JP2002008028A - Pattern matching method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pattern matching method and device therefor for performing a precise and high-speed processing at a low cost by an operation with a small number of points arranged on a template about only the partial feature shape conformed to a pattern to be recognized. SOLUTION: A rough search processing part 103 automatically generates initial optimum rough search template data having both data for feature shape pattern and prohibited feature shape pattern by an initial rough search template set for a specified pattern necessary for image recognition of information such as position of a subject from a taken image memory part 102. The data are then subjected to addition and reduction of necessary operation points by a rough search template optimization part 107 to form an optimum rough search template data. The rough search processing part performs a matching by use of the data to attain the purpose. Further, a comprehensive search processing part 123 further comprises a precise search processing part, and the image recognition can be performed with a data satisfying both rough and precise matchings.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a consumer and industrial image processing method and apparatus, and more particularly, to a visual recognition apparatus or the like which requires high-speed and high-precision positioning such as an electronic component mounting machine or a semiconductor assembly. The present invention relates to a method and an apparatus for performing high-speed and high-precision pattern matching in order to recognize necessary information such as the position of an image.

[0002]

2. Description of the Related Art One of the conventional techniques in pattern matching is a template matching method. In this method, template data can be taught only by pointing the corresponding partial image (template). However, when accuracy is a requirement, the template size becomes large, and large-scale image calculation hardware is indispensable.

The following equation 1 shows a template matching operation equation at a certain operation point (coordinate).

[0004]

(Equation 1) T in Equation 1 is a template shown on the left side of FIG.
i · j indicates a position on the template T. Therefore, T
(I, j) indicates the gray value at the position of i · j on the template T.

[0005] G is a target image shown on the right side of FIG.
Similarly, x and y indicate the position of a certain operation point, and i and j indicate the offset position on the template T from the certain operation point. Therefore, G (x + j, y + j) is the image (x + i) · (y +
The shading value at the position j) is shown.

[0006] In addition, Θ generally indicates matching operations such as addition, multiplication, exclusive OR, and normalized correlation. The template matching is usually performed for all points on the image. Therefore, the Θ operation is performed by the number of O (MNXY) times. Here, M / N is the template size, X
Y is the image size. If a large template is used to ensure accuracy, M · N increases. If the amount of calculation is roughly calculated for a 500 × 500 image using a 50 × 50 template, the processing capacity required for the hardware is approximately 5000 M
IPS (Million instructions Per Second).
Currently, the fastest microprocessor is about 500 MIPS
(Approximate value), dedicated hardware is indispensable.

FIG. 6 shows a conventional example in which the operation cost of the ordinary template matching as described above is improved. O is calculated by thinning out the calculation positions set in the template T.
(MN) is reduced. Also, O (X · Y) is reduced by thinning out calculation points on the image G. Depending on the decimation, the matching operation with the fastest microprocessor can be performed without using dedicated hardware.

[0008]

However, in the above-described thinning method, the correct answer is overlooked because the actual target point is located at the center between the thinning operation points. Alternatively, template T
Due to the regular thinning, unique feature points are thinned, and the number of erroneous recognition increases. Such dangers are likely to occur.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a low-cost, high-speed, and high-precision pattern matching process using a small number of points arranged only on a partial feature shape that matches a pattern to be recognized on a template. It is an object to provide a method and an apparatus thereof.

[0010]

According to the pattern matching method of the present invention, an image data obtained by capturing an object is subjected to spatial primary in eight directions by using a specific partial pattern necessary for image recognition of information on the object. Differentiation is performed, and a rough search pattern matching is performed using an 8-bit search template with a point having a large differential value set to 1 on each primary differentiation plane, and the detected rough matching pattern is evaluated for a correct answer and an incorrect answer. From the incorrect answer pattern, the feature shape data included in the correct answer pattern but not included in the incorrect answer pattern and the prohibited feature shape data included in the incorrect answer pattern and not included in the correct answer pattern are extracted, and the correct answer, Feature shape pattern and prohibited feature shape pattern that have been optimized to correct for excess or deficiency of detection information for detecting incorrect pattern The optimal rough search template having the position data of the selected feature is set, and the matching with the image data is calculated only at each set position of the feature shape pattern and the prohibited feature shape pattern on the optimal rough search template, and the prohibited feature shape is calculated. One of the features is to detect image data matched with a characteristic shape pattern so that necessary information of a target object is image-recognized while eliminating further processing operations of image data matched with a pattern.

