JP3496530B2 - Pattern identification device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern specifying device, and more particularly to a pattern specifying device for specifying a pattern by specifying a characteristic image corresponding to a pattern existing in the image.

[0002]

2. Description of the Related Art A pattern recognition technique for detecting various patterns in an image picked up by a television camera or the like and estimating their existence, position, size, posture, etc. is applied in a wide range of fields as a basic technique of image recognition. There is. One of the methods of pattern recognition is generalized Hough transform. This method has the following features. -Applicable to figures of arbitrary shape.

Recognition is possible only by scanning the image once.

Since it is a voting system, it is unlikely to be affected by false feature points and missing feature points due to noise.

A calculation procedure for obtaining the position of the pattern using the generalized Hough transform will be described.

That is, a reference table storing the characteristic value (density value) of an edge portion (outline portion) of a predetermined reference image (reference edge image) and each pixel position of the reference image is prepared.
Using the position of each pixel stored in this lookup table,
The position (barycentric position) of the unknown image is determined by address-converting each point of the edge portion of the unknown image (unknown edge image) corresponding to the input pattern and creating a plurality of reference edge images in the two-dimensional parameter space. Vote and decide the peak position with the most votes as the position of the unknown image. If the posture (rotation angle) of the unknown input image is unknown, the rotation angle is obtained by further increasing the parameter by one and voting in the three-dimensional parameter space.

[0007]

However, the above generalized Hough transform has the following problems.

That is, only one pattern (pattern having the same contour shape) can be recognized by one scanning. Specifically, the position of the unknown image cannot be determined unless address conversion is performed on each point of the input unknown image edge portion to create a plurality of reference edge images in the two-dimensional parameter space.

Further, if the parameters such as rotation increase in addition to the position as described above, the dimension of the parameter space increases,
This increases the voting space in proportion to the power. Therefore, the calculation time and the required memory capacity are required, and it takes time to search the peak position.

Further, as a set of patterns to be recognized, only a set that can be generated by affine transformation such as parallel movement, rotation and enlargement of a certain shape pattern can be handled.

The reason why the above problems occur is that the generalized H
The ohugh transform represents the diversity of patterns represented by parameters such as position and rotation in an analog manner, and the shape of one pattern is represented by a multidimensional parameter space of position and rotation. It depends.

The present invention has been made in view of the above facts, and it is an object of the present invention to propose a pattern specifying device capable of specifying various patterns in a short time.

[0013]

The present invention will be described below with reference to FIG. That is, in order to achieve the above-mentioned object, the present invention provides an input means 11 for inputting an image and a first feature that corresponds to a pattern existing in the image and represents the feature of the pattern from the input image. Extraction means 12 for extracting a characteristic image defined for each of a plurality of portions, and a plurality of candidate images expected as the characteristic images represent the characteristics of the candidate image for each of a plurality of portions corresponding to the plurality of portions. Second
For each of the portions of the extracted feature image that is determined with the features of the first feature and the second feature of the portion of the candidate image that includes the first feature of each of the portions of the feature image. Of the candidate image, and based on the result of the comparison, for each part of the feature image, the candidate image is compared with the candidate image corresponding to the second feature close to the first feature. By voting means 14 for voting the degree of contribution that represents the probability of being an image, and by identifying the candidate image corresponding to the characteristic image based on the degree of contribution voted by the voting means 14 for the candidate image, The specifying means 15 for specifying the pattern is provided.

That is, the input means 11 inputs an image, and the extraction means 12 extracts a characteristic image from the input image.

Here, the characteristic image corresponds to a pattern existing in the image input by the input means 11 and to be specified, and a plurality of first characteristics representing the characteristic of the pattern. Is defined for each part of.

The "portion" referred to here is each pixel of the characteristic image, each pixel of the characteristic image, each block including a plurality of pixels as one block, and a geometric portion (for example, a corner portion or a curved portion). Etc.)

