JP2009186243A - Discriminator, discrimination method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a discriminator, a discrimination method and a program capable of adjusting a discrimination rule so that a desired discrimination result can be acquired to a wrongly determined determination object, while suppressing an influence on another discrimination result. <P>SOLUTION: This discriminator 100 for discriminating to which assembly among a plurality of kinds a determination object belongs has a discrimination plane on which a point in a space which is a mapping destination is linearly separated, when learning characteristic information is expressed as the point in the space by characteristic information and the point is mapped, based on the learning characteristic information having a known discrimination result and learning discrimination result information; a discrimination function calculation part 112 for calculating the discrimination plane so that a distance to a point having the shortest distance to the discrimination plane becomes maximum, and using the result as a discrimination function; a discrimination function correction part 120 for adjusting an influence coefficient to the discrimination function of wrongly discriminated characteristic information wherein discrimination result information by the discrimination function is different from a known discrimination result; and a discrimination part 130 for discriminating a discrimination object based on characteristic information of the discrimination object having an unknown discrimination result and a corrected discrimination function. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a discriminating apparatus, a discriminating method, and a program for determining pass / fail or a plurality of data obtained in a production line or the like.

  In recent years, for example, production lines have been automated in production plants, and various inspections have been performed in order to perform quality control of products produced in the production plants. Advances in inspection technology make it possible to measure the details of defects such as product defects in detail, and it is important to determine what kind of defect the detected defect is based on the measurement data It is. Note that the assumed data often includes data of a plurality of measurement items as elements.

  For example, in a steel plate production line, an inspection device that detects defects such as wrinkles on the surface of a steel plate by performing predetermined image processing on a captured image of the steel plate by a camera arranged on the production line is used. In this inspection apparatus, an image is binarized, and an area (blob, ON (= 1) area) estimated to represent a defect in the image is extracted as label candidate data (an example of measurement data) by labeling. To do. Then, in the inspection apparatus, for example, if the area is a harmful defect or harmless noise from the feature information (feature amount) of the area, such as height, width, area, and luminance distribution, Further, the type (defect type) of the defect and the degree of harm (severity level, grade) are determined. For example, when an unacceptable defect has occurred or a defect has occurred due to a defect in the production line itself, there are many cases in which early countermeasures are required. Can greatly affect quality.

  In order to execute such discrimination processing, it is necessary to define the discrimination rule. However, it is generally necessary to construct a huge number of complex discrimination rule groups in order to realize performance discrimination processing that can withstand practical use. It is not realistic for the user to construct the discriminant rule group himself because the load on the user at the time of construction becomes enormous and the maintainability becomes very poor.

  Therefore, for example, in order to derive a discrimination rule to be used in a discrimination device, a necessary number of learning data including feature information and a discrimination result (correct information) by a discrimination operator of the feature information is prepared, and based on the learning data A learning function has been developed to automatically construct discrimination rules.

  As such learning functions, for example, production systems, neural networks, decision trees, and the like have been studied. For example, as shown in Non-Patent Document 1 below, these learning functions are applied to actual problems.

JP 2003-344300 A “Pattern recognition”, Morio Onoe (translation), 2001 “Statistical Learning Theory”, V.Vapnik, 1998

On the other hand, SVM (Support Vector Machine Method) is one of the methods that have been attracting attention as a learning function in recent years (see Non-Patent Document 2 above). This SVM has the following features, for example.
(1) The optimization calculation using the evaluation function set so as to balance the correct answer with the over-learning which is a decrease in which the error with respect to other data increases due to the fact that it corresponds too much to the learning data, and each Based on the data points, a discrimination boundary between harmful defects and harmless noise is derived by calculation in a space (feature amount space) spanned by each feature amount of feature information as a coordinate axis.
(2) The prepared learning data is mapped from the feature space to another high-dimensional space by a virtually defined mapping function, and is discriminated by the hyperplane in the high-dimensional space (original In space, it is identified by a curved surface).
(3) A discriminant boundary (hyperplane) obtained as a result of optimization calculation is expressed as a discriminant function by some linear combination of prepared learning data.

  This learning function by SVM has an excellent characteristic that it can learn a complicated discrimination rule based on the above-described features and the like, and at the same time, is difficult to fall into overlearning.

  However, the calculation process of the discrimination rule by the SVM has many black box side aspects for the user, and it is often difficult to adjust the discrimination rule flexibly so as to exhibit a desired discrimination performance. For example, if an optimization calculation is performed by setting a discriminant function necessary for calculation by SVM and parameters included in the discriminant function, a discriminant rule adjusted to some extent can be calculated. However, for example, when a part of the learning data includes errors or misleading data, the determination rule expected by the user may not always be calculated. That is, for example, there may be a case where a determination rule that leads to an erroneous determination result is calculated for feature information that cannot allow an erroneous determination result to be derived.

  In some cases, it is possible for the user to adjust parameters or the like in order to calculate a desired determination rule based on the determination rule calculated by the SVM. For example, the user's desire can be reflected to some extent by the user adjusting the parameters of the evaluation function and the mapping function later. However, the adjustment of the discrimination rule by these parameter adjustments is an indirect adjustment, and there are many cases where sufficient adjustment cannot be performed. That is, these parameters are optimization calculation parameters, and the determination result based on the determination rule cannot be directly adjusted.

  The reason why the discrimination rule may not be sufficiently adjusted by adjusting the parameters may be due to incomplete learning data such as lack of feature information or contradiction of correct answers. However, enormous maintenance costs and time are required to correct learning data and perform relearning. In the first place, it is often difficult to prepare perfect learning data in terms of accuracy of correct answers. Therefore, in the application of practical problems (practical use), the desired discrimination results can be obtained without sacrificing enormous maintenance costs and time, even if the optimality of discrimination rules is sacrificed to some extent using actual learning data. May be desired.

  In Patent Document 1, a hyperplane in the high-dimensional space corresponding to a discrimination rule is translated so that a desired discrimination result can be obtained so that erroneously determined feature information can be correctly discriminated. A technique for adjusting a discrimination rule is disclosed. However, translating the hyperplane in this way may have a significant effect on the feature information near the hyperplane (near the boundary). In particular, in the SVM, the feature information in the vicinity of the hyperplane determines the discrimination rule, so that there is a possibility that the feature information in the vicinity of the hyperplane is erroneously determined.

  The present invention has been made in view of the above problems, and an object of the present invention is to determine the suitability and quality of data composed of a plurality of feature amounts and elements such as bag data. When deriving a discrimination rule by SVM using the learning data of the data, the other judgments in a plurality of data are obtained so that a desired discrimination result can be obtained for a predetermined judgment target such as an erroneously judged data point. An object of the present invention is to provide a discriminating apparatus, a discriminating method, and a program capable of adjusting discriminant rules more flexibly than before while suppressing the influence on the discrimination result of data points.

  In order to solve the above-described problem, according to an aspect of the present invention, a discriminant function having as input data feature information composed of feature quantities characterizing a discriminant to which of a plurality of types of discriminants belong. , And a discrimination device that discriminates based on a result that is an output value of the discriminant function, wherein the discrimination result represents a plurality of known learning feature information and a discrimination result associated with each of the learning feature information Based on the learning discrimination result information, each of the learning feature information is represented as a point on the space by the feature information, and when the point on the space is mapped by a virtually defined mapping function, The discriminant plane is calculated using the support vector machine method so that the distance between the discriminant plane that linearly separates the points according to the discriminant results and the point with the shortest distance to the discriminant plane is maximized. A discriminant function calculation unit for deriving the discriminant function, and feature information having a known discriminant result, the discriminant result information obtained by inputting the feature information into the discriminant function, and a known discriminant result of the feature information For the misclassification feature information different from the above, the influence coefficient representing the degree of influence of the misclassification feature information on the discrimination function is adjusted so that the discrimination result information matches the known discrimination result, A discriminant function correcting unit that corrects the discriminant function; and a discriminating unit that discriminates a discriminant whose discrimination result is unknown using the discriminant function that has been corrected using the feature information of the discriminant as input data. A discriminating device is provided.

  According to this configuration, the discriminant function using the learning function based on the support vector machine method can be derived by the discriminant function calculation unit based on the learning feature information and the learning discriminant result information. Then, the discriminant function correcting unit can adjust the influence coefficient of the misclassification feature information in which the discrimination result is incorrect. In other words, the discriminant function correcting unit can directly adjust a term in the discriminant function that is associated with erroneous discrimination feature information. Therefore, the discriminant function can be modified so that a correct discriminant result can be derived even for misclassified feature information.

The discriminant function is a sum of kernel function values of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and the respective influence coefficients. And the discriminant function correcting unit performs a predetermined process so that the discriminant result of the discriminant result is correct when any of the learning feature information is erroneous discriminant feature information. The discriminant function may be modified by adjusting the influence coefficient of the misclassification feature information.
According to this configuration, the discriminant function can be modified so that a correct discriminant result can be derived for the learning feature information.

The discriminant function is a sum of kernel function values of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and the respective influence coefficients. And the discriminant function correcting unit newly adds the misidentification feature information misidentified by the discriminator to the discriminant function formed by one or more of the learning feature information, and the misclassification The discriminant function may be corrected by a predetermined process so that the magnitude of the influence coefficient of the feature information is adjusted and the discrimination result of the discrimination result is correct.
According to this configuration, the discriminant function can be modified so that a correct discrimination result can be derived even with respect to erroneous discrimination feature information that is not feature information for learning.