[0011] According to this, for the image data obtained by imaging the object, a predetermined calculation location and a predetermined calculation position are set in advance by a search template set for a specific pattern necessary for image recognition of information such as the position of the object. By performing the coarse search pattern matching based on the pattern matching threshold, a rough matching pattern can be obtained earlier by an amount corresponding to the roughening of the calculation location and the pattern matching threshold.
The obtained rough matching pattern includes a correct answer pattern and an incorrect answer pattern by roughening the pattern matching threshold. However, correct or incorrect patterns can be easily evaluated by humans. From the evaluated correct and incorrect answer patterns,
By extracting the feature shape data included in the correct answer pattern but not included in the incorrect answer pattern and the prohibited feature shape data included in the incorrect answer pattern and not included in the correct answer pattern, the number of operation data can be thinned out. By performing optimization processing to correct the excess or deficiency of detection information for detecting correct and incorrect patterns, the number of operation data is small, but enough feature shape patterns and prohibitions to detect correct and incorrect patterns It is possible to set an optimal coarse search template having the position data of the characteristic shape pattern. Therefore, by performing a matching operation with the image data only at a small number of setting positions based on the feature shape pattern and the prohibited feature shape pattern on the optimal rough search template, it is possible to further operate the image data matched with the prohibited feature pattern. In spite of the omission, the detection of the matching image data with the feature pattern and the recognition of the necessary information of the target image can be achieved at low cost, at high speed and with high accuracy.

In the optimizing process, it is preferable to add or remove position data which is indispensable to detection of correct and incorrect patterns in feature shape pattern matching with the optimal coarse search template, or which is obstructive. The template is preferably generated from correct and incorrect answer patterns collected by repeatedly performing rough search pattern matching on a plurality of objects.

For a fine feature portion or a portion that is very similar to a non-recognized portion that is likely to be erroneously detected by the feature shape pattern matching with the optimum rough search template, a fine search template matching this portion is used as compared with the case of the coarse search pattern matching. If a fine search pattern matching is performed at a dense calculation location and image data satisfying both the feature shape pattern matching and the fine search pattern matching is detected so that necessary information of the object is recognized, an optimal coarse search template is obtained. While making use of the above features by the feature shape pattern matching described above, it is possible to secure the necessary accuracy of the detection by performing a fine search pattern matching only on a portion that is erroneously detected in the feature shape pattern matching.

The pattern matching apparatus according to the present invention comprises an image storage means for storing input image data of an imaged object, and a search matching a specific pattern necessary for image recognition of necessary information of the object. A coarse search pattern teaching unit that teaches a template, and a coarse search pattern matching performed on image data from the image storage unit based on a predetermined calculation location and a pattern matching threshold set in advance by the taught coarse search template. A rough search processing means for storing and displaying the matching pattern, and a correct pattern based on the evaluation of the correct and incorrect answers of the detected rough matching pattern by a person and an incorrect pattern, which are included in the correct pattern and included in the incorrect pattern. Feature pattern data and correct answer patterns included in incorrect answer patterns Optimization that extracts and displays prohibited feature shape data that is not included in the data and corrects the feature data and prohibited feature data based on the teaching of excess or deficiency of information for detecting correct and incorrect answer patterns based on the data. A coarse search template optimization processing unit that generates an optimal coarse search template having position data of the processed characteristic shape pattern and the prohibited characteristic shape pattern,
The coarse search processing means performs pattern matching calculation with the image data only at each set position of the feature shape pattern and the prohibited feature shape pattern on the optimal rough search template, and as a result of the calculation, the image data matched with the prohibited feature pattern. One feature is that image data matching a feature pattern is detected so that necessary information of a target object is image-recognized while omitting any further processing operations.

