JPH07311816A - Pattern recognizing method - Google Patents

Abstract

PURPOSE:To provide a pattern recognizing method which can clearly discrimi nate and recognize with high accuracy even many patterns that have their entire shapes approximate to each other. CONSTITUTION:When a subject pattern is recognized by the pattern matching, the subject pattern and a template are divided into plural blocks corresponding to each other. Then, the relative value of each block is calculated and the minimum relative value is extracted. The minimum block relative value are compared with each other among plural reference patterns, and the subject pattern is discriminated based on the result of this comparison.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method for discriminating a target pattern based on matching of a target pattern such as a character image with various reference patterns.

[0002]

2. Description of the Related Art Conventionally, a method of recognizing characters and the like by pattern matching in which a target pattern and various reference patterns (templates) are matched is generally known.

As a basic method of this pattern matching, the degree of coincidence (for example, a correlation value) of the entire pattern between the target pattern and the reference pattern is obtained, and the above-mentioned degree of coincidence of each of the various reference patterns is compared to determine the degree of coincidence. It is determined that the largest reference pattern corresponds to the target pattern.

However, the data of the character image which is the target pattern includes various error factors such as faint characters, the influence of illumination at the time of image pickup, and image noise. Since there are variations due to such error factors, it is difficult to accurately distinguish and recognize a pattern whose overall shape approximates only by examining the correlation value of the entire pattern.

Therefore, in order to improve the accuracy of pattern recognition, there has been considered a method of performing pattern matching only for the characteristic portion when there are several types of patterns having similar overall shapes. This method will be described with reference to FIG. 4 by taking the case of recognizing alphabet characters as an example.

As a result of pattern matching of the entire shape of the target character on the screen using templates of various characters, a correlation value is obtained when the "C" template is used and when the "O" template is used. Suppose that are almost the same. In this case, as shown in FIG. 4A, for example, if the range forming the character pattern is divided into nine areas of vertical and horizontal three columns, “O” for “C” (or “C” for “O”). ) Characteristic portion 100 is a portion of the right-side interruption indicated by hatching. Therefore, this part 100
If only pattern matching is performed, it is possible to clearly distinguish between “C” and “O”.

[0007]

However, even with the method described above, when there are many characters having similar overall shapes, it becomes difficult to extract the characteristic portion for discrimination.

That is, as shown in FIG. 4B, as a result of pattern matching of the entire shape, "C""O" is obtained.
When the correlation values using the templates of “Q”, “U”, and “D” are almost the same, for example, the characteristic part of “C” is the part 101 on the right middle stage with respect to “O”, "Q"
To the right middle and lower parts 102, "U" to the upper parts and the right middle part 103, "D" to the left parts and the right middle part 104, and so on. Therefore, in order to distinguish these characters, the part 1 corresponding to the logical sum of the characteristic parts for the above individual characters is used.
It is necessary to perform 10 pattern matchings, but this portion 110 occupies most of the entire pattern,
There is not much difference from the case where the entire pattern matching is performed. Therefore, the characteristics of each character are blurred, and it becomes difficult to clearly identify each character.

The present invention solves such a problem, and provides a pattern recognition method capable of clearly discriminating these patterns and recognizing them accurately even when there are many patterns whose overall shapes are similar. The purpose is to do.

[0010]

The invention according to claim 1 is
In the pattern recognition method for determining which reference pattern the target pattern corresponds to based on the matching of the target pattern and various reference patterns, the target pattern and the reference pattern are divided into a plurality of corresponding blocks, each of which The degree of coincidence of both patterns is obtained for each block, the minimum value of the degree of block coincidence, which is the smallest value among the degrees of coincidence of each block, is extracted, and the minimum value of the degree of block coincidence is extracted for each of the plurality of reference patterns. After that, the target pattern is determined based on the comparison of the minimum block coincidence values.

According to a second aspect of the present invention, in the first aspect of the present invention, the overall degree of coincidence between the target pattern and the reference pattern is calculated for each of the various reference patterns, and the overall degree of coincidence among the various reference patterns is determined. Those patterns that are high and are close to each other are selected, and the reference pattern having the largest block matching degree minimum value is determined from the selected reference patterns to determine the target pattern.