Further, the discriminant function calculating unit calculates the discriminant surface for each combination of two sets in the plurality of types of sets, and each of the discriminant surfaces is used as a discriminant function. May be determined to belong to the set determined by the largest number of the discriminant functions as a result of discrimination for each of the two sets based on each of the discriminant functions.
According to this configuration, when discriminating from a plurality of types of sets, the discriminant function calculating unit can calculate a discriminant function for each combination of two sets in the plurality of types of sets. Then, the discriminating object whose discrimination result is unknown can be discriminated by the discriminating unit as belonging to the set most frequently discriminated using the discriminant functions. That is, for example, when the number of sets is N, the discriminant function calculating unit can calculate _N C ₂ discriminant functions. Then, the determination unit calculates _N C ₂ determination results using the _N C ₂ determination functions, and determines that the determination target belongs to the set having the largest number of determination results. When it is determined that the discrimination target belongs to one set, the one set can be discriminated based on the discrimination results of N-1 discriminant functions at the maximum.

Further, the discriminant function correcting unit may correct each of the discriminant functions until at least a known discrimination result of the erroneous discrimination feature information matches a discrimination result by the discriminator.
According to this configuration, each of the discriminant functions can be corrected by the discriminant function correcting unit. This correction is performed until the known discrimination result matches the discrimination result by the discrimination unit. That is, as described above, the discriminating unit discriminates the discrimination target in the set discriminated by the most discriminant functions. Therefore, the discriminant function correcting unit corrects each discriminant function so that at least a set represented by known discriminant results is discriminated by the most discriminant functions. That is, for example, it is possible to modify at least the number of discriminant functions that maximizes the number of discriminant functions that use the set of known discriminant results as the discriminant result.

  Further, the discrimination target is a defective state generated in the product to be manufactured in a manufacturing plant that manufactures the product, and the feature information may include one or more measurement values obtained as a result of inspecting the defective state. Good.

  In order to solve the above-described problem, according to another aspect of the present invention, input data includes feature information including feature quantities that characterize a discrimination target to which of a plurality of types of sets the discrimination target belongs. A discriminant function using a discriminant function, and discriminating the discriminant result corresponding to each of the plurality of learning feature information and the learning feature information. Based on the learning discrimination result information that represents the result, each of the learning feature information is represented as a point in the space by the feature information, and when the point in the space is mapped by a virtually defined mapping function, The discriminant plane is determined using the support vector machine method so that the distance between the discriminant plane that linearly separates the points according to discrimination results in space and the point with the shortest distance to the discriminant plane is maximized. A discriminant function calculation step for calculating and deriving the discriminant function; and feature information having a known discriminant result, and the discriminant result information obtained by inputting the feature information into the discriminant function and the known feature information For misclassification feature information with different discrimination results, an influence coefficient representing the degree of influence of the misclassification feature information on the discrimination function is adjusted so that the discrimination result information matches the known discrimination result A discriminant function correcting step for correcting the discriminant function; and a discriminating step for discriminating a discriminant whose discrimination result is unknown by using the discriminant function corrected using the characteristic information of the discriminant as input data. There is provided a discrimination method characterized by comprising:

  The discriminant function is a sum of kernel function values of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and the respective influence coefficients. Yes, in the discriminant function correction step, if any of the learning feature information is misidentified feature information for the discriminant result of the discriminant function, the discriminant function corrects the discriminant result in a predetermined process so that the discriminant result is correct. You may adjust the magnitude | size of the influence coefficient of misidentification characteristic information.

  The discriminant function is a sum of kernel function values of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and the respective influence coefficients. Yes, in the discriminant function correcting step, the discriminant feature information erroneously discriminated by the discriminator is newly added to the discriminant function formed by one or more of the learning feature information, and the misclassification The discriminant function may be corrected by a predetermined process so that the magnitude of the influence coefficient of the feature information is adjusted and the discrimination result of the discrimination result is correct.

  In the discriminant function calculating step, the discriminant plane is calculated for each combination of two sets in the plurality of types of sets, and each discriminant plane is used as a discriminant function. In the discriminant step, the discriminant result is unknown. May be determined to belong to the set determined by the largest number of the discriminant functions as a result of discrimination for each of the two sets based on each of the discriminant functions.

  In the discriminant function correcting step, each of the discriminant functions may be corrected until at least the known discriminant result of the erroneous discriminant feature information matches the discriminant result in the discriminating step.

  Further, the discrimination target is a defective state generated in the product to be manufactured in a manufacturing plant that manufactures the product, and the feature information may include one or more measurement values obtained as a result of inspecting the defective state. Good.

  In order to solve the above-described problem, according to another aspect of the present invention, a computer is provided with feature information including feature amounts that characterize a discrimination target to which of a plurality of types of sets the discrimination target belongs. A program for using a discriminant function as input data and executing a discriminating procedure for discriminating based on a result that is an output value of the discriminant function, wherein the computer has a plurality of learning feature information whose discrimination results are known. Based on the learning discrimination result information indicating the discrimination result associated with each of the learning feature information, each of the learning feature information is represented as a point on the space by the feature information, and the mapping function defined virtually When a point in the space is mapped, the distance between the discriminating surface that linearly separates the point according to the discrimination result in the mapping destination space and the point with the shortest distance to the discriminating surface is the maximum. A discriminant function calculation procedure for deriving the discriminant function by calculating the discriminant plane using a support vector machine method, and feature information having a known discriminant result, wherein the feature is included in the discriminant function For misidentification feature information in which discrimination result information that is a result of inputting information differs from a known discrimination result of the feature information, the misdiscrimination feature information is set so that the discrimination result information matches the known discrimination result. Adjusts the influence coefficient that expresses the degree of influence of the discriminant function on the discriminant function, corrects the discriminant function, and uses the discriminant whose discrimination result is unknown as the input data. A discriminating procedure for discriminating using the corrected discriminant function is provided.

  As described above, according to the present invention, in order to determine the suitability or pass / fail of data composed of a plurality of feature amounts or elements such as bag data, a discrimination rule is determined by SVM using the learning data of the data. In order to obtain a desired determination result for a predetermined determination object such as a mis-determined data point, the influence on the determination result of other data points in a plurality of data is suppressed while deriving from It is also possible to adjust the discrimination rules flexibly.

  Exemplary embodiments of the present invention will be described below in detail with reference to the accompanying drawings. In addition, in this specification and drawing, about the component which has the substantially same function structure, duplication description is abbreviate | omitted by attaching | subjecting the same code | symbol.

  The discriminating apparatus according to each embodiment of the present invention can discriminate which of a plurality of types of sets (also referred to as classes) a data point to be discriminated belongs based on feature information of the discriminating target. . Here, the “feature information” is various information characterizing the discrimination target, and is composed of a plurality of types of feature amounts.

  In addition, the determination device can be configured in various embodiments that determine a determination target for various types of data from which feature information is obtained. Below, the case where a discrimination device is applied to the surface flaw inspection apparatus is demonstrated for convenience of explanation. For example, the discrimination target of the discrimination device is candidate data for wrinkles (also referred to as wrinkle candidates) detected by the surface wrinkle inspection device, and a plurality of sets for determining whether or not a wrinkle candidate belongs is the type of wrinkle ( (It is also referred to as “Kaikin”).

  However, the surface flaw inspection device is an example of a device to which the discrimination device is applied, and does not limit the application destination of the discrimination device. The discriminating device is also applicable to, for example, a device that discriminates a failure state of a manufacturing device of a manufacturing plant that manufactures a product, a device that classifies manufactured products, a device that discriminates a defective state that has occurred in a manufactured product, etc. can do. That is, the discriminating device can measure various states by a predetermined measuring device, and discriminates various discriminating objects classified into a plurality of types according to measurement data (feature information) obtained by the measurement. It can be applied to the device. This discriminating device is very effective particularly when applied to a device that discriminates a defective state of a product manufactured in a manufacturing plant that manufactures a product. Furthermore, the discriminating apparatus is applied not only to the measured numerical value but also to various places that perform pattern recognition such as character recognition such as OCR (Optical Character Reader), voice recognition, fingerprint, retina, face identification, etc. be able to. In order to explain this discrimination device, first, an outline of a surface flaw inspection device to which the discrimination device is applied will be described below.

<Surface inspection device>
FIG. 1 is an explanatory diagram for explaining an outline of a configuration of a surface flaw inspection apparatus for a steel sheet, which is an example to which a determination apparatus according to each embodiment of the present invention is applied. FIG. 2 is an explanatory diagram for explaining the outline of the operation of the surface flaw inspection apparatus.

  As shown in FIG. 1, the surface defect inspection apparatus 1 includes a light source device 11, an imaging device 12, an image processing device 13, a determination device 14, a determination result storage device 15, a display device 16, and a control device 17. And have.

  The surface flaw inspection apparatus 1 is arranged in a steel plate manufacturing plant (an example of a manufacturing plant), and whether or not wrinkles such as irregular defects, scratches, and dents are formed on the surface of the steel plate F manufactured in this plant. Inspect. Therefore, the surface flaw inspection apparatus 1 irradiates light from the light source device 11 to the steel plate F that passes through the roll R as shown in FIG. This light may be, for example, parallel light. And the surface flaw inspection apparatus 1 images the reflected light reflected by the steel plate F with the imaging device 12. The captured image includes reflected light from the eyelid in the region T across the width direction of the surface of the steel sheet F irradiated with light. This reflected light is imaged as the brightness (intensity, gradation) in the captured image.

  The image processing device 13 performs image processing on the captured image captured by the image capturing device 12 and analyzes the result, and labels a region of a wrinkle candidate (that is, a region estimated to represent a wrinkle) from the distribution of brightness in the captured image. Extract by At this time, a label (identification information) is attached to each bag candidate. Then, the image processing device 13 calculates feature information including features such as the shape and brightness of the region representing the wrinkle candidate in the image. This feature information includes values representing one or more types of features, and each value is also referred to as a “feature amount”. Examples of the feature amount include height, depth, width, length, area, luminance distribution, color distribution, and formation position of the ridge. Then, the image processing device 13 outputs the wrinkle candidate data in which the wrinkle candidate label is associated with the feature information to the determination device 14.