Thus, the rough search pattern matching process using the taught search template and the correct and incorrect answer patterns based on the correct and incorrect answer evaluations of the rough matching pattern detected by the rough search pattern matching are performed. , The above-mentioned specific shape data and the prohibited specific shape data are obtained, and the above-mentioned optimum coarse search template is generated by adding and deleting the position data by an optimization process based on the teaching of the excess and deficiency of the information. Automatically performing the feature shape pattern matching according to the above, and the above-described pattern matching method can be easily and accurately achieved.

A fine search teaching means for teaching a fine search template suitable for a fine feature portion which is likely to be erroneously detected in the feature shape pattern matching by the optimum coarse search template or a portion very similar to the recognition target portion, and a fine search template taught. A fine search processing means for performing a fine search pattern matching at a denser calculation position than in the case of the coarse search pattern matching, and detecting a pattern satisfying both the characteristic shape pattern matching and the fine search pattern matching to detect an object; If the necessary information is image-recognized, pattern matching with the necessary accuracy is automatically and easily performed by fine search pattern matching in parts where misdetection is likely to occur in feature shape pattern matching while utilizing the features of feature shape pattern matching. Accurately achieved It is.

Further objects and features of the present invention will become apparent from the following detailed description and drawings. Each feature of the present invention can be used alone or in various combinations in combination as far as possible.

[0018]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of a pattern matching method and apparatus according to the present invention will be described with reference to FIGS. 1 to 4 together with examples to provide an understanding of the present invention.

FIG. 1 shows one embodiment of the pattern matching method of the present invention. Image data 1 shown in FIG.
a is data captured by the camera 100 as an imaging system shown in FIG.
, And performs spatial primary differentiation in eight directions with a specific partial pattern required for image recognition of information such as the position of the target object 1, and finds a point having a large differential value on each primary differential plane. Rough search pattern matching is performed using an 8-bit search template set to 1 ″.

As a result, the rough matching pattern detected as a match with the rough search template 2 includes the specific pattern and a pattern similar to the specific pattern because the pattern matching threshold is relaxed. More than one is often detected.

Therefore, as shown in FIG. 1A, a person evaluates the correct / incorrect answer for the matching pattern detected by the rough search pattern matching, and
And incorrect answer patterns H1, H2,.

If there is no correct answer pattern T, the rough search pattern matching may be repeated until the correct answer pattern T is obtained, while sequentially changing the calculation locations of the rough search pattern matching. Even in this case, the probability of obtaining the correct pattern T earlier than in the case of performing the fine search pattern matching for all the pixels of the image data 1a is higher,
Work costs are greatly reduced.

On the other hand, the specific patterns of the individual objects 1 do not always match exactly, and often have variations. In order to deal with this, in the present embodiment,
For a plurality of objects 1, to be more specific, coarse search pattern matching is performed on as many objects 1 as necessary to obtain the necessary pattern matching accuracy and image recognition accuracy, and the obtained correct patterns T each time are obtained. Incorrect answer pattern H1, H
Collect 2, ...

From the obtained correct patterns T and incorrect patterns H1, H2,... As shown in FIG.
, H2,... And the correct answer pattern T included in the incorrect answer patterns H1, H2,.
... and the prohibited feature shape data not included in.

More specifically, an OR image of all the correct patterns and all the incorrect patterns is calculated, and the characteristic shape data included only in the specific pattern of the object 1 is calculated based on the difference between the OR image and each correct pattern. Is extracted, and forbidden feature shape data included only in a pattern similar to a specific pattern of the object 1 is extracted based on an image difference between the OR image and each candidate pattern. As a result, characteristic shape data and prohibited characteristic shape data corresponding to a specific pattern having a variation in each object 1 are obtained.

For the sake of simplicity, the number of detection positions is extremely small (for example, about 1/30), and a specific example is shown only by the relationship between the correct answer pattern T and the incorrect answer pattern H1 shown in FIG. In other words, both OR images are obtained as P1 shown in FIG. Depending on whether or not there is a difference between the OR image P1 in FIG. 1B and the correct pattern T in FIG. 1A, characteristic shape data P2 included only in the specific pattern shown in FIG. 1C is obtained. . Also, the OR of FIG.
Prohibition feature shape data P3 included only in the specific pattern shown in FIG. 1D is obtained based on the presence or absence of a difference between the image and the incorrect pattern H1 in FIG.

