JP2017102906A - Information processing apparatus, information processing method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To determine an appropriate parameter of a discriminator even when sufficient data of defective items cannot be used.SOLUTION: An information processing apparatus comprises: receiving means that receives a plurality of pieces of learning data used for creating a discriminator for determination that determines whether object data is specific category data or non-specific category data; first data evaluation means that determines first likelihood indicating the probability that the learning data are the specific category data; and parameter determination means that determines parameters of the discriminator for determination on the basis of the first likelihood of each of the plurality of pieces of learning data.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an information processing apparatus, an information processing method, and a program.

  2. Description of the Related Art Conventionally, a method using a large number of feature quantities is known as one of methods for automating the appearance inspection of whether a product manufactured in a factory or the like is a non-defective product or a defective product. This method extracts a large number of feature values such as average and maximum pixel values from multiple good and defective images for learning, and classifies good and defective products on the feature space consisting of the extracted feature values. The discriminator to be learned is learned, and the discriminator is used to determine whether the test object is a non-defective product or a defective product.

  In such a defective product detection method using image processing, in order to learn an appropriate discriminator, non-defective product data and defective product data are required as learning data. On the other hand, Patent Document 1 discloses a technique for excluding non-defective product data from non-defective product data in a data set given as learning data.

  In addition, a sufficient number of non-defective product data can be prepared at the start of an actual inspection process, but there is a case where a defective product generation rate is low and a sufficient number of data cannot be prepared. On the other hand, a one-class classifier model that can learn a classifier from only one label data is also known. The one class discriminator learns a feature space expressing good product data, and determines whether the product is good or defective based on whether or not it belongs to the learned space.

JP 2011-70635 A

  However, even when the one-class identification model is used, it is necessary to use defective product data or manually determine the hyperparameters in order to determine the hyperparameters necessary for learning the classifier. For this reason, since there is not enough defective product data, it may be difficult to learn an appropriate classifier. Further, when the user determines a hyper parameter, it is difficult to determine an appropriate hyper parameter.

  The present invention has been made in view of such problems, and an object of the present invention is to determine appropriate parameters of a discriminator even when sufficient defective product data cannot be used.

  Therefore, the present invention is an information processing apparatus that accepts a plurality of learning data used to generate a discriminator for determination that determines whether target data is specific category data or non-specific category data. Based on reception means, first data evaluation means for obtaining a first likelihood indicating the certainty that the learning data is the specific category data, and the first likelihood of each of a plurality of learning data And parameter determining means for determining parameters of the discriminator for determination.

  According to the present invention, even if sufficient defective product data cannot be used, it is possible to determine appropriate parameters of the discriminator.

It is a figure which shows the hardware constitutions of information processing apparatus. It is a figure which shows the software structure of information processing apparatus. It is a flowchart which shows a learning process. It is a flowchart which shows a learning data classification process. It is a flowchart which shows a parameter determination process. It is a flowchart which shows a determination process.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a diagram illustrating a hardware configuration of the information processing apparatus 100 according to the present embodiment. The information processing apparatus 100 learns a discriminator that identifies correct answer data and incorrect answer data using a learning data set including a plurality of learning data given as correct answer data. The information processing apparatus 100 further uses the learned discriminator to determine whether the target data to be determined is correct data or incorrect data. The discriminator only needs to identify specific category data that is data classified into a certain category and non-specific category data that is data other than the specific category data. It is not limited to correct data.

  In the present embodiment, a case where the information processing apparatus 100 is used for an appearance inspection of a product in a factory or the like will be described as an example. In other words, a non-defective photographed image (defective product data) becomes correct data, and a non-defective photographed image (defective product data) becomes incorrect data. Then, the information processing apparatus 100 uses the discriminator for determination obtained by learning, with the captured image of the actual inspection target as target data, whether the target data is good product data or defective product data. Determine. Thereby, it can be determined whether the inspection target object indicated by the target data is a non-defective product or a defective product.