  The determination device 14 is an example of a determination device according to the present embodiment, and acquires wrinkle candidate data including a wrinkle candidate label and feature information from the image processing device 13. Then, the discriminating device 14 discriminates the species of the cocoon candidate according to the discrimination logic described below. The discriminator 14 calculates discrimination result information as a discrimination result. This “discrimination result information” is information indicating the discrimination result of the wrinkle candidate. As a specific example, it takes a positive or negative value and indicates the discrimination result by its sign (for example, if it is positive, If it is negative, it indicates a harmless defect.) The discrimination result information includes not only discrimination result information calculated by the discrimination device 14 but also discrimination result information representing the discrimination result with respect to feature information whose discrimination result is known.

  Examples of the species that the discriminating device 14 discriminates include, for example, harmful species such as rubbing and missing flaws (those that have a large effect on product quality and are not acceptable), and their harmfulness (severity, grade, etc.) ), And harmless soot (acceptable soot, mere noise, etc.). In this embodiment, the soot type is a generic term for these soot types and their classifications. For convenience of explanation, the following description will be made on a discriminating device for discriminating which of the two types of bag is based on the feature information as the discriminating device according to the first embodiment of the present invention. Then, as a discrimination device according to the second embodiment of the present invention, a discrimination device that discriminates which of the two or more types of grapes belongs to based on feature information will be described.

  Then, the determination device 14 outputs to the determination result storage device 15 and the display device 16 the bag label / feature information that is the input value and the determination result information that is the determination result. The discrimination result storage device 15 records labels, feature information, and discrimination result information in association with each other. The display device 16 displays the label, the feature information, and the discrimination result information, for example, whether the soot has been detected, whether the soot is harmful or harmless, or which type of soot if it is harmful The user is notified of information such as whether it is harmful or not. The light source device 11, the imaging device 12, the image processing device 13, the discrimination device 14, the discrimination result storage device 15, and the display device 16 are controlled by the control device 17 and perform the above-described operations.

An outline of the operation of the surface defect inspection apparatus 1 will be described with reference to FIG.
First, step S01 is processed, light is emitted from the light source device 11, and the imaging device 12 images the irradiated portion of the steel sheet F and outputs a captured image having a luminance distribution (or color distribution). Next, the process proceeds to step S03.

  In step S03, the image processing device 13 extracts a wrinkle candidate from a luminance distribution in the captured image for a region having a possibility of a wrinkle. Then, the process proceeds to step S05. In step S05, the image processing apparatus 13 determines the height, depth, width, length, area, luminance distribution, color of the wrinkle region for the region in the image representing the wrinkle candidate. The feature information of the wrinkles such as the distribution / formation position is calculated, and the process proceeds to step S07.

  In step S07, the discrimination device 14 discriminates the species from the feature information and calculates discrimination result information. In step S09, the display device 16 displays the label, feature information, discrimination result information, and the like, and the discrimination result information is stored in the discrimination result storage device 15. Note that the display device 16 may display, for example, a determination result included in the determination result information (that is, a determination result based on a sign). Further, the process proceeds to step S011.

  In step S11, the control device 17 determines whether or not the production line is stopped. When the production line stops, the surface flaw inspection apparatus 1 stops its operation manually or by a command from a process computer that controls the production line. On the other hand, when the production line is not stopped, the surface flaw inspection apparatus 1 repeats the above steps S01 to S09 again.

  In the above, the surface flaw inspection apparatus 1 which is an application example of the discrimination apparatus according to each embodiment of the present invention has been described. The surface wrinkle inspection device 1 has the discriminator 14 so that it can discriminate the species of the detected wrinkle candidate. Next, each embodiment of the discrimination device 14 will be described below.

<Determination device according to the first embodiment>
FIG. 3 is an explanatory diagram for explaining the configuration of the determination device according to the first embodiment of the present invention.

  The discriminating apparatus 100 according to the first embodiment of the present invention shown in FIG. 3 determines which one of the two types of soot species (set) the cocoon candidate (discrimination target) belongs to based on the feature information of the cocoon candidate. Determine. Note that the discriminating apparatus 100 according to the present embodiment corresponds to the discriminating apparatus 14 of the surface flaw inspection apparatus 1 shown in FIG. As described above, the feature information includes, for example, feature amounts such as the height, depth, width, length, area, luminance distribution, and formation position of the ridge. In the multidimensional space spanned with each feature quantity as a coordinate axis, the candidate data point is a vector having each feature value of feature information as a component or a point on the space having each feature value as coordinates. Can be expressed. This multidimensional space is also referred to as “feature amount space (feature information space)”. It is not uncommon for the dimensions of the multi-dimensional space in the steel plate wrinkle inspection to be over 100 dimensions.

  For convenience of explanation, “a set of harmful wrinkles” and “a set of harmless wrinkles” will be described as examples of the two types of wrinkles to be discriminated by the discriminating apparatus 100. That is, in the determination apparatus 100 according to the present embodiment, it is determined whether the wrinkle candidate belongs to a harmful wrinkle set or a harmless wrinkle set.

  As shown in FIG. 3, the discriminating apparatus 100 includes a discriminant function creating unit 110, a discriminant function correcting unit 120, and a discriminating unit 130.

(Discriminant function creation unit 110)
The discriminant function creating unit 110 creates a discriminant function used for discriminating a wrinkle candidate. The discriminant function is a function that returns (calculates) discriminant result information using the characteristic information of the wrinkle candidate as an argument (input value, input data), and is also called a discriminant function. As shown in FIG. 3, the discriminant function creating unit 110 includes a learning data storage unit 111 and a discriminant function calculating unit 112 in order to create the discriminant function.

The learning data storage unit 111 stores a plurality of learning data input in advance by the user.
“Learning data” is correct data for making a learning function create a discriminant function. The learning data includes feature information and correct discrimination result information of the feature information. In order to distinguish the feature information and discrimination result information of the wrinkles included in this learning data from the feature information and discrimination result information of other wrinkle candidates that are discrimination targets, here, “learning feature information” and “learning feature information” It is also called “discrimination result information”. The learning feature information included in the learning data and the learning discrimination result information are associated with each other. That is, the learning feature information and the learning discrimination result information are associated one-to-one.

  Usually, this learning data is based on the characteristic information of each rod and the result of determining the rod type to which the rod belongs (correct answer) for a large number of rods actually detected by the steel sheet inspection operator in the steel plate production line or the like. Often created. As described above, as the learning feature information, the discrimination result information such as the feature information of the cocoon candidate detected by the experiment conducted in the past or the actual inspection, the feature information of the cocoon candidate determined artificially is known. Feature information can be used. Note that the learning feature information and the learning discrimination result information are created by determining the discrimination result information first by the inspection operator and creating the feature information included in the discrimination result information (species). Also good.

In the following, the learning feature information of the learning data i is represented by x _i which is a vector defined in the feature space (multidimensional space), and the learning discrimination result information of the vector x _i is a scalar value. This is represented by _yi . The feature information that is a variable input to the discriminant function is represented by a vector x, and the discriminant result information calculated as a result is represented by a scalar y. Note that each element of the vectors x _i and x represents a feature amount. Incidentally, i for identifying the defects that are subjected to the learning feature information x _i learning determination result information y _i is a positive integer of 1 to n, n represents the number of learning data. Further, a set of harmful moths and a set of harmless moths are represented by class X1 and class X2, respectively.

The discriminant function calculation unit 112 acquires a plurality of learning data from the learning data storage unit 111. Then, a discriminant function is calculated based on the learning feature information x _i and the learning identification information y _i included in each of the learning data. The discriminant function calculation unit 112 has a learning function based on SVM (Support Vector Machine), and calculates a discriminant function using this SVM. The SVM referred to here may be any SVM such as a linear SVM or a non-linear SVM.

A method for calculating the discriminant function by the SVM will be briefly described as follows.
First, the discriminant function calculating unit 112 represents the learning feature information x _i of each learning data i as a point (vector x _i ) on the feature amount space. FIG. 4 shows an example in which each learning data is represented as a point on the feature amount space. ○ (open circles), □ (open squares), ● (black circles), ■ (black squares), etc. represent the points in the learning data. For convenience of explanation, FIG. 4 shows a two-dimensional space as a feature quantity space, and shows a first feature quantity and a second feature quantity as two elements constituting the feature information. The actual feature amount space has dimensions as many as the feature amounts constituting the feature information.

  Then, the discriminant function calculating unit 112 linearly maps each point according to the discriminant result in the mapping destination space (hereinafter also referred to as “mapped space”) when the points are mapped by the virtually defined mapping function φ. A discriminant surface (second condition) that is a discriminating surface (hyperplane) to be separated (first condition) and has a maximum distance to a point having the shortest distance to the discriminant surface is calculated.

This discriminant plane is represented by a curved surface when it is mapped again (that is, inversely mapped) to the feature amount space of the mapping source. In FIG. 4, since the feature amount space is represented in two dimensions, the curved surface is represented by a curve (determination boundary B12). The discriminant surface in the mapping space and the curved surface or curve in the feature amount space are collectively referred to herein as “discrimination boundary B12”. The discrimination boundary B12 represents a boundary line (a boundary surface in three or more dimensions) that separates the class X1 (harmful bag set) and the class X2 (harmless bag set). FIG 4 represents the point of learning data belonging to the class X1 (vector x _i) with ○ and ● (circle), the training data belonging to the class X2 point (vector x _i), □ and ■ at (squares) expressed.

Then, the discriminant function calculating unit 112 derives and sets a discriminant function g (x) representing the discriminant plane using the learning data. This discriminant function is expressed in the form of the following equation. The coefficients of the right side of this equation with the number of learning data x _i determined by processing described below. The determined discriminant function is then used to output the output value as the discrimination result using the feature information of the wrinkle candidate as input data.

  If the feature information (vector x) of the discrimination target (疵) to be discriminated is substituted into the discriminant function g (x), the discriminant function g (x) outputs discrimination result information y. For example, the determination result information y is a positive value when the determination target is determined to be the class X1, and is a negative value when the determination target is determined to be the class X2. That is, the discriminant function g (x) is a function that can discriminate two types of sets based on the sign of the calculated value.

