JP3475422B2 - Image recognition device, image recognition method, and recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image recognition apparatus, an image recognition method, and a recording medium having a program for executing the image recognition method recorded therein, and more particularly, it has a plurality of categories and each category for input data. In image recognition that calculates the degree of matching (or similar information), the recognition result that the degree of matching is high cannot be determined whether it is correct or incorrect, "we accept because the recognition result is confident that it is correct" , "Reject because the recognition result is not sure that it is correct", "Reject because the recognition result is undetermined", or "Reject because the recognition result is incorrect", etc.
The present invention relates to an image recognition device, an image recognition method, and a recording medium that improve the acceptance correct rate and make the false recognition rate close to 0 [%].

[0002]

2. Description of the Related Art Conventionally, in an image recognition system having a plurality of categories and provided with means for calculating the degree of coincidence (or similar information) of each category with respect to input data, the recognition result of image recognition is It could not be determined until a person checked whether was the correct answer or a false recognition.

That is, the operation method in the conventional image recognition system is performed by the following procedure. That is, first, the image recognition system is made to recognize the recognition target data, and secondly, the recognition result is displayed on the screen of a computer or the like, and a person matches the recognition target with the recognition result and confirms the result. Thirdly, the erroneously recognized result is corrected, and fourthly, the recognition result is finalized after the matching confirmation of the recognition target is completed to the end.

[0004]

However, in the operation method of the conventional image recognition system described above, although the labor saving is achieved as compared with the case where the whole sentence is manually input, the second manual recognition is performed. There is a problem that it takes a lot of time to confirm the matching between the object and the recognition result.
For example, when recognizing hundreds or thousands of forms at a time, it is difficult to say that it is automated if a person checks each one after recognizing it, so the introduction of an image recognition system is postponed. There were many things.

Further, there is also a problem with the recognition performance, and if all the results recognized by the conventional image recognition system are accepted without checking, the correct answer rate is about 80 to 95 [%] at present. In other words, 5 to 20% is accepted as a correct answer even though it is erroneously recognized.

The present invention has been made by paying attention to such a conventional problem, and an object thereof is to have a plurality of categories and the degree of coincidence of each category with input data (or information similar thereto). In the image recognition that calculates, the recognition result that cannot be determined whether the result with a high degree of coincidence is correct or incorrect is "accepted because the recognition result is confident that it is correct", "recognition result is not correct" Therefore, it will be rejected "," Rejected because the recognition result is undetermined ", or" Rejected because the recognition result is incorrect ", etc., to improve the acceptance rate and correct the false recognition rate. [%]
To provide an image recognition device, an image recognition method and a recording medium which are close to the above.

[0007]

In order to solve the above-mentioned problems, the invention according to claim 1 of the present application has recognition objects of a plurality of categories, and conforms in pattern matching between a registered image pattern and an input image pattern. In an image recognition device that recognizes an input image based on a degree (distance), data having the same first candidate category as a recognition result is collected from data including a recognition result of image recognition and its accompanying information, and the collected data A first pre-stage knowledge creating means for extracting a high-presence rate of similar categories in specific candidate ranks below the second candidate and holding the same as first pre-stage knowledge indicating the relationship of the similar categories to the first candidate category; when similar category specification candidate ranking recognition result as follows first candidate category Oyo <br/> beauty second candidate input image is obtained, the first candidate catheter It is held in Li and the first pre-knowledge similar category of second candidate following JP <br/> constant candidate ranking
The image recognition apparatus is characterized by further comprising a first pre-stage determination means for determining whether or not the category of the first candidate is “correct answer” by comparison with the established association.

The invention according to claim 2 of the present application is
It has recognition targets of multiple categories and inputs registered image patterns.
Goodness of fit in pattern matching with the image pattern (distance
Image recognition device that recognizes the input image based on
Data including the recognition result of image recognition and its accompanying information.
Have the same first candidate category as the recognition result.
Collect data, and from the collected dataFirst candidate category
From the goodness of fit (distance) inCategory of the first candidate
Calculates the degree of goodness of fit (distance) that can be determined as "correct answer"
Second pre-stage knowledge creating means for holding as second pre-stage knowledge
And as the recognition result of the input imageFirst candidateGot the category of
WhenOf the first candidateGoodness of fit (distance
The distance and the second knowledgeRetainedGoodness of fit
By comparing with the size of (distance),First candidate catego
Li"Correct answer"Whether or notA second front-stage judging means for judging
An image recognition device characterized in that it is provided.

The invention according to claim 3 of the present application is
An image recognition apparatus that has a plurality of categories of recognition targets and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, and includes data including a recognition result of the image recognition and its accompanying information. From the data, the data having the same first candidate category as the recognition result is collected, and from the collected data, from the difference in the matching degree (distance) between the category of the first candidate and the similar category of the specific candidate ranks below the second candidate , A third pre-stage knowledge creating unit that calculates the magnitude of the goodness-of-fit (distance) difference that can determine the category of the first candidate as “correct answer” and holds it as the third pre-stage knowledge, and the first candidate as the recognition result of the input image. categories and the second candidate
When similar categories with the following specific candidate ranks are obtained,
First candidate category and adaptability of the kind <br/> similar categories following specific candidate ranking second candidate (distance) difference in size between said held in the third front knowledge fit (distance) the size of the difference The image recognition apparatus is characterized by further comprising a third pre-stage determination means for determining whether or not the category of the first candidate is “correct answer” by comparison with the above.

The invention according to claim 4 of the present application is
An image recognition apparatus that has a plurality of categories of recognition targets and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, and includes data including a recognition result of the image recognition and its accompanying information. From the data whose recognition result is “correct” in a certain category, the axis when clustering the set of the above data is used as the upper candidate, and the category with a high appearance rate is used as the axis and appears in the lower candidate. With the category having a high rate as the axis, the category having a high appearance rate in the upper candidates and the category having a high appearance rate in the lower candidates are combined as an axis,
A first axis extracting means for extracting a category having a high appearance rate in the middle candidates as the axis, categories having a high correlation coefficient as the axis, or categories having a low correlation coefficient as the axis; First post-stage knowledge creation in which clustering of the set of data is performed based on the axes extracted by the first axis extraction means to generate OK clusters of the category, and the shape of the generated OK cluster is retained as first post-stage knowledge. And a data including a recognition result and its accompanying information as a recognition result of the input image, it is checked whether or not the data is included in the OK cluster belonging to the category having the first latter knowledge. If it is included in the OK cluster, the category to which the OK cluster belongs is determined to be the “correct answer” as the recognition result of the input image, and any OK cluster is included. Rare in the case not is in the image recognition apparatus characterized by comprising a first second-stage determination means determines that the "undetermined".

The invention according to claim 5 of the present application is
The first post-stage determination means calculates a certainty factor that is an index indicating how much certainty the "correct answer" can be called when the data is included in an OK cluster belonging to a certain category. A means is provided, and when the data enters an OK cluster belonging to a plurality of different categories, which category is determined as "correct answer" is determined based on the certainty factor. The image recognition device described in 1.

The invention according to claim 6 of the present application is
An image recognition apparatus that has a plurality of categories of recognition targets and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, and includes data including a recognition result of the image recognition and its accompanying information. From the data whose recognition result is "misidentified" in a certain category, the axis when clustering the set of the above data is used as the upper candidate, and the category with a high appearance rate is used as the axis and appears in the lower candidate. With the category having a high rate as the axis, the category having a high appearance rate in the upper candidates and the category having a high appearance rate in the lower candidates are combined as an axis,
A second axis extracting means for extracting a category having a high appearance rate as a middle candidate as the axis, categories having a high correlation coefficient as the axis, or categories having a low correlation coefficient as the axis; Second post-stage knowledge creation in which clustering of the set of data is performed based on the axes extracted by the second axis extraction means to generate NG clusters of the category, and the shape of the generated NG cluster is held as second post-stage knowledge. And a data including the recognition result and its accompanying information as the recognition result of the input image, whether or not the data is included in the NG cluster belonging to the category having the second latter knowledge is checked. If it is included in the NG cluster, the category to which the NG cluster belongs is determined to be “incorrect” as the recognition result of the input image, and which NG cluster is included Included when not is the image recognition apparatus characterized by comprising a second second-stage determination means determines that the "correct answer".

The invention according to claim 7 of the present application is
An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, which includes recognition objects of a plurality of categories, and data including a recognition result of the image recognition and its accompanying information. Data having the same first candidate category as the recognition result is collected, and those having a high appearance rate of the similar categories in the specific candidate ranks of the second candidate and below are extracted from the collected data to obtain the first candidate category. A first pre-stage knowledge creation step that holds the first pre-stage knowledge that indicates the relationship of similar categories, and a first candidate category as the recognition result of the input image.
When similar categories of specific rank candidates below Gori and the second candidate are obtained, the category of the first candidate and the similar category of the specific candidate rank are compared with the relation held in the first pre-stage knowledge. Whether the category of the first candidate is “correct”
And a first pre-stage determination step for determining whether or not there is an image recognition method.

The invention according to claim 8 of the present application is
It has recognition targets of multiple categories and inputs registered image patterns.
Goodness of fit in pattern matching with the image pattern (distance
In the image recognition method that recognizes the input image based on
Data including the recognition result of image recognition and its accompanying information.
Have the same first candidate category as the recognition result.
Collect data, and from the collected dataFirst candidate category
From the goodness of fit (distance) inFirst candidate category
Calculates the degree of goodness of fit (distance) that can be determined as "correct answer"
Then, the second pre-stage knowledge creation step that is held as the second pre-stage knowledge is
And the recognition result of the input imageFirst candidateCategories
When is obtained, theFirst candidateGoodness of fit in the category
The size of (distance) and the knowledge of the second pre-stageRetainedConform
By comparing with the magnitude of the degree (distance),First candidate category
Gori"Correct answer"Whether or notSecond pre-stage determination step for determination
An image recognition method characterized by comprising:

The invention according to claim 9 of the present application is
An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, which includes recognition objects of a plurality of categories, and data including a recognition result of the image recognition and its accompanying information. From the data, the data having the same first candidate category as the recognition result is collected, and from the collected data, from the difference in the matching degree (distance) between the category of the first candidate and the similar category of the specific candidate ranks below the second candidate , A third pre-stage knowledge creating step of calculating the magnitude of a goodness-of-fit (distance) difference that can determine the category of the first candidate as "correct answer", and retaining it as the third pre-stage knowledge, and the first candidate as the recognition result of the input image. categories and the second
When similar categories following specific candidate ranking candidates are obtained, the first candidate of the category and the second candidate following specific candidate sequence
The degree of conformance (distance) difference with the rank category and the third
A third pre-stage determination step of determining whether or not the category of the first candidate is “correct” by comparing with the magnitude of the goodness-of-fit (distance) difference held in the pre-stage knowledge. The image recognition method is characterized by.

Further, the invention according to claim 10 of the present application has recognition objects of a plurality of categories, and recognizes the input image based on the matching degree (distance) in the pattern matching between the registered image pattern and the input image pattern. In the image recognition method, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and from the collected data, the second candidate and the following specific candidates are identified. A first pre-stage knowledge creating step of extracting a high-similarity appearance rate of similar categories in the candidate ranks and holding the same as first pre-stage knowledge indicating a relationship of the similar categories to the first candidate category, a recognition result of image recognition and its attendant From the data including information, data having the same first candidate category as the recognition result is collected, and the first data is collected from the collected data .
Fit in complementary category from (distance), the first candidate
From the data including the second pre-stage knowledge creation step of calculating the degree of goodness of fit (distance) that can determine the category of "correct answer" and holding it as the second pre-stage knowledge, the recognition result of image recognition and its accompanying information. , Collecting data having the same first candidate category as the recognition result, and collecting the first data from the collected data .
The categorization of the first candidate is calculated based on the difference in the degree of matching (distance) between the candidate category and the similar categories in the ranks of the second and lower candidates .
To calculate the size of the goodness of fit (distance) difference can confirm with the Li "correct answer", a third preceding knowledge generating step of holding the third preceding knowledge, catheterization of the first candidate as a recognition result of the input image
When similar categories of specific rank candidates below Gori and the second candidate are obtained, the category of the first candidate and the similar category of the specific candidate rank are compared with the relation held in the first pre-stage knowledge. Whether the category of the first candidate is “correct”
A first pre-determination step of determining whether said the determined input image to "correct" in the first pre-determination step, the size and the second pre knowledge of fit in the category of the first candidate (distance) by comparison with the size of the retained fitness (distance), the category of the first candidate and the second pre-determination step of determining whether "correct", and "correct answer" at the second pre-determination step for the determined input image, category of the first candidate and the second candidate following specific climate
Whether or not the category of the first candidate is “correct” by comparing the magnitude of the difference in fitness (distance) with the category of the complementary rank with the magnitude of the difference in fitness (distance) held in the third pre-stage knowledge
A third pre-stage determination step for determining
The image recognition method is characterized by

The invention according to claim 11 of the present application has a plurality of categories of recognition targets, and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the image recognition method, data including the recognition result of image recognition and its accompanying information is collected as a sample, and the recognition result is obtained in a certain category.
From the data “correct answer”, the axis when clustering the set of data is set as the upper candidate by using the category with a high appearance rate as the upper candidate as the axis and the category with a higher appearance rate by the lower candidate as the axis. A category with a high appearance rate and a category with a high appearance rate for lower candidates are combined as an axis, a category with a high appearance rate for a middle candidate is used as the axis, categories with high correlation coefficients are used as the axis, or a correlation is obtained. A first axis extracting step of extracting with categories having a low number as the axes, and clustering of the data set based on the axes extracted in the first axis extracting step,
Generated as an OK cluster of the category and generated O
When a first post-stage knowledge creating step of holding the shape of the K cluster as the first post-stage knowledge and data including the recognition result and its accompanying information as the recognition result of the input image are obtained,
It is checked whether or not the data is included in the OK cluster belonging to the category having the first latter knowledge, and if the data is included in the OK cluster, the category to which the OK cluster belongs is included in the input image. "Correct answer" as recognition result
The image recognition method is characterized by further comprising a first post-stage determination step of determining “undetermined” when the image is not included in any of the OK clusters.