Therefore, the feature shape data P2 is useful for detecting a correct pattern T corresponding to a specific pattern of the object 1 by pattern matching with the image data 1a, and the prohibited feature shape data P3 is used for matching with the image data 1a. The pattern matching is useful for detecting, as noise, an incorrect answer pattern H1 which is similar to the specific pattern T of the target object 1 and is likely to be erroneously detected.

Therefore, in the present embodiment, the initial optimum coarse shape including the characteristic shape data and the prohibited characteristic shape data, and showing the operation points by black circles and solid triangles in FIG. A search template T1 is set.

By performing pattern matching with the image data 1a using the initial optimum coarse search template T1 set by such a method, the prohibited feature shape data P
3 are incorrect answer patterns H1, H2,...
It is low cost to detect a specific pattern matched with the characteristic shape pattern P2, that is, the correct answer pattern T and perform image recognition of necessary position information of the target object 1 while omitting further operations for High speed and high accuracy.

However, the characteristic shape data P2 and the forbidden characteristic shape data P3 of the set initial optimal coarse search template T1 lack position data indispensable for fulfilling the role of the respective detections or include position data that is disturbing. There are times when they do.

In order to cope with this, in the present embodiment, the excess or deficiency of the information is evaluated for each of the data P2 and P3 of the feature shape and the prohibited feature shape which are the respective calculation points of the initial optimum coarse search template T1. Corresponding to this evaluation result, calculation points P2a and P3a indicated by broken-line circles and triangles that are insufficient for the initial optimum coarse search template T1 are added, and the calculation points that cause problems are deleted. The above-described pattern matching with the image data 1a is performed by using the obtained optimum rough search template T2, so that it is possible to cope with a case where the position data of the initial optimum rough search template T1 is excessive or insufficient.

Additional calculation points P2a and P3a for the characteristic shape patterns P2 and P2a and the prohibited characteristic shape patterns P3 and P3a set in the optimal rough search template T2.
Are the feature shape data P2 and the prohibited feature shape data P3
Is extracted, and a portion that becomes an empty calculation point can be distributed to both sides and discretely set.

Such an optimal coarse search template T2
Is optimally set for the accuracy of pattern matching.
Correct answer pattern T, as well as incorrect answer pattern H1,
Are also preferable as the number of detections is small, and the correct answer pattern T and the incorrect answer patterns H1, H2,.
It is preferable to gradually change the matching degree threshold value in the coarse search pattern matching for detecting, and use the result obtained when convergence is performed so as to minimize the number of detections.

Here, the operation for obtaining the optimal coarse search template T2 is performed by averaging the coincidence between the collected image of the correct answer pattern T ··· and the optimal coarse search template T2 and the incorrect answer patterns H1, H2, ···. A thinning operation in which the difference between the image of the part and the average degree of coincidence between the optimal coarse search template T2 is maximized. In the actual pattern matching process using the optimal rough search template T2, this condition may be broken due to a difference in the production lot of the same kind of target object 1, and NG may occur in the detection. In such a case, it can be dealt with by setting the optimum rough search template T2 again as shown in the processing flow of FIG. In addition,
The object 1 in FIG. 2 is a gear meshing portion in a video tape recorder or the like.

By the way, the specific pattern of the object 1 includes a minute feature portion which is easily erroneously detected by the feature shape pattern matching and a portion which is very similar to the recognition target portion.
In such a case, a fine search template that matches the pattern of those parts is set, and the fine search template is used to perform fine search pattern matching with a denser calculation point than in the case of feature shape pattern matching,
A feature pattern that satisfies both the feature shape pattern matching and the fine search pattern matching, that is, a correct answer pattern T, is detected, and necessary information such as the position of the object 1 can be image-recognized. Thus, while utilizing the features obtained by the feature shape pattern matching, it is possible to perform the precise search pattern matching only on the portions that are erroneously detected in the feature shape pattern matching, thereby ensuring the required accuracy. The number of calculation locations in the fine search pattern matching using the fine search template may be set more than in the case of the feature shape pattern matching. However, depending on the required accuracy, the calculation accuracy can be maximized by calculating for all pixels. it can. Even in this case, since fine search pattern matching is performed only for a part of a specific pattern, the number of calculation points is not particularly large and the pattern matching operation does not take a long time.

