed by the similarity calculator 7. First, in step (hereinafter abbreviated as ST) 1,
Initialization of variables i and r is performed. These variables i and r are stored in the memory 4. The addition of Expression 5 in ST2 includes a subtraction unit 10, a direction difference evaluation unit 11, a product-sum operation unit 12
This is performed according to the addition order in FIG.

In ST3, the addition is performed in accordance with the equation (5).
The similarity r is set to a predetermined threshold θ ₁(Here, 0.
8) It is determined whether or not (threshold value θ)₁Greater than
In ST4, the similarity r in the middle is set in advance.
Threshold θ_Two(Here 0.6) or not
(Threshold θ_TwoIs smaller than). S
When the determinations of T3 and T4 are both YES, that is, 0.6 <r <
In the case of 0.8, the variable i is set to i + 1 (ST5),
Whether the number i is greater than the model image size mx × my
(ST6), and if YES, similarity to the next ST7
Move to the calculation of the degree R, and if NO, perform the processing of ST2 to ST6.
repeat.

If NO in ST3 or ST4, that is, r
If a> 0.8 or r <0.6, the similarity R exceeds the threshold theta _1, or because it would have been guaranteed not to exceed the threshold theta _2, the addition at which time , And the process proceeds to the calculation of the similarity R in ST7. The addition termination determination is performed by the addition termination determination unit 13 based on the threshold values θ ₁ and θ ₂ stored in the memory 4.

In ST7, the division unit 14 calculates the intermediate similarity r at the point in time when the addition is terminated by the weight sum calculation unit 9.
The similarity R is calculated by dividing by the sum of the density gradient intensities of the model images calculated in the above, and the processing is terminated. In this similarity calculation process, specifically, when ST2 is added in accordance with the addition order in FIG. 11, the determination in ST3 becomes NO when the addition is performed 40 times, and the addition is stopped. When the image 1 in FIG. 4 is a model image and the image 3 in FIG. 9 having a small similarity to the model image is the input image 2, the density gradient direction of the image 3 is as shown in FIG. When the addition of ST2 is performed according to the addition order of 11, the determination of ST4 becomes NO when the addition is performed 12 times, and the addition is stopped.

On the other hand, in the case of the conventional apparatus (weighted evaluation values are not added in accordance with the corresponding pixels in the order of increasing density gradient strength of the model image), if image 1 is a model image and image 2 is an input image 1, , ST2 is 60
At this time, the determination in ST3 is NO, and the addition is stopped. If the image 1 is a model image and the image 3 is an input image 2, the determination in ST4 becomes NO when the addition of ST2 is performed 25 times, and the addition is stopped.

As described above, by performing the addition in a predetermined order, the number of additions is reduced from 60 to 40 in the former case, and the number of additions is reduced from 25 to 12 in the latter case.
Therefore, since the addition is discontinued earlier than in the conventional apparatus, the similarity can be calculated at a higher speed. The speeding up of the similarity calculation becomes more remarkable as the sizes mx and my of the model image become larger.

In the above embodiment, the addition process is terminated when the similarity on the way exceeds the preset threshold value or when it is guaranteed that the similarity does not exceed the preset threshold value. The addition process may be terminated at the time when only the value is guaranteed or when only the value that does not exceed the preset threshold value is guaranteed. Of course, the model image and the input image described in the embodiment are merely examples, and various marks, characters, figures, and the like can be targeted regardless of their size (number of pixels), and are used for recognizing those patterns. can do. Also, the preset threshold value may be changed as appropriate.

[0036]

As described above, according to the first aspect of the present invention,
When calculating the similarity, the similarity calculation device adds the weighted evaluation values in accordance with the corresponding pixels in descending order of the density gradient strength of the model image. The addition processing is discontinued earlier than the apparatus described above, and the similarity can be calculated at high speed. Especially,
The larger the model image is, the more remarkable the calculation of the similarity becomes.

According to the second aspect, since the similarity can be normalized, the similarity can be evaluated as an absolute value for matching with a plurality of models.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a similarity calculation device according to an embodiment.

FIG. 2 is a block diagram showing a configuration of a similarity calculation unit in the similarity calculation device of the embodiment.

FIG. 3 is a flowchart illustrating a similarity calculation process in the similarity calculation device of the embodiment.

FIG. 4 is a diagram illustrating an example of a model image (image 1).

FIG. 5 is a diagram showing an example of an input image 1 (image 2) having a large similarity to a model image.

FIG. 6 is a diagram showing a density gradient intensity of the model image of FIG. 4;

FIG. 7 is a diagram illustrating a density gradient direction of the model image of FIG. 4;

8 is a diagram showing a density gradient direction of the input image 1 of FIG.

FIG. 9 is a diagram illustrating an example of an input image 2 (image 3) having a small similarity to a model image.

FIG. 10 is a diagram illustrating a density gradient direction of the input image 2 of FIG. 9;

11 is a diagram illustrating an addition order in which the density gradient intensities of the model image in FIG.

FIG. 12 is a block diagram illustrating a configuration of a similarity calculation device according to a conventional example.

FIG. 13 is a diagram showing the size (mx × my) of a model image and the order of addition in the conventional device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Address / data bus 2 CPU 3 Image memory 4 Memory 5 Density gradient direction calculation part 6 Density gradient strength calculation part 7 Similarity calculation part 8 Addition cutoff threshold value calculation part 13 Addition cutoff determination part 14 Division part 15 Density gradient strength alignment Part (concentration gradient strength extraction means)

Claims (5)

[Claims] 1. A density gradient direction calculating means for determining a density gradient direction at each pixel of a model image and an input image; a density gradient intensity calculating means for determining a density gradient intensity at each pixel of a model image; Density gradient intensity extraction means for extracting the density gradient direction of the model image from the density gradient direction of the model image and the density gradient direction of the input image, density gradient direction evaluation value calculation means for obtaining a value for evaluating the difference in the density gradient direction, Weighting means for weighting the evaluation value with a density gradient intensity value of the pixel obtained from the model image;
An adding means for calculating the intermediate similarity by adding the weighted evaluation values in accordance with the corresponding pixels in order of the density gradient intensity of the model image, and comparing the intermediate similarity during the addition with a preset threshold value A similarity calculation device comprising: an addition discontinuation determination unit that terminates the addition based on the similarity; and a unit that calculates the similarity based on the halfway similarity at the time when the addition is terminated. And means for calculating the similarity by dividing the intermediate similarity at the time when the addition has been discontinued by the addition discontinuation determining means by the sum of the density gradient intensities of the model images. 2. The similarity calculating device according to claim 1, wherein the similarity calculating device is a means. 3. The similarity calculating apparatus according to claim 1, wherein said addition discontinuation judging means terminates the addition at a point in time when it is guaranteed that the intermediate similarity exceeds a preset threshold value. 4. The similarity calculating apparatus according to claim 1, wherein said addition discontinuation judging means terminates the addition when it is guaranteed that the intermediate similarity does not exceed a preset threshold value. 5. The similarity degree according to claim 1, wherein said addition termination determination means terminates the addition when it is guaranteed that the halfway similarity exceeds or does not exceed a preset threshold value. Calculation device.