  The information processing apparatus 100 includes a CPU 101, a ROM 102, a RAM 103, an HDD 104, a display unit 105, an input unit 106, and a communication unit 107. The CPU 101 reads the control program stored in the ROM 102 and executes various processes. The ROM 102 stores an operating system (OS), processing programs, device drivers, and the like. The RAM 103 is used as a temporary storage area such as a main memory and a work area for the CPU 101. The HDD 104 stores various information such as image data and various programs. Note that the functions and processing of the information processing apparatus 100 to be described later are realized by the CPU 101 reading a program stored in the ROM 102 or the HDD 104 and executing this program.

  The display unit 105 displays various information. The input unit 106 includes a keyboard and a mouse, and accepts various operations by the user. A communication unit 107 performs communication processing with an external apparatus such as an image forming apparatus via a network.

  FIG. 2 is a diagram illustrating a software configuration of the information processing apparatus 100. The information processing apparatus 100 includes a reception unit 201, a feature amount extraction unit 202, a classification unit 203, a parameter determination unit 204, a learning unit 205, and an identification unit 206. The accepting unit 201 accepts input of learning data and determination target data. Here, the learning data is data for which it is known whether the data is good product data or defective product data. On the other hand, the determination target data is data in which it is unknown whether it is non-defective product data or defective product data, and is data that is a determination target of which data. The feature amount extraction unit 202 extracts feature amounts of learning data and determination target data. The classification unit 203 evaluates the probability that each learning data is non-defective data based on the feature amount, and classifies the learning data set into two data sets according to the evaluation result. The parameter determination unit 204 estimates the parameters of the discriminator for determination. The learning unit 205 learns a determination classifier.

  Note that the information processing apparatus 100 may include each functional unit illustrated in FIG. 2 as a hardware configuration. In this case, the information processing apparatus 100 may have a calculation unit and a circuit corresponding to each functional unit.

  FIG. 3 is a flowchart showing learning processing by the information processing apparatus 100. In S301, the reception unit 201 receives a learning data set. An image as learning data included in the learning data set is a learning image in which an object to be inspected is captured by an imaging device. An object captured as a learning image is an object that is known in advance to be a non-defective product. In the present embodiment, the information processing apparatus 100 accepts an image input from an external apparatus such as an imaging apparatus. However, as another example, the information processing apparatus 100 may include an HDD 104 or the like of its own apparatus in advance. You may read the learning data set memorize | stored in the memory | storage part.

Next, in S302, the feature amount extraction unit 202 extracts a plurality of types of feature amounts determined in advance from each learning data. Examples of the feature amount include an average luminance value, variance, skewness, kurtosis, mode value, entropy, and the like. Examples of the feature quantity include a texture feature quantity using Co-Occurrence Matrix, a local feature quantity using SIFT, and the like. References 1 and 2 shown below can be referred to for the texture feature quantity using Co-Occurrence Matrix and the local feature quantity using IFT, respectively.

Reference 1
Robert M. Haralick, K. Sharnmugam, and Itshak Dinstein, "Texture Features for Image Classification", IEEE Transactions on System, Man and Cybernatic, Vol.6, pp. 610-621, 1973.

Reference 2
Lowe, David G, "Object Recognition from Local Scale-invariant Features", Proceedings of the International, Conference on Computer Vision 2 pp. 1150-1157, 1999.

  It is assumed that the feature quantity extraction unit 202 extracts a plurality of predetermined feature quantities from among these feature quantities. Then, the feature quantity extraction unit 202 obtains a feature vector in which a plurality of extracted feature quantities are arranged in order as a final feature quantity. Note that the type of feature quantity to be extracted is recorded in the ROM 102 or the like as a setting file or the like. Further, it is assumed that the CPU 101 can change the contents of the setting file based on a user operation via the input unit 106.

  Next, in S303, the classification unit 203 classifies the learning data set into two data sets, a good product data set and a defective product candidate data set (classification process), and assigns a label indicating the good product data or the defective product candidate data. . This process will be described in detail later with reference to FIG. Next, in S304, the parameter determination unit 204 determines a parameter of the classifier based on the feature amount obtained in S303 and the label given in S303. This process will be described in detail later with reference to FIG.