[Description of SVM (Support Vector Machine)]
First, specifically, the process of calculating the discriminant function g (x) by the discriminant function calculating unit 112, that is, the SVM will be described as follows. Here, a discrimination target whose discrimination result information is not known is referred to as an “unknown discrimination target”, and its feature information is referred to as “unknown feature information”. The known discrimination result information including the learning discrimination result information is also referred to as “teacher discrimination result information (teacher data)”.

  The SVM uses a discriminant function (discriminant function) represented by the following formula (1). This discriminant function is derived by the discriminant function calculation unit 112 using a plurality of learning data.

  Here, the vector w represents a coefficient vector to be determined using learning data, and its dimension is n. A vector x represents an input vector (that is, feature information x to be discriminated), and a scalar b represents a scalar constant to be determined using learning data.

(I) First Condition As described above, when the determination result information (scalar y) calculated from the result g (x) obtained by substituting the vector x is positive, it is determined that “the vector x is class X1”. When negative, “vector x is class X2” is determined. Similarly, the scalar yi for the vector x _i of the learning data i also represent a class X1 if positive, representing the class X2 if it is negative. When this scalar y _i is formulated, the following equation (2) is obtained.

  At this time, the constraint condition that the SVM should learn is such that the following expression (3) is satisfied for i = 1, 2,.

Here, the scalar ξ _i (≧ 0) represents a slack variable that becomes non-zero when the vector x _i is erroneously determined, and is a term for allowing erroneous determination by a discriminant function. Then, the above-described equations (2) and (3) are put together, and the constraint conditions that the SVM should learn are described by the following equations (4) and (5).

(II) a second condition the other hand, discriminant function calculation section 112, a determination boundary B12 represented by g (x) = 0, the distance to the discrimination boundary B12 is the distance between the shortest vector _{x i} The discriminant function g (x) is derived and determined so as to be maximized. This distance is also called a margin and is represented by 1 / || w ||. Therefore, the discriminant function calculation unit 112 derives the discriminant function g (x) so as to maximize this margin. This margin maximization can be reduced to the minimization problem of minimizing the evaluation function G (w) expressed by the following equation (6). Further, as described above, the scalar ξ _i is a term for allowing misidentification, so this equation (6) means a discriminant function that reduces misidentification and maximizes the margin. That is, the first term on the right side of Equation (6) is the reciprocal of the margin, and the second term represents the total sum of misclassification data allowed for each learning data X _i . The purpose is to minimize both terms, but since the dimensions are different, a linear sum obtained by multiplying the second term by the weight parameter C is defined as an evaluation function G (w).

In order to solve this optimization problem, a dual problem D ₀ with the evaluation function G (w) as the main problem P ₀ is derived. That is, the discriminant function calculation unit 112 solves this minimization problem by the Lagrange undetermined constant method. When Lagrangian variables λ _i (≧ 0) and γ _i (≧ 0) are introduced, the Lagrangian function L is expressed by the following equation (7).

Necessary and sufficient conditions for solving the Lagrangian function, that is, necessary and sufficient conditions for solving the main problem P ₀ are the following equations (8) to (10) from the KKT (Karush-Kuhn-Tucker) condition. Become.

Assuming that the optimized coefficient vector w is w ^* , Expression (8) shifts to Expression (11) below.

If the optimized scalar constant b is b ^* , the Lagrangian function L can be described as the function F in the following equation (12) from the equation (7).

That is, the function F of the equation (12) becomes an evaluation function of the dual problem D ₀ , and by maximizing the evaluation function F, the optimized coefficient W ^* and the optimized scalar constant b ^* are obtained. The discriminant function g (x) can be derived.

  The second term in the above formula (12) is calculated as the following formula (13) using the above formula (9), and the third term in the above formula (12) is calculated by the above formula (10). Is calculated as in the following formula (14).

Therefore, the evaluation function F of the dual problem D ₀ is expressed by the following equation (15). This equation (15) is maximized. In equation (15), the Lagrangian variable γ disappears from the variable on the right side.

  The constraint condition for maximizing this equation (15) is 2 of the following equation (16) derived from the above equation (9) and the following equation (17) derived from the above equation (10) and γ ≧ 0. It becomes an expression.

In summary, W ^* expressed as a function of λ _i as a result of substituting the feature information (vector xi) of each learning data i as input data into the right side of equation (11) is substituted into the right side of equation (15). However, an optimized Lagrangian variable λ _i, i.e., λ _i , so that Equation (15) takes a maximum value by a known exploratory convergence calculation method under the conditions of Equation (16) and Equation (17). ^{* Is} determined.

Then, the discriminant function calculating unit 112 calculates w ^* from the determined λ _i ^* and the above equation (11). Further, the optimized scalar constant b ^* is set to w = w ^* using the equation (4) (equal sign part), and the feature information vector x _i and the discrimination result information y _{i of} a certain learning data i are obtained. Substitute and calculate. Therefore, the discriminant function g (x) (the above formula (1)) as a result of learning the learning data is represented by the following formula (18) with the symbol “(•, •)” as an inner product of vectors.

As can be seen from equation (18), in the first term, only x _i that is an optimized Lagrangian variable λ _i ^* > 0 forms a discriminant function g (x), and λ _i ^* = 0. x _i does not contribute to the value of the discriminant function g (x). Therefore, the vectors x _i which form the discriminant function g (x) is referred to as a "support vector".

The outline of this support vector will be described with reference to FIG. FIG. 4 is a schematic diagram of the feature amount space in the present embodiment. In the vector x _i shown in FIG. 4, symbols ●, point indicated by ■ (black circles or squares) represents the support vectors. The support vector is a vector x _i close to the discrimination boundary B12. In other words, discrimination function g (x) is represented by the formula (18) by a linear combination of the vectors x _i which is close to the discrimination surface in all the training data. In addition, since the coefficient of each term including the vector x _i is an optimized Lagrangian variable λ _i ^* , this non-zero variable λ _i ^* is determined by each vector for the discriminant function g (x). It can be seen that this represents the degree of influence given by _xi . This variable λ _i ^* is also referred to herein as an influence coefficient.

Then, the nonlinear SVM, when mapped onto the mapping space using mapping function phi (x) is virtually assumed discrimination surface for linearly separating the vectors x and vector x _i on the mapping space Will be calculated. Therefore, consider mapping the like vector x and the vector x _i by mapping function is virtually assumed phi (x), than the characteristic amount space into high-dimensional virtual mapping space. When this mapping function φ (x) is used, the discriminant function g (x) is expressed by the above equation (18) in the feature amount space. In the mapping space, a new discriminant function g (x) is expressed by the following equation ( 19). The symbol “(φ (x _i ), φ (x))” represents an inner product in the mapping space.

  In SVM, it is known that a mapping function φ (x) in which a function K (x, x ′) like the following formula (20) exists can be constructed. This function K (x, x ') is called a kernel function, and is a function that returns the inner product (similarity) of x and x' on the mapping space. The kernel function K (x, x ′) is a well-known function in the SVM method, and the right side of Expression (19) can be calculated using this function.

  By using this kernel function K, the discriminant function g (x) expressed by the above equation (19) further becomes the following equation (21).

  Using the SVM learning function as described above, the discriminant function calculation unit 112 can derive and determine the discriminant function represented by Expression (21). The detailed description of the learning function by SVM is described in Non-Patent Document 2.

(Operation of Discriminant Function Creation Unit 110)
An operation of creating a discriminant function by the discriminant function creating unit 110 will be described. FIG. 5 is an explanatory diagram for explaining the flow of the discriminant function creating operation and the correcting operation at the time of creation by the discriminating apparatus according to the first embodiment of the present invention.

  As shown in FIG. 5, first, in step S <b> 101, the discriminant function calculation unit 112 acquires a plurality of learning data from the learning data storage unit 111. In step S103 (discriminant function calculating step), the discriminant function calculating unit 112 performs each calculation described in the description of the SVM, thereby calculating a discriminant function (the above formula (21)). Thereafter, the calculated discriminant function is corrected so as not to make an erroneous discrimination by a user's instruction or the like. The corrective operation at the time of creation (after step S201) will be described later.

(Discriminant function correcting unit 120)
Referring back to FIG. 3, the description returns to another configuration included in the determination device 100.
The discriminant function correcting unit 120 acquires the discriminant function g (x) created by the discriminant function creating unit 110. Then, when there is data for a wrinkle candidate (feature information x) for which the discrimination result by the discriminant function is incorrect, the discriminant function correcting unit 120 calculates a correct discrimination result for the feature information of the wrinkle candidate. Modify the discriminant function. The feature information in which the discrimination result is incorrect is also referred to as “misidentification feature information” in order to distinguish it from other feature information x. Further, the case where the discrimination result is incorrect means a case where the sign of the discrimination result information calculated by the discrimination function is incorrect (reverse). Note that the feature information for learning of some of the learning data may become misclassification information, or the feature amount information of another cocoon candidate that is not the learning data, that is, unknown feature information is misclassified. It can be information. The discriminant function correcting unit 120 can correct the discriminant function with respect to both the learning feature information and the unknown feature information that are misidentified feature information. These specific learning data, other wrinkle candidates, and their feature information are usually specified by the user. The reason why an erroneous determination result may be included in a large number of learning data is, for example, in a subtle gray zone that makes it difficult for an inspection worker to make an incorrect determination or to distinguish the feature information itself. It depends on things.

  For this purpose, the discriminant function correcting unit 120 includes an important learning data selecting unit 121, a discriminant function confirming unit 122, a coefficient calculating unit 123, a discriminant function adjusting unit 124, and an error data input unit 125.