Further, in the present invention, each of a plurality of candidate images expected as characteristic images is defined. Since each candidate image is expected as the characteristic image,
It is set together with the second feature representing the feature of the candidate image for each of the plurality of portions corresponding to the plurality of portions of the feature image. The second feature corresponds to the first feature.

Here, the characteristic image is a contour image of a pattern existing in the input image, and the first feature and the second feature are the amount of change in density and the direction of change in density of the contour portion of the contour image. (Whether the density increases or decreases) may be included. That is, the first feature and the second feature may be the amount of change in the density of the pixel and the changing direction of the density, the average value of the amount of change in the density of each block, and the changing direction of the maximum density. Note that geometric features (for example,
Corners, curves, etc.).

By the way, the candidate image may be determined in advance and stored in the storage means, or may be created each time the characteristic image is extracted by the extraction means 12.

The voting unit 14 includes, for each part of the extracted feature image, the first feature of each part of the feature image and the second part of the candidate image including the part corresponding to the first feature.
And the characteristics of, and based on the result of the comparison,
For each part of the characteristic image, the contribution degree representing the probability that the candidate image is the characteristic image is voted for the candidate image corresponding to the second characteristic close to the first characteristic.

Here, the contribution may be a value normalized such that the cumulative value of the contribution when voting is made in all parts of the characteristic image is a constant value in all patterns.

Then, the specifying means 15 specifies the pattern by specifying the candidate image corresponding to the characteristic image based on the contribution degree voted by the voting means 14 to the candidate image.

In this way, each of the plurality of candidate images expected as the characteristic image is determined and extracted for each of the plurality of portions corresponding to the plurality of portions of the characteristic image together with the second characteristic representing the characteristic of the candidate image. For each part of the feature image, the contribution degree representing the probability that the candidate image is the feature image is voted for the candidate image corresponding to the second feature close to the first feature of each part of the feature image. , The pattern is specified by specifying the candidate image corresponding to the characteristic image based on the voted contribution degree, so that the pattern can be specified from various candidate images in one processing, and various patterns can be specified. It is possible to specify the pattern in a short time.

Here, a conversion means for converting the candidate image specified by the specifying means 15 into a required format may be further provided.

Further, as described above, it further comprises storage means for storing the characteristic of each portion of each of the plurality of patterns as the second characteristic, and the voting means 14 has the second characteristic stored in the first characteristic and the storage means. You may make it compare with the characteristic of.

Further, there is further provided cumulative storage means for cumulatively storing the contribution degrees voted by the voting means 14 for the candidate images in correspondence with the candidate image identification information for identifying the candidate images. The candidate image that is the characteristic image may be specified based on the candidate image identification information corresponding to the largest cumulative value that is cumulatively stored in the storage unit corresponding to the candidate image identification information.

[0027]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the following, an example of recognizing a plurality of objects having different shapes, which is difficult by the conventional generalized Hough transform, will be shown.

As shown in FIG. 1, the pattern specifying apparatus according to this embodiment includes an input means 11 for inputting an image. The input unit 11 can be composed of, for example, a monochrome TV camera. Incidentally, this input means 11
Is installed with the optical axis oriented horizontally at a certain height from the surface of the earth, as shown in FIG. Then, the recognition of the passenger car B1 and the pedestrian B2 crossing in front of the input means 11 will be described as an example (see FIG. 3). Two objects (B
1, B2) from the input means 11 is in the range of Xm to Ym (X <Y), and the traveling directions of the passenger car B1 and the pedestrian B2 are all perpendicular to the optical axis of the camera. Can be reduced to the problem of recognizing two shapes with different positions.

The input means 11 is provided with a contour extracting means 21 for extracting a contour portion (see also FIG. 4) in which the grayscale changes drastically from the image (monochrome grayscale image) input by the inputting means 11. Means 12 are connected. Sobel, which is a general method of image processing, is used to extract the contour portion.
It can be realized by a filter.