Next, in S305, the learning unit 205 learns a discriminator for determination based on the feature amount obtained in S303 and the parameter determined in S304. This completes the learning process. In this embodiment, a one class support vector machine (SVM) is used as the discriminator. For the one class SVM, the following documents can be referred to.

Vapnik, V. (1995). "Support-vector networks". Machine Learning 20 (3): 273.

In the present embodiment, One class SVM is used as a discriminator. However, the discriminator is not limited to the embodiment as long as it is a discriminating model capable of class classification. As another example, a Mahalanobis distance, a projection distance method which is a kind of subspace method, a neural network, or the like may be used as a discriminator.

FIG. 4 is a flowchart showing detailed processing in the learning data set classification processing (S303) described with reference to FIG. Here, it is assumed that the learning data set received in S302 is D = {d ₁ , d ₂ , d ₃ ,... D _N }. N is the number of learning data included in the learning data set. Further, each learning data is assumed to be d _i = {x _i , l _i } (1 ≦ i ≦ N). Here, x _i is a feature vector, and l _i is a label given to each learning data. In the present embodiment, since all the learning data is non-defective data, a label (l _i = + 1) indicating that it is non-defective data is given to the received learning data. Note that if no label is assigned to the learning data, the accepting unit 201 assigns (sets) a label to each learning data.

を設定する。本実施形態では、識別器としてＯｎｅ ｃｌａｓｓ ＳＶＭを用いるため、ハイパーパラメータの候補集合Φは、予め用意し、情報処理装置１００のＨＤＤ１０４等に記憶しておいてもよいし、任意のハイパーパラメータφで学習した結果から更新してもよい。Ｏｎｅ Ｃｌａｓｓ ＳＶＭでは、誤分類を許容する範囲を決定するＣパラメータや、カーネルとしてＲＢＦを用いた場合は、ＲＢＦカーネルのγパラメータなどがハイパーパラメータとなる。その他、Ｏｎｅ ｃｌａｓｓ ＳＶＭの他に識別器として部分空間法を用いた場合は、ハイパーパラメータは部分空間次元数となり、ニューラルネットワークを用いた場合は隠れ層や出力層のノード数となる。また、入力特徴量の次元数に対して次元削減を実施した場合、削減後の次元数を決定する部分をハイパーパラメータとしても良い。例えば、次元削減を実施するためにPrincipal Component nalysis（ＰＣＡ）を用いた場合、削減後の次元数を寄与率から決定することがある。この場合、寄与率を複数パターン用意し、ハイパーパラメータの候補集合に含めて計算しても良い。次元削減の方法はＰＣＡに限定するものではなく、その他の方法を用いても構わない。以下、ハイパーパラメータを単にパラメータと称する。 In S401, the classification unit 203 determines the hyperparameters necessary for learning the classifier.

Set. In this embodiment, since One class SVM is used as a discriminator, a hyperparameter candidate set Φ may be prepared in advance and stored in the HDD 104 or the like of the information processing apparatus 100, or an arbitrary hyperparameter φ. You may update from the learning result. In the One Class SVM, a hyperparameter such as a C parameter that determines a range in which misclassification is allowed, or a γ parameter of the RBF kernel when the RBF is used as a kernel. In addition, when the subspace method is used as a discriminator in addition to the One class SVM, the hyperparameter is the number of subspace dimensions, and when the neural network is used, the hyperparameter is the number of nodes in the hidden layer or the output layer. In addition, when dimension reduction is performed on the number of dimensions of the input feature quantity, a part that determines the number of dimensions after reduction may be a hyper parameter. For example, when Principal Component Analysis (PCA) is used to implement dimension reduction, the number of dimensions after reduction may be determined from the contribution rate. In this case, a plurality of patterns of contribution rates may be prepared and included in the hyperparameter candidate set. The dimension reduction method is not limited to PCA, and other methods may be used. Hereinafter, the hyper parameter is simply referred to as a parameter.

