JP6659120B2 - Information processing apparatus, information processing method, and program - Google Patents

Description

  The present invention relates to an information processing apparatus, an information processing method, and a program, and is particularly suitable for use in creating an ensemble discriminator.

  As a method often used in the field of object recognition using machine learning, there is a recognition method using an ensemble classifier. The recognition method using the ensemble classifier is a method for improving the accuracy of supervised learning by combining a plurality of classifiers with low accuracy (weak classifiers). As a method of performing recognition using an ensemble classifier, there are methods such as Bagging and Randomized Trees.

  Basically, the performance of the ensemble classifier improves as the number of weak classifiers constituting the ensemble classifier increases. On the other hand, even when the number of weak classifiers forming the ensemble classifier is the same, the performance of the ensemble classifier varies depending on the combination of a plurality of weak classifiers forming the ensemble classifier. Therefore, it is desired to create an ensemble classifier with higher performance by configuring an ensemble classifier with an appropriate combination of weak classifiers.

  As a method of creating an ensemble classifier, there is a method of evaluating the created ensemble classifier based on pre-labeled sample data and repeating learning of the ensemble classifier until the evaluation result satisfies a predetermined condition. . Here, there is a technique that uses labeled data as sample data used for evaluating the ensemble classifier. The label is also called teacher data, Ground Truth, or the like, and is information indicating the correct content (correct answer) of the identification result in the classifier. For example, class information indicating the class to which the identification target belongs in the sample data is assigned as a label. As this type of technology, Patent Literature 1 discloses a technology for learning an ensemble classifier using AdaBoost. In Patent Document 1, first, a plurality of weak classifiers to be combined are selected from the learned weak classifiers. Next, an ensemble discriminator obtained by combining a plurality of selected weak discriminators is evaluated using a sample image to which a label is added as an input. The selection of the weak classifiers constituting the ensemble classifier and the evaluation of the ensemble classifier are repeated until the classification accuracy of the ensemble classifier becomes equal to or larger than the threshold.

Japanese Patent No. 5123759

Gall, J., "Hough forests for object detection, tracking, and action recognition", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 11, pp. 2188-2202, 2011.

  In Patent Literature 1, a part of data used for learning is used as sample data for evaluation of an ensemble classifier. It is considered that the data used for the learning of the classifier desirably includes only the object to be identified in order to enhance the learning accuracy. On the other hand, the purpose of the sample data is to confirm the generalization performance for unknown data input to the classifier at the time of classification. For this reason, it is desirable that the sample data is data that is actually input to the discriminator at the time of discrimination such that a plurality of objects are miscellaneously included. Therefore, rather than using the data used for learning as sample data, it is better to separately prepare a set of data (for example, actual photographed image data) of the same type as the data used at the time of identification as sample data. Can be evaluated. As a result, the performance of the ensemble classifier can be improved.

  However, it is not always possible to assign a label to data of the same type as the data used at the time of identification. In order to label data of the same type as the data used for identification, for example, it is necessary for a human to determine, for each sample data, whether or not there is an identification target in the sample data, a class to which the identification target belongs in the sample data, and the like. There is. Then, it is necessary to set information indicating the determined result as a label in the sample data. However, when the total number of sample data sets is enormous, it is not easy to manually add class information or the like to all sample data.

  For example, in Hough Forest used in Non-Patent Document 1, a partial region on an image is used as input data. Then, when the input data includes the identification target, the weak classifier outputs an offset to the center of gravity of the identification target object. On the other hand, when the input data does not include the identification target, the weak classifier determines that the sample is a negative sample. That is, in the Hough Forest, it is necessary to perform the following in order to add class information and the like to the sample data. First, it is determined whether or not each partial region includes an object to be identified, for all partial regions of the sample data. Then, when an object to be identified is included, an offset to the position of the center of gravity of the object to be identified is calculated. However, it is not realistic to perform this task manually for all sample data sets.

Therefore, if it is not easy to prepare sample data with a label, it is not easy to evaluate the performance of the created ensemble discriminator according to the method of Patent Document 1.
The present invention has been made in view of such a problem, and even without using labeled sample data, a combination of weak discriminators useful for discriminating an identification target from a plurality of weak discriminators. The purpose is to be selectable.