[0012]

According to the first aspect of the present invention, even if there are a plurality of reference patterns whose overall shape is similar to the target pattern, according to the reference pattern corresponding to the target pattern, the variation in the degree of coincidence between the blocks is large. However, the minimum value of the block matching degree does not become significantly small, whereas the other patterns other than the reference pattern have a portion different from the target pattern in at least one of the blocks. Since the values are extremely small, these are clearly discriminated.

According to the invention of claim 2,
First, the matching degree of the entire pattern is checked to select a reference pattern having a similar overall shape to the target pattern, and then the one having the largest block matching degree minimum value is determined from the selected reference patterns. By doing so, the pattern can be easily and accurately discriminated.

[0014]

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

FIG. 1 is a functional block diagram showing an example of a character recognition device to which the pattern recognition method according to the present invention is applied. In this figure, a character recognition device 1 includes an image input unit 2 for picking up an image of a character, such as a CCD camera.
An image memory 3 for storing the input image, and an image determination unit 4 for reading out the data in the image memory 3 and determining characters by processing such as pattern matching. The image determination unit 4 has an overall correlation calculation unit 5, a block correlation calculation unit 6, a similar character detection unit 7, a character determination unit 8 and an address generation unit 9.

The overall correlation calculation unit 5 and the block correlation calculation unit 6 collate the target pattern read from the image memory 3 with the template representing the reference pattern of various characters, and based on this collation, for each template, respectively. , A correlation value indicating the degree of coincidence between the target pattern and the template is obtained. In this case, the overall correlation calculating section 5 obtains an overall correlation value which is a correlation value of the entire pattern. on the other hand,
The block correlation calculation unit 6 divides the target pattern and the template into a plurality of corresponding blocks, and a block correlation value (block matching degree) that is a correlation value for each block.
The block minimum correlation value (block matching degree minimum value) that is the minimum value among the block correlation values of the plurality of blocks is extracted, and the extraction of this block minimum correlation value is performed for each template. I am supposed to do it.

The similar character detecting section 7 is based on the data of the overall correlation value obtained by the overall correlation calculating section 5,
When there are a plurality of those having a large overall correlation value and having a similar overall correlation value, the character of the template is selected, that is, the similar character (including the same character) to the character of the target pattern is selected. Specifically, as will be described in detail later, the maximum value of the overall correlation values obtained based on the above-mentioned matching using various templates is equal to or greater than a preset threshold value, and the overall correlation values other than this. If there is one that is close to the above maximum value, the characters of these templates are selected.

Further, the character determining section 8 determines a character based on the calculation data by the overall correlation calculating section 5 or the block correlation calculating section 6. Specifically, as will be described in detail later, when the maximum value of the overall correlation values is equal to or greater than a preset threshold value, and there is no other overall correlation value that approximates the maximum value, , The template character corresponding to the maximum overall correlation value is the recognition result,
That is, it is determined that this character corresponds to the character of the target pattern. On the other hand, when a plurality of similar characters are selected by the similar character detection unit 7, the block correlation minimum value for the plurality of characters is extracted from the calculation data of the block correlation calculation unit 5, and the block correlation minimum values The one with the maximum size is selected, and the characters in the template corresponding to this are used as the recognition results.

The address generator 9 generates the address of the target pattern in the image read out from the image memory 3, and when there are a plurality of target patterns (a plurality of characters), the target pattern is generated for one target pattern. When the recognition process as described above is completed, the address is updated so as to shift to the process for the next target pattern.

FIG. 2 shows an example of a method of dividing a block by the block correlation value calculation section 6 and a concrete example of a block correlation value when one target pattern is matched with various templates. .

In the example shown in this figure, the target pattern 10 and each of the templates 11 to 17 are divided into upper and lower three blocks. In this figure, Ba, Bb, and Bc are block correlation values obtained for each block based on matching between the target pattern 10 and the templates 11 to 17, and B.
1 to B7 are the minimum block correlation values when the templates 11 to 17 are used. In addition, in this figure, the character of the target pattern 10 is “O”, whereas “C” “O”
Block correlation values Ba and B when templates 11 to 17 of each character of “Q”, “U”, “D”, “X”, and “Y” are used
Specific values of b and Bc are illustrated. Although these numerical values are merely examples, they are shown so that the tendency of the block correlation value can be easily understood. The above correlation value is 1.0 in the case of perfect match.