  Next, in S402, the classification unit 203 learns (generates) a discriminator using the parameter φ set in S401 and the learning data set D. The discriminator to be learned here is a discriminator for learning used for classifying the learning data set. In the present embodiment, the learning data set is classified using the same discriminator as the discriminating discriminator used in the determination process. However, as another example, a different type of discriminator from the discriminating discriminator is used. A discriminator may be used.

を求める。ここで、帰属度は、良品データ（正解データ）であることの確からしさを示す、学習用の識別器に依存した尤度の一例である。また、Ｓ４０３の処理は、学習用の識別器に依存した尤度を求めるデータ評価処理の一例である。 Next, in S403, the classification unit 203 performs learning data identification processing. Specifically, the classification unit 203 uses the classifier learned in S403, and the degree of belonging to the good product class with respect to the learning data x _i

Ask for. Here, the degree of attribution is an example of the likelihood depending on the learning discriminator indicating the certainty of being good product data (correct data). The process of S403 is an example of a data evaluation process for obtaining a likelihood that depends on a learning classifier.

なお、Ｓ４０４の処理は、学習用の識別器に依存した複数の尤度に基づいて、学習データの尤度を求めるデータ評価処理の一例である。 Next, in S404, the classification unit 203 performs a voting process as shown in (Expression 1) by a comparison process between the degree of attribution s _i and the threshold value T _v . In other words, the classifying unit 203 votes for learning data having a degree of belonging s _i smaller than the threshold value T _v .

Note that the process of S404 is an example of a data evaluation process for obtaining the likelihood of learning data based on a plurality of likelihoods depending on the learning discriminator.

ここで、Ｓは、帰属度ｓ_iの集合である。すなわち、Ｓ＝｛ｓ₁，ｓ₂，ｓ₃，…ｓ_N｝である。また、

は、帰属度集合Ｓに含まれるデータを降順にソートしたときの帰属度Ｓの順位を返す関数である。 In the present embodiment, voting is performed on data with the degree of attribution s _i smaller than the threshold value T _v , but the voting process is not limited to this. As another example, the classification section 203 may be voted in proportion to the data number N included in the threshold T _v, rather than training dataset D. In this embodiment, the vote value is 1. However, as another example, the classification unit 203 may determine the vote value by weighting. For example, the classification unit 203 may vote a value proportional to the degree of attribution as shown in (Expression 2). As another example, the classification unit 203 may determine a value from the ranks with all the degrees of belonging in the learning data set, as shown in (Expression 3).

Here, S is a set of attribution degrees s _i . That is, S = {s ₁ , s ₂ , s ₃ ,... S _N }. Also,

Is a function that returns the rank of the attribute S when the data included in the attribute set S is sorted in descending order.

  In step S <b> 405, the classification unit 203 confirms whether there is an unselected parameter. If there is an unselected parameter (Yes in S405), the classification unit 203 advances the process to S401. In step S <b> 401, the classification unit 203 selects an unselected parameter from the parameter candidate set Φ, sets it, and continues the subsequent processing. If there is no unselected parameter (No in S405), the classification unit 203 advances the process to S406.

In S406, the classification unit 203 determines whether the learning data x _i is non-defective product data or defective product candidate data based on the voting result. Specifically, the classification section 203, by (Equation 4), v _i is determined that the defective candidate data when less than the threshold T _h, the learning data x _i, the defective candidate data label (l _i = 0). On the other hand, v _i is judged to be good data in the case of less than the threshold value T _h, imparts label non-defective data in the learning data _{x i (l i = + 1} ). As a result, the learning data set is classified into two data sets, a good product data set and a defective product candidate data set. That is, this process is an example of a classification process for classifying a plurality of learning data into two data sets.

  FIG. 5 is a flowchart showing detailed processing in the parameter determination processing (S304) described with reference to FIG. In the present embodiment, the parameter determination unit 204 determines parameters using a cross-validation method. In the present embodiment, the cross-validation method is used to determine the hyperparameter, but a method other than the cross-validation method may be used.