The invention according to claim 12 of the present application has a plurality of categories of recognition targets, and recognizes the input image based on the matching degree (distance) in the pattern matching between the registered image pattern and the input image pattern. In the image recognition method, data including the recognition result of image recognition and its accompanying information is collected as a sample, and the recognition result is obtained in a certain category.
From the data that is regarded as “misidentification”, the axis when clustering the set of the above data is set as the upper candidate by using the category having a high appearance rate as the upper candidate as the axis and the category having the higher appearance rate by the lower candidate as the axis. A category with a high appearance rate and a category with a high appearance rate for lower candidates are combined as an axis, a category with a high appearance rate for a middle candidate is used as the axis, categories with high correlation coefficients are used as the axis, or a correlation is obtained. A second axis extracting step of extracting with a category having a low number as the axis, and clustering of the set of data based on the axis extracted in the second axis extracting step,
Generated as an NG cluster of the category and generated N
When a second post-stage knowledge creating step for holding the shape of the G cluster as the second post-stage knowledge and data including the recognition result and its accompanying information as the recognition result of the input image are obtained,
It is checked whether or not the data is included in the NG cluster belonging to the category having the second latter knowledge, and if the data is included in the NG cluster, the category to which the NG cluster belongs is included in the input image. The image recognition method is characterized by comprising a second post-stage determination step of determining "incorrect answer" as a recognition result, and determining "correct answer" when it is not included in any NG cluster.

In the invention according to claim 13 of the present application, in the first latter-stage determination step, when the data is included in an OK cluster belonging to a certain category, with certainty of confidence, "correct answer" is obtained. If the data is in an OK cluster belonging to a plurality of different categories, which category is calculated based on the certainty factor. ”
12. It is determined whether or not the answer is "correct".
The image recognition method described in 1.

The invention according to claim 14 of the present application has a plurality of categories of recognition targets, and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In a recording medium in which a program for causing a computer to execute the image recognition method is recorded, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the collection is performed. From the data
A first pre-stage knowledge creating step of extracting a high-similarity appearance rate of similar categories in specific candidate ranks less than or equal to the second candidate , and retaining the same as first pre-stage knowledge indicating the relationship of the similar categories to the first candidate category; As a recognition result of the image, the category of the first candidate and the specific candidates below the second candidate
When similar categories of ranks are obtained, by comparing the similar categories of the first candidate category and specific candidate ranks equal to or lower than the second candidate with the relationships held in the first pre-stage knowledge,
First to determine whether the category of the first candidate is “correct”
A recording medium, characterized in that the recording medium comprises a first-stage determination step.

Further, the invention according to claim 15 of the present application has a plurality of categories of recognition targets, and recognizes the input image based on the matching degree (distance) in the pattern matching between the registered image pattern and the input image pattern. In a recording medium in which the image recognition method is recorded as a program for causing a computer to execute, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the collection is performed. From the goodness of fit (distance) in the first candidate category ,
The category of the first candidate to calculate the size of the goodness of fit may be determined as "correct" (distance), a second pre-knowledge generating step of holding the second pre-knowledge, the first candidate as a recognition result of the input image when the category is obtained by comparison with the size of <br/> category definitive fit of the first candidate (distance) measurement and the second pre-knowledge held the fit (distance), the first A second pre-stage determination step of determining whether or not the candidate category is “correct answer” is provided.

The invention according to claim 16 of the present application has recognition objects of a plurality of categories, and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In a recording medium in which the image recognition method is recorded as a program for causing a computer to execute, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the collection is performed. The first candidate category and the second and lower candidate ranks
From the similarity (distance) with the similar category of
The third pre-stage knowledge creation step of calculating the magnitude of the goodness-of-fit (distance) difference that can determine the category of "correct answer" as the third pre-stage knowledge, and the first result as the recognition result of the input image.
When similar category of auxiliary categories and second candidate following specific candidate ranking is obtained, and the category of the first candidate second
The category of the first candidate is determined to be “correct answer” or “unknown” by comparing the magnitude of the difference in fitness (distance) of the specific candidate ranks below the candidate with the magnitude of the difference in fitness (distance) in the third preceding knowledge. And a third pre-stage determination step for making a determination.

The invention according to claim 17 of the present application has a plurality of categories of recognition targets, and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In a recording medium in which the image recognition method is recorded as a program for causing a computer to execute, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the collection is performed. A first pre-stage knowledge creating step of extracting a high-presence occurrence rate of similar categories in specific candidate ranks below the second candidate from the obtained data, and retaining the same as first pre-stage knowledge indicating the relationship of similar categories to the first candidate category And the data including the recognition result of the image recognition and the accompanying information, the categorization of the same first candidate as the recognition result. Data were collected with adaptability in category of the first candidate from the collected data from the (distance), and calculate the magnitude of the goodness of fit may be determined to "correct" the category of the first candidate (distance), Data having the same first candidate category as the recognition result is collected from the second pre-stage knowledge creating step held as the second pre-stage knowledge and the data including the recognition result of the image recognition and its accompanying information, and the One candidate category and the second and following candidates
From the difference in the degree of matching (distance) with the similar category of the specific candidate order, the magnitude of the difference in the degree of matching (distance) that can determine the category of the first candidate as the “correct answer” is calculated and stored as the third pre-stage knowledge. Third pre-stage knowledge creation step, and identification of the first candidate category and second candidate and below as the recognition result of the input image
When similar category of candidate ranking is obtained by comparison with related held in the first pre-knowledge similar category of the first candidate category and the second candidate following specific candidate ranking, the first candidate category There a first pre-determination step of determining whether "correct", the determination input image to "correct" in the first pre-determination step, the first candidate
Of the goodness of fit (distance) in the category of
By comparison with the size of the preceding knowledge held the fit (distance), a second pre-determination step of determining whether the category of the first candidate is "correct", in the second pre-determination step "correct answer for "a determined input image, the first climate
By comparing the magnitude of the goodness- of- fit (distance) difference between the complementary category and the similar category of the second candidate or lower with a specific candidate rank with the magnitude of the goodness- of- fit (distance) difference held in the third preceding knowledge, And a third pre-stage determination step of determining whether or not one candidate category is “correct answer”.

The invention according to claim 18 of the present application has a plurality of categories of recognition targets, and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. Data in which a recording result is recorded as a program for causing a computer to execute the image recognition method, data including the recognition result of the image recognition and its accompanying information is sampled, and the recognition result is "correct" in a certain category. From the above, the axis when clustering the data set is a category with a high appearance rate in the upper candidates, a category with a high appearance rate in the lower candidates as the axis, and a category with a high appearance rate in the upper candidates and a lower rank. Correlation coefficient with the category with a high appearance rate as candidates as the axis and the category with a high appearance rate as the middle candidate as the axis A first axis extracting step of extracting categories having a high category as the axis or categories having a low correlation coefficient as the axis, and clustering of the data set based on the axes extracted in the first axis extracting step. Data including the recognition result and its accompanying information as the recognition result of the input image, and the first second-stage knowledge creating step of generating the OK cluster of the category and retaining the shape of the generated OK cluster as the first second-stage knowledge. When it is obtained, it is checked whether or not the data is included in the OK cluster belonging to the category having the first latter knowledge, and if it is included in the OK cluster, the category to which the OK cluster belongs Is determined as the “correct answer” as the recognition result of the input image, and is determined as “undetermined” if it is not included in any OK cluster. In recording medium characterized by comprising a first second-stage determination step.

The invention according to claim 19 of the present application has a plurality of categories of recognition targets, and recognizes the input image based on the matching degree (distance) in the pattern matching between the registered image pattern and the input image pattern. Data in which a data including the recognition result of image recognition and its accompanying information is sampled on a recording medium in which the image recognition method is recorded as a program to be executed by a computer, and the recognition result is "misidentification" in a certain category. From the above, the axis when clustering the data set is a category with a high appearance rate in the upper candidates, a category with a high appearance rate in the lower candidates as the axis, and a category with a high appearance rate in the upper candidates and a lower rank. Correlation coefficient with the category with a high appearance rate as candidates as the axis and the category with a high appearance rate as the middle candidate as the axis A second axis extracting step of extracting with categories having a high category as the axis or categories having a low correlation coefficient as the axis, and clustering of the data set based on the axes extracted in the second axis extracting step. A second post-stage knowledge creating step of generating the NG cluster of the category and holding the shape of the generated NG cluster as the second post-stage knowledge, and the data including the recognition result and its accompanying information as the recognition result of the input image. When it is obtained, it is checked whether or not the data is included in the NG cluster belonging to the category having the second latter knowledge, and if it is included in the NG cluster, the category to which the NG cluster belongs Is determined as "incorrect answer" as the recognition result of the input image, and is determined as "correct answer" if it is not included in any NG cluster. In recording medium characterized by comprising a second second-stage determination step.

Further, in the invention according to claim 20 of the present application, in the first latter-stage determination step, when the data is included in an OK cluster belonging to a certain category, with a certain degree of confidence, the "correct answer" is given. If the data is in an OK cluster belonging to a plurality of different categories, which category is calculated based on the certainty factor. ”
20. It is determined whether the answer is "correct".
The recording medium described in 1.

According to the invention described in claim 1, 7 or 14, the first pre-stage knowledge creating means (first pre-stage knowledge creating step) converts the data including the recognition result of the image recognition and its accompanying information. , The same as the recognition result 1st
First, a collection of data having candidate categories, a collection of data having a high appearance rate of similar categories in a specific candidate rank of the second candidate and below, from the collected data, showing a relation of the similar categories to the first candidate category. The first-stage determination means (first first-stage determination step) holds the knowledge as the first-stage knowledge.
As the recognition result of the input image, the first candidate category and the first candidate
When a similar category with a specific candidate rank of 2 candidates or less is obtained, the category of the first candidate and the similar category with a specific candidate rank of 2 or less candidates and the association held in the first pre-stage knowledge are obtained. By comparison, the category of the first candidate is “correct”
Whether or not it is decided.

Here, "data" refers to the recognition result in the image recognition system and the information accompanying it. A "category" refers to a group of the same characters. Further, the image recognition system to which the present invention is applied, for example, of the similar categories, the one having the highest degree of matching (distance) in the pattern matching (matching calculation) between the registered image pattern and the input image pattern is recognized. The recognition result is output as a recognition result (first candidate), and "correct answer", "cannot be convinced correct", and "incorrect answer" are determined for the recognition result. Depending on the calculation method of the goodness of fit, the one having the lowest goodness of fit (distance) may be output as the recognition result.
Further, in the present invention, the data including the recognition result of the image recognition system and the information associated therewith is obtained, and the recognition result confirmation process is performed based on the data. The “accompanying information” is a candidate for a similar category. It is the distance to the category dictionary (registered image pattern) calculated by the candidate.

That is, the structure of the present invention is arranged in the image recognition system after the image recognition means.

According to the present invention, it is not necessary for a person to check what has been confirmed as a correct answer, and it suffices to check only those that cannot be determined as a correct answer, thus reducing the false recognition rate of acceptance. Therefore, labor saving in the image recognition system can be further achieved.

Further, according to the present invention as set forth in claim 2, 8 or 15, the second pre-stage knowledge creating means (second pre-stage knowledge creating step) converts the data including the recognition result of the image recognition and the accompanying information thereof. Data having the same first candidate category as the recognition result is collected, and the first data is collected from the collected data .
Fit in the category of the candidate from (distance), the first climate
The magnitude of the goodness of fit (distance) that can determine the complementary category as "correct answer" is calculated and held as the second pre-stage knowledge, and the second pre-stage determination means (second pre-stage determination step) recognizes the input image recognition result. when the category of the first candidate is obtained as, the first
By comparison with the size of one fit in the candidate category (distance) measurement and the goodness of fit held in the second pre-knowledge (distance), the category of the first candidate is whether "correct"
So that to determine. As a result, it is not necessary for a person to check what has been confirmed as a correct answer, and it suffices to check only those that cannot be determined as a correct answer. It is possible to further reduce labor.

Further, according to the invention described in claim 3, 9 or 16, the third pre-stage knowledge creating means (the third pre-stage knowledge creating step) is used to extract the data including the recognition result of the image recognition and the accompanying information thereof. Data having the same first candidate category as the recognition result is collected, and the first data is collected from the collected data .
The categorization of the first candidate is calculated based on the difference in the degree of matching (distance) between the candidate category and the similar categories in the ranks of the second and lower candidates .
Is calculated as the magnitude of the difference in fitness (distance) that can be determined as the “correct answer”, and is stored as the third pre-stage knowledge complement. In the third pre-stage determination means (third pre-stage determination step), As a recognition result, the category of the first candidate and the identification signs below the second candidate
When similar category of auxiliary rank is obtained, and similar categories of mosquito <br/> categories and the following specific candidate ranking second candidate of the first candidate
Moderate (distance) of the holding on the size and the third front knowledge of the difference between
By comparison with the size of the fit (distance) difference, said first
It is determined whether or not the candidate category is “correct answer”. As a result, it is not necessary for a person to check what has been confirmed as a correct answer, and it suffices to check only those that cannot be determined as a correct answer. It is possible to further reduce labor.