The extraction means 12 includes a candidate image collation device 10
0 is connected. The candidate image matching device 100 includes a contribution degree storage unit 13 that stores a contribution degree that indicates the likelihood that a plurality of candidate images expected as the contour images (feature images) are the contour images. The contribution degree storage means 13 includes a parameterization means 31, a contribution degree calculation means 32, and a contribution degree memory 33.

The candidate image collation apparatus 100 comprises a voting means 1 connected to the extraction means 12 and contribution degree storage means 13.
4 and the specifying means 15 connected to the voting means 14 are further provided. In addition, the identification unit 15 is the voting result memory 5
1, a peak search means 52, and a patterning means 53.

Next, the operation of this embodiment will be described.

First, the process of storing the contributions of the plurality of candidate images in advance in the contribution storage unit 13 will be described.

That is, a plurality of candidate images predicted by the parameterizing means 31 as feature images corresponding to the patterns existing in the image input by the input means 11 are
It is parameterized by generating it on a virtual plane and making each candidate image correspond to a parameter value. It should be noted that when generating a plurality of candidate images, each of the basic contour patterns is CA
Using a contour image generated by D or inputting and outputting an actual image to the feature extraction means, based on that, from the range of possible values of the distance to the object and the position in the traveling direction of the object,
By moving and enlarging the contour in parallel, candidate images corresponding to all patterns appearing on the characteristic image are generated.

The rate of change in parallel movement and enlargement is finally determined from the number of parameters that are determined by the resolution required for recognition and the constraints of the calculation scale. Further, the parameterization is performed by assigning numbers to the generated candidate images, for example, 1, 2, 3, 4 ... In addition, 1, 2, 3, assigned to the candidate image
The numbers such as 4 ... are hereinafter referred to as parameter values. FIG. 5 shows an example of a plurality of candidate images generated by the difference in the traveling direction position in the case of the passenger car B1, and FIG. 6 shows an example of a plurality of candidate images generated by the difference in the distance to the object. . Although the above-mentioned method of generating candidate images is an automatic method, it is also possible to manually generate all candidate images.

Further, in general, the candidate image represented by one parameter has a range of a certain value in the distance and the position in the traveling direction, so that the contour image is a candidate image with overlapping contour lines. The parameterization method may be a multidimensional parameterization method using a plurality of parameters other than the one-dimensional parameter method described above.

The contribution degree calculation means 32 previously calculates the contribution degree used by the voting means 4 for voting based on the candidate image generated by the parameterization means 31, and writes it in the contribution degree memory 33. That is, the contribution degree calculation means 32 executes the contribution degree storage processing routine shown in FIG. 7 when the parameterization means 31 generates a plurality of candidate images.

In step 62 of FIG. 7, a variable G for identifying a pixel on a virtual plane where a candidate image exists is initialized (G ←
0) In step 64, the variable G is incremented by 1, and in step 66, all candidate images including the pixel G are captured.

In step 68, a variable P for identifying all the captured candidate images is initialized (P ← 0). In step 70, the variable P is incremented by 1, and in step 72
Then, the contribution degree in the pixel G of the candidate image P is set.

Here, the degree of contribution represents the probability that the candidate image is a characteristic image. In addition, the contribution degree is a value that is normalized such that the cumulative value of the contribution degree when voting is performed in all pixels of the characteristic image is a constant value in all candidate images, as will be described later. For example, when the fixed value is 1, and the total number of pixels of a certain candidate image is 1500, the contribution of each pixel of this candidate image is 1/1500. That is, when the contribution degree 1/1500 is voted for all the pixels of this candidate image, the cumulative value of the contribution degree becomes a constant value = 1.

In step 74, the characteristic value (second characteristic), the parameter value, and the contribution degree of the pixel G of the candidate image P are stored in the contribution degree memory 33 corresponding to the pixel G. The feature value is the intensity of the contour portion (edge portion) (the amount of change in density and the direction of the gradient (dark → light, dark ← light).