Here, a general cross-validation method will be described. In the cross-validation method, each of the non-defective product data and the defective product data in the learning data set D is divided into K groups. Then, one of the divided K groups is used for learning, and the remaining one is used for evaluation (verification). That is, the good data set D _OK , which is a good data group, is divided into K groups, one of which is D _{OK (1)} and the remaining group is D _{OK (K-1)} . Similarly, the defective product data set _DNG , which is a defective product data group, is divided into K groups, one of which is _{DNG (1)} and the remaining group is _{DNG (K-1)} . In order to evaluate an arbitrary parameter, the discriminator is learned using the non _- defective product group D _{OK (K-1)} and the defective product group D _{NG (K-1)} . Then, the degree of separation of the remaining good product group D _{OK (1)} and defective product group D _{NG (1)} is calculated using the learned classifier. This process is repeated by replacing the learning group and the evaluation group, and the parameters are selected. Thereby, it is possible to select a parameter that most separates the non-defective product data set and the defective product data set.

  However, when the parameter for classifying the non-defective product data and the defective product candidate data according to the present embodiment is selected by the above method, the discrimination boundary is between the non-defective product data and the defective product candidate data. For this reason, there is a possibility that the defective product candidate data is determined as defective product data although the user gives it as good product data. Therefore, in the present embodiment, by performing the processing described below, it is determined that the non-defective product data is more likely to be a non-defective product than the defective product candidate data, not the parameters for classifying the non-defective product data and the defective product candidate data. Select parameters like

First, in S501, parameter determination section 204, the training data set D, based on the label assigned in S303, and divided into a defective candidate data sets D _NGC good data set D _OK, each data set K Divide into groups. Next, in S502, the parameter determination unit 204 selects parameter candidates. Next, in S503, the parameter determination unit 204 selects one group D _{OK (1)} as an evaluation group from the K groups of the non-defective product data set D _OK . Similarly, the parameter determination unit 204 selects one group D _{NGC (1)} from the K groups of the defective product candidate data set D _NGC as an evaluation group.

Next, in S504, the parameter determining unit 204 uses the non-defective product group D _{OK (K-1)} other than the evaluation group and the defective product candidate group D _{NGC (K-1)} other than the evaluation group to determine the classifier. To learn. That is, the parameter determination unit 204 regards both good product data and defective product candidate data as good product data, and learns a learning classifier (learning process). Next, in S505, the parameter determining unit 204 evaluates the effectiveness of the parameters used for learning in S504 using the evaluation groups D _{OK (1)} and D _{NGC (1)} selected in S503 (parameter evaluation). processing). In the present embodiment, Area Under the Curve (AUC) is used as the evaluation value. That is, the parameter determination unit 204 calculates the evaluation value C (φ) by the following formula.

C (φ) = AUC (D _{OK (1)} , D _{NGC (1)} )

In the present embodiment, AUC is used as the evaluation value, but the present invention is not limited to this. The evaluation value only needs to be a value that can evaluate the degree of separation between the two classes, and may be an AIC, a BIC, or the like.

  In step S506, the parameter determination unit 204 checks whether there is a group that is not selected as the evaluation group. If there is an unselected group (Yes in S506), the parameter determination unit 204 advances the process to S503. In step S503, the parameter determination unit 204 selects an unselected group as an evaluation group, and performs subsequent processing. As described above, the parameter determination unit 204 repeats the processes of S502 to S505 while changing the evaluation group. On the other hand, when all the groups have been selected as evaluation groups (No in S506), parameter determination unit 204 advances the process to S507.

  In step S507, the parameter determination unit 204 confirms whether there are unselected parameter candidates. If there is an unselected parameter candidate (Yes in S507), the parameter determination unit 204 advances the process to S502. In step S502, the parameter determination unit 204 selects an unselected parameter candidate, and performs subsequent processing. Thus, the parameter determination unit 204 calculates an evaluation value for each parameter candidate. On the other hand, if all parameter candidates have been selected (No in S507), the parameter determination unit 204 advances the process to S508.