In addition, according to the invention described in claim 10 or 17, the first pre-stage knowledge (relation of similar category to the first candidate category) is performed in advance by the first pre-stage knowledge producing means (first pre-stage knowledge producing step). The second pre-stage knowledge creating step (second pre-stage knowledge creating step) produces the second pre-stage knowledge (the degree of goodness of fit (distance) at which a candidate of a specific rank can be determined as "correct answer") to the third pre-stage knowledge creating By means (third pre-stage knowledge creating step), third pre-stage knowledge (size of difference in fitness (distance) at which the first candidate can be determined as "correct answer") is created, and first pre-stage determination means ( First front stage determination step), second front stage determination means (second front stage determination step)
And the third pre-stage determination means (third pre-stage determination step) uses the first, second and third pre-stage knowledge to determine whether or not the recognition result of the input image is a "correct answer" . I am trying. As a result, the false recognition rate of acceptance can be reduced, and labor saving in the image recognition system can be further achieved. In addition, it is possible to automate the reading of forms. That is, for example, when one form is used as the recognition target area, if all confirmed results are correct answers, it is possible to save the results without checking, and conversely, results other than those that can be convinced to be correct answers When even one is included, it is possible to perform an operation such that the input of a person is collectively requested later without saving the recognition result of the entire form.

According to the invention as defined in claim 4, 11 or 18, the first axis extracting means (first axis extracting step).
Enables to collect data including the recognition result of image recognition and its accompanying information, and the recognition result can be obtained in a certain category.
The axis for clustering the set of data is extracted from the data that are “correct answer”, and the first second-stage knowledge creating unit (first second-stage knowledge creating step) extracts the first axis extracting unit (first axis extracting step). ) Clustering of the data set based on the axis extracted in step 1) to generate an OK cluster of the category, hold the shape of the generated OK cluster as first second-stage knowledge, and determine the first second-stage determination means (first In the latter-stage determination step), when data including the recognition result and its accompanying information is obtained as the recognition result of the input image, it is determined whether the data is included in the OK cluster belonging to the first knowledge category. If it is included in the OK cluster, it is determined that the category to which the OK cluster belongs is “correct” as the recognition result of the input image. If it is not included in the cluster, it is determined to be “undetermined.” In addition, for the axis extraction by the first axis extraction means (first axis extraction step), for example, a category with a high appearance rate is selected as a top candidate. As an axis, a category having a high appearance rate in the lower candidates is used as an axis, a category having a high appearance rate in the upper candidates and a category having a high appearance rate in the lower candidates are used as an axis, and a category having a high appearance rate in the middle candidate is used as the axis. As an axis, a method of extracting categories having a high correlation coefficient as the axis, or extracting categories having a low correlation coefficient as the axis can be considered. Since it becomes unnecessary to check only the items that could not be confirmed as correct answers, it is possible to reduce the false recognition rate of acceptance and reduce the image recognition system. It can be further achieved that labor saving.

According to the invention of claim 6, 12 or 19, second axis extracting means (second axis extracting step).
Enables to collect data including the recognition result of image recognition and its accompanying information, and the recognition result can be obtained in a certain category.
An axis for clustering the set of data is extracted from the data which is regarded as "misidentification", and the second second-stage knowledge creating unit (second second-stage knowledge creating step) extracts the second axis extracting unit (second axis extracting step). ) Clustering the set of data based on the axis extracted in), generating as an NG cluster of the category, and holding the shape of the generated NG cluster as second second-stage knowledge, a second second-stage determining means (second In the latter-stage determination step), when data including the recognition result and its accompanying information is obtained as the recognition result of the input image, it is determined whether the data is included in the NG cluster belonging to the second knowledge category. If it is checked, if it is included in the NG cluster, the category to which the NG cluster belongs is determined as the recognition result of the input image.
It is determined that the answer is "wrong", and if it is not included in any NG cluster, it is determined as "correct".
For the axis extraction in the second axis extraction means (second axis extraction step), a category having a high appearance rate in the top candidates is used as the axis.
A category with a high appearance rate in the lower candidates is used as the axis, a category with a high appearance rate in the upper candidates and a category with a high appearance rate in the lower candidates are combined as an axis, and a category with a high appearance rate in the middle candidate is used as the axis. A method of extracting categories having a high correlation coefficient as the axis, or extracting categories having a low correlation coefficient as the axis can be considered.

As a result, it is not necessary for a person to check what has been confirmed as a correct answer, and it suffices to check only those that cannot be determined as a correct answer, so that the acceptance error recognition rate can be reduced. It is possible to further save labor in the image recognition system.

Further, according to the invention described in claim 5, 13 or 20, in the first second-stage determination means (first second-stage determination step), the data belonging to a certain category is O.
When included in the K cluster, the certainty factor, which is an index indicating how much confidence can be said to be the “correct answer”, is calculated by the certainty factor calculation means (certainty factor calculation step), and the data is stored in plural numbers. When an OK cluster belonging to another category is entered, it is determined which category is the “correct answer” based on the certainty factor. For example, when the magnitude of the certainty factor is equal to or less than a predetermined value, the “correct answer” is determined. I am not sure. " As a result, the OK of a certain category
Data that has entered the cluster is determined to belong to that category. Conversely, if it does not belong to any OK cluster, it is regarded as "unknown" (not belonging to any category) data, but it is determined to be data that belongs to a certain category. If it is attached, by calculating the confidence level of the recognition result,
The recognition result will be classified into three types, "unknown", "correct answer", and "we cannot be convinced to be correct". Furthermore, if at least one of the NG clusters is entered, it will be regarded as "incorrect" data. The recognition result is determined, and the recognition result is classified into three of “correct answer”, “incorrect answer”, and “I cannot be sure of correct answer”.

[0038]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of an image recognition apparatus, an image recognition method and a recording medium of the present invention will be described in detail below with reference to the drawings.

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

In FIG. 1, the image recognition apparatus of the present embodiment roughly comprises an OCR recognition processing unit 101 and a recognition result confirmation unit 102. The recognition result confirmation unit 102 includes a knowledge generation unit 103 and a preceding stage. Judgment unit (first front stage judgment means, second front stage judgment means and third front stage judgment means) 1
15, rear stage determination unit (first rear stage determination unit and second rear stage determination unit) 116, rule determination unit 117, recognition determination unit 11
8 and a reject unit 119.
In addition, the knowledge generation unit 103 includes a pre-stage knowledge creating unit (first pre-stage knowledge creating unit, second pre-stage knowledge creating unit, and third pre-stage knowledge creating unit) 111, a pre-stage knowledge base (first pre-stage knowledge, second pre-stage knowledge). And third pre-stage knowledge) 112, a post-stage knowledge creating unit (first post-stage knowledge creating means and second post-stage knowledge creating means) 113, and a post-stage knowledge base (first post-stage knowledge and second post-stage knowledge) 114. is there.

FIG. 2 is an explanatory view for explaining the outline of the overall processing of the image recognition apparatus of this embodiment. As shown in the figure, the image recognition apparatus of this embodiment has recognition targets of a plurality of categories, and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. However, the recognition result determination unit 1
02, and the OCR recognition processing unit 101
The recognition determination unit 118 determines that the recognition result is “undetermined” by performing the determination of the recognition result and the associated information 151b from the preceding determination unit 115 and the determination unit 116 based on the upstream knowledge base 112 and the downstream knowledge base 114. (Incorrect answer), “correct answer”, or “I am not sure that the answer is correct”.

A brief description will be given of each of the components that make up the image recognition apparatus of this embodiment and the transfer data between the components. In the image recognition apparatus of the present embodiment, in order to simplify the description, the character images to be recognized are numbers (0 to 9) and three types of symbols ("(", ")", ".
-"), A total of 13 types of characters are handled.

First, the OCR recognition processing unit 101 performs a fitness calculation on each character dictionary among similar characters. Here, the fitness calculation is performed in pattern matching between the registered image pattern and the input image pattern,
Similar characters are arranged in descending order of suitability (hereinafter referred to as distance), and the one with the highest distance is output as the OCR recognition result.

The recognition results 151a and 151b are data including the recognition result of image recognition and its accompanying information, and the data is composed of 13 candidates per character and the distance to each candidate dictionary. (See Figure 6). In particular, the knowledge generation recognition result 151a is acquired by the OCR recognition processing unit 101 before the actual operation of the sample image used for knowledge generation, and the knowledge generation unit 1
Passed to 03. Further, in particular, the recognition result 151b is supplied from the OCR recognition processing unit 101 to the pre-stage determination unit 115 on a character-by-character basis during actual system operation.

As described above, the pre-stage knowledge creating unit 111 is configured to include the first pre-stage knowledge creating unit, the second pre-stage knowledge creating unit, and the third pre-stage knowledge creating unit. In the first pre-stage knowledge creating unit, Data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the second data is collected from the collected data .
In addition, the one with a high appearance rate of similar categories in the following specific candidate ranks is extracted, and this is registered in the upstream knowledge base 112 as the first upstream knowledge indicating the relationship of the similar categories to the first candidate category.

In the second pre-stage knowledge creating means, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the first candidate is selected from the collected data . from goodness of fit in the category (distance), a category of the first candidate to calculate the size of the goodness of fit may be determined as "correct" (distance), and registers the preceding knowledge base 112 as second pre-knowledge.

Furthermore, in the third pre-stage knowledge creating means, from the data including the recognition result of the image recognition and its accompanying information,
Data having the same first candidate category as the recognition result is collected, and from the collected data, the first candidate category and the second candidate are collected.
Complementary, the magnitude of the goodness of fit (distance) difference that can determine the category of the first candidate as the “correct answer” is calculated from the goodness of fit (distance) difference with similar categories of specific candidate ranks, It is registered in the upstream knowledge base 112 as the upstream knowledge.

In this embodiment, there are 13 categories of characters to be recognized, and all 13 of the data output from the OCR recognition processing unit 101 are always included in the candidate characters. Of the second and third candidate character data, a character with a high appearance rate (first knowledge) is acquired, and
The minimum value (second pre-stage knowledge) at which the correct distance can be determined from the candidate distance data, and the minimum value (third pre-stage knowledge) at which the correct answer can be determined from the data of the difference between the first candidate distance and the second candidate distance are calculated. The knowledge created here will be used by the pre-stage determination unit 115. There are thousands of categories for characters other than 13 numbers and symbols, but in this case, the data output from the OCR recognition processing unit 101 is the top tens of the categories with a large distance. From these, in the same way, it is likely that the ones that appear most frequently will be selected as the Xth (X can be arbitrarily selected) candidates.

Next, the second-stage knowledge creating unit 113 is configured to include the first second-stage knowledge creating unit and the second second-stage knowledge creating unit as described above, and the first axis included in the first second-stage knowledge creating unit. The extraction means collects a sample of data including the recognition result of image recognition and its accompanying information, and extracts the axis when clustering the data set from the data whose recognition result is "correct" in a certain category, First
The latter-stage knowledge creating unit performs clustering on a set of data based on the extracted axes to generate an OK cluster of the category, and the shape of the generated OK cluster is first
It is registered in the second-stage knowledge base 114 as the second-stage knowledge.

Further, the second axis extracting means included in the second latter-stage knowledge creating means collects a sample of data including the recognition result of the image recognition and its accompanying information, and the recognition result is regarded as "misidentification" in a certain category. From the existing data, an axis for clustering the data set is extracted, and the second post-stage knowledge creating unit performs clustering of the data set based on the extracted axis to generate an NG cluster of the category. The shape of the NG cluster is registered in the second-stage knowledge base 114 as the second-stage second-stage knowledge.

In the axis extraction in the first and second axis extracting means, the category having a high appearance rate is used as the upper candidate (the category which is easily misrecognized is selected), and the category having a high appearance rate is selected as the lower candidate. As the axis (select a category that is difficult to misrecognize), combine the category with a high appearance rate in the upper candidates and the category with a high appearance rate in the lower candidates (select a category that is difficult to misidentify), The category with a high appearance rate as a candidate is selected as the axis (a category that is not easily misidentified or difficult is selected), the categories with a high correlation coefficient are used as the axis, or the categories with a low correlation coefficient are used as the axis. A method of extracting may be considered.

That is, an axis is selected to generate a cluster, and the axis data is used to generate a cluster. If it is a number, there are 13 categories, and 13 candidates
The items are output in descending order of distance to each category. From among them, the axis having the available price characteristic for the recognition confirmation is selected. In the present embodiment, the top three characters having a high appearance rate (that is, the characters that are likely to be erroneously recognized) as the top candidates are selected as the cluster use axes. And the recognition result 1
Clustering was performed using 51a. The created cluster shape is an ellipsoid. Also, per category,
There will be a plurality of OK clusters and a plurality of NG clusters.

Next, the pre-stage determination unit 115 is configured to include the first pre-stage determination unit, the second pre-stage determination unit, and the third pre-stage determination unit as described above.
As the recognition result of the input image, the first candidate category and the first candidate
When similar categories with a specific candidate rank less than or equal to 2 candidates are obtained, the first candidate category and the similar categories with a specific candidate rank less than or equal to the second candidate are compared with the association held in the first pre-stage knowledge to determine the first category. 1 Candidate category is determined to be “correct answer” or “unknown”.