  Whether or not the discriminating feature information for learning of the learning data is misidentified feature information is confirmed by the operations of the important learning data selecting unit 121 and the discriminant function confirming unit 122. The coefficient calculation unit 123 and the discrimination function adjustment unit 124 correct the discrimination function. This correction of the discriminant function for the learning data is called “correction at the time of creation” (adjustment of error determination of learning data).

  On the other hand, when an unknown discrimination target with an unknown discrimination result is erroneously discriminated, that is, when the unknown feature information is misclassification feature information, the user uses the error data input unit 125 to identify the unknown feature information of the unknown discrimination target and its Teacher discrimination result information (correct discrimination result information) is input to the coefficient calculation unit 123. Note that the unknown feature information that is the misidentification feature information and the teacher discrimination result information are also referred to as “error data”. Then, the discriminant function is corrected by the coefficient calculator 123 and the discriminant function adjuster 124 for the error data. This correction of the discriminant function for the error data is referred to as “correction after creation” (adjustment of error determination of data that is not learning data).

More specifically, each of these components will be described.
The important learning data selection unit 121 selects and acquires particularly important learning data (hereinafter also referred to as “important learning data”) from the learning data recorded in the learning data storage unit 111 in the correction at the time of creation. . The important learning data means learning data that cannot be erroneously discriminated. Examples of important learning data include the following learning data. However, these important learning data examples do not limit the important learning data.
(1) Learning data set by the user (2) Learning data with high importance calculated by a predetermined calculation, such as learning data in which a predetermined feature amount exceeds a threshold value (3) Corresponding to each learning data Highly important learning data recorded in the learning data storage unit 111

  The discriminant function confirmation unit 122 confirms whether or not a discriminant function capable of performing accurate discrimination has been calculated in the correction at the time of creation using the important learning data. More specifically, the discriminant function confirmation unit 122 substitutes the learning feature information of the important learning data into the discriminant function calculated by the discriminant function calculator 112. Then, determination result information is calculated. Then, the discrimination function confirmation unit 122 determines whether or not the discrimination result indicated by the calculated discrimination result information is equal to the correct discrimination result indicated by the learning discrimination result information of the important learning data, that is, whether or not the signs match. Check. Note that the numerical values of the discrimination result information do not need to match. In other words, the discriminant function confirmation unit 122 confirms whether or not the learning feature information of the important learning data is misidentified feature information. The discriminant function confirming unit 122 outputs the discriminant function to the discriminator 130 when the discriminant result is correct, and when the discriminant result is incorrect, the discriminant function and the important learning data from which the incorrect result is derived, Is output to the coefficient calculation unit 123.

The coefficient calculation unit 123 performs a new influence coefficient (x _j ) on the misclassification feature information (x _j ) so that a correct discrimination result is derived for the misclassification feature information of the important learning data as an operation in correction at the time of creation. λ _j ′) is calculated. Then, the coefficient calculating unit 123 outputs the calculated influence coefficient and discriminant function to the discriminant function adjusting unit 124.

The discriminant function adjusting unit 124 calculates the influence coefficient (λ _j ) of the misclassified feature information (x _j ) of the important learning data forming the discriminant function as the operation in the correction at the time of creation. The discriminant function is adjusted by changing to the influence coefficient (λ _j ′). Then, the discriminant function adjusting unit 124 sets the adjusted discriminant function as a discriminant function used by the discriminator 130 for actual discrimination.

  On the other hand, the error data input unit 125 inputs error data to the coefficient calculation unit 123 according to a user instruction in the correction after creation. The error data input unit 125 is controlled by the control device 17 or the user. And when actually determining the wrinkle candidate, if a wrinkle candidate whose determination is wrong is found, the error data input unit 125 is operated by the control device 17 or the user, and the unknown feature information of the wrinkle candidate and Teacher discrimination result information (correct discrimination result information) is input to the coefficient calculation unit 123 as error data.

  The coefficient calculation unit 123 acquires a discriminant function used for discrimination by the discriminator 130 as an operation in correction after creation. Then, the coefficient calculation unit 123 calculates a new influence coefficient (λ ′) that leads to a correct determination result for the erroneous determination feature information (x ′) of the error data. Note that the process of calculating the influence coefficient by the coefficient calculation unit 123 is different between the correction at the time of creation and the correction after the creation. Then, the coefficient calculating unit 123 outputs the calculated influence coefficient and discriminant function to the discriminant function adjusting unit 124.

  The discriminant function adjusting unit 124 discriminates by adding (incorporating) erroneous discriminating feature information (vector x ′) of error data to the discriminant function g (x) as a support vector as an operation in correction after creation. Adjust the function. At that time, the discriminant function adjustment unit 124 uses the influence coefficient (λ ′) calculated by the coefficient calculation unit 123 as the influence coefficient of the error data. Then, the discriminant function adjusting unit 124 sets the adjusted discriminant function as a discriminant function used by the discriminator 130 for actual discrimination.

  Before describing the correcting operation by the discriminant function correcting unit 120, the discriminating unit 130 which is another configuration of the discriminating apparatus 100 will be described.

(Determination unit 130)
The discriminating unit 130 discriminates a cocoon candidate (discrimination target) using the discriminant function set by the discriminant function adjusting unit 124. That is, the determination unit 130 substitutes the feature information x of the eyelid candidate into the determination function g (x), and calculates the determination result information y. Then, the determination unit 130 determines whether the candidate for the eye belongs to a harmful eyelid set (class X1) or harmless depending on whether the sign of the calculated determination result information y is positive or negative. It is discriminated whether it belongs to a set of bags (class X2), and the discrimination result is output. At this time, the determination unit 130 may output not only the determination result but also the feature information x and the determination result information y.

(Determination operation of the discrimination unit 130)
The determination operation by the determination unit 130 will be described with reference to FIG.
FIG. 6 is an explanatory diagram for explaining the discrimination operation by the discrimination device according to the first embodiment of the present invention.

  As shown in FIG. 6, the determination unit 130 processes the following steps S301 to S305 as determination steps. More specifically, when the determination unit 130 acquires unknown feature information of an unknown determination target (feature information of a candy candidate to be determined), the determination unit 130 processes step S301. In step S <b> 301, the determination unit 130 substitutes the unknown feature information x for the set determination function g (x). Then, the process proceeds to step S303, where the calculation by the discriminant function is performed to calculate the discriminant result information y. After the process of step S303, the process proceeds to step S305, and the determination unit 130 outputs a determination result corresponding to the determination result information.

(Regarding the correction operation by the discriminant function correction unit 120)
Here, the correction of the discriminant function by the discriminant function correcting unit 120 will be described in more detail.

When the important learning data is erroneously determined or error data is generated, the erroneous determination feature information of the important learning data or the error data can also be expressed as a point on the feature amount space. This misjudgment feature information is represented by a vector (a point in space) x ′ or x _j and shown in FIGS. 7A and 7B (j = 1 to n). Incidentally, the determination of false, <a place that should have been determined to belong to (0, incorrectly class X1 (y a vector x 'or x _j class X2 y)> had been determined to belong to 0) Means the case and vice versa. FIG 7A and 7B, the discrimination function (determination boundary B12) is 'where or x _j is must separate the feature space to belong to a class X2, vector x' vectors x or x _j belong to the class X1 As shown, the feature space is divided. When such a discriminant function is calculated, if the vector _xj is important learning data, the discriminant function confirmation unit 122 detects this important learning data. On the other hand, when the vector x ′ is error data, the error data input unit 125 inputs this error data. Then, the discriminant function confirmation unit 122 corrects the discriminant function by correction at the time of creation or correction after creation.

[About modification at the time of creation]
This correction at the time of creation will be described in more detail.
Before explaining the correction at the time of creation, the relationship between the support vector constituting the calculated discriminant function and the learning data will be explained.

  In SVM, the following formulas (22) and (23) are established from the complementary condition of KKT (Karush-Kuhn-Tucker).

If the learning data (x _i ) is not a support vector, λ _i ^* = 0 as described above. Therefore, the following formula (24) is derived from the above formula (10), and the following formula (25) is further derived from the formula (24) and the above formula (23).

  From this equation (25) and the above equation (4), the following equation (26) is derived.

From equation (26), the calculation result of the discrimination function becomes negative if y _i is negative, if y _i is positive it can be seen that the calculation results positive discriminant function. That is, it can be seen that the discriminant function calculated by the discriminant function calculator 112 can always make a correct discrimination with respect to learning data (x _i ) that has not become a support vector.

On the other hand, when the learning data (x _i ) is a support vector, λ _i ^* ≠ 0 as described above. Therefore, the following formula (27) is derived from the above formula (22).

The lambda _i ^* Since taking values until 0~C the above equation (17), a case analysis on the lambda _i ^* in the following case A and case B.

Case A) When 0 <λ _i ^* <C, the following equation (28) is derived from the above equation (10), and γ _i ≠ 0. Therefore, the following equation (29) is derived from the equation (23).

  Therefore, the following equation (30) is derived from the equation (29) and the above equation (27).

From equation (30), the calculation result of the discrimination function becomes negative if y _i is negative, if y _i is positive it can be seen that the calculation results positive discriminant function. In other words, the discriminant function calculated by the discriminant function calculator 112 can surely perform correct discrimination even for learning data in which 0 <λ _i ^* <C among learning data (x _i ) that has become a support vector. I understand.

Case B) When λ _i ^* = C On the other hand, in the case of Case B, the following equation (31) is derived from the above equation (10).

Therefore, ξ _{i is} further divided into case B1 and case B2 within the range of the above equation (5).

Case B1) When 0 ≦ ξ _i ≦ 1, the following equation (32) is derived from the above equation (27).

Therefore, it can be seen from equation (32) that when y _i is negative, the calculation result of the discriminant function is also negative, and when y _i is positive, the calculation result of the discriminant function is also positive. That is, the discriminant function calculated by the discriminant function calculating unit 112 is always the learning data (λ _i ^* = C, 0 ≦ ξ _i ≦ 1) among the learning data (x _i ) that is the support vector. It can be seen that correct discrimination can be made.