In step 76, it is determined whether the variable P for identifying the captured candidate image is equal to the total number P ₀ of captured candidate images. If there are a plurality of candidate images partially including the pixel G, the total number P ₀ is the total number of candidate images partially including the pixel G.

If the variable P is not equal to the total number P ₀ , then
Returning to step 70, the above processing (steps 70 to 7)
Perform 6). If the variable P is equal to the total number P ₀ , then
In step 78, it is determined whether the variable G for identifying the pixel on the virtual plane where the candidate image exists is equal to the total number of pixels G ₀ on the virtual plane where the candidate image exists. If the variable G is not equal to the total number of pixels G ₀ , the process returns to step 64 to execute the above processing (steps 64 to 78). If the variable G is equal to the total number of pixels G ₀ , this routine is executed. To finish.

Here, the contribution memory 33 corresponds to each pixel (each pixel in which the candidate image exists), the characteristic value (second characteristic), the parameter value of the candidate image to be voted, and the contribution degree. This is a structure capable of storing a plurality of sets. Then, as a result of the above processing, as shown in FIG. 9, the contribution degree to the candidate image 10 is set successively for the candidate image 3 and the contribution degree memory 33 is set.
When the data is stored in the candidate image 10, the data set is similarly stored in the contribution degree memory corresponding to the pixel of the contour portion of the candidate image 10. Both voting data sets of image 10 are stored. More specifically, for example, as shown in FIG. 10, the pixel G ₁ (x, y) = (2, 2) has the candidate image 2
Since only part of (parameter value 2) exists, pixel G
_The feature value, the parameter value (2), and the contribution degree are stored corresponding to ₁ . Since a part of the candidate image 2 (parameter value 2) and a part of the candidate image 3 (parameter value 3) exist in the pixel G ₂ (x, y) = (5, 6), the pixel G ₂ is in the pixel G ₂ . Correspondingly (feature value, parameter value (2), and contribution),
(Feature value, parameter value (3), and contribution degree) are stored. For example, it is stored as follows.

(X, y) = (2, 2): (feature value, parameter value (2), and contribution degree) ────────────────────── ───────────────────────────────── = (5,6): (feature value, parameter value (2), And contribution), (feature value, parameter value (3), and contribution) ─────────────────────────── A process of identifying a pattern existing in the generated image will be described.

The image obtained by photographing, for example, by the input means 11 is inputted to the extracting means 12. The contour extracting means 21 of the extracting means 12 selects a point having a spatial peak of the edge strength as a contour point from the edge strength and the gradient direction calculated by the Sobel filter. Thereby, the contour image (feature image) shown in FIG. 4 from the input image of FIG.
Is output to the candidate image matching device 100. The contour image is defined using the edge strength of the contour and the gradient direction. Further, the edge strength of the part without the contour is zero. When the contour image is input by the contour extracting means 21, the voting means 14
Executes the voting processing routine shown in FIG.

That is, in step 80 of FIG. 11, a variable C (C ← 0) for identifying each pixel of the input contour image is initialized, and in step 82, the variable C is incremented by 1,
In step 84, the edge strength, the gradient direction, the parameter value, which are stored in the contribution memory corresponding to the pixel C position,
And a set of contributions are captured for each candidate image.

In step 86, a variable P for identifying each candidate image in the captured set is initialized, and in step 8
At 8, the variable P is incremented by 1.

At step 90, it is judged whether or not the edge strength and the gradient direction at the pixel C of the contour image and the candidate image are the same. If they are not the same, the process proceeds to step 94. If they are the same, in step 92, the specifying means 15
The parameter value and the contribution degree are output, and the process proceeds to step 94.