In addition, the second front stage judging means uses the first candidate.
By comparison with the size of the goodness of fit (distance) measurement and retained fit the second pre-knowledge (distance) in the category, it is determined that the category of the first candidate "correct" or "unknown". Furthermore, in the third front determination means, the first candidate
Categories and holding the magnitude and third front knowledge of fitness (distance) difference between similar categories following specific candidate ranking second candidate
By comparison with the magnitude of has been fit (distance) difference, it is determined that the category of the first candidate "correct" or "unknown".

In the present embodiment, the characters of the second candidate and the third candidate in the recognition result 151b and the preceding knowledge base 1
Compare with 12 knowledge. For example, if the first candidate is “1”, the second candidate and the third candidate are respectively “
The characters "6" and ")" appear at a high rate. That is, when the characters appearing in the second and third candidates are "6" and ")" in the first pre-stage knowledge, the first candidate is the first candidate. But"
The person who possesses the knowledge that the correct answer is "1", and first checks the character types of the second and third candidates of the recognition result 151b.

Next, with reference to the second pre-stage knowledge (the lowest value that can be determined to be the correct answer from the distance data of the first candidate) held in the pre-stage knowledge base 112, a value equal to or greater than the minimum value is the distance of the first candidate. Check whether it has been calculated, and if it is at least the minimum value, it is regarded as "correct". Further, since it is possible that the first candidate and the second candidate have very close distances to each other, it may be erroneously recognized. Therefore, the third upstream knowledge (first candidate distance and second The minimum value that can be determined to be the correct answer based on the data of the difference in the candidate distances) is referred to, and it is checked whether the value is equal to or more than the minimum value.
The result (confirmed data 155) is passed to the recognition confirmation unit 118 only for the data determined to be “correct” in all of the above three checks.

Therefore, the definite data 155 is the data that the pre-stage determination unit 115 has determined to be correct, and the unknown data 154 is the data that the pre-stage determination unit 115 has not determined to be correct.

Next, as described above, the rear-stage determination section 116 is configured to include the first rear-stage determination means and the second rear-stage determination means, and the first rear-stage determination means recognizes the recognition result and the recognition result as the recognition result of the input image. When the data including the accompanying information is obtained, it is checked whether the data is included in the OK cluster belonging to the category having the first second-stage knowledge, and if the data is included in the OK cluster, the OK is determined. The category to which the cluster belongs is determined as "correct answer" as the recognition result of the input image, and if it is not included in any OK cluster, it is determined as "undetermined". Further, the second latter-stage determination means checks whether or not the data is included in the NG cluster belonging to the category having the second latter-stage knowledge, and if the data is included in the NG cluster, the NG cluster is The category to which the user belongs belongs as the recognition result of the input image.
If the answer is "wrong answer" and it is not included in any NG cluster, it is judged as "correct answer".

That is, the data 154 which is determined to be unknown by the pre-stage determination unit 115 is subjected to the determination using the clusters generated in each category. The subsequent determination unit 116 is
Knowledge of the K clusters and the NG clusters is supplied from the post-stage knowledge base 114, and a two-stage judgment is performed to check whether or not to enter the OK cluster and to check whether or not to enter the NG cluster.

When the data is included in the OK cluster belonging to a certain category in the first latter-stage determination means,
The certainty factor, which is an index indicating how much certainty can be said to be the “correct answer”, is calculated by the certainty factor calculation means included in the latter-stage determination unit 116, and the data falls within the OK cluster belonging to a plurality of different categories. If so, which category is set as the “correct answer” is determined based on the certainty factor. For example, when the degree of certainty is equal to or less than a predetermined value, it is determined that "the correct answer cannot be obtained".

The confidence factor calculation method in the confidence factor calculation means calculates how far the data is located from the center of gravity of the OK cluster, and if it is closer to the center of gravity, a higher confidence factor is calculated. Based on the first method of calculating a low confidence degree if the distance is far, and based on the distribution of the data forming the OK cluster, a high confidence degree is set when the data exists in the high density area, and a low confidence area is set in the low density area. 2), there is a second method for calculating a low certainty factor.

If the second post-stage determination means includes the data in an NG cluster belonging to a certain category, the confidence factor updating means included in the post-stage determination section 116 calculates that the data is included in the OK cluster. A process of lowering the determined certainty factor is performed.

Therefore, the definite data 156 is data for which the certainty factor is included in at least one OK cluster, and the undecided data 157 is the post-stage determination unit 11.
6 is the data when it was not included in any OK cluster, and the recognition result could not be determined in any category.

Next, the reject unit 119 is a unit that receives the data 157 that has been undetermined by the post-stage determination unit 116 or the data 159 that has been determined by the rule determination unit 117 to be inaccurate.

The rule determination section 117 is a section for determining, based on the certainty factor calculated by the post-stage determination section 116, whether to reject the data as uncertain as the correct answer or accept it as the final data. Actually, “determined” and “not sure that the answer is correct” are classified by comparing the value with the value set by the user.

Therefore, the fixed data 158 is data having a higher certainty factor than the value set by the user in the rule determination unit 117, and the data 159 that cannot be surely the correct answer is higher than the certain value set by the user in the rule determination unit 117. Is small data.

Further, the recognition confirming unit 118 is a unit for receiving the data 155, 158 for which the recognition result can be confirmed by the pre-stage judging unit 115 or the rule judging unit 117.

With the components described above, the image recognition apparatus of this embodiment performs the following operations. FIG. 3, FIG. 4 and FIG. 5 show schematic flow charts for explaining the schematic operation of the image recognition apparatus of this embodiment.

First, FIG. 3 is a schematic flowchart of the knowledge creating process. In the knowledge creation process, first step 301
Then, for the sample data, the OCR recognition processing unit 101
Thus, the recognition result 151a for knowledge generation is collected.

Next, in step 302, the knowledge generation unit 1
The pre-stage knowledge base 1 by the pre-stage knowledge creation unit 111 of 03.
12 are built. That is, the first pre-stage knowledge creating unit can determine the first pre-stage knowledge (the association of the similar category with the first candidate category) and the second pre-stage knowledge creating unit can determine the second pre-stage knowledge (the candidate of the specific rank is "correct answer"). The degree of goodness of fit (distance) is created and registered by the third knowledge making means as the third knowledge of knowledge (the magnitude of the difference of goodness of fit (distance) at which the first candidate can be determined as "correct answer"). To be done.

Further, in step 303, the latter-stage knowledge base 114 is constructed by the latter-stage knowledge creating unit 113 of the knowledge generating unit 103. That is, the first second-stage knowledge creating unit creates and registers the first second-stage knowledge (OK cluster shape), and the second second-stage knowledge creating unit creates and registers the second second-stage knowledge (NG cluster shape).

Next, FIG. 4 and FIG. 5 are schematic flow charts of processing during system operation. First, step 4
In 01, the OCR recognition processing unit 101 performs recognition processing on the actual recognition target image, and the recognition result 151b is supplied to the pre-stage determination unit 115 of the recognition result determination unit 102.

Next, the processing by the pre-stage determination section 115 is performed. First, in step 402, the first front stage determination means of the front stage determination section 115 determines the first candidate category and the first candidate .
The similar category of the specific candidate ranks of two candidates or less is compared with the relation held in the first-stage knowledge, and if the knowledge is the same, the category of the first candidate is determined as “correct answer” and step 4
If it is different from the knowledge, it is judged as "unknown" and the process proceeds to step 501 in FIG.

Next, at step 403, the second pre-stage determination means compares the magnitude of the goodness of fit (distance) in the category of the first candidate with the magnitude of the goodness of fit (distance) in the second pre-stage knowledge, If the value is larger than the knowledge distance, the first candidate is determined to be "correct" and step 40
4. If it is less than the knowledge distance, it is judged as "unknown" and the process proceeds to step 501 in FIG.

Further, in step 404, the third pre-stage determination means determines the magnitude of the goodness-of-fit (distance) difference between the category or the similar category of the specific candidate rank and the magnitude of the goodness-of-fit (distance) difference in the third pre-stage knowledge. And if the value is larger than the magnitude of the difference in knowledge distance, the first
The candidate is determined to be "correct" and the process proceeds to step 508 in FIG. 5, and if it is less than or equal to the knowledge distance difference, it is determined to be "unknown" and the process proceeds to step 501 in FIG.

Next, the processing proceeds to the latter stage judging section 116 in FIG. First, in step 501, the post-stage determination unit 11
The first second-stage determination unit 6 checks whether the data is included in the OK cluster belonging to the category having the first second-stage knowledge. Here, if it is not included in the OK cluster, it is rejected in step 509.

In step 501, if the data is included in an OK cluster belonging to a certain category, the procedure proceeds to step 502 or step 503 to show how much confidence the index is. The certainty factor is calculated by the certainty factor calculation means.

In the confidence factor calculation by the confidence factor calculating means, as in step 502, the distance at which the data is located from the center of gravity of the OK cluster is calculated. If it is far from the center of gravity, a low confidence factor is calculated, or, as in step 503, based on the distribution of the data forming the OK cluster, a high confidence factor is calculated when the data exists in a high density area. If it exists in a low density area, low confidence is calculated for each of them.

In this embodiment, as shown in FIG. 5, in step 502, the certainty factor based on the distance from the cluster centroid is obtained, in step 503 the certainty factor based on the intra-cluster coordinate value is obtained, and in step 504, these two calculations are performed. The determined certainty factor is added to obtain the final certainty factor.

Next, in step 505, the second latter-stage determination means checks whether or not the data is included in the NG cluster belonging to the category having the second latter-stage knowledge. Here, if it is not included in the NG cluster, the process proceeds to step 507.

In step 505, if the data is included in the NG cluster belonging to a certain category, the process proceeds to step 506, and the certainty factor updating means
The certainty factor is adjusted so that the certainty factor calculated by being included in the OK cluster is lowered.

Further, in step 507, the rule determination unit 117 compares the certainty factor calculated by the latter determination unit 116 with the value set by the user, and if the certainty factor is less than or equal to the user set value, the step is performed. In step 509, the data is rejected as "uncertain that the answer is correct", and if the certainty factor is larger than the user set value, in step 508 it is accepted as the "correct answer" confirmation data.

Next, the function and operation of each component constituting the image recognition apparatus of this embodiment, the method used, and the like will be described in detail below with reference to the drawings.

First, the recognition result output from the OCR recognition processing unit 101 will be described with reference to FIG. The data output for each character to be recognized is the distance between the candidate character in the upper row of each line and each candidate in the lower row of each line, as shown in each row of FIG.

The description will be as follows, taking the data of the first row as an example. That is, the first candidate is "0"
Therefore, the distance of “0” is 234. The second candidate is "9"
Therefore, the distance of "9" is 214. …. The 13th candidate is “)”, and the distance of “)” is 135. here,
The larger the distance, the higher the reliability of the recognized character. That is, regarding the data in the first row, OC
The R recognition processing unit 101 has recognized the recognition target character as “0”.

Next, the construction of the upstream knowledge base 112 by the upstream knowledge creation unit 111 will be described. First, the first
In the first-stage knowledge creating means, the recognition result 151a (see FIG. 6)
The candidate characters appearing in the second candidate and the third candidate of the data of are counted. Among them, the two candidate characters with the highest number of appearances are saved as the first-stage knowledge.

Next, in the second and third front-stage knowledge creating means, the first result is obtained from the data of the recognition result 151a (see FIG. 6).
The second and third pre-stage knowledge for judging "confirmed" or "unconfirmed" of the recognition result by using two features of the size of the candidate distance and the size of (first candidate distance-second candidate distance) Threshold parameter).

First, the data used for parameter determination is selected from the knowledge creation data (performed for each category). Specifically, the steps from a) to d) are performed.

A) (first candidate distance-second candidate distance)
The number of correct samples for 0, 1, 2, ...

(B) Check the ratio of the number of correct samples and the number of erroneously determined samples.

C) Find the distance difference α at which the ratio of the number of erroneously determined samples and the number of correct samples is equal to or more than a certain value (for example, 40% or more).

D) (first candidate distance-second candidate distance) is α
All the following data are excluded from the object of parameter determination in (2) as the object of reject.

Next (secondly), the first candidate distance necessary for rejecting misjudgment for each character (category) and the threshold value of the distance difference between the first candidate and the second candidate are found. Specifically, the steps a) to c) are performed.

A) Check the first candidate distance and (first candidate distance-second candidate distance) of the misjudgment sample.

(B) Check the first candidate distance and (first candidate distance-second candidate distance) of the correct sample.

C) The thresholds of the first candidate and (first candidate distance-second candidate distance) that can reject all misjudgments and minimize the rejection of correct answers are determined.

The parameters obtained as described above are shown in FIG. In the figure, the "first candidate distance" is the second upstream knowledge, the "first-second candidate distance" is the third upstream knowledge,
“Characters with a high appearance rate in the second and third candidates” are the first preliminary knowledge.

Next, the determination method by the upstream determination unit 115 will be described. The pre-stage determination unit 115 includes the knowledge generation unit 10
The recognition result 151b of the OCR recognition processing unit 101 is divided into "confirmed" and "unconfirmed" by using the preceding knowledge base 112 based on the distance data of the candidate character created in 3. Here, the rules for dividing into “confirmed” and “unconfirmed” are as described below.