The pattern matching apparatus according to the present embodiment has a configuration as shown in FIG. 3 for realizing the above-described pattern matching method. The target object 1 is supplied to, or conveyed or mounted on, a circuit board to manufacture an electronic circuit board, an electronic component to be mounted, a gear meshing portion in a video tape recorder, or the like, a spring hooking portion. And so on.

The camera 100 includes an image pickup device, an image pickup circuit, and an image pickup optical system, and is installed in a process portion where necessary information such as the position and type of the object 1 is to be image-recognized. Image data 1a of a grayscale image of the object 1 captured by the camera 100
For example, the grayscale image input unit 101 converts the digital image data into digital image data 110 in a unit to be subjected to pattern matching, and inputs the digital image data 110 to the grayscale image storage unit 102.
22.

The grayscale image storage unit 102 stores the input image data and outputs it in response to requests from the coarse search processing unit 103 and the fine search processing unit 104. The coarse search processing unit 103 calculates the spatial first derivative of the image data 111 from the grayscale image storage unit 102 in eight directions based on the coarse search template data 113 taught through the coarse search teaching unit 105 based on the external input 131 relating to the search template. A point having a large differential value is set to 1 on each primary differential plane to obtain an 8-bit image. Similarly, spatial primary differentiation in eight directions is performed, and coarse search pattern matching is first performed using an 8-bit search template in which a point having a large differential value is set to 1 on each primary differentiation plane. The rough search processing unit 103 also displays a matching pattern detected by the rough search template data 113 on a monitor via the video controller 121, and evaluates the correct answer pattern T,... And the incorrect answer patterns H1, H2,. To serve.

The evaluation result 131 made by the person is taught to the rough search teaching unit 105, and the rough search teaching unit 105 outputs the data of the taught correct patterns T... And the incorrect answer patterns H1, H2,. From the data, the correct answer pattern T
Are included in the incorrect answer patterns H1, H2,..., And are included in the incorrect answer patterns H1, H2,.
The initial optimal coarse search template T1 having the forbidden feature shape data P3 obtained by thinning out incorrect patterns not included in the correct pattern T... Is automatically generated by the above operation.

The automatically generated initial optimum coarse search template data T1 is used for the coarse search template optimization processing unit 10.
7 and displayed on a monitor 122 through a video controller 121.

The display on the monitor 122 is used to evaluate whether the number of calculation points for feature detection set in the initial optimum coarse search template data T1 is excessive or insufficient. The coarse search template optimization processing unit 107 adds indispensable calculation points to the automatically generated initial optimum coarse search template data T1 based on the input of the evaluation result 132 of the excess or deficiency of the calculation points, and deletes disturbing calculation points. Is performed to automatically generate the optimal coarse search template data T2, which is input to the coarse search processing unit 103 through the coarse search teaching unit 105.

Here, the coarse search processing unit 103 thereafter performs low-cost, high-speed, and high-precision pattern matching at a small number of calculation points by coarse pattern matching with the image data 1a using the given optimum coarse search template T2. Can be achieved automatically.

On the other hand, the fine search processing unit 104 receives the fine search template data 1 taught from the fine search teaching unit 106 based on the external input 133 relating to the fine search template.
14, the image data 1a from the grayscale image storage unit 102
Then, fine pattern matching is performed on all pixels in the fine search pattern matching target area or on calculation points that are set more secretly than in the case of coarse search pattern matching.

Thus, the comprehensive search processing unit 12 composed of the coarse search processing unit 103 and the fine search processing unit 104
3 detects image data that satisfies both the coarse search pattern matching and the fine search pattern matching so that necessary information of the object 1 can be image-recognized. In addition,
FIG. 2 is a flowchart schematically illustrating a processing flow involving generation of the above-described optimal coarse search template.

The automatic generation of the optimized rough search template is performed, for example, by the simulated annealing method (S
A: Simulated Annealing (simulated slow cooling method) is used. An example of this is shown in FIG.