  In S508, the parameter determination unit 204 selects an appropriate parameter using a plurality of evaluation values obtained for each parameter candidate φ by repeating the processes of S502 to S505. For example, the parameter determination unit 204 obtains an average value of a plurality of evaluation values obtained for each parameter candidate, and selects a parameter φ that maximizes the average value. As another example, the parameter determination unit 204 may select a parameter φ having a maximum minimum value among a plurality of evaluation values of each parameter candidate. As another example, the parameter determination unit 204 may obtain a median value of a plurality of evaluation values of each parameter candidate and select a parameter that maximizes the median value. This completes the parameter determination process.

  FIG. 6 is a flowchart illustrating the determination process performed by the information processing apparatus 100. The determination process uses the determination discriminator obtained by the learning process described with reference to FIG. 3 and the like, and determines whether the captured image of the object to be inspected is non-defective product data or defective product data It is processing to do. In step S601, the reception unit 201 receives a captured image of an object to be inspected, that is, target data. In the present embodiment, the receiving unit 201 receives target data from the imaging device. As another example, the information processing device 100 is a target stored in advance in a storage unit such as the HDD 104 of the own device. Data may be read out.

  In step S <b> 602, the feature amount extraction unit 202 extracts a plurality of types of feature amounts determined in advance from the target data. The type and number of feature values extracted here are the same as the type and number of feature values extracted in S302. As another example, in S602, the feature amount extraction unit 202 extracts only feature amounts that can classify target data into non-defective product data and defective product data by using a discriminator obtained by learning processing. It is good as well.

  Next, in S603, the identification unit 206 identifies whether the target data is non-defective product data or defective product data based on the feature amount extracted in S602 by using the classifier obtained by the learning process. To do. This is the end of the determination process.

  As described above, in this embodiment, the parameter determination unit 204 learns the learning data of both the good product group and the defective product candidate group as good product data, and learns the discriminator. On the other hand, in the evaluation of the learned discriminator, the parameter determination unit 204 regards the non-defective product group learning data as non-defective product data and the defective product candidate group learning data as defective product data, and calculates the degree of separation (evaluation value). To do. For this reason, in the discriminating device for determination learned by the selected parameter, the learning data given by the user as non-defective product data is determined to be non-defective. However, the learning data classified into the defective product candidate data set is determined to have a lower probability of being non-defective data than the learning data classified into the good product group.

  If the discriminator is trained by considering only the training data of the non-defective product data set that does not include the defective product candidate data set as the non-defective product data, a parameter that separates the good product data set from the defective product candidate data set is selected. The For this reason, a discriminator that determines target data belonging to a defective product candidate data set as defective product data may be learned. On the other hand, in this embodiment, the parameter determination unit 204 learns the discriminator using not only the learning data of the non-defective product data set but also the learning data of the defective product candidate data set as the non-defective product data. Thus, learning is performed such that the learning data classified into the defective product candidate data set is determined to have a low probability of being non-defective data compared to the learning data classified into the good product group. Can do. That is, it is possible to determine an appropriate parameter of the discriminator for determination from only a learning data set that is known in advance to be non-defective data.

  Note that the information processing apparatus 100 according to the present embodiment performs both the learning process and the determination process, but instead, the information processing apparatus 100 may perform only the learning process. In this case, the discriminator obtained by the learning process is set in a device that performs a determination process different from the information processing apparatus 100. And in the apparatus which performs a determination process, a determination process is performed.

(Second Embodiment)
Next, the information processing apparatus 100 according to the second embodiment will be described. The information processing apparatus 100 according to the second embodiment learns a discriminator that discriminates between correct answer data and incorrect answer data using a learning data set including correct answer data and a small amount of incorrect answer data. Also in the second embodiment, a case where the information processing apparatus 100 is used for an appearance inspection of a product in a factory or the like will be described as an example. That is, the correct answer data is a non-defective photographed image (non-defective product data). The incorrect answer data is a photographed image (defective product data) of a defective product.

  When learning data as defective product data is provided in an amount sufficient for learning of a discriminator for determination, it is possible to learn the discriminator for discrimination using both good product data and defective product data. it can. However, when there is a small amount of defective product data, the discriminator is over-learned with a small amount of defective product data, and the separation accuracy between good product data and defective product data may be reduced. On the other hand, the information processing apparatus 100 according to the second embodiment, like the information processing apparatus 100 according to the first embodiment, uses learning data given as good product data as a good product data set and defective product candidate data. Sort into sets and process. Hereinafter, the difference between the information processing apparatus 100 according to the second embodiment and the information processing apparatus 100 according to the first embodiment will be described.