An information processing apparatus according to the present invention is an information processing apparatus for performing a process for selecting a plurality of weak classifiers forming an ensemble classifier, and an input for inputting a candidate of a weak classifier forming the ensemble classifier. Means and the same data is given to at least two of the weak discriminators among the candidates input by the input means, and information obtained when each of the weak discriminators performs discrimination processing on the data is compared. And deriving means for deriving an index representing the diversity of the classification result between the at least two weak classifiers based on the comparison result for each of a plurality of combinations of the at least two weak classifiers; , derived by the deriving means, on the basis of the index representing the identification result of the diversity between the at least two of said weak classifier, the ensemble identification For each number of candidates of the plurality of weak classifiers constituting the vessel, and having a processing means for selecting a combination of said plurality of weak classifiers.

  According to the present invention, it is possible to select a combination of weak classifiers useful for identifying a classification target from a plurality of weak classifiers without using sample data to which a label is assigned.

FIG. 2 is a diagram illustrating a configuration of an information processing device. It is a flowchart which shows the 1st example of a weak discriminator diversity index calculation process. FIG. 9 is a diagram illustrating a first example of a process of calculating a variation degree of an identification result. FIG. 11 is a diagram illustrating a second example of the process of calculating the degree of variation in the identification result. It is a figure showing the 3rd example of a calculation process of a classification result variation degree. It is a figure showing the 4th example of a calculation process of a classification result variation degree. It is a figure showing the 5th example of a calculation process of a classification result variation degree. It is a flowchart which shows the process of a weak discriminator combination selection part. It is a figure showing a diversity index between weak discriminators. It is a figure explaining processing of a weak discriminator combination selection part. FIG. 9 is a diagram illustrating a flow of identification using a decision tree. It is a flowchart which shows the 2nd example of a weak discriminator diversity index calculation process. It is a figure which shows the process of calculating a similar result variation degree. It is a figure showing the composition of an ensemble discriminator construction system. It is a flowchart which shows the process of an ensemble discriminator construction system.

Hereinafter, a first embodiment will be described with reference to the drawings.
(First embodiment)
First, a first embodiment will be described.
The ensemble classifier is a classifier that performs highly accurate classification by combining a plurality of weak classifiers. However, the capacity of the ensemble classifier increases in proportion to the number of weak classifiers. Therefore, when the capacity of the ensemble classifier is limited due to the execution environment, it is necessary to create the ensemble classifier by limiting the number of weak classifiers constituting the ensemble classifier in advance.

  On the other hand, in order to construct an ensemble classifier having high generalization performance, diversity of individual weak classifiers is important. Therefore, when creating an ensemble classifier using a combination of a plurality of weak classifiers, create a larger number of weak classifiers and select a combination of weak classifiers with a large variety from among the weak classifiers. Preferably, an ensemble discriminator is created. This is because a higher-performance ensemble classifier can be created.

  The purpose of the present embodiment is to create an ensemble discriminator composed of a predetermined number of weak discriminator groups, using sample data without labels such as class information, from a large number of prepared weak discriminator sets. It is to select an appropriate combination of weak classifiers. In the present embodiment, an index indicating the diversity of the classification results between weak classifiers is derived using the degree of variation in the classification results of each weak classifier when such sample data is input. Then, a high-performance ensemble classifier is created by selecting a combination of weak classifier groups for forming an ensemble classifier using such an index. In the present embodiment, a case where the data to be handled is image data will be described as an example. Further, in the present embodiment, a case will be described as an example where the classifier identifies the class to which the identification target included in the input data belongs.

  FIG. 1 is a diagram illustrating an example of a configuration of an information processing apparatus 100 according to the present embodiment. The information processing apparatus 100 according to the present embodiment includes a weak discriminator diversity index calculation unit 110, a weak discriminator combination selection unit 120, and an ensemble discriminator generation unit 130. The information processing apparatus 100 is realized by using, for example, a device including a CPU, a ROM, a RAM, an HDD, and various interfaces and dedicated hardware.

The weak discriminator diversity index calculation unit 110 receives the weak discriminator set 11 and the sample data group 12 as inputs, and calculates an index indicating the diversity of the classification results between weak discriminators. In the following description, an index indicating the diversity of the classification results between weak classifiers will be referred to as a weak classifier diversity index as needed.
The weak discriminator combination selector 120 selects a predetermined number of weak discriminators based on the weak discriminator diversity index calculated by the weak discriminator diversity index calculator 110. The predetermined number is the number of weak classifiers constituting the ensemble classifier.
The ensemble classifier generation unit 130 generates an ensemble classifier using the weak classifiers selected by the weak classifier combination selection unit 120.
Hereinafter, an example of the operation of the information processing apparatus 100 will be described in detail.