The incorrect answer patterns H1, H2,..., H in FIG.
The calculation between the candidate pattern n and the evaluation pattern T as the correct answer pattern calculates the degree of coincidence between the two patterns.
This is one of the general-purpose approximate solutions to the optimization problem. S
A specific example of the predetermined procedure of the method A is as follows. (1) Initialization of evaluation pattern calculation point sequence, temperature, slow cooling stage (2) Calculation of energy (degree of coincidence between candidate pattern and evaluation pattern) (3) Add perturbation (change position of calculation point) (4) ) Calculation of energy (5) Resin in state (replace perturbation pattern with evaluation pattern) (6) Repeat specified number of times from (3) (7) Slow cooling (update of temperature) (8) Certain conditions (temperature, energy ) Is repeated until (3) is satisfied. Each operation point is extracted from the evaluation pattern and the candidate pattern, and is converged to a rough search template including the optimum operation points by using the SA method. Conditions for obtaining convergence in the SA method are shown in Table 1 below as specifications for automatically generating a rough search template.

[0047]

[Table 1] There are a plurality of conditions under which convergence can be obtained. Further, the perturbation-slow cooling stop condition is a condition in which even if the perturbation is performed, the degree of coincidence cannot be further improved. In the case of Table 1, the coincidence threshold is set from the initial threshold Ts = 30.0 to the final threshold Tf =
Increased to 0.1. The setting of the convergence conditions for this is done using experiments and experiences from each position,
It is sufficient to reduce the threshold value by 1/2 and repeat the following processes from (3) until the limit cannot be further improved. The upper limit of the threshold update frequency is considered to be about 300 times, but is not limited thereto. The calculation points are extracted in a matrix from the reference pixel in the pixel group to the vertical and horizontal directions.
The calculation pixels may be set in the directions, or the calculation pixels may be set in eight directions obtained by adding the diagonal directions to the calculation pixels.

In the case where the final threshold value is 0.1, the number of final calculation points is reduced to 32 points for a template of 50 × 50 pixels, and a large number of thinnings are performed to obtain a highly accurate pattern matching result.

[0049]

According to the present invention, the excess or deficiency of the detected information is corrected in the characteristic shape data and the prohibited characteristic shape data of the specific part relating to the correct answer pattern and the incorrect answer pattern similar thereto and easily detected by mistake. The position data of the feature shape pattern and the forbidden feature shape pattern which are small in the number of operation data but sufficient for detecting the correct answer pattern and the incorrect answer pattern only by performing the optimization processing are set in one optimal rough search template. By performing a matching operation with the image data only at a small number of setting positions based on the feature shape pattern and the prohibited feature shape pattern on the optimal rough search template, further operation of the image data matched with the prohibited feature shape pattern can be performed. While omitting, detect a specific pattern that matches the feature shape pattern and Making the necessary information to allow the image recognition is achieved with high speed and high accuracy at low cost.

For a fine feature portion that is likely to be erroneously detected in the feature shape pattern matching with the optimal rough search template and a portion that is very similar to the non-recognized portion, a fine search template matching this portion is used to achieve a higher density than in the coarse search pattern matching. Perform precise search pattern matching at the appropriate calculation locations, detect image data that satisfies both the feature shape pattern matching and the fine search pattern matching, and enable image recognition of the necessary information of the target object.
While making use of the above-mentioned features obtained by the feature shape pattern matching in the optimal rough search template, it is possible to secure the necessary accuracy of the detection by performing the fine search pattern matching only on the portion that is erroneously detected in the feature shape pattern matching.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram showing one example of a pattern matching method according to an embodiment of the present invention.

FIG. 2 is a flowchart schematically showing a processing flow involving generation of an optimal coarse search template in the pattern matching method of FIG. 1;

FIG. 3 is a block diagram showing one example of a pattern matching apparatus according to an embodiment of the present invention that realizes the pattern matching method of FIG. 1;

FIG. 4 is a flowchart showing a processing flow of the SA method employed in the processing flow of FIG. 2;

FIG. 5 is an explanatory diagram schematically showing a template matching method in a conventional pattern matching method.