First (procedure 1), the first candidate is referred to, and the parameter used for checking is selected from the first pre-stage knowledge (see FIG. 7). Next (procedure 2), it is checked whether the second candidate character and the third candidate character match the first pre-stage knowledge (those with a high appearance rate). If there is data that does not match, it is treated as "unknown" data. Next (procedure 3),
If the first candidate distance is smaller than the second pre-stage knowledge (the first candidate distance parameter in FIG. 7) for the data that was not made "unknown" data in step 2, it is made "unknown" data. Further (procedure 4), for data that is not "unknown" data in procedure 3, (first candidate distance-second candidate distance) is checked. That is, if (first candidate distance-second candidate distance) is smaller than the third pre-stage knowledge ((first candidate distance-second candidate distance) parameter in FIG. 7), "unknown" data is set.

FIG. 8 shows the above procedure as a flowchart. That is, first, in step 801, the OCR is performed.
The first candidate is acquired from the recognition result 151b of the recognition processing unit 101. Next, in step 802, the “first candidate” is referred to and “a candidate character having a high appearance rate in the second and third candidates” is extracted from the first preliminary knowledge. Then, in step 803,
It is determined whether the second and third candidates of the recognition result 151b match the second and third candidates of the first preliminary knowledge. If they do not match, the process proceeds to step 808, and the recognition result 1
The first candidate of 51b is set to "unknown".

Next, if they match in step 803, the process proceeds to step 804, in which the "first candidate distance" parameter α and the "first-first distance" are calculated from the second and third pre-stage knowledge.
The second candidate distance difference ”parameter β is acquired. Then, in step 805, the first candidate distance of the recognition result 151b is compared with the “first candidate distance” parameter α, and if the first candidate distance is smaller than the “first candidate distance” parameter α, the process proceeds to step 808. , The first candidate of the recognition result 151b is "unknown".

If the first candidate distance is larger than the "first candidate distance" parameter α in step 805,
In step 806, the (first candidate distance-second candidate distance) of the recognition result 151b is compared with the "first-second candidate distance difference" parameter β to calculate (first candidate distance-second candidate distance). Is smaller than the “first-second candidate distance difference” parameter β, the process proceeds to step 808, and the first candidate of the recognition result 151b is set to “unknown”.

If (first candidate distance-second candidate distance) is larger than the "first-second candidate distance difference" parameter β in step 806, the process proceeds to step 807 to select the first candidate of the recognition result 151b. Confirm the "correct answer".

The "first candidate distance" used in FIG.
The parameter α and the “first-second candidate distance difference” parameter β can be changed from the outside. That is, the parameters shown in FIG. 7 are standard parameters used for each character (category), and by varying this, the ratio of the number of confirmed and unconfirmed data can be adjusted.

That is, the "first candidate distance" parameter α
The “first-second candidate distance difference” parameter β can be adjusted by variably setting Kd, the first fixed value, and the second fixed value in the following equation.

Α = first candidate distance + first fixed value × Kd (−1.0 ≦ Kd ≦ 1.0) β = (first candidate distance−second candidate distance) + second fixed value × K
d (−1.0 ≦ Kd ≦ 1.0) Here, the method for determining the first fixed value and the second fixed value is shown in FIG.
When the fluctuation range of ± 2 is provided with reference to the parameter of, the first fixed value = the second fixed value = 2. Further, when the parameters in FIG. 7 are used as reference values, the larger the parameters α and β are, the conditions in FIG. 8 (step 805,
Since the number of data that can pass 806) is small, Kd
The closer to 1.0, the more severe the setting. Conversely, the closer Kd is to -1.0, the looser the setting.

Next, a method of constructing the latter-stage knowledge base 114 by the latter-stage knowledge creating unit 113 will be described. In the present embodiment, the top three characters having a high appearance rate (that is, the characters that are likely to be erroneously recognized) are selected as the top candidates as the cluster use axis, and the cluster is formed in the three-dimensional space.
That is, the candidate character having a high appearance rate in FIG. 7 is selected as the axis. The process of performing clustering using these axes will be briefly described below.

First (procedure 1), the OCR recognition processing unit 101
The recognition result (for knowledge generation) 151a is received from and the data is classified into categories (13 pieces). Next (procedure 2), the data classified by category is classified into OCR recognition result correct data (hereinafter referred to as "group A") and error recognition result (hereinafter referred to as "group B"). Next (procedure 3), a data group of "candidate characters having a high appearance rate in the second and third candidates" of each category shown in FIG. 7 is extracted from the data of group A, and these are extracted with a high appearance rate. Let A, B and C be in order. Taking category "0" as an example, 0 (A), 6 (B), 9 (C)
The group A data of the category is extracted.

Next (procedure 4), for the data group A,
The data of axis A, axis B and axis C are extracted and clustering is performed. The cluster created here is a.
Next (procedure 5), data of axis A, axis B, and axis C is extracted and clustered for the data group of B. Let the cluster created here be b. Further (procedure 6), for the data group of C, axis A, axis B, and axis C
Data is extracted and clustering is performed. Let the cluster created here be c. Next (procedure 7), data of axis A, axis B, and axis C is extracted from the data of group B to perform clustering.

Here, since there are 13 categories, the procedure 1
-Procedure 6 and procedure 7 will be performed 13 times, respectively. Also, a plurality of clusters will be generated for each of group A and group B. Incidentally, steps 3 to 6
"OK cluster" (a, b, c
Is included) and the cluster created in step 7 is called an “NG cluster”.

Next, the clustering method (hyperellipsoidal method) used in this embodiment will be described. Unlike the normal clustering method that merges clusters that are close to each other in distance to generate clusters, clustering by the hyperellipsoid method is linear (meaning how a linear model can be formed). And the continuity (meaning whether or not the distribution of data forming the cluster is continuous) can be taken into consideration to generate the cluster.

In normal clustering, the shape cannot be changed once the clusters are merged only by the criterion of distance. That is, there is no standard for separating the parts forming the cluster after the merger. However, in hyperelliptic clustering, linearity and continuity are taken into account,
By changing the cluster to be merged and moving the cluster configuration data after the merger, it is possible to construct a cluster group according to the defined norm.

Taking the cluster generation shown in FIG. 9 as an example, when importance is attached to the degree of continuity, the clusters A and B having a close distance are merged to generate a cluster having a high configuration data density. If linearity is emphasized, clusters B and C are
Will be merged, and if necessary, we will move one point,
A cluster having a data distribution that is easy to construct a linear model is generated. The standard of hyper-ellipsoidal clustering is to minimize the sum of the volumes (super products) of each cluster.

Next, the flow of the hyperelliptic clustering process will be briefly described below. The outline of the hyperelliptic method clustering processing is as shown in the flowchart in FIG. 10. As described above, this process is performed for each character (category) to create a plurality of clusters / categories.

The processing steps will be briefly described below.
First, in step 1001, various parameters are set. Next, the initial cluster is 1 [data / cluster]
And the clusters having a short inter-cluster distance are merged to form a processing cluster group. The conventional clustering method (step 1002) called the Ward method is used here.

Next, in step 1003, after a plurality of [data / cluster] (including 1 [data / cluster] in some cases), the cluster volume sum, which is the evaluation criterion for the hyperellipsoid, increases. The clusters that minimize? Are merged (step 1005). However, the distance between the cluster centroids to be merged must be δu or less (step 1004).

Next, after the merger, without changing the number of clusters,
Move configuration data units between clusters. Here, the data to be moved is the data in which the amount of decrease in the cluster volume sum, which is the evaluation criterion for the hyperellipsoid, is maximum (step 1
007-1010). However, the cluster between which the data is moved must have a processing between cluster centroids of δm or less (step 1008).

Finally, the end judgment is made. Whether the number of clusters is the designated final cluster number (step 100
6) Alternatively, when the reference increase amount is equal to or greater than the specified value (step 1011), the process ends. Otherwise, steps 1005-1010 are repeated.

The parameter δu used in the above description
And δm have the following meanings. That is, δ
u is a merging range limiting parameter. When δu is increased, linearity is emphasized, and when δu is decreased, continuity is emphasized. Further, δm is a movement range limiting parameter. When δm is increased, linearity is emphasized, and when δm is decreased, continuity is emphasized. These parameters δu and δm
Cannot be uniquely determined, so it is necessary to set it by trial and error.

FIG. 11 shows a data format when a file is output as post-stage knowledge as a result of executing the above clustering.

Next, the relationship between the ellipsoid and each parameter will be described. The eigenvectors of the covariance matrix generated by the cluster configuration data determine the spreading direction of the ellipsoid, and the ellipsoid radius in that direction corresponds to the square root of the eigenvalues of the covariance matrix. For example, if one data consists of 13 components and hyper-ellipsoidal clustering is performed using all components, 13 ×
Thirteen covariance matrices are formed. Then, 13 eigenvectors and 13 eigenvalues can be obtained. However, as shown in the cluster data format of FIG. 11, the covariance matrix is not stored as the latter knowledge.

FIG. 12 is an explanatory view for explaining the relationship between the ellipsoid and the eigenvalues and eigenvectors. Further, the ellipsoidal formula and the ellipsoidal volume can be represented by the following formulas.

[0123]

[Equation 1]

[Equation 2] Next, the determination method by the latter-stage determination unit 116 will be described. As described above, the latter-stage determination unit 116 is the former-stage determination unit 1
With respect to the unknown data 154 which has been undecided in 15, the undecided data is checked again. The check is performed depending on whether or not the target data is included in the subsequent knowledge base 114 (OK cluster) created by the subsequent knowledge creating unit 113. If it is included in the cluster of a certain category, it has a certainty factor calculation means for calculating the certainty factor (reliability).

FIG. 13 shows a flow chart for explaining the flow of processing in the latter-stage judging section 116. First, in step 1301, the first candidate of the confirmation target data 154 is acquired, and the axis data (three-dimensional) used by the first candidate is extracted. For example, if the first candidate is “0”, then from FIG.
Select the 0, 6, and 9 axes.

Next, in step 1303, it is checked whether all of the “OK clusters” among the clusters possessed as the latter knowledge are included in the inside.

In step 1303, if it is included in a certain cluster, the process proceeds to step 1304,
The certainty factor based on the distance from the cluster centroid is calculated, and subsequently, in step 1305, the certainty factor based on the intra-cluster coordinate values is calculated. Then, in step 1306, these two certainty factors are added to obtain the final certainty factor. The above process is performed for all clusters (steps 1302 to 130).
7).

Here, the calculated certainty factor is the degree of belonging to the data group (category) used when creating the cluster. That is, the degree of belonging to A when entering the cluster a created in the latter knowledge, the degree of belonging to B when entering the cluster b, and the degree of belonging to C when entering the cluster c. It becomes a degree.

A specific example will be described for the category "0". The OK cluster in the category "0" is
0 for each of the data groups of 0, 6, and 9 categories
A cluster is created using axes 6, 6 and 9. Of these, if it enters a cluster created with a data group of 0, the degree of belonging is 0. In addition, when a cluster created by the data group 6 is included, the degree of belonging to 6 is entered, and when a cluster created by the group of 9 data is entered, the degree of belonging to 9 is entered.

Next, in step 1308, it is determined whether or not there is a cluster contained inside. If it is included in a plurality of clusters, in step 1309, the category of the cluster having the highest certainty factor is extracted, and the data group used when generating the cluster is used as the recognition result.

Next, in step 1312, of the clusters owned as knowledge, it is checked whether all the NG clusters owned by the category (recognition result) obtained in step 1309 are included in the inside. . If N
If it is included in the G cluster, step 13
In step 13, the certainty factor obtained in step 1309 is
Fine-tune it according to a rule so that it becomes smaller.
If it is not included inside, the result of step 1309 is confirmed.

The point to be noted here is step 1309.
This is because the recognition result by the OCR recognition processing unit 101 is overridden and the recognition result can be replaced to obtain the correct answer.

Next, the check of "OK cluster" (step 1303) will be described in detail. Whether or not the data to be determined is included in the cluster is calculated by converting the three axes used in clustering into new coordinates with the eigenvectors of the cluster and then substituting into the following ellipsoidal expression representing the ellipsoidal cluster. Is less than or equal to the covering distance of the ellipsoidal cluster. In the following equation, γ is the coordinate of three axes converted to the coordinate of the eigenvector.

[0133]

[Equation 3] If the distance is less than the covering distance, it is included in the cluster, and the confidence factor calculation process (step 1
304-1306). On the contrary, if the distance is equal to or more than the covering distance, the process proceeds to check another OK cluster.

Next, the confidence factor calculation process (step 130)
4 to 1306) will be described in detail. First, when it is determined that the recognition result is included in the cluster, the certainty factor of the distance from the cluster centroid to be calculated (step 130).
4) will be described.

To briefly explain the outline, in the file (see FIG. 11) secured when the cluster knowledge is generated,
dis_max, dis_median, dis_mi
The membership function is generated by using n (hereinafter referred to as distance point data), and the fitness to the function is calculated. The distance point data used here is data in which the distance from the center is calculated for all data forming the cluster, and the maximum, median, and minimum values among them are stored. The reason why the median value is used instead of the average value is that the data does not fluctuate much due to some distant values. However, at this time, the value of dis_max is only stored and not used.

The formula of the membership function adopted is as follows. The distance is calculated by inserting the distance of the recognition target character into the expression dis. Note that t> 0 and dis>
If dis_min, the degree of conformity = 1.

[0137]

[Equation 4] Further, FIG. 14 shows the membership function used for calculating the certainty factor of the distance from the center. Here, FIG.
The outline of No. 4 will be described. When the distance from the cluster center of the recognition target data is equal to or less than dis_min, the goodness of fit = 1 is given. The inflection point of the membership function is dis_medi
It is an. The magnitude of the calculated goodness of fit can be changed by changing the value of t. That is, the greater t is, the greater the calculation suitability is.