In the present embodiment, a weak classifier set 11 composed of a plurality of weak classifiers that have been learned in advance and a sample data group 12 composed of a plurality of sample data are prepared as preliminary data.
The weak classifier in the ensemble classifier of the present embodiment is a classifier that performs classification with higher accuracy than a classifier that outputs a solution at random, and performs learning using learning data to which class information to be classified is added. Is performed. In the present embodiment, it is assumed that the learning of the weak classifier has been performed in advance. However, the information processing apparatus 100 may input the learning data group and perform learning of the weak classifier using the learning data group before the processing by the weak discriminator diversity index calculation unit 110 is performed. Good. Since the learning method of the weak classifier can be realized by, for example, a known method using Adaboost, a detailed description thereof is omitted here.

The sample data is different from the learning data and is an image similar to an image actually input to the classifier at the time of identification. The image similar to the image actually input to the classifier at the time of identification is an image to which no label (class information in the present embodiment) is added.
Further, in the present embodiment, candidates for the ensemble classifier (combinations of the same number of weak classifiers as the number of weak classifiers constituting the ensemble classifier) are evaluated using sample data, and the final result is determined based on the result. Create a unique ensemble classifier. By doing so, a higher performance ensemble discriminator can be created. It is desirable that a large amount of sample data be prepared in order to reduce the influence of some characteristic sample data and increase the reliability of evaluation for the classifier.

Next, an example of the processing of the weak discriminator diversity index calculating unit 110 will be described with reference to the flowchart of FIG.
In step S201, the weak discriminator diversity index calculating unit 110 selects one unselected sample data from the sample data group 12.
Next, in step S202, the inter-weak classifier diversity index calculating unit 110 inputs the sample data selected in step S201 to all weak classifiers belonging to the weak classifier set 11 prepared in advance. Then, in each of those weak classifiers, the class to which the sample data belongs is identified. The weak classifier diversity index calculation unit 110 acquires the identification result of the sample data. By this processing, each weak classifier independently outputs the classification result of the selected sample data.

  Next, in step S203, the weak discriminator diversity index calculating unit 110 calculates a weak discriminator pair including two weak discriminators based on the result of the identification of the sample data in each weak discriminator calculated in step S202. Is calculated. Here, the discrimination result variation degree is an index indicating the degree to which the discrimination results of the two weak discriminators are different. FIG. 3A to FIG. 3E show a plurality of examples of the calculation method of the discrimination result variation degree.

  FIG. 3A is a schematic diagram illustrating an example of a calculation process of the variation degree of the identification result when the weak classifier outputs a single class information as the identification result. In this case, the degree of variation in the identification result can be expressed in binary. That is, as shown in FIG. 3A, when the identification results of the weak classifier A and the weak classifier B when the sample data i are input match, the weak classifier diversity index calculation unit 110 Derives 0 (zero) as the degree of variation in the identification result. On the other hand, as shown in FIG. 3A (b), when the identification results of the weak classifier A and the weak classifier B when the sample data i are input do not match, the weak classifier diversity index calculating unit 110 Derives “1” as the discrimination variation degree.

  FIG. 3B is a schematic diagram illustrating an example of a calculation process of the variation degree of the identification result when the weak identifier outputs a plurality of class information as the identification result. In this case, the identification result is regarded as a set, and the class information is regarded as an element of the set. The weak classifier diversity index calculation unit 110 calculates an inter-set similarity, which is a similarity between sets, and calculates a value obtained by subtracting the calculated inter-set similarity from 1 (= 1−inter-set similarity). It is derived as the degree of variation in the classification result. As a method of representing the similarity between sets, for example, there is a method of representing the similarity using a Jacquard coefficient or a dice coefficient. The Jacquard coefficient is a value obtained by dividing the number of common elements of the sets X and Y by the total number of elements in at least one of the sets X and Y. The dice coefficient is a value obtained by dividing the number of common elements of the sets X and Y by the average of the number of elements of each of the sets X and Y. That is, the Jacquard coefficient J and the dice coefficient D of the set X and the set Y are expressed by the following equations (1) and (2), respectively.