Case B2) When 1 <ξ _i Similarly, the following equation (33) is derived from the above equation (27).

Therefore, from the equation (33), the calculation result of the discrimination function is positive if y _i is negative, y _i is the case of a positive calculation result of the discrimination function is understood to be a negative. That is, the discriminant function calculated by the discriminant function calculation unit 112 is always wrong for the learning data (λ _i ^* = C, 1 <ξ _i ) among the learning data (x _i ) that is the support vector. It turns out that discrimination is performed.

That is, it can be seen from the relationship between the support vector and the learning data described above that the learning data that is erroneously determined is a support vector and that λ _i ^* = C and 1 <ξ _i .

Therefore, as shown in FIG. 5, the discriminant function correcting unit 120 performs the following correcting operation at the time of creation. Note that learning feature information (misidentification feature information) of important learning data that is erroneously determined is a vector x _j (j is an integer from 1 to n), and learning determination result information (a) indicating the correct determination result ( Let _{j j} be the teacher discrimination result information). In this case, determination result information y for calculating judged by vectors _{x j} function g (x) is is represented by the following formula (34).

  First, when the discriminant function correcting unit 120 acquires the discriminant function from the discriminant function creating unit 110, the discriminant function correcting unit 120 processes step S201 as shown in FIG. In step S <b> 201, the important learning data selection unit 121 selects and acquires important learning data from the learning data storage unit 111. Then, the process proceeds to step S203.

In step S203, the discrimination function confirmation unit 122 calculates the determination result information y the training feature information x _j important training data is substituted into the discriminant function g (x). Then, the process proceeds to step S205.

In step S205, the discriminant function confirmation unit 122 confirms whether or not the discriminant function performs correct discrimination. In other words, sign determination result information y is, checks whether to match the sign of the learning determination result information y _j. If the signs match, the discrimination function confirmation unit 122 determines that the discrimination result is correct, and the process proceeds to step S211. On the other hand, if the signs of the two do not match, the discriminant function confirmation unit 122 determines that the result is an incorrect discrimination result, and proceeds to step S207.

In step S207, the coefficient calculation unit 123 calculates the influence coefficient λ _j ′ so that the discrimination result information y matches the learning discrimination result information y _j (so that the signs match). More specifically, the coefficient calculation unit 123 calculates the influence coefficient λ _j ′ so as to satisfy the following formula (35). Then, the process proceeds to step S209.

In step S209 (determination function modifying step), discriminant function adjusting portion 124, using the above calculated influence coefficients lambda _{j 'to} adjust the influence coefficient lambda _j, modifying the discriminant function. More specifically, the important learning data subjected to the error determination as described above is a support vector of the discriminant function. Therefore, the discrimination function adjusting portion 124, erroneous influence coefficients lambda _j of the determined critical learning training feature information x _j of the data is changed to the coefficient calculation unit 123 has calculated influence coefficients lambda _{j '.} Then, the discriminant function adjusting unit 124 corrects the discriminant function g (x) created by the discriminant function creating unit 110 into a new discriminant function g ₁ (x) represented by the following formula (36). Then, the process proceeds to step S211.

In step S211, the discriminant function adjusting unit 124 sets the discriminant function used by the discriminator 130 for discrimination to the corrected discriminant function g ₁ (x). Then, the discriminant function creation operation and correction operation end.

Changing the influence coefficient λ _j ′ in this way corresponds to an operation for directly changing the discrimination result in the discrimination function. That is, in the general SVM, since the upper limit is determined for the influence coefficient λ by the above equation (17), there is a possibility of calculating a discriminant function for erroneously determining important learning data. However, the determination apparatus 100 according to the present embodiment adjusts the influence coefficient λ beyond the above equation (17) only for erroneously determined important learning data. Therefore, as conceptually shown in FIG. 7A or FIG. 7B, judgment apparatus 100, as determined boundary B12 correctly compartment also learning data x _j it is determined error, curved discrimination boundary B12 (Curve B-1 7B), or a region separated by the discrimination boundary B-2 can be formed. Therefore, the discriminating apparatus 100 uses a discriminant function (that is, a discriminant rule) so as to obtain a desired discriminant result for a judgment object that has been erroneously judged while suppressing the influence on the discrimination result of other feature information. Can be adjusted.

[Modification after creation]
The correction at the time of creation in the correction operation by the discriminant function correction unit 120 has been described above.
Next, with reference to FIG. 8, the correction after creation, which is another correction operation, will be described in detail. FIG. 8 is an explanatory diagram for explaining a correction operation after creation of a discriminant function by the discriminating apparatus according to the first embodiment of the present invention.

  In SVM, if there is error data that is not learning data, and this error data is unacceptable data, it is necessary to re-learn to correctly determine the error data, that is, the discriminant function is set again. It needs to be recreated. In order to create the discriminant function again, the discriminant function calculation unit 112 needs to add the error data to the learning data and execute the optimization calculation again.

  However, it is not uncommon for the surface defect inspection apparatus 1 to have several thousand pieces of learning data, and the amount of feature information of each piece of learning exceeds several hundred pieces. Therefore, repeating this relearning every time error data is generated requires enormous maintenance costs and time.

  Therefore, the discrimination device 100 according to the present embodiment performs the following operation by having the above configuration. Note that the error data is a vector x ', and teacher discrimination result information indicating the correct discrimination result is y'. In this case, the discrimination result information y calculated by the discriminant function based on the vector x ′ is expressed by the following equation (37).

  First, a case where the discrimination result information is y = g (x ′) <0 while the teacher discrimination result information is y ′ = + 1 will be described.

  As shown in FIG. 8, first, step S <b> 401 is processed, and the error data input unit 125 inputs error data to the coefficient calculation unit 123. That is, the error data input unit 125 inputs the unknown feature information x ′ (error discrimination feature information) of the error data and the teacher discrimination result information y ′ that is the correct discrimination result to the coefficient calculation unit 123. Then, the process proceeds to step S403.

  In step S403, the coefficient calculation unit 123 sends out the influence coefficient λ ′ so that the discrimination result information y matches the teacher discrimination result information y ′ (so that the signs match). More specifically, the coefficient calculation unit 123 calculates the influence coefficient λ ′ so as to satisfy the following formula (38). Then, the process proceeds to step S405.

In step 405, the discriminant function adjusting unit 124 adds the erroneous discrimination feature information x ′ of the error data to the discriminant function g (x) as a support vector. More specifically, the discriminant function adjusting unit 124 adds a term described by the misclassification feature information x ′ to the discriminant function g (x) used by the discriminator 130, so that the discriminant function g ( x) is corrected to a discriminant function g ₂ (x) represented by the following formula (39). Then, the process proceeds to step S407.

In step S407, the discriminant function adjusting unit 124 sets the discriminant function used by the discriminator 130 for discrimination to the corrected discriminant function g ₂ (x). Then, the discriminant function creation operation and correction operation are completed.

  As a result, the discrimination result information calculated by the discrimination unit 130 for the error data x ′ is positive as shown in the following formula (40).

Even when the teacher determination result information is y ′ = − 1 and the determination result information is y = g (x ′)> 0, the correction can be made in the same manner. At this time, in step S403, the coefficient calculation unit 123 calculates the influence coefficient λ _j ′ so as to satisfy the following equation (41). As a result, the discrimination result information y calculated by the discrimination unit 130 with respect to the unknown feature information x ′ is negative as shown in the following formula (42).

  As described above, calculating the influence coefficient λ ′ and introducing the error data as a support vector into the discriminant function also belongs to an operation of directly changing the discriminant result in the discriminant function. The SVM discriminant function before correction is represented by a linear combination of support vectors, which are feature information for learning of learning data. Therefore, generally, in order to obtain a correct discrimination result for error data that is not learning data, it is necessary to calculate a discrimination function again. However, the discriminating apparatus 100 according to the present embodiment can adjust the discriminant function by adding error data different from the learning data as a support vector to the discriminant function. Therefore, it is possible to adjust the discriminant function (that is, the discriminant rule) so as to obtain a desired discriminant result with respect to the judgment target that has been erroneously judged while suppressing the influence on the discriminant result of other feature information. .

Second Embodiment
Heretofore, the determination device 100 according to the first embodiment of the present invention has been described.
Next, a determination device 200 according to the second embodiment of the present invention will be described with reference to FIG. FIG. 9 is an explanatory diagram for explaining a configuration of a determination device according to the second embodiment of the present invention.

  The discriminating apparatus 200 according to the present embodiment shown in FIG. 9 discriminates which of the three or more types of cocoon types (sets) the cocoon candidate (discrimination target) belongs to based on the feature information of the cocoon candidate. To do. Note that the discrimination device 200 according to the present embodiment corresponds to the discrimination device 14 of the surface flaw inspection apparatus 1 shown in FIG.

The discriminating apparatus 200 also discriminates species by using the SVM discriminant function similar to that of the discriminating apparatus 100 according to the first embodiment. As described above, the discrimination function of the SVM represents discrimination of two types of sets (classes) by the sign of discrimination result information. Therefore, the discriminating apparatus 200 uses _N C _two discriminant functions when discriminating N types of classes (N> 2). That is, the discriminating apparatus 200 uses a discriminant function for discriminating two types of classes out of N types of classes by the number of combinations of classes (the number of discriminant boundaries). Then, the determination device 200 determines which class it belongs to by majority vote. That is, the discriminating apparatus 200 substitutes the unknown feature information x of the unknown discrimination target into a plurality of discriminant functions, and discriminates that the unknown discrimination target belongs to the class indicated most by the discrimination result information y. Here, when one or two or more discriminant functions indicate one class, the number of discriminant functions indicating the class is also referred to as “votes” acquired by the class. When the plurality of classes are X1, X2, X3,..., The number of votes obtained for each class is n1, n2, n3,. That is, the determination device 200 determines that “the unknown determination target x belongs to the class Xk with the largest number of votes nk (k = 1 to N)”. Since one class is discriminated by N-1 discriminant functions, the maximum number of votes obtained for one class is N-1.