In step 94, the session in which the variable P is fetched
Total number of candidate images in₀Determine if equal to
To do. The variable P is the total number P₀If not equal to
Return to step 88 to perform the above processing (step 88 to step
94) is executed. The variable P is the total number P₀Is equal to
In step 96, the variable C is the total number of pixels C of the contour image.
₀To determine whether or not. Variable C is the total number of pixels C₀To
If they are not equal, the process returns to step 82 and the above process is performed.
(Step 82 to step 96) is executed and the variable C is
Total number of pixels C₀If this is equal to
It

That is, by the above processing, the contour image calculated by the contour extracting / extracting means 21 is input, and when the contour exists, the data set stored in the corresponding pixel position of the contribution degree memory 33 is read out to extract the contour image. The actual edge strength,
The gradient direction is compared with the stored edge strength and gradient direction values, and if they are equal, the contribution degree is voted in the voting result memory 51 corresponding to the parameter value of the corresponding candidate image. As a result, the identifying unit 15 cumulatively adds the contribution degrees corresponding to the parameter values in the voting result memory 51, as shown in FIG.

The edge strength and the value in the gradient direction may be represented by discrete values such as binary values to simplify the comparison operation. The voting result memory 51 stores the cumulative addition value for the number of parameters of the candidate image, and is initialized to 0 before recognition.

Two candidate images 15 and 16 as shown in FIG.
14 is set in advance in the contribution degree memory, the voting means 14 and the voting result memory 51 in the example in which the input image of FIG. 14 is given.
The operation of will be described. A contour image output from the input image of FIG. 14 by the feature extraction means 2 is as shown in FIG. Then, the voting unit 14 votes for a candidate image in which the contour of the contour image exists. In this case, the outline portion B1 of the candidate image 15
And the contour of the horizontal line H exists, the voting result memory corresponding to the candidate image 15 has a large cumulative result almost equal to that obtained when the candidate image is present without being missing, and the cumulative stored value is substantially the same. It becomes the above constant value. Candidate image 1
For 6, the voting value is obtained at the portion where the contour of the horizontal line H and the 16 candidate images overlap, but as a result, only a small cumulative value is obtained, that is, the cumulative storage value is a value that is significantly different from the above-mentioned fixed value. Becomes The peak searching means 52 inputs the cumulative value for each parameter accumulated in the voting result memory 51, and outputs one or a plurality of parameters having a large cumulative value. In the examples of FIGS. 13 to 15, the candidate image 15 having a large cumulative value is output.

The patterning means 53 is the peak searching means 5
Based on the parameter value having a large cumulative value output from No. 2, conversion to information necessary for the system application is performed. In the case of the embodiment, the type, distance, and position in the traveling direction of the object are converted into that format if necessary, and if output as an outline on the image is required, conversion is performed into that format. The information necessary for the conversion is obtained by generating candidate images by the parameterization means and storing the information of the candidate images corresponding to the parameters when the parameters are associated.

When the peak searching means 52 outputs a plurality of parameters, when converting the information into information required by the system application, the information is expressed as the cumulative value of the contributions corresponding to the plurality of parameter values. A weighted intermediate value may be used, or only a part of information excluding uncertain information may be output. When outputting as the outline of a candidate image, an intermediate candidate image is generated and output, or a part of a common image is generated and output,
You may.

For example, when a parameter pattern having a similar cumulative value is obtained, the information required by the system application is the pattern position, and the pattern represented by each parameter is (0, 1). , (0,
If it is 2), the weights are the same, so the information is output as (0, 1.5) which is the average value of each position.

As described above, the candidate image for the characteristic image is specified.

As described above, in the present embodiment,
Since the degree of freedom of the problem that can be dealt with can be expanded by previously generating and storing the candidate image that appears on the image by the contribution degree calculation unit 32, recognition of a plurality of objects having different shapes that cannot be handled by the generalized Hough transform and A recognition target that undergoes transformations other than affine transformations such as perspective transformation can also be applied.

Table 1 shows a comparison between the generalized Hough transform and the present embodiment.

[0060]

[Table 1] As shown in Table 1, the object that can be recognized by one scanning is a candidate image having the same shape in the generalized Hough transform, but in the present embodiment, it is a plurality of types of candidate images. Be different.