  In the example shown in FIG. 3B, when the sample data i is input, the weak classifier A identifies that the class to which the sample data i belongs is one of the class P, the class Q, and the class R. On the other hand, the weak classifier B identifies that the class to which the sample data i belongs is any one of the class P, the class R, the class S, and the class T. Therefore, the common elements of the classification result of the weak classifier A and the classification result of the weak classifier B are the class P and the class R. On the other hand, the elements in at least one of the classification result of the weak classifier A and the classification result of the weak classifier B are class P, class Q, class R, class S, and class T.

  As described above, when the Jacquard coefficient J is used as the similarity between sets, the similarity between sets of the weak classifier A and the weak classifier B is 0.4 (= 2/5) according to the equation (1). Therefore, when the Jacquard coefficient J is used as the similarity between sets, the weak discriminator diversity index calculating unit 110 calculates the discrimination result variation between the weak discriminators A and B as 0.6 (= 1−1). 0.4) is derived. On the other hand, when the dice coefficient D is used as the inter-set similarity, the inter-set similarity between the weak discriminators A and B is 0.57 (= 2 / 3.5) according to the equation (2). Therefore, when the dice coefficient D is used as the similarity between sets, the inter-weak discriminator diversity index calculating unit 110 calculates the discrimination result variation between the weak discriminators A and B as 0.43 (1-0. 57) is derived.

  FIG. 3C is a schematic diagram illustrating an example of a calculation process of the variation degree of the classification result when the weak classifier outputs the appearance likelihood of each class as the classification result. In this case, the discrimination result variation degree can be defined by the difference H of the likelihood histogram having each class as an element. That is, the discrimination result variation degree is expressed by the following equation (3).

  In the example shown in FIG. 3C, when the sample data i is input, the weak classifier A has a class P appearance frequency of 0.5, a class Q appearance frequency of 0.25, and a class R appearance frequency of 0.5. 25 is identified. On the other hand, the weak classifier B classifies that the appearance frequency of the class P is 0.2, the appearance frequency of the class R is 0.6, and the appearance frequency of the class T is 0.2. Therefore, the absolute value of the likelihood difference of class P, the absolute value of the likelihood difference of class Q, the absolute value of the likelihood difference of class R, and the absolute value of the likelihood difference of class T are 0.3 (= 0.5-0.2), 0.25, 0.35 (= 0.6-0.25), 0.2. Therefore, the inter-weak discriminator diversity index calculating unit 110 determines from the equation (3) that the variation degree of the discrimination result between the weak discriminators A and B is 0.55 (= (0.3 + 0.25 + 0.35 + 0. 2) / 2) is derived.

  Returning to the description of FIG. 2, in step S204, the inter-weak classifier diversity index calculation unit 110 determines whether there is a weak classifier pair for which the calculation of the degree of variation in the classification result has not been performed. As a result of this determination, when there is a weak discriminator pair for which the calculation of the degree of discrimination result variation has not been performed, the process returns to step S203, and the weak discriminator diversity index calculation unit 110 discriminates the weak discriminator pair. The degree of result variation is calculated. By repeating Step S203 and Step S204, the degree of variation in the classification result with respect to the sample data selected in Step S201 is calculated for all the weak classifier pairs in the weak classifier set 11.

  Next, in step S205, the weak discriminator diversity index calculating unit 110 determines whether or not unselected sample data exists in the sample data group 12. As a result of this determination, if there is unselected sample data, the process returns to step S201, and the weak inter-identifier diversity index calculating unit 110 selects the sample data. By repeating the loop from step S201 to step S205, the discrimination result variation degree of the weak discriminator pair is calculated for all the sample data of the sample data group 12.

  Next, in step S206, the inter-weak classifier diversity index calculation unit 110 calculates the average value of the degree of variation in the classification result in all the sample data of each weak classifier pair calculated in the loop from step S201 to step S205. . Then, the inter-weak classifier diversity index calculation unit 110 sets the average value of the variation degrees of the classification results in all the sample data of each weak classifier pair as the diversity index of each weak classifier pair.

  Through the above processing, the weakness classifier diversity index calculation unit 110 sets an index representing the diversity of the classification results between weak classifiers. In addition, the weakness classifier diversity index calculation unit 110 of the present embodiment may calculate the diversity index of the weak classifier pair based on “the degree of variation in the identification result of the sample data”. Therefore, the method of calculating the diversity index of the weak classifier pair and the definition of the classification result variation degree are not limited to those described above.