  As described above, the determination device 200 determines three or more classes by performing two-class determination in the determination device 100 according to the first embodiment a plurality of times. Therefore, since the determination device 200 has many configurations and operations that are common to the determination device 100 according to the first embodiment, the following description will focus on differences from the determination device 100 according to the first embodiment.

  As illustrated in FIG. 9, the determination device 200 includes a determination function correction unit 220 and a determination unit 230 instead of the determination function correction unit 120 and the determination unit 130 included in the determination device 100 according to the first embodiment. Note that the discriminant function creating unit 110 included in the discriminating apparatus 200 calculates a plurality of discriminant functions with the same configuration as that of the discriminating apparatus 100 according to the first embodiment.

  The discriminating unit 230 is an example of a discriminating unit, and substitutes unknown feature information of an unknown discrimination target into these discriminant functions using a plurality of discriminant functions as described above. Then, the determination unit 230 calculates the number of votes for each class Xk. Then, the determination unit 230 outputs the class Xk having the largest number of votes nk as the determination result. Other operations of the determination unit 230 are the same as those of the determination unit 130 according to the first embodiment.

  The discriminant function correcting unit 220 further includes a vote number checking unit 226 and a discriminant function selecting unit 227 in addition to the configuration of the discriminant function correcting unit 120 according to the first embodiment. The discriminant function correction performed by the discriminant function correcting unit 220 on one discriminant function is the same as the discriminant function correction performed by the discriminant function correcting unit 120 according to the first embodiment. However, it is not always necessary to correct all the discriminant functions so that the discriminator 230 corrects the discriminant functions so as to derive a correct discriminant result. One or more discriminant functions are selected and the discriminant function is selected. May be corrected. In order to select the discriminant function to be corrected, the discriminant function correcting unit 220 includes the vote number checking unit 226 and the discriminant function selecting unit 227. Hereinafter, each configuration and operation will be described focusing on the selection operation of the discriminant function to be corrected.

  In the correction at the time of creation (correction for important learning data), if the discriminant function confirmation unit 122 confirms all the discriminant functions one by one at the time of creation, the discriminant function that performed the wrong discrimination is corrected each time. It is easy to modify all discriminant functions. However, in the correction after correction (correction for error data), the discriminant function has already been created, and it takes maintenance cost and time to check every discriminant function and correct all discriminant functions. is necessary. Therefore, selecting and correcting the discriminant function is particularly effective in correction after creation. Therefore, in the following, the discriminant function selection operation will be described using the modification after the creation as an example. However, it goes without saying that the selection function of the discriminant function in the correction at the time of creation can be performed in the same manner as in the case of the correction after creation described below, and the cost and time at the time of construction of the discrimination rule can be saved. .

  The vote count confirmation unit 226 acquires error data from the error data input unit 125. Using the acquisition of the error data as a trigger, the vote number confirmation unit 226 obtains the vote number nk of each class Xk from the determination unit 230. Then, the vote number confirmation unit 226 determines how many discriminant functions need to be corrected in order to discriminate the correct class Xt (t = 1 to N) indicated by the teacher discrimination result information of the error data, that is, the vote It is confirmed how many times nt is increased to determine whether the error data is the correct class Xt.

  In the confirmation of the number of votes nk by the vote number confirmation unit 226, there are a maximum vote mode and a minimum vote mode. The maximum vote mode is a mode for adjusting the discrimination function so that the class Xt is selected with the maximum number of votes nt, and the minimum vote mode is determined so that the class Xt is selected with the minimum number of votes nt. This mode adjusts the function.

  In the maximum vote mode, class Xt is selected with the maximum number of votes N-1 by correcting all the discriminant functions for discriminating between correct class Xt and other classes Xo (o = 1 to N, o ≠ t). Is done. Therefore, the vote count checking unit 226 sets the number of discriminant functions Δn that needs to be corrected to N−1.

  On the other hand, in the minimum vote mode, first, the vote number confirmation unit 226 confirms the number nk of votes nk of each class Xk, and identifies the class Xo that obtains the number of votes no greater than the number of votes nt of the correct class Xt ( nt <no). Then, the vote number confirmation unit 226 sets the number of necessary discriminant functions Δn to 1. That is, in this minimum vote mode, Δn is set to 1 until the vote number nt of the correct class Xt becomes larger than the vote number no of other classes, and Δn is set when the required vote number is reached and nt> no. Set to 0.

  In the correction at the time of creation, the vote number confirmation unit 226 starts the operation when the confirmation result by the discrimination function confirmation unit 122 is incorrect instead of the error data. Then, the vote count confirmation unit 226 calculates the vote count nk by accumulating the discrimination result information calculated when the discrimination function confirmation unit 122 confirms the discrimination function. In addition, the vote count confirmation unit 226 identifies the correct class Xk from the learning discrimination result information acquired from the discrimination function confirmation unit 122.

  The discriminant function selection unit 227 obtains the number of discriminant functions Δn to be corrected and the correct class Xk from the vote count confirmation unit 226. The discriminant function selection unit 227 further acquires, from the discriminator 230, N−1 discriminant functions whose class Xk is the class to be discriminated and the discrimination result information y of the discriminant function. Then, the discriminant function selection unit 227 ranks (sorts) the acquired discriminant functions in ascending order of the absolute value of the discrimination result information y. Further, the discriminant function selection unit 227 outputs as many discriminant functions as the number Δn to the coefficient calculator 123, and causes the coefficient calculator 123 and the discriminant function adjuster 124 to correct the discriminant function. At this time, the discriminant function selection unit 227 outputs the discriminant functions in ascending order. That is, the discriminant function selection unit 227 outputs the upper discriminant function. In addition, the discriminant function selection unit 227 outputs a discriminant function other than the discriminant function output once in the correction of one. That is, the already corrected discriminant function is not output.

  Therefore, in the maximum vote mode, N-1 discriminant functions whose class Xk is the class to be discriminated are corrected. On the other hand, in the minimum vote mode, the discrimination function is corrected in ascending order of the absolute value of the discrimination result information y until the correct class Xt is selected.

  In the modification at the time of creation, the discriminant function selection unit 227 acquires, from the discriminant function confirmation unit 122, N-1 discriminant functions having the class Xk as a class to be discriminated and the discriminant result information y of the discriminant function. .

The discriminant function correcting operation in the minimum vote mode will be described with reference to FIG.
FIG. 10 is an explanatory diagram for explaining the operation of the determination device according to the second embodiment of the present invention.

  In this minimum vote mode, the vote number confirmation unit 226 that has acquired error data or the like from the error data input unit 125 sets the number Δn of discriminant functions that need to be corrected to 1. Then, step S501 is processed, and the discriminant function selection unit 227 obtains N−1 discriminant functions that are discriminating the correct class Xt, and the discriminant function is obtained with a small absolute value of the discriminant result information y. Sort in order. Then, the process proceeds to step S503.

  In step S503, the discriminant function selection unit 227 outputs one uncorrected and higher-order discriminant function to the coefficient calculator 123, and causes the coefficient calculator 123 and the discriminant function adjuster 124 to correct the discriminant function. . After the process of step S503, the process proceeds to step S505.

  In step S505, the determination unit 230 calculates the number of votes ni for each class Xi using the corrected determination function. Then, the vote number confirmation unit 226 confirms whether or not the number nt of the correct class Xt has reached a sufficient number of votes (necessary number of votes) for selecting the class Xt. That is, it is confirmed whether or not nt> no. If the required number of votes has not been reached, step S503 is repeated, and if the required number of votes has been reached, the operation is terminated.

  By such an operation, in the minimum vote mode, the number of discriminant functions to be corrected can be minimized. Therefore, the time and cost required for correcting the discriminant function can be reduced. On the other hand, a small absolute value of the discrimination result information y calculated by the discrimination function means that it is close to the discrimination boundary B12. Therefore, by correcting from a discriminant function having a small absolute value of y, it is possible to correct a discriminant function having a low probability of discrimination. That is, in the minimum vote mode, it is possible to minimize the influence due to the correction of the discriminant function.

  As mentioned above, although preferred embodiment of this invention was described in detail, referring an accompanying drawing, it cannot be overemphasized that this invention is not limited to this example. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

  For example, the series of processes described in the above embodiments may be executed by dedicated hardware, but may be executed by software. When the series of processes is performed by software, the above series of processes can be realized by causing a general-purpose or dedicated computer as shown in FIG. 11 to execute the program.

  FIG. 11 is an explanatory diagram for explaining a configuration example of a computer that realizes a series of processes by executing a program. The execution of a program for performing a series of processes by a computer will be described as follows.

  As shown in FIG. 11, for example, the computer includes a central processing unit (CPU) 302, a hard disk drive (HDD) 303, and a read only memory (ROM) connected via a bus 301 and an input / output interface 306 and the like. 304, a recording device such as a RAM (Random Access Memory) 305, a communication device 307 connected to a network 308 such as a LAN (Local Area Network) or the Internet, and an input device 309 such as a mouse / keyboard (not shown) , Flexible disk, various optical discs such as CD (Compact Disc), MO (Magneto Optical) disc, DVD (Digital Versatile Disc), etc., a drive 310 for reading / writing from / to a removable recording medium 311 such as a magnetic disc or a semiconductor memory, a monitor, etc. Display device 16-Audio output device such as speakers and headphones An output device such as such as 12.

  The CPU 302 executes various processes in accordance with a program recorded in the recording device, a program received via the network 308, a program read from the removable recording medium 311, etc. Realized. At this time, the CPU 302 may perform various processes based on information and signals input from the input device 309 as necessary. Then, the processing result may be recorded on the recording device, the removable recording medium 311 or the like as necessary, may be output to the output device, or may be transmitted to the network 308.