According to the example of this figure, the overall shape is "O".
Since the characters “C”, “Q”, “U”, and “D” are similar to each other, the whole case when the templates 11 to 15 of “C”, “O”, “Q”, “U”, and “D” are used. The correlation value (average value of the block correlation values) is large and approximates each other. However, looking at the block correlation values Ba, Bb, and Bc, the template 12 having the same “O” as the target pattern 10
When using, the block correlation values Ba, Bb, and Bc vary somewhat due to the above-mentioned error factors, but the block correlation minimum value B2 is relatively large, while
Each template 11, 1 of "C""Q""U""D"
In the case of using Nos. 3, 14 and 15, the template 12 having the block correlation minimum values B1, B3, B4 and B5 of “O” is used because there is a portion different from “O” in any of the blocks. It is much smaller than it was. Therefore, the character "O" can be discriminated by comparing the block correlation minimum values.

When the templates 16 and 17 of "X" and "Y" which are dissimilar to "0" are used, the overall correlation value becomes small.

FIG. 3 is a flow chart showing an example of the pattern recognition method performed by the character recognition device 1. The procedure shown in this flowchart will be described below.

First, an image is read from the image memory 3 (step S1). Next, one target pattern in the above image is compared with templates of various characters. In this case, each character is divided into a plurality of blocks as shown in FIG. 2 and pattern matching is performed. For each of the templates, the block correlation minimum (B
, B2, ..., And the overall correlation value (A1, A2, ...) Is calculated (step S2).

Then, the respective overall correlation values (A1, A2,
..), one maximum total correlation value (Ai) is selected, and it is determined whether or not this maximum total correlation value (Ai) is larger than a preset threshold value (step S3).
S4). Here, if the maximum overall correlation value (Ai) is less than or equal to the threshold value, it is determined that the character cannot be specified, and the process proceeds to step S8 described below without specifying the character.

If it is determined that the maximum overall correlation value (Ai) is greater than the threshold value, then there is another overall correlation value (including the same value) that approximates the maximum overall correlation value (Ai). It is determined whether or not there is, that is, whether or not there is another overall correlation value whose difference from the maximum overall correlation value (Ai) is less than or equal to a predetermined value (step S5). Then, when it is determined that there is no other overall correlation value that approximates the maximum overall correlation value (Ai), the character of the template having the maximum overall correlation value (Ai) is set as the recognition result (step S6). ).

When it is determined in step S5 that there is an overall correlation value (Aj, Ak ...) That approximates the maximum overall correlation value (Ai), these maximum and approximate overall correlation values ( Each block correlation minimum value (Bi, B) according to the template corresponding to Ai, Aj, Ak ...
j, Bk ...) are compared, the maximum value is selected from these block correlation minimum values (Bi, Bj, Bk ...), and the character of the template having this maximum value is taken as the recognition result (step S7).

After the processing in step S6 or step S7 or the determination that the character cannot be specified is performed in step S4, the address for pattern matching is updated (step S8), and then all the objects within the processing range are updated. It is determined whether or not the process for the pattern is completed (step S9), and if not completed, the process returns to step S2, and step S2 is performed for the next target pattern.
The process of 2 or less is repeated.

The recognition method as described above will be described with reference to a specific example shown in FIG. If the recognition target character is “O”, the template 11 of the character and the characters similar thereto, that is, the characters “C”, “O”, “Q”, “U”, and “D” according to the example shown in FIG. The overall correlation value for ~ 15 is high,
In addition, the overall correlation value in the case of using the templates 16 and 17 of characters (“X” and “Y” in the example of FIG. 2) that are not similar to “O” and are similar to each other is low. Therefore, "C", "O", "Q", "U", "D" in the processing of steps S4 and S5.
The overall correlation value in the case of the template of is selected. In this case, since the total correlation value varies to some extent due to the error included in the image data, the total correlation value in the case of the “O” template 12 is not necessarily the maximum, but at least approximates to the maximum total correlation value. Some of what you do is included.