  Here, a modified example of the process of the weak discriminator inter-unit diversity index calculating unit 110 will be described. As in Hough Forest, a weak classifier diversity index calculating process in a method of detecting an object to be identified by voting to the center of gravity of the object to be identified by inputting a partial region of the image as an input is shown in FIG. This will be described using the example shown in FIG. 3E. FIGS. 3D and 3E are schematic diagrams showing first and second examples of the process of calculating the variation degree of the identification result when the identification result output from each weak classifier is an offset to the center of gravity of the object to be identified. . Here, when a sample image i is input, the weak classifier A outputs offsets to the centers of gravity of three objects to be identified as identification results, and the weak classifier B outputs five identification objects as identification results. The offset to the center of gravity of the object is output.

  In the diversity index calculation process in this case, the data selected in step S201 shown in FIG. 2 is a partial area of the sample data. All partial regions of all sample data in the sample data group 12 are evaluated in a loop from step S201 to step S205.

  The discrimination result variation degree can be expressed by, for example, a variation in offset position, which is a discrimination result for each weak discriminator. As an example of a method of indexing the variation of the offset position, the entire region of the sample data is divided into regions of the same size, and the center of gravity of the object to be identified is determined according to the frequency of voting for each region. There is a method of setting the likelihood. Specifically, first, the inter-weak classifier diversity index calculating unit 110 divides the entire region of the sample data including the input partial region into regions of the same size. In the examples shown in FIGS. 3D and 3E, the sample data i is equally divided into 24 regions. Next, the inter-weak classifier diversity index calculation unit 110 checks a region including the offset position of the classification result in each weak classifier, and calculates the likelihood of the region where the center of gravity of the object to be classified is located. In the following description, the likelihood of the region where the center of gravity of the object to be identified is located will be referred to as the voting destination region likelihood as necessary.

  The voting destination region likelihood can be represented by, for example, the ratio of the offset position included in each region to the total number of offset positions as the identification result. In the example shown in FIG. 3D, in the weak classifier A, three votes are cast in one vote in the area 4, one vote in the area 18, and one vote in the area 23. Therefore, the voting destination region likelihoods of the region 4, the region 18, and the region 23 are each 0.33. On the other hand, in the example shown in FIG. 3E, in the weak discriminator B, one vote is cast in the area 10, one vote in the area 14, two votes in the area 18, and one vote in the area 23, for a total of five votes. Therefore, the voting destination region likelihood of the region 10, the region 14, and the region 23 is 0.2, and the voting destination region likelihood of the region 18 is 0.4.

  The weak discriminator diversity index calculation unit 110 calculates the voting destination region likelihood of each weak discriminator as described above. Then, the weak discriminator diversity index calculating unit 110 calculates, for example, the difference H between the likelihood histograms described with reference to FIG. The degree of variation in the classification result of the classifier B is calculated (see equation (3)). The weak discriminator diversity index calculation unit 110 calculates the discrimination result variability for all the partial regions of all the sample data, and averages the average values thereof between the weak discriminators A and B. Sex index.

  As a result of the classification by the weak classifier, the partial region of the sample data may be determined to be a region where no object to be classified is present (in Hough Forest, the input data is a negative sample). The area where it is determined that the object to be identified does not exist is an area that does not contribute to the detection of the center of gravity of the object to be identified during the identification processing by the weak classifier. Therefore, in the calculation of the degree of variation in the classification result in step S203, if any weak classifier determines that no object to be classified is present in the partial area of the input sample data, the diversity between weak classifiers is determined. The index calculation unit 110 performs the following. That is, the weak discriminator diversity index calculation unit 110 does not calculate the discrimination result variation degree of the weak discriminator pair including the weak discriminator in this partial area, and counts the calculation of the diversity index between discriminators ( In the example described above, it is excluded from the calculation target of the average value of the degree of variation of the identification result. By doing so, the diversity index between classifiers (weak classifier diversity index) can be calculated more appropriately.

Next, an example of the processing of the weak discriminator combination selection unit 120 will be described in detail with reference to the flowchart shown in FIG. 4 and the schematic diagrams shown in FIGS. 5A and 5B.
Assuming that the number of elements of the weak classifier set 11 is M, one weak classifier diversity index is calculated for each weak classifier pair. The calculated total number of weak classifier diversity indices is expressed by the following equation (4).