  A learning process performed by the information processing apparatus 100 according to the second embodiment will be described with reference to FIG. In the second embodiment, in S301, a learning data set including learning data given as good product data and a small amount of learning data given as defective product data is accepted. Hereinafter, the learning data given as good product data is referred to as good product learning data. The learning data given as defective product data is referred to as defective product learning data.

It should be noted that a label (l _i = + 1) indicating a group of non-defective data is assigned to the given learning data that is non-defective data included in the learning data set. Further, a label (l _i = −1) representing a group of defective product data is given to the learning data given as defective product data. Note that if no label is assigned to the learning data, the accepting unit 201 assigns (sets) a label to each learning data.

  Next, in the feature amount extraction process (S302), the feature amount extraction unit 202 performs a process of extracting feature amounts in the same manner as described in the first embodiment with respect to non-defective learning data. In the subsequent learning data classification process (S303), the classification unit 203 uses the non-defective learning data as a processing target, and converts the non-defective learning data into the non-defective data set and the non-defective data set as described in the first embodiment. Classify into good product candidate data sets.

The subsequent parameter determination process (S304) will be described with reference to FIG. In the second embodiment, the parameter determination unit 204 evaluates parameter candidates by using not only defective data learning data but also defective data learning data. Specifically, in S504, the parameter determination unit 204 learns the discriminator using the non _- defective product group D _{OK (K-1)} and the defective product candidate group D _{NGC (K-1)} as good product data. Next, in S505, the parameter determination unit 204 evaluates the parameter candidates by calculating the degree of separation using the defective product candidate group D _{NGC (1)} and the defective product data group D _NG as defective product data. The remaining configuration and processing of the information processing apparatus 100 according to the second embodiment are the same as the configuration and processing of the information processing apparatus 100 according to the first embodiment.

  As described above, the information processing apparatus 100 according to the second embodiment, when the learning data for defective products is not sufficient, is similar to that described in the first embodiment, The discriminator for determination is learned by regarding the part as defective product data. For this reason, it is possible to determine an appropriate parameter that is not overlearning of learning data of defective products.

を算出する。パラメータ決定部２０４は、さらに、良品グループＤ_OK(1)と不良品データ群Ｄ_NGとの分離度

を算出する。そして、パラメータ決定部２０４は、（式５）により２つの分離度の積Ｌ'を、評価値として用いてもよい。

また、他の例としては、パラメータ決定部２０４は、（式６）に示すように、上記２つの分離度の線形和を評価値として用いてもよい。

Note that the information processing apparatus 100 may determine parameters using defective product data, and specific processing for that purpose is not limited to the embodiment. For example, in S505, the parameter determination unit 204 determines the degree of separation between the non-defective product group D _{OK (1)} and the defective product candidate group D _{NGC (1).}

Is calculated. The parameter determination unit 204 further determines the degree of separation between the good product group D _{OK (1)} and the defective product data group D _NG.

Is calculated. Then, the parameter determination unit 204 may use the product L ′ of the two separation degrees as an evaluation value according to (Equation 5).

As another example, the parameter determination unit 204 may use a linear sum of the two degrees of separation as an evaluation value, as shown in (Expression 6).

また、パラメータ決定部２０４は、（式９）又は（式１０）に示すように、分離度の線形和を評価値として用いてもよい。

As another example, considering that the defective product candidate group is learning data given as non-defective product data, the parameter determination unit 204 has a separation degree as shown in (Expression 7) or (Expression 8). The product may be used as the evaluation value.

Moreover, the parameter determination part 204 may use the linear sum of a separation degree as an evaluation value, as shown in (Formula 9) or (Formula 10).

  As mentioned above, according to each embodiment mentioned above, even if it is a case where sufficient inferior goods data cannot be used, an appropriate parameter of a discriminator can be determined.