  In this specification, the steps described in the flowcharts are executed in parallel or individually even if they are not necessarily processed in time series, as well as processes performed in time series in the described order. Including processing to be performed. Further, it goes without saying that the order can be appropriately changed even in the steps processed in time series.

It is explanatory drawing for demonstrating the outline | summary of a structure of the surface flaw inspection apparatus which is an example to which the discrimination | determination apparatus concerning each embodiment of this invention was applied. It is explanatory drawing for demonstrating the outline | summary of operation | movement of the surface flaw inspection apparatus which is an example to which the discrimination | determination apparatus concerning each embodiment of this invention was applied. It is explanatory drawing for demonstrating the structure of the discrimination | determination apparatus concerning 1st Embodiment of this invention. It is explanatory drawing for demonstrating an example which represented the discrimination | determination object which the discrimination | determination apparatus concerning each embodiment of this invention discriminate | determines as a point on feature-value space. It is explanatory drawing for demonstrating the production | generation operation | movement of the discriminant function by the discriminating device which concerns on 1st Embodiment of this invention, and the correction operation at the time of creation. It is explanatory drawing for demonstrating the discrimination | determination operation | movement by the discrimination device which concerns on 1st Embodiment of this invention. It is explanatory drawing for demonstrating correction operation | movement of the discriminant function by the discriminating apparatus which concerns on each embodiment of this invention. It is explanatory drawing for demonstrating correction operation | movement of the discriminant function by the discriminating apparatus which concerns on each embodiment of this invention. It is explanatory drawing for demonstrating correction operation | movement after preparation of the discriminant function by the discriminating apparatus which concerns on 1st Embodiment of this invention. It is explanatory drawing for demonstrating the structure of the discrimination | determination apparatus concerning 2nd Embodiment of this invention. It is explanatory drawing for demonstrating operation | movement of the discrimination | determination apparatus which concerns on 2nd Embodiment of this invention. And FIG. 14 is an explanatory diagram for explaining a configuration example of a computer that realizes a series of processes by executing a program.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Surface wrinkle inspection apparatus 11 Light source apparatus 12 Imaging apparatus 13 Image processing apparatus 14,100,200 Discriminating apparatus 15 Discrimination result storage apparatus 16 Display apparatus 17 Control apparatus 110 Discriminant function creation part 111 Learning data storage part 112 Discriminant function calculation part 120, 220 Discriminant Function Correction Unit 121 Important Learning Data Selection Unit 122 Discrimination Function Confirmation Unit 123 Coefficient Calculation Unit 124 Discrimination Function Adjustment Unit 125 Error Data Input Unit 130, 230 Discrimination Unit 226 Number of Votes Confirmation Unit 227 Discrimination Function Selection Unit 301 Bus 302 CPU
303 HDD
304 ROM
305 RAM
306 Input / output interface 307 Communication device 308 Network 309 Input device 310 Drive 311 Removable recording medium 312 Audio output device F Steel plate R Roll T Surface area

Claims (13)

Using a discriminant function that uses feature information composed of feature quantities that characterize the discriminant as input data, it is discriminated based on the result that is the output value of the discriminant function. A discrimination device,
Based on a plurality of learning feature information whose discrimination results are known and learning discrimination result information representing the discrimination results associated with each of the learning feature information, each of the learning feature information is represented by a point in the space based on the feature information. When a point in the space is mapped by a virtually defined mapping function, the determination surface that linearly separates the point according to the determination result in the mapping destination space, and the distance to the determination surface is the shortest A discriminant function calculator for calculating the discriminant surface using a support vector machine method and deriving the discriminant function so that the distance to the point is maximized;
Discrimination result information for feature information having a known discrimination result, wherein the discrimination result information is a result of inputting the feature information to the discrimination function and the known discrimination result of the feature information is different. And a discriminant function correcting unit that corrects the discriminant function by adjusting an influence coefficient representing the degree of influence of the misclassified feature information on the discriminant function so that the known discriminant result matches
A discrimination unit that discriminates a discrimination target whose discrimination result is unknown using the corrected discrimination function using the feature information of the discrimination target as input data;
A discrimination apparatus characterized by comprising: The discriminant function is a sum obtained by multiplying the kernel function value of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and multiplying each of the influence coefficients,
The discriminant function correcting unit performs the discriminant in a predetermined process so that the discriminant result of the discriminant is correct when any of the learning feature information is mis-discriminated feature information. The discriminating apparatus according to claim 1, wherein the discriminant function is modified by adjusting a magnitude of an influence coefficient of feature information. The discriminant function is a sum obtained by multiplying the kernel function value of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and multiplying each of the influence coefficients,
The discriminant function correcting unit newly adds misdiscrimination feature information erroneously discriminated by the discriminator to the discriminant function formed by one or more of the learning feature information, and the misclassification feature information. The discriminating apparatus according to claim 1, wherein the discriminant function is corrected by a predetermined process so that the discriminant result of the discriminant result is correct by adjusting the magnitude of the influence coefficient of the discriminant. The discriminant function calculating unit calculates the discriminant surface for each combination of two sets in the plurality of types of sets, and each of the discriminant surfaces is used as a discriminant function,
The discriminating unit discriminates that a discriminant whose discrimination result is unknown belongs to a set discriminated by the most discriminant functions as a result of discriminating each discriminant for each of the two sets based on the discriminant functions. The discrimination device according to claim 1, wherein   5. The discriminant function correcting unit corrects each of the discriminant functions until at least a known discrimination result of the misclassification feature information matches a discrimination result by the discriminator. Discriminating device. The determination target is a defective state generated in the product to be manufactured in a manufacturing plant that manufactures the product,
6. The determination apparatus according to claim 1, wherein the feature information includes one or more measurement values obtained as a result of inspecting the defective state. Using a discriminant function that uses feature information composed of feature quantities that characterize the discriminant as input data, it is discriminated based on the result that is the output value of the discriminant function. A discrimination method,
Based on a plurality of learning feature information whose discrimination results are known and learning discrimination result information representing the discrimination results associated with each of the learning feature information, each of the learning feature information is represented by a point in the space based on the feature information. When a point in the space is mapped by a virtually defined mapping function, the determination surface that linearly separates the point according to the determination result in the mapping destination space, and the distance to the determination surface is the shortest A discriminant function calculating step for deriving the discriminant function by calculating the discriminant surface using a support vector machine method so that the distance to the point is maximized;
Discrimination result information for feature information having a known discrimination result, wherein the discrimination result information is a result of inputting the feature information to the discrimination function and the known discrimination result of the feature information is different. And a discriminant function correcting step for correcting the discriminant function by adjusting an influence coefficient representing the degree of influence of the misclassified feature information on the discriminant function so that the known discriminant result and the known discriminant result match.
A discrimination step of discriminating a discrimination target whose discrimination result is unknown using the corrected discrimination function using the feature information of the discrimination target as input data;
A discrimination method characterized by comprising: The discriminant function is a sum obtained by adding each of the influence coefficients to the kernel function value of the support vector machine method using each of the learning feature information and the feature information to be discriminated as input data,
In the discriminant function correcting step, if any of the learning feature information is misidentified feature information with respect to the discriminant result of the discriminant function, the misclassification is performed in a predetermined process so that the discriminant result of the discriminant result is correct. 8. The discrimination method according to claim 7, wherein the discriminant function is modified by adjusting a magnitude of an influence coefficient of feature information. The discriminant function is a sum obtained by multiplying the kernel function value of the support vector machine method using each of the learning feature information and the discriminating target feature information as input data and multiplying each of the influence coefficients,
In the discriminant function correcting step, the discriminant feature information misidentified by the discriminator is newly added to the discriminant function formed by one or more of the learning feature information, and the misclassified feature information 9. The discrimination method according to claim 7, wherein the discriminant function is corrected by a predetermined process such that the influence coefficient of the discrepancy is adjusted and the discrimination result of the discrimination result becomes a correct answer. In the discriminant function calculating step, the discriminant surface is calculated for each combination of two sets in the plurality of types of sets, and each of the discriminant surfaces is used as a discriminant function,
In the discrimination step, the discrimination target whose discrimination result is unknown is discriminated for each of the two sets based on each of the discriminant functions, and as a result, is discriminated to belong to the set discriminated by the most discriminant functions. The discrimination device according to any one of claims 7 to 9, wherein   11. The discriminant function correction step corrects each of the discriminant functions until at least a known discrimination result of the misclassification feature information matches a discrimination result in the discrimination step. Discriminating device. The determination target is a defective state generated in the product to be manufactured in a manufacturing plant that manufactures the product,
The determination method according to claim 7, wherein the feature information includes one or more measurement values obtained as a result of inspecting the defect state. Based on the result that is the output value of the discriminant function, using a discriminant function that uses as input data the feature information composed of the feature quantity that characterizes the discriminant object, to which of the multiple types of sets the discriminant object belongs to A program for executing a determination procedure for determining
On the computer,
Based on a plurality of learning feature information whose discrimination results are known and learning discrimination result information representing the discrimination results associated with each of the learning feature information, each of the learning feature information is represented by a point in the space based on the feature information. When a point in the space is mapped by a virtually defined mapping function, the determination surface that linearly separates the point according to the determination result in the mapping destination space, and the distance to the determination surface is the shortest A discriminant function calculation procedure for deriving the discriminant function by calculating the discriminant surface using a support vector machine method so that the distance from the point is maximized;
Discrimination result information for feature information having a known discrimination result, wherein the discrimination result information is a result of inputting the feature information to the discrimination function and the known discrimination result of the feature information is different. And a discriminant function correction procedure for correcting the discriminant function by adjusting an influence coefficient representing the degree of influence of the misdiscrimination feature information on the discriminant function so that the known discriminant result matches
A discrimination procedure for discriminating a discrimination target whose discrimination result is unknown using the corrected discrimination function with the feature information of the discrimination target as input data,
A program for running

Images