Further, with respect to the set of candidate images, the generalized Hough transform is a set formed by affine transforming a certain shape on a plane, whereas in the present embodiment, all the transforms are included. Is different.

Further, in the generalized Hough transform, one shape is stored, whereas in the present embodiment, a set of candidate images is generated, each candidate image is parameterized, and each pixel of each candidate image is converted. The difference is that the degree of contribution is calculated and stored.

Further, in the generalized Hough transform, a complicated operation is executed to calculate the contribution, and a large amount of calculation is required when the pattern is specified. In contrast, in the present embodiment, the contribution is read and the pattern is calculated. The difference is that the calculation amount is small at the time of identification.

Further, in the above-described embodiment, the point that the information required for recognition is stored in advance in the contribution degree memory 33 and the parameter space formed by the voting result memory 51 can be made into a low dimension such as one dimension. Since the peak search means 52 can be realized as a low-dimensional peak search, the recognition process can be executed at high speed.

Further, since it is a voting-type recognition method, it has robustness against noise and some pattern missing which are general advantages of the voting-type recognition method.

In the above-described embodiment, the feature image feature and the candidate image feature are compared for each pixel of the feature image. However, the present invention is not limited to this, and the feature image is scattered. The feature image of the feature image and the feature of the candidate image may be compared for each pixel, each block including a plurality of pixels as one block, and each geometric portion (for example, a corner portion or a curved portion). Note that the feature for comparing the feature image and the candidate image for each geometric portion may be, for example, a corner or a curve.

In the above-described embodiment, the passenger car B
Although the recognition of 1 and the pedestrian B2 has been described as an example, the present invention is not limited to this, and may be similarly applied to recognition of various kinds of objects expected in advance. The distance and the moving direction of the object from the input means 11 are not limited to the above example.

[0068]

As described above, according to the present invention, each of a plurality of candidate images expected as a characteristic image represents the characteristic of the candidate image for each of a plurality of portions corresponding to a plurality of portions of the characteristic image. For each part of the extracted feature image, the probability that the candidate image is the feature image for the candidate image corresponding to the second feature close to the first feature of each part of the feature image. Since the pattern is specified by voting the contribution degree that represents, and specifying the candidate image corresponding to the feature image based on the voted contribution degree, the pattern is specified from various candidate images in one processing. You can
It is possible to identify various patterns in a short time,
Has the effect.

[Brief description of drawings]

FIG. 1 is a diagram showing constituent features of the present invention.

FIG. 2 is a block diagram of a pattern specifying device according to the present embodiment.

FIG. 3 is a diagram showing an input image.

FIG. 4 is a diagram showing extracted feature images.

FIG. 5 is a diagram showing various patterns created by moving a basic pattern in parallel.

FIG. 6 is a diagram showing various patterns created by enlarging a basic pattern.

FIG. 7 is a flowchart showing a contribution degree storage processing routine.

FIG. 8 is an explanatory diagram illustrating setting a contribution degree.

FIG. 9 is another explanatory diagram for explaining setting the contribution degree.

FIG. 10 is still another explanatory diagram illustrating setting the contribution degree.

FIG. 11 is a flowchart showing a control routine executed by a voting unit.

FIG. 12 is a diagram showing a table for cumulatively storing voted contributions.

FIG. 13 is a diagram showing stored patterns.

FIG. 14 is a diagram showing a specific target according to the present embodiment.

FIG. 15 is a diagram showing a characteristic image extracted from the input image of FIG.