  The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to such specific embodiments, and various modifications can be made within the scope of the gist of the present invention described in the claims.・ Change is possible.

(Other examples)
The present invention supplies a program that realizes one or more functions of the above-described embodiments to a system or apparatus via a network or a storage medium, and one or more processors in a computer of the system or apparatus read and execute the program This process can be realized. It can also be realized by a circuit (for example, ASIC) that realizes one or more functions.

DESCRIPTION OF SYMBOLS 100 Information processing apparatus 202 Feature-value extraction part 203 Classification part 204 Parameter determination part 205 Learning part 206 Identification part

Claims (14)

Receiving means for receiving a plurality of learning data used to generate a discriminator for determination that determines whether the target data is specific category data or non-specific category data;
First data evaluation means for obtaining a first likelihood indicating the certainty that the learning data is the specific category data;
An information processing apparatus comprising: parameter determination means for determining a parameter of the determination discriminator based on the first likelihood of each of a plurality of learning data.   The information processing apparatus according to claim 1, wherein the first data evaluation unit obtains the first likelihood based on a feature amount of the learning data. First learning means for generating a plurality of learning discriminators using each of a plurality of parameters and the learning data;
Second data for obtaining a second likelihood depending on the learning classifier, which indicates the certainty that the learning data is the specific category data, using each of the plurality of learning classifiers. An evaluation means,
The first data evaluation unit obtains the first likelihood of the learning data based on a plurality of second likelihoods obtained by the second data evaluation unit for the learning data. The information processing apparatus according to claim 1.   The information processing apparatus according to claim 3, wherein the first learning unit generates the learning discriminator using each of the plurality of parameters and a feature amount of the learning data. . Based on the first likelihood, the plurality of learning data is a first data set, and a second data set that is less likely to be the specific category data than the first data set; And a classification means for classifying
The information processing apparatus according to claim 1, wherein the parameter determination unit determines the parameter based on the first data set and the second data set. Second learning means for generating learning discriminators using each of a plurality of parameters by regarding learning data of both the first data set and the second data set as correct answer data;
Parameter evaluation means for evaluating each parameter used for learning by the second learning means, assuming that the learning data of the first data set is correct data and the learning data of the second data set is incorrect data And
The information processing apparatus according to claim 5, wherein the parameter determination unit determines a parameter of the discriminator for determination from the plurality of parameters based on an evaluation result by the parameter evaluation unit.   7. The information processing according to claim 6, wherein the parameter evaluation unit evaluates the parameter based on a degree of separation between learning data of the first data set and learning data of the second data set. apparatus. The accepting means further accepts incorrect answer learning data given as incorrect answer data,
The information processing apparatus according to claim 6, wherein the parameter evaluation unit further evaluates the parameter using incorrect answer learning data.   The parameter evaluation means includes a degree of separation between learning data of the first data set and learning data of the second data set, a degree of separation between learning data of the first data set and learning data of incorrect answers, The information processing apparatus according to claim 8, wherein the parameter is evaluated based on the information.   The information processing apparatus according to claim 9, wherein the parameter evaluation unit evaluates the parameter based on a degree of separation between learning data of the second data set and incorrect learning data.   The parameter evaluation means is configured to determine the parameter based on the degree of separation between the learning data of the first data set and the learning data for incorrect answers and the degree of separation between the learning data for the second data set and learning data for incorrect answers. The information processing apparatus according to claim 8, wherein the information processing apparatus is evaluated.   The information processing according to any one of claims 1 to 11, further comprising third learning means for learning the discriminator for determination based on the parameter determined by the parameter determination means. apparatus. An information processing method executed by an information processing apparatus,
An accepting step for receiving a plurality of learning data used for generating a discriminator for determination that determines whether the target data is specific category data or non-specific category data;
A first data evaluation step for obtaining a first likelihood indicating the certainty that the learning data is the specific category data;
And a parameter determining step for determining a parameter of the discriminator for determination based on the first likelihood of each of a plurality of learning data.   The program for functioning a computer as each means of the information processing apparatus of any one of Claims 1 thru | or 12.

Images