  For example, when M = 5, the weak discriminator diversity indices of ten weak discriminator pairs are calculated. FIG. 5A is a diagram illustrating an example of the weak discriminator diversity index when M = 5. In FIG. 5A, the numerical value at the intersection of each row and each column represents the weakness classifier diversity index of each weak classifier pair. For example, the weak discriminator diversity index of the weak discriminator of ID2 and the weak discriminator of ID 4 is 0.749. The weak discriminator combination selection unit 120 selects L weak discriminators from the M weak discriminator sets based on the inter-weak discriminator diversity index.

In step S401, the weak discriminator combination selection unit 120 does not calculate the sum value described later from the Comb (M, L) combination patterns for selecting L from the M weak discriminator sets. Select one combination pattern. L is the number of weak classifiers forming the ensemble classifier.
Next, in step S402, the weak discriminator combination selection unit 120 sums up the Comb (L, 2) inter-weak discriminator diversity indexes of the combination pattern of the L weak discriminators selected in step S401. Calculate the value.

  Next, in step S403, the weak discriminator combination selection unit 120 determines that there are L weak discriminator patterns for which the sum of the Comb (L, 2) weak discriminator diversity indexes has not been calculated. It is determined whether or not to perform. If the result of this determination is that there are L weak discriminator patterns for which the total value of the Comb (L, 2) weak discriminator diversity indexes has not been calculated, the process returns to step S401. Then, the weak discriminator combination selection unit 120 selects a combination pattern for which the sum of the Comb (L, 2) weak discriminator diversity indexes has not been calculated. By repeating the loop from step S401 to step S403, the sum value of the diversity index between weak discriminators is calculated for all the combination patterns of Comb (M, L).

Then, in step S404, the weak discriminator combination selection unit 120 searches the L combination patterns for a combination in which the sum of the weak discriminator diversity indices calculated in step S402 is the largest, and determines the weakest one. Select a discriminator combination pattern.
Through the above processing, it is possible to select an appropriate combination of weak classifiers constituting the ensemble classifier based on the diversity index between weak classifiers.

  FIG. 5B is a diagram illustrating an example of processing of the weak classifier combination selection unit 120 when M = 5, L = 3, that is, a combination of three weak classifiers is selected from a set of five weak classifiers. is there. In this case, since Comb (5, 3) = 20, there are a total of 20 combinations patterns. The weak discriminator combination selection unit 120 calculates the sum of the weak discriminator diversity indices for all the combination patterns.

  For example, the leftmost table in FIG. 5B shows the sum of the weak discriminator diversity indices for the combination patterns of the weak discriminators having IDs of 2, 3, and 4. The diversity index of the weak discriminator pair of ID 2 and ID 3 is 0.675, the diversity index of the weak discriminator pair of ID 2 and ID 4 is 0.749, and the ID of ID 3 and ID 4 is The diversity index is 0.034. Therefore, their total value is 1.458. The weak discriminator combination selection unit 120 individually calculates the sum of the weak discriminator diversity indices for each of all the combination patterns by repeating the above processing for the 20 combination patterns. Then, the weak discriminator combination selection unit 120 selects a combination pattern in which the sum total is maximum. In the example shown in FIG. 5B, the sum of the weak discriminator diversity indices for the combination pattern of the weak discriminators having IDs of 0, 2 and 3 is 2.364, and among all the combination patterns, Since the maximum value is obtained, the combination pattern is selected.

  The ensemble classifier generation unit 130 generates the ensemble classifier 13 using the L weak classifiers selected by the weak classifier combination selection unit 120. Note that a method of generating the ensemble classifier 13 using a plurality of weak classifiers can be realized by a known technique, and a detailed description thereof will be omitted here.

  As described above, in the present embodiment, the sample data is input to the weak classifier, the classification result of the classification target in the sample data is obtained, and the degree of variation of the classification result in the weak classifier pair is calculated. The calculation of the degree of variation of the classification result in such a weak classifier pair is performed for each of the possible weak classifier pairs and a plurality of sample data. Then, a weakness classifier diversity index indicating the diversity of the classification results of the weak classifier pairs is calculated from the degree of variation of the classification results of the weak classifier pairs when the plurality of sample data are used. Then, the combination of the predetermined number of weak discriminators is selected based on the diversity index between weak discriminators in the weak discriminator pair included in the combination pattern of the predetermined number of weak discriminators constituting the ensemble