[Explanation of symbols]

11 Input means 12 Extraction means 14 voting means 15 Identification means

Continuation of front page (56) Reference JP-A-5-165968 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G06T ⁷ /00-7/60 G06T 1/00 JISST file (JOIS)

Claims (7)

(57) [Claims] 1. Input means for inputting an image, and a first feature corresponding to a pattern existing in the image and representing the feature of the pattern is determined for each of a plurality of parts from the input image. Extracting means for extracting the identified feature image, each of the plurality of candidate images expected as the feature image represents a feature of the candidate image for each of a plurality of portions corresponding to the plurality of portions, and the first feature For each part of the extracted feature image that is defined with a corresponding second feature, corresponds to that part of the candidate image that includes the first feature of each part of the feature image and the part corresponding to that part. Second part
And the characteristics of, and based on the result of the comparison,
For each part of the characteristic image, with respect to a candidate image corresponding to a second characteristic close to the first characteristic, a voting means for voting a contribution degree indicating the likelihood that the candidate image is the characteristic image, Specification for identifying a candidate image corresponding to the characteristic image based on the cumulative value of contributions voted to the candidate image by the voting means, and outputting candidate image identification information for identifying the identified candidate image When outputting a plurality of candidate image identification information, the specifying means converts the candidate image identification information into predetermined information, and uses the cumulative value of contribution rates corresponding to the plurality of candidate image identification information. An intermediate value obtained by weighting the information is output, and the candidate image is a feature image extracted by the extracting means.
By translating or enlarging the pattern corresponding to the image
Is generated by a plurality of pixels of the characteristic image as one block
Each block, each pixel of the characteristic image, and the jump of the characteristic image
Any of the skipped pixels or the geometric part of the characteristic image
The pattern specifying device , wherein the first and second characteristic values are the strength of the contour portion . 2. Input means for inputting an image, and a first feature corresponding to a pattern existing in the image and representing the feature of the image is determined for each of a plurality of parts from the input image. Extracting means for extracting the identified feature image, each of the plurality of candidate images expected as the feature image represents a feature of the candidate image for each of a plurality of portions corresponding to the plurality of portions, and the first feature Corresponding to the portion of the candidate image including the first characteristic of each portion of the characteristic image and the portion corresponding to the portion, which is defined together with the corresponding second characteristic. The second feature of the part is compared, and based on the result of the comparison, for each part of the feature image, the candidate image is compared with the candidate image corresponding to the second feature close to the first feature. The image that represents the certainty that the image is the characteristic image. And a specifying unit that specifies a candidate image corresponding to the characteristic image based on a cumulative value of contribution degrees voted by the voting unit for candidate images, the specifying unit Outputs a common image portion of the candidate images when specifying a plurality of candidate images. The candidate images are characteristic images extracted by the extracting means.
By translating or enlarging the pattern corresponding to the image
Is generated by a plurality of pixels of the characteristic image as one block
Each block, each pixel of the characteristic image, and the jump of the characteristic image
Any of the skipped pixels or the geometric part of the characteristic image
The pattern specifying device , wherein the first and second characteristic values are the strength of the contour portion . 3. The pattern specifying device according to claim 1, further comprising a conversion unit that converts the candidate image specified by the specifying unit into a required format. 4. The feature image is a contour image of a pattern existing in the input image, and the first feature and the second feature are variation amounts of density of a contour portion of the contour image. The pattern specifying device according to claim 1, further comprising: 5. The contribution is a value normalized in advance so that the cumulative value of the contribution when voting is made in all parts of the characteristic image becomes a constant value in all candidate images. The pattern identification device according to any one of claims 1 to 4. 6. The storage device further includes a storage unit that stores a characteristic of each portion of each of the plurality of candidate images as the second characteristic, and the voting unit stores the first characteristic and the storage unit. The pattern specifying device according to claim 1, wherein the pattern specifying device compares the second characteristic. 7. A cumulative storage unit that cumulatively stores the contribution degrees voted by the voting unit for the candidate images in correspondence with the candidate image identification information, wherein the specifying unit is a candidate in the cumulative storage unit. 7. The candidate image that is the characteristic image is specified based on the candidate image identification information corresponding to the largest cumulative value that is cumulatively stored corresponding to the image identification information. Pattern identification device.