FIG. 6 is an explanatory diagram schematically showing a thinning template matching method in a conventional pattern matching method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Object 1a Image data 2 Search template 100 Imaging system 101 Gray-scale image input unit 102 Gray-scale image storage unit 103 Coarse search processing unit 104 Fine search processing unit 105 Coarse search teaching unit 106 Fine search teaching unit 107 Coarse search template optimization processing unit 122 Monitor 123 Comprehensive search processing unit T Correct pattern H1, H2 Incorrect pattern P1 OR image P2 Feature shape data P3 Prohibited feature shape data T1 Initial optimal coarse search template T2 Optimal coarse search template P2a, P3a Additional calculation points

Claims (6)

[Claims] An image data obtained by imaging an object is subjected to spatial primary differentiation in eight directions by using a specific partial pattern necessary for recognizing the information of the object, and a point having a large differential value is determined for each primary. On the differential plane, rough search pattern matching is performed using an 8-bit search template set to “1”, correct / incorrect evaluation is performed on the detected rough matching pattern, and the evaluated correct and incorrect patterns are converted into correct patterns. The feature shape data included and not included in the incorrect answer pattern and the prohibited feature shape data included in the incorrect answer pattern and not included in the correct answer pattern are extracted, and the detection information for detecting the correct / incorrect answer pattern is extracted from them. Optimized rough search template with position data of feature shape pattern and prohibition feature shape pattern that have been optimized to correct excess or deficiency A matching operation with the image data is performed only at each set position of the feature shape pattern and the prohibited feature shape pattern on the optimal rough search template, and further processing of the image data matched with the prohibited feature shape pattern is performed. A pattern matching method for detecting image data matched with a characteristic shape pattern while eliminating operations, and recognizing an image of necessary information of the target object. 2. The optimization process according to claim 1, wherein the feature shape data and the prohibited feature shape data are added with position data indispensable for detecting a correct answer / incorrect answer pattern in feature shape pattern matching to remove unnecessary position data. Item 1. The pattern matching method according to Item 1. 3. The pattern matching method according to claim 1, wherein the optimum rough search template is generated from correct and incorrect answer patterns collected by repeatedly performing rough search pattern matching on a plurality of objects. 4. A fine feature portion which is likely to be erroneously detected by a feature shape pattern matching using an optimal rough search template or a portion which is very similar to a recognition target portion is compared with a feature shape pattern matching by a fine search template matching the portion. Performs fine search pattern matching at dense calculation locations, detects image data that satisfies both the feature shape pattern matching and the fine search pattern matching, and recognizes necessary information of the target object.
The pattern matching method according to claim 3. 5. An image storage means for storing input image data of an imaged object, and a spatial partial derivative in eight directions in a specific partial pattern required for image recognition of necessary information of the object. A coarse search template teaching means for performing a coarse search pattern matching using an 8-bit search template with a point having a large differential value set to 1 on each primary differential plane and teaching a search template matching the pattern, and a coarse search template taught. Coarse search processing means for performing a coarse search pattern matching on the image data from the image storage means based on a predetermined calculation location and a pattern matching threshold set in advance and storing and displaying a detected rough matching pattern; Correct putter based on evaluation of correct / incorrect answer And extracting and displaying the feature shape data included in the correct answer pattern but not included in the incorrect answer pattern and the prohibited feature shape data included in the incorrect answer pattern and not included in the correct answer pattern, and the correct answer based on them. Optimization with correction processing of feature data and prohibition feature data based on teaching of excess and deficiency of information for detection of patterns and incorrect answer patterns Optimization with feature shape pattern and prohibition feature shape pattern position data A coarse search template optimizing processing unit for generating a coarse search template, wherein the coarse search processing means is configured to compare the image data only at each set position of the characteristic shape pattern and the prohibited characteristic shape pattern on the optimal coarse search template. Performs a pattern matching operation and results in an image that matches the prohibited feature pattern While omitting the further processing operations on over data, image recognizing pattern matching apparatus necessary information of the object by detecting the image data to feature pattern matching. 6. A fine search teaching means for teaching a fine search template suitable for a fine feature portion or a portion very similar to a recognition target portion, which is likely to be erroneously detected in the feature shape pattern matching using an optimum coarse search template, A fine search processing means for performing a fine search pattern matching with image data at a denser calculation location than in the case of the coarse search pattern matching by using a template, and a pattern satisfying both the feature shape pattern matching and the fine search pattern matching is provided. The pattern matching apparatus according to claim 5, wherein the pattern matching apparatus detects and recognizes necessary information of the target image.