The "distance" used in the post-stage judging section 116 is a consideration of the axial length of the ellipsoid in the direction of the eigenvector, and the relationship between the membership function and the ellipsoid is the conformity of FIG. And ellipse. A two-dimensional ellipse is used for the explanation and has a spread in the x and y axis directions. The goodness of fit calculated here is referred to as the "certainty from the cluster centroid".

Next, the calculation of the certainty factor from the intra-cluster coordinates (step 1305) will be described. As described above, the post-stage knowledge base 11 by the post-stage knowledge creating unit 113 is used.
In the construction of No. 4, the frequency of the coordinate values of the data forming the cluster is calculated. The confidence factor assigned according to the frequency is freq in the stored knowledge shown in FIG.
iency. A description will be added in the case of a two-dimensional ellipsoid.

When a cluster is formed by the distribution of data as shown in FIG. 16, the function for obtaining the certainty factor is set so as to take a high value in the vicinity of frequently used coordinate values.
This indicates to which coordinate value the data belonging to this cluster is concentrated. FIG.
Shows that they are concentrated in the first quadrant. That is, when the data to be determined is included in the first quadrant,
Calculate high confidence. On the contrary, a low certainty factor is calculated in the third and fourth quadrants.

The peak value of the certainty factor function is obtained by dividing 1 by the number of axes used (number of cluster constituent axes).
That is, if a cluster is composed of two axes, the peak value of the certainty factor function of each axis is 0.5, and if it is three axes, it is 0.33. Therefore, when the certainty factors calculated for each axis are added together, the maximum value is set to 1.

Next, the maximum confidence factor extraction processing (step 1
309) will be described. There are cases where data to be confirmed is included in multiple clusters. In that case, the data group (category) used when generating the cluster that gives the maximum certainty factor among all the calculated certainty factors is valid at this time. As described above, the first candidate of the recognition result by the OCR recognition processing unit 101 may be invalid at this point.

Next, NG cluster check (step 1
308) will be described. When a valid category is calculated in the maximum confidence factor extraction process (step 1309), an NG cluster check is performed. As described above, the NG cluster is a cluster generated by collecting only the data erroneously recognized by the OCR recognition processing unit 101. The cluster generation method is similar to the OK cluster.

The NG clusters to be checked are all NG clusters belonging to the category calculated as a valid category. In other words, if 2 is determined to be a valid cluster, the check is performed on all NG clusters created by collecting data that is actually 2 but has been erroneously recognized by another category.

Next, the process of lowering the certainty factor (step 1
313) will be described. If a certain NG cluster is entered in the NG cluster check (step 1308),
A process of lowering the certainty factor calculated in the maximum certainty factor extraction process (step 1309) is performed. Like the OK cluster, the NG cluster is also created with the candidates that are likely to appear in the second and third candidates as the axis. That is, the OK cluster and the NG cluster exist in the same space (same axis).

FIG. 17 shows the positional relationship between OK clusters and NG clusters of a certain category. Second-stage knowledge base 11
As described in the construction method of No. 4, the character (A, B, C) that frequently appears in the second and third candidates of the category A is used as the axis, and the axis of the data group in which the correct answer data of the character A is collected. "OK cluster a" composed of A, axis B, and axis C
And “OK cluster b” composed of axis A, axis B, and axis C of a data group in which correct answer data of the letter B is collected,
"OK cluster c" composed of axis A, axis B, and axis C of the data group that collects correct answer data of the letter C, and axis A, axis B, and axis of the data group in which the correct answer is A but is mistaken for B C
The “NG cluster” composed of is shown in a two-dimensionally projected manner. However, the “OK cluster c” is not shown for the sake of simplicity.

As shown in FIG. 17, OK clusters a and O
If there is data to be confirmed in the area where the K cluster b overlaps, and the certainty factor of the OK cluster b is maximized in the OK cluster check (step 1303), A
It is assumed that there is data that has been erroneously recognized as B. At this time, since the data is also included in the NG cluster,
By the function shown on the axis inside the NG cluster in the figure, the calculated result is multiplied by the certainty factor of B to adjust. That is, in the region R4 where the overlap of the OK clusters a and b is large, the certainty factor of B is not so low, but the other region R
It becomes considerably lower for 1, R2 and R3.

Next, the processing by the rule determination section 117 will be described in detail. The rule determination unit 117 performs a process of determining whether the data is determined data or reject determination data as undetermined data, based on the certainty factor calculated by the certainty factor calculation unit of the subsequent determination unit 116. is there. FIG. 18 shows a flowchart of processing by the rule determination unit 117.

The content of the process is simple. In step 1801, if the certainty factor calculated by the certainty factor calculating means exceeds the predetermined value γ, the process proceeds to step 1802 to make the finalized data, and if the value is less than that, Step 1803
Proceed to step (4) and make the judgment data that cannot be convinced to be the correct answer, and set it as the rejection target data. The predetermined value used here can be set by the user and can be set arbitrarily. The larger the predetermined value, the smaller the fixed data.

As described above, the image recognition of this embodiment is performed.
According to the identification device, the knowledge generation unit 111
Then, the first pre-stage knowledge is created by the first pre-stage knowledge creating means.
(Relationship of similar category to complementary category)
2nd pre-stage knowledge by knowledge creation means
The degree of goodness of fit (distance) that can be determined as “correct answer”
The third-stage knowledge (first candidate is "correct"
The degree of goodness of fit (distance) that can be determined as a "solution"
Created in the previous knowledge base 112,
In the knowledge creation unit 113, the first second-stage knowledge creation means
1st second-stage knowledge (shape of OK cluster), second second-stage knowledge
Second-stage knowledge (shape of NG cluster) by knowledge creation means
Are created in the post-stage knowledge base 114, respectively.
Determination unit (first front stage determination unit, second front stage determination unit, and
3 pre-stage determination means) 115, these pre-stage knowledge bases 1
Input image using 12 first, second and third pre-knowledge
About the recognition result of the image, "correct answer" or "unknown" (which
Classification of the recognition results into two (data that does not belong to any category)
Of the input image that was confirmed as “Unknown” here.
For the recognition result, the second-stage determination unit (first second-stage determination unit or
And second second-stage determination means) 116, the second-stage knowledge base
Using the first and second latter knowledge of 114,
Data that has entered the OK cluster of
"OK" data, and to which OK cluster
If you don't enter, "Unknown" (does not belong to any category
Data) and then at least NG clusters
If 1 is included, it is determined as "incorrect answer" data and 3
Re-check by separating and confirming recognition results
I'm trying to do. This will reduce the false recognition rate
Can be lowered, labor saving in image recognition system
Can be made even more automated, and the form can be automatically read.
Can be realized. In addition, the recognition of the image recognition system
Replacing the recognition result with a "correct answer" recognition result
As a result, the recognition performance can be improved.

[0151]

As is apparent from the above description, according to the present invention, the first pre-stage knowledge creating means (first pre-stage knowledge creating step) is used to extract the data including the recognition result of image recognition and the accompanying information thereof. Data having the same first candidate category as the recognition result is collected, and second data is collected from the collected data .
In addition, the ones with a high appearance rate of similar categories in the following specific candidate ranks are extracted and held as first-stage knowledge that indicates the relationship of similar categories to the first candidate category.
In the first-stage determination means (first first-stage determination step), as the recognition result of the input image, the category of the first candidate and the second candidate
When a lower similar category with a specific candidate rank is obtained, by comparing the category of the first candidate and the similar category with a specific candidate rank lower than or equal to the second candidate with the relation held in the first pre-stage knowledge, Whether 1 candidate category is "correct"
Since it is decided to judge, it is not necessary for a person to check what was confirmed as a correct answer, and it suffices to check only those that could not be confirmed as a correct answer. Therefore, it is possible to provide the image recognition device, the image recognition method, and the recording medium that can further save the labor in the image recognition system.

Further, according to the present invention, by the second pre-stage knowledge creating means (second pre-stage knowledge creating step), from the data including the recognition result of the image recognition and its accompanying information, the category of the first candidate which is the same as the recognition result is obtained. Collect data with
Fit in the category of the first candidate from data collected the from (distance), and calculate the magnitude of the goodness of fit may be determined to "correct" the category of the first candidate (distance), held as second pre-knowledge Then, in the second pre-stage determination means (second pre-stage determination step), the magnitude of the goodness of fit (distance) in the category of the first candidate and the goodness of fit (distance) held in the second pre-stage knowledge as the recognition result of the input image. By comparison with the size of
And determining that the category of the first candidate "correct" or "unknown", also by the third front knowledge generating means (3 front knowledge generating step), the recognition result of the image recognition and data including the accompanying information From this, data having the same first candidate category as the recognition result is collected, and from the collected data, the category of the first candidate and the rank of the specific candidate below the second candidate
From the similarity (distance) with the similar category of
The degree of fitness (distance) difference that can determine the category of "correct answer" is calculated and held as the third pre-stage knowledge, and the third pre-stage determination means (third pre-stage determination step) recognizes the input image. As a result, the category of the first candidate and the features below the second candidate
When similar categories of constant candidate ranking is obtained, the size and the third front knowledge of fitness (distance) difference between similar categories of the first candidate <br/> category and the following specific candidate ranking second candidate
By comparison with the magnitude of the held goodness of fit (distance) difference,
Since it is determined whether or not the category of the first candidate is “correct answer”, it is not necessary for a person to check what has been confirmed as a correct answer, and only to check those that cannot be determined as a correct answer. As a result, it is possible to provide an image recognition device, an image recognition method, and a recording medium that can reduce the acceptance error recognition rate and can further reduce the labor in the image recognition system.

Further, according to the present invention, the first pre-stage knowledge creation means (first pre-stage knowledge creation step) previously creates the first pre-stage knowledge (relationship of the similar category to the first candidate category) with the second pre-stage knowledge creation. By means (second pre-stage knowledge creation step), the second pre-stage knowledge (a candidate with a specific rank is “correct”)
The degree of goodness of fit (distance) that can be decided as follows is the goodness of fit (distance) that can be decided as third corrective knowledge (the first candidate is “correct answer”) by the third preliminary knowledge creating means (third preliminary knowledge creating step). ). The magnitude of the difference) is created respectively, and the first front stage determination means (first front stage determination step), the second front stage determination means (second front stage determination step), and the third front stage determination means (third front stage determination step) In the step), using the knowledge of the first, second and third pre-stages, the recognition result of the input image is
Since it is decided whether or not the answer is "correct", the false recognition rate of acceptance can be reduced, the labor saving in the image recognition system can be further promoted, and the image recognition apparatus capable of automating the reading of the form, An image recognition method and a recording medium can be provided.

Further, according to the present invention, the first axis extracting means (first axis extracting step) samples and collects the data including the recognition result of the image recognition and its accompanying information, and the recognition result is "correct" in a certain category. The axis for clustering the set of data is extracted from the data described as ", and the first second-stage knowledge creating unit (first second-stage knowledge creating step) extracts the first axis-extracting unit (first axis-extracting step).
Clustering the set of data based on the axis extracted in step S6 to generate an OK cluster of the category,
When the shape of the generated OK cluster is held as the first post-stage knowledge and the first post-stage determination means (first post-stage determination step) obtains the data including the recognition result and the accompanying information as the recognition result of the input image, It is checked whether or not the data is included in the OK cluster belonging to the category having the first latter-stage knowledge. If the data is included in the OK cluster, the category to which the OK cluster belongs is recognized as the recognition result of the input image. As a result, it is determined as "correct answer", and when it is not included in any of the OK clusters, it is determined as "undetermined". Therefore, the false recognition rate of acceptance can be reduced, and labor saving in the image recognition system can be achieved. It is possible to provide an image recognition device, an image recognition method, and a recording medium that can further achieve the above.

Further, according to the present invention, the second axis extracting means (second axis extracting step) samples and collects the data including the recognition result of the image recognition and its accompanying information, and the recognition result in a certain category is "erroneous recognition". The axis for clustering the set of data is extracted from the data described as ", and the second second-stage knowledge creating means (second second-stage knowledge creating step) extracts the second axis-extracting means (second axis extracting step).
Clustering the set of data based on the axis extracted in step S6 to generate an NG cluster of the category,
When the shape of the generated NG cluster is held as second second-stage knowledge, and the second second-stage determination means (second second-stage determination step) obtains data including the recognition result and its accompanying information as the recognition result of the input image, It is checked whether or not the data is included in an NG cluster belonging to the category having the second latter knowledge, and if it is included in the NG cluster, the category to which the NG cluster belongs is recognized as the input image. As a result, it was judged as “incorrect” and which NG
If it is not included in the cluster, it is determined to be the “correct answer”, so the false recognition rate for acceptance can be reduced, and the image recognition device, the image recognition method, and the image recognition method that can further reduce the labor in the image recognition system. A recording medium can be provided.