Then, these "C""O""Q"
The block correlation minimum values B1 to B5 in the case of using the templates 11 to 15 of “U” and “D” are compared, and in this case,
Even if the character is similar to the recognition target character, the correlation value of the block to which the part different from the recognition target character belongs is sufficiently small. Therefore, the block correlation minimum in the case of using the template 12 of “O” corresponding to the recognition target character The value B2 is certainly larger than the other block correlation minimum values. Therefore, by selecting the largest one among these block correlation minimum values B1 to B5 in the process of step S7, the character "O" corresponding to the recognition target can be correctly identified.

Further, when the maximum overall correlation value among the above-mentioned overall correlation values is larger than the threshold value and there is no approximation thereto (step S4 is YES, step S5 is NO).
In the case of), since the characters other than the character having the maximum overall correlation value are not similar to the recognition target character, the character having the maximum overall correlation value is set as the recognition result.

In the above embodiment, the target pattern and each template are divided into upper and lower three blocks to obtain the block correlation value. However, the number of blocks and the dividing method are not limited to this, and similar blocks of similar characters can be obtained. The division may be performed so that the different portion is sufficiently reflected in the block correlation value, and the variation in the block correlation value does not become extremely large due to the error factor of the image data.

In the above embodiment, the similar character is selected based on the comparison of the overall correlation values, and the block correlation minimum value is compared for the selected ones. However, all the similar characters are not selected. It is also possible to select the largest block correlation minimum value and use that character as the recognition result.

[0035]

As described above, according to the pattern recognition method of the present invention, the target pattern and the reference pattern are divided into a plurality of blocks corresponding to each other, the degree of coincidence of each block is calculated, and the minimum value of the degree of block coincidence is calculated. Is extracted and the target pattern is determined based on the comparison of the block matching degree minimum values of a plurality of reference patterns, so that there are many reference patterns whose overall shape is similar to the target pattern. Even in this case, when the reference pattern having a portion different from the target pattern is used, the minimum value of the block matching degree is remarkably reduced. Therefore, it is clarified that this case and the case of using the reference pattern corresponding to the target pattern. Can be distinguished. Therefore, the target pattern can be accurately determined.

In particular, first, the degree of coincidence of the whole is obtained using various reference patterns, and the ones having a high degree of coincidence and being close to each other are selected from the various reference patterns, and from the selected reference patterns. If the block matching degree minimum value is specified to be the largest, pattern determination can be performed easily and accurately.

[Brief description of drawings]

FIG. 1 is a functional block diagram showing an example of a character recognition device to which a pattern recognition method of the present invention is applied.

FIG. 2 is an explanatory diagram showing an example of a block division method and a specific example of a block correlation value.

FIG. 3 is a flowchart showing an example of the method of the present invention.

4A and 4B are explanatory diagrams showing a conventional method and its problems.

[Explanation of symbols]

 4 Image determination unit 5 Overall correlation calculation unit 6 Block correlation calculation unit 7 Similar character detection unit 8 Character determination unit

Claims (2)

[Claims] 1. A pattern recognition method for determining which reference pattern a target pattern corresponds to based on a comparison between a target pattern and various reference patterns, wherein the target pattern and the reference pattern are divided into a plurality of corresponding blocks. Divide and obtain the degree of coincidence of both patterns for each block, extract the minimum value of the block coincidence degree that is the smallest value of the coincidence degree of each block, and extract the minimum value of this block coincidence degree A pattern recognition method, characterized in that after performing each reference pattern, the target pattern is determined based on comparison of respective block matching minimum values. 2. The degree of coincidence between the target pattern and the reference pattern is calculated for each of the various reference patterns, and the ones having a high degree of overall coincidence and being close to each other are selected from among the various reference patterns and selected. 2. The pattern recognition method according to claim 1, wherein the target pattern is determined from the reference patterns having the largest block matching degree minimum value.