According to the present invention, the first second-stage knowledge (shape of OK cluster) is previously stored in the second second-stage knowledge creating unit (second second-stage knowledge creating unit) by the first second-stage knowledge creating unit (first second-stage knowledge creating step). The second post-stage knowledge (shape of the NG cluster) is created by the knowledge creating step), and the first post-stage determining means (first post-stage determining step) and the second post-stage determining means (second post-stage determining step) These first
Using the knowledge of the second post-stage, it is determined that the data obtained in the OK cluster of a certain category is the “correct answer” data belonging to the category in the recognition result of the input image. After the data is set to "unknown" (does not belong to any category), and if at least one of the NG clusters is entered, it is determined to be "incorrect" data, and the recognition results are classified into three and confirmed. Therefore, the false recognition rate of acceptance can be reduced, and labor saving in the image recognition system can be further achieved.
Further, it is possible to automate the reading of the form, and further, since the recognition result of the image recognition system can be negated and replaced with the “correct” recognition result, an image recognition device and an image recognition method that can improve the recognition performance. And a recording medium can be provided.

Further, according to the present invention, the first pre-stage knowledge (relation of the similar category to the first candidate category) is previously created by the first pre-stage knowledge creating means (first pre-stage knowledge creating step) to the second pre-stage knowledge creating step. The second pre-stage knowledge ( the degree of goodness of fit (distance) that can determine the category of the first candidate as “correct answer”) by the means (second pre-stage knowledge creating step) is defined as the third pre-stage knowledge creating unit (third pre-stage knowledge). In the creation step), the third pre-stage knowledge (the magnitude of the difference in fitness (distance) that can determine the first candidate as the “correct answer”) is used as the first post-stage knowledge creating means (first
The first post-stage knowledge (OK cluster shape) is created by the post-stage knowledge creating step, and the second post-stage knowledge (NG cluster shape) is created by the second post-stage knowledge creating unit (second post-stage knowledge creating step). In the first front stage determination means (first front stage determination step), the second front stage determination means (second front stage determination step) and the third front stage determination means (third front stage determination step), these first, second and third Using the previous knowledge, the recognition result of the input image is classified into two, "correct answer" or "unknown" (data that does not belong to any category), and is confirmed. Regarding the recognition result of the input image obtained by the first rear-stage determination means (first rear-stage determination step) and the second rear-stage determination means (second rear-stage determination step), the first and second rear-stage knowledge is used to obtain a certain result. The entered data to the Gori of OK cluster belonging to the category "correct" data and decisions with,
On the contrary, if the data does not belong to any OK cluster, the data is set to "unknown" (does not belong to any category), and then N
"Incorrect answer" when entering at least one of G clusters
Since it is decided that the data is recognized and the recognition result is divided into three to be confirmed and the check is performed again, the false recognition rate of acceptance can be reduced and the labor saving in the image recognition system can be further achieved. In addition, it is possible to automate the reading of forms, and since the recognition result of the image recognition system can be negated and replaced with the “correct” recognition result, an image recognition device and an image that can improve recognition performance can be obtained. A recognition method and a recording medium can be provided.

Further, according to the present invention, in the first second-stage determination means (first second-stage determination step), when the data is included in the OK cluster belonging to a certain category, with certainty of confidence, "correct answer" The certainty factor, which is an index indicating whether or not it can be said, is calculated by the certainty factor calculation means (certainty factor calculation step), and when the data enters an OK cluster belonging to a plurality of different categories, the certainty factor is calculated. Since it was decided based on which category the "correct answer" should be set, data that has entered an OK cluster of a certain category is determined as data belonging to that category, and conversely, if it does not belong to any OK cluster, Although it is considered as "unknown" data (which does not belong to any category), when it is determined that the data belongs to a certain category, the confidence level of the recognition result can be calculated. By this, the recognition result can be classified and determined into three, "Unknown", "Correct answer", and "I can't be sure of correct answer". Furthermore, if at least one of the NG clusters is entered, it is determined as "wrong answer" data. "Unknown", "
It is possible to provide an image recognition device, an image recognition method, and a recording medium capable of classifying and recognizing recognition results into four categories, “correct answer”, “incorrect answer”, and “not sure that the answer is correct”.

Further, according to the present invention, in the second latter stage judging means (second latter stage judging step), when the data is included in the NG cluster belonging to a certain category, the confidence factor updating means (confidence factor updating) By determining the confidence level, which is an index indicating how much confidence can be said to be the “correct answer”, the data entered in the OK cluster of a certain category is determined as the data belonging to that category. On the contrary, if the data does not belong to any OK cluster, it is regarded as "unknown" data (which does not belong to any category), but if it is determined that the data belongs to a certain category, the confidence level of the recognition result is calculated. As a result, "unknown", "correct answer", and "I can't be sure of correct answer"
The recognition result can be separately classified and confirmed, and further, if at least one of the NG clusters is entered, it is determined as "incorrect answer" data or the calculated certainty factor can be lowered.

Summarizing the above, according to the present invention, as a result of the image matching having a plurality of categories, the degree of matching is high in the image recognition for calculating the degree of matching (or similar information) of each category with respect to the input data. Is not correct or incorrect, accepts the recognition result because it is confident that the recognition result is correct, rejects because the recognition result is not certain that it is correct, and rejects because the recognition result is inconclusive. , "And" Rejected because the recognition result is incorrect "
It is possible to provide an image recognition device, an image recognition method, and a recording medium that improve the acceptance correct rate and make the false recognition rate close to 0 [%].

[Brief description of drawings]

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

FIG. 2 is an explanatory diagram illustrating an overview of overall processing performed by the image recognition apparatus according to the embodiment.

FIG. 3 is a schematic flowchart of knowledge creation processing in the image recognition apparatus of the embodiment.

FIG. 4 is a schematic flowchart (No. 1) of processing during system operation in the image recognition apparatus of the embodiment.

FIG. 5 is a schematic flowchart (part 2) of processing during system operation in the image recognition apparatus of the embodiment.

FIG. 6 is an explanatory diagram illustrating a recognition result output from an OCR recognition processing unit.

FIG. 7 is an explanatory diagram illustrating parameters of first, second, and third pre-stage knowledge.

FIG. 8 is a flowchart illustrating a determination process performed by a former determination unit.

FIG. 9 is an explanatory diagram illustrating cluster generation.

FIG. 10 is a flowchart illustrating an outline of a hyperelliptic method clustering process.

FIG. 11 is an explanatory diagram exemplifying a data format when a file is output as second-stage knowledge.

FIG. 12 is an explanatory diagram illustrating a relationship between an ellipsoid and eigenvalues and eigenvectors.

FIG. 13 is a flowchart illustrating a processing flow in a latter-stage determination unit.

FIG. 14 is an explanatory diagram illustrating a membership function used in the calculation of the certainty factor of the distance from the center.

FIG. 15 is an explanatory diagram showing a relationship between a goodness of fit and an ellipse.

FIG. 16 is an explanatory diagram showing a relationship between an ellipse and a certainty factor function.

FIG. 17 is an explanatory diagram exemplifying a positional relationship between an OK cluster and an NG cluster of a certain category.

FIG. 18 is a flowchart of processing by a rule determination unit.

[Explanation of symbols]

101 OCR recognition processing unit 102 recognition result confirmation unit 103 knowledge generation unit 111 first-stage knowledge creation unit (first first-stage knowledge creation unit, second
1st-stage knowledge creation means and 3rd-stage knowledge creation means) 112 1st-stage knowledge base (1st-stage knowledge, 2nd-stage knowledge and 3rd-stage knowledge) 113 2nd-stage knowledge creation unit (1st 2nd-stage knowledge creation means and 2nd 2nd-stage knowledge) Creating means) 114 Second-stage knowledge base (first second-stage knowledge and second second-stage knowledge) 115 First-stage determination section (first front-stage determination means, second front-stage determination means and third front-stage determination means) 116 Second-stage determination section (first second-stage determination) Means and second latter-stage determination means) 117 rule determination section 118 recognition confirmation section 119 rejection section 151a, 151b recognition results 152,153 knowledge 154 unknown data 155, 156, 158 confirmed data 157 unconfirmed data 159 data that cannot be confirmed to be the correct answer

Continuation of the front page (56) Reference JP-A-4-88490 (JP, A) JP-A-9-128487 (JP, A) JP-A-1-277983 (JP, A) JP-A-8-329194 (JP , A) JP-A-11-15969 (JP, A) JP-A-63-8884 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G06K 9 / 00-9 / 82

Claims (20)

(57) [Claims] 1. An image recognition apparatus which has a plurality of categories of recognition targets and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, and a recognition result of the image recognition and its recognition result. Data having the same first candidate category as the recognition result is collected from the data including the accompanying information, and those having a high appearance rate of similar categories in the specific candidate ranks of the second candidate and below are extracted from the collected data, First pre-stage knowledge creating means for holding as first pre-stage knowledge showing the relationship of similar categories to the first candidate category, and the first candidate category and the first candidate category as the recognition result of the input image .
When a similar category with a specific candidate rank of 2 candidates or less is obtained, the category of the first candidate and the similar category with a specific candidate rank of 2 or less candidates and the association held in the first pre-stage knowledge are obtained. By comparison, the category of the first candidate is “correct”
An image recognition apparatus comprising: a first front-stage determination means for determining whether or not 2. A recognition result of image recognition and its recognition result in an image recognition device having recognition objects of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. Data having the same first candidate category as the recognition result is collected from the data including the accompanying information, and the first candidate is calculated from the matching degree (distance) in the first candidate category from the collected data . The second pre-stage knowledge creating means for calculating the magnitude of the goodness of fit (distance) that can determine the category of "correct answer" as the second pre-stage knowledge, and the first candidate category as the recognition result of the input image are obtained. The degree of matching (distance) in the category of the first candidate when
The size and compared with the size of the retained fit the second pre-knowledge (distance), the category of said first candidate
An image recognition apparatus comprising: a second front-stage determining means for determining whether or not is a "correct answer". 3. A recognition result of image recognition and its recognition result in an image recognition device having recognition objects of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. Data having the same first candidate category as the recognition result is collected from the data including the accompanying information, and the first candidate category and the first candidate category are collected from the collected data .
From the goodness- of- fit (distance) difference with the similar category of the specific candidate rank of two candidates or less, the magnitude of the goodness-of-fit (distance) difference that can determine the category of the first candidate as “correct answer” is calculated, and the third front stage The third pre-stage knowledge creating means that holds the knowledge, and the first candidate category and the first candidate as the recognition result of the input image .
When similar categories 2 candidates following specific candidate ranking is obtained, the first candidate of the category and the second candidate following specific candidate sequence
Whether or not the category of the first candidate is the “correct answer” by comparing the magnitude of the difference in fitness (distance) with the similar category of rank and the magnitude of the difference in fitness (distance) held in the third preceding knowledge.
An image recognition apparatus, comprising: a third front-stage determination means for determining. 4. An image recognition apparatus which has a plurality of categories of recognition targets and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern and a recognition result of the image recognition and its recognition result. Data that includes incidental information is sampled, and from the data whose recognition result is "correct" in a certain category, the axis when clustering the above-mentioned data set is the axis with the category with a high appearance rate as the top candidate. , The lower candidate is a category with a high appearance rate, the upper candidate is a category with a high appearance rate and the lower candidate is a category with a high appearance rate, and the middle candidate is a category with a high appearance rate , The first axis extractor that extracts categories with high correlation coefficient as the axis, or categories with low correlation coefficient as the axis And clustering the set of data on the basis of the axes extracted by the first axis extracting means to generate an OK cluster of the category, and retaining the shape of the generated OK cluster as the first second-stage knowledge. When data including the recognition result and its accompanying information as the recognition result of the input image is obtained, whether the data is included in the OK cluster belonging to the category having the first latter knowledge If it is included in the OK cluster, the category to which the OK cluster belongs is determined as “correct answer” as the recognition result of the input image, and if it is not included in any OK cluster, An image recognition apparatus comprising: a first rear-stage determination unit that determines "undetermined". 5. The first second-stage determination means sets a certainty factor, which is an index indicating how much confidence can be said to be a “correct answer” when the data is included in an OK cluster belonging to a certain category. A certainty factor calculating means for calculating is provided, and when the data enters an OK cluster belonging to a plurality of different categories, which category is determined as the “correct answer” is determined based on the certainty factor. The image recognition device according to claim 4. 6. An image recognition apparatus which has a plurality of categories of recognition targets and recognizes an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, and a recognition result of the image recognition and its recognition result. Data that includes incidental information is sampled, and from the data whose recognition result is "misidentified" in a certain category, the axis when clustering the above-mentioned data sets the category with a high appearance rate as the top candidate as the axis. , The lower candidate is a category with a high appearance rate, the upper candidate is a category with a high appearance rate and the lower candidate is a category with a high appearance rate, and the middle candidate is a category with a high appearance rate , A second axis extractor for extracting categories with high correlation coefficient as the axis, or categories with low correlation coefficient as the axis And clustering the set of data based on the axes extracted by the second axis extracting means to generate NG clusters of the category, and holding the shape of the generated NG cluster as second post-stage knowledge. When the data including the recognition result and its accompanying information as the recognition result of the input image is obtained by the latter-stage knowledge creating means, whether or not the data is included in the NG cluster belonging to the category having the second latter-stage knowledge. If any of the NG clusters is included in the NG cluster, the category to which the NG cluster belongs is determined to be “incorrect” as the recognition result of the input image.
An image recognition apparatus, comprising: a second latter-stage determination unit that determines “correct answer” if it is not included in the cluster. 7. An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, the recognition result of the image recognition and its recognition result. Data having the same first candidate category as the recognition result is collected from the data including the accompanying information, and those having a high appearance rate of similar categories in the specific candidate ranks of the second candidate and below are extracted from the collected data, a first front knowledge creation step of holding the first pre-knowledge to related similar categories for the first candidate category, the category of the first candidate as a recognition result of the input image and the
When similar categories 2 candidates following specific rank candidate is obtained by comparison with related held in the first pre-knowledge similar category of the first candidate category and a specific candidate ranking, the first candidate category A first pre-stage determination step of determining whether or not is a “correct answer”, and the image recognition method. 8. An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, the recognition result of the image recognition, and the recognition result thereof. from data including accompanying information, from the recognition result as collected data with the categories of the same first candidate, Contact <br/> Keru fit from the collected data to the category of the first candidate (distance), the first candidate The second pre-stage knowledge creation step of calculating the degree of goodness of fit (distance) that can determine the category of "correct answer" as the second pre-stage knowledge, and the first candidate category as the recognition result of the input image are obtained. when it is, it fits in the category of the first candidate (distance)
The size and compared with the size of the retained fit the second pre-knowledge (distance), the category of said first candidate
And a second pre-stage determination step of determining whether or not is a "correct answer". 9. An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, the recognition result of the image recognition and its recognition result. Data having the same first candidate category as the recognition result is collected from the data including the accompanying information, and the first candidate category and the first candidate category are collected from the collected data .
From the goodness- of- fit (distance) difference with the similar category of the specific candidate rank of two candidates or less, the magnitude of the goodness-of-fit (distance) difference that can determine the category of the first candidate as “correct answer” is calculated, and the third front stage The third pre-stage knowledge creation step that holds as knowledge, and the first candidate category and the first candidate as the recognition result of the input image .
When similar categories 2 candidates following specific candidate ranking is obtained, the first candidate of the category and the second candidate following specific candidate sequence
The degree of conformance (distance) difference with the rank category and the third
A third pre-stage determination step of determining whether or not the category of the first candidate is “correct” by comparing with the magnitude of the goodness-of-fit (distance) difference held in the pre-stage knowledge. An image recognition method characterized by: 10. An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, the recognition result of the image recognition, and the recognition result thereof. Data having the same first candidate category as the recognition result is collected from the data including the accompanying information, and those having a high appearance rate of similar categories in the specific candidate ranks of the second candidate and below are extracted from the collected data, From the first pre-stage knowledge creating step for holding the first pre-stage knowledge indicating the relation of the similar category to the first candidate category, and the data including the recognition result of image recognition and its accompanying information, the same first candidate category is selected as the recognition result. having collected data, <br/> Keru fit from the collected data to the category of the first candidate from (distance) of the first candidate Categories were calculated size of the goodness of fit may be determined as "correct" (distance) from the data including a second pre knowledge generating step of holding the second pre-knowledge, the recognition result of the image recognition and the accompanying information, Data having the same first candidate category as the recognition result is collected, and the first candidate category and the first candidate category are collected from the collected data .
From the goodness- of- fit (distance) difference with the similar category of the specific candidate rank of two candidates or less, the magnitude of the goodness-of-fit (distance) difference that can determine the category of the first candidate as “correct answer” is calculated, and the third front stage The third pre-stage knowledge creation step that holds as knowledge, and the first candidate category and the first candidate as the recognition result of the input image .
When similar categories 2 candidates following specific rank candidate is obtained by comparison with related held in the first pre-knowledge similar category of the first candidate category and a specific candidate ranking, the first candidate category a first pre-determination step determines whether but "correct" or for determining input image to "correct" in the first pre-determination step, the size of the fit in the category of the first candidate (distance) by comparison with the size of the retained fit the second pre-knowledge (distance), catheterization of the first candidate
Gori and a second pre-determination step of determining whether "correct", the determination input image to "correct" in the second pre-determination step, of the first candidate category and the following second candidate
By comparing the magnitude of the difference in fitness (distance) with the category of the specific candidate rank with the magnitude of the difference in fitness (distance) held in the third preceding knowledge, the category of the first candidate is “correct”.
And a third pre-stage determination step for determining whether or not the image recognition method. 11. An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, the recognition result of the image recognition and its recognition result. Data that includes incidental information is sampled, and from the data whose recognition result is "correct" in a certain category, the axis when clustering the above-mentioned data set is the axis with the category with a high appearance rate as the top candidate. , The lower candidate is a category with a high appearance rate, the upper candidate is a category with a high appearance rate and the lower candidate is a category with a high appearance rate, and the middle candidate is a category with a high appearance rate , 1st axis extraction to extract categories with high correlation coefficient as the axis, or categories with low correlation coefficient as the axis And step performs clustering of the set of data based on the axis extracted by said first shaft extraction step, OK of the categories
When a data including a recognition result and its accompanying information is obtained as a recognition result of the input image, a first second-stage knowledge creating step of generating the cluster as an OK cluster and retaining the shape of the generated OK cluster as the first second-stage knowledge, Is included in the OK cluster belonging to the category having the first latter knowledge, and if it is included in the OK cluster, the category to which the OK cluster belongs is recognized as the recognition result of the input image. And a first post-stage determination step for determining “correct answer” and determining “undetermined” if the cluster is not included in any OK cluster. 12. An image recognition method for recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern, the recognition result of the image recognition, and the recognition result thereof. Data that includes incidental information is sampled, and from the data whose recognition result is "misidentified" in a certain category, the axis when clustering the above-mentioned data sets the category with a high appearance rate as the top candidate as the axis. , The lower candidate is a category with a high appearance rate, the upper candidate is a category with a high appearance rate and the lower candidate is a category with a high appearance rate, and the middle candidate is a category with a high appearance rate , 2nd axis extraction to extract categories with high correlation coefficient as the axis, or categories with low correlation coefficient as the axis Step and, based on the axis extracted by said second shaft extraction step performs clustering of said set of data, NG of the categories
When a data including a recognition result and its accompanying information is obtained as a recognition result of the input image, the data is generated as a cluster and holds the shape of the generated NG cluster as the second knowledge. Is included in the NG cluster belonging to the category having the second latter knowledge, and if included in the NG cluster, the category to which the NG cluster belongs is recognized as the recognition result of the input image. Is judged as “incorrect” and which NG
An image recognition method comprising: a second post-stage determination step of determining “correct answer” if the cluster is not included in the cluster. 13. The first post-stage determination step is performed to determine a certainty factor, which is an index indicating how much confidence can be said to be a “correct answer” when the data is included in an OK cluster belonging to a certain category. A certainty factor calculation step for calculating, wherein when the data enters an OK cluster belonging to a plurality of different categories, which category is determined as the “correct answer” is determined based on the certainty factor. The image recognition method according to claim 11. 14. A program for causing a computer to execute an image recognition method having recognition targets of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the recording medium in which is recorded, data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the specific candidate ranks of the second candidate and below are collected from the collected data . In the first pre-stage knowledge creation step of extracting the one having a high appearance rate of the similar category in 1) and retaining it as the first pre-stage knowledge showing the relation of the similar category to the first candidate category, and the first candidate as the recognition result of the input image . Category and number
When a similar category with a specific candidate rank of 2 candidates or less is obtained, the category of the first candidate and the similar category with a specific candidate rank of 2 or less candidates and the association held in the first pre-stage knowledge are obtained. By comparison, the category of the first candidate is “correct”
A first pre-stage determination step of determining whether or not the recording medium is provided. 15. A program for causing a computer to execute an image recognition method having recognition objects of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the recording medium recorded as, the data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the collected data is classified into the first candidate category. /> from kicking fit (distance), and the category of the first candidate to calculate the size of the goodness of fit may be determined as "correct" (distance), second pre-knowledge generating step of holding the second pre-knowledge , when the category of the first candidate is obtained as a recognition result of the input image, the size and the second pre knowledge of the first candidate category definitive fitness (distance) Lifting been fit (distance)
Of the comparison between the size, the recording medium that comprises a said second pre determination step of determining whether the first candidate category is "correct" or the features a. 16. A program for causing a computer to execute an image recognition method having recognition objects of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the recording medium recorded as, the data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the first candidate category and the first candidate category are collected from the collected data .
From the goodness- of- fit (distance) difference with the similar category of the specific candidate rank of two candidates or less, the magnitude of the goodness-of-fit (distance) difference that can determine the category of the first candidate as “correct answer” is calculated, and the third front stage The third pre-stage knowledge creation step that holds as knowledge, and the first candidate category and the first candidate as the recognition result of the input image .
When similar categories 2 candidates following specific candidate ranking is obtained, the first candidate of the category and the second candidate following specific candidate sequence
By comparison with the size of the goodness of fit (distance) difference in magnitude and said third front knowledge of position of fitness (distance) difference, said first
And a third pre-stage determination step of determining whether or not the candidate category is “correct”. 17. A program for causing a computer to execute an image recognition method having recognition targets of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the recording medium recorded as, the data having the same first candidate category as the recognition result is collected from the data including the recognition result of the image recognition and its accompanying information, and the specific candidate ranks of the second candidate and below are collected from the collected data . The first pre-stage knowledge creation step of extracting a high-similarity appearance rate of similar categories in 1) and holding it as first pre-stage knowledge showing the relationship of the similar categories to the first candidate category, the recognition result of image recognition and its accompanying information. From the included data, data having the same first candidate category as the recognition result is collected, and the collected data is collected. Et Contact <br/> Keru fit into the category of the first candidate from (distance), and calculate the magnitude of the goodness of fit may be determined to "correct" the category of the first candidate (distance), second pre-knowledge a second pre-knowledge generating step of holding the recognition result of the image recognition and the data including the accompanying information, collects data having the category of the same first candidate as a recognition result, the first candidate from the collected data category And the
From the goodness- of- fit (distance) difference with the similar category of the specific candidate rank of two candidates or less, the magnitude of the goodness-of-fit (distance) difference that can determine the category of the first candidate as “correct answer” is calculated, and the third front stage The third pre-stage knowledge creation step that holds as knowledge, and the first candidate category and the first candidate as the recognition result of the input image .
When a similar category with a specific candidate rank of 2 candidates or less is obtained, the category of the first candidate and the similar category with a specific candidate rank of 2 or less candidates and the association held in the first pre-stage knowledge are obtained. By comparison, the category of the first candidate is “correct”
A first pre-stage determination step for determining whether or not the first pre-stage determination step is performed, and for the input image determined as “correct answer” in the first pre-stage determination step, the degree of goodness of fit (distance) in the category of the first candidate and the second pre-stage by comparison with the size of the retained fitness knowledge (distance), catheterization of the first candidate
Gori and a second pre-determination step of determining whether "correct", the determination input image to "correct" in the second pre-determination step, of the first candidate category and the following second candidate
Whether the category of the first candidate is the “correct answer” by comparing the magnitude of the difference in fitness (distance) with the similar category of the specific candidate rank and the magnitude of the difference in fitness (distance) held in the third previous knowledge. A third pre-stage determination step of determining whether or not the recording medium is provided. 18. A program for causing a computer to execute an image recognition method having recognition objects of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the recording medium recorded as, a sample of data including the recognition result of image recognition and its accompanying information is sampled, and the axis when clustering the set of the data from the data whose recognition result is "correct" in a certain category , The category with high appearance rate in the upper candidates is the axis, the category with high appearance rate in the lower candidates is the axis, and the category with high appearance rate in the upper candidates and the category with high appearance rate in the lower candidates is the axis , Categories with a high appearance rate in the middle candidates as the axis, categories with high correlation coefficients as the axis, There performs the shaft lower category between correlation coefficients, a first shaft extracting, clustering of said set of data based on the axis extracted by said first shaft extraction step, OK of the categories
When a data including a recognition result and its accompanying information is obtained as a recognition result of the input image, a first second-stage knowledge creating step of generating the cluster as an OK cluster and retaining the shape of the generated OK cluster as the first second-stage knowledge, Is included in the OK cluster belonging to the category having the first latter knowledge, and if it is included in the OK cluster, the category to which the OK cluster belongs is recognized as the recognition result of the input image. And a first post-stage determination step of determining “correct answer” as “correct answer” and determining as “undetermined” when not included in any OK cluster. 19. A program for causing a computer to execute an image recognition method having recognition objects of a plurality of categories and recognizing an input image based on a matching degree (distance) in pattern matching between a registered image pattern and an input image pattern. In the recording medium recorded as, a sample of data including the recognition result of image recognition and its accompanying information is collected, and the axis when clustering the set of the data from the data whose recognition result is "misidentified" in a certain category , The category with high appearance rate in the upper candidates is the axis, the category with high appearance rate in the lower candidates is the axis, and the category with high appearance rate in the upper candidates and the category with high appearance rate in the lower candidates is the axis , Categories with a high appearance rate in the middle candidates as the axis, categories with high correlation coefficients as the axis, There is, as a shaft of the lower category between correlation coefficients, and the second shaft extracting performs clustering of the set of data based on the axis extracted by said second shaft extraction step, NG of the categories
When a data including a recognition result and its accompanying information is obtained as a recognition result of the input image, the data is generated as a cluster and holds the shape of the generated NG cluster as the second knowledge. Is included in the NG cluster belonging to the category having the second latter knowledge, and if included in the NG cluster, the category to which the NG cluster belongs is recognized as the recognition result of the input image. Is judged as “incorrect” and which NG
A recording medium, comprising: a second latter-stage determination step of determining a “correct answer” if it is not included in the cluster. 20. The first latter-stage determination step includes a certainty factor, which is an index indicating how much certainty can be said to be a “correct answer” when the data is included in an OK cluster belonging to a certain category. A certainty factor calculation step for calculating, wherein when the data enters an OK cluster belonging to a plurality of different categories, which category is determined as the “correct answer” is determined based on the certainty factor. The recording medium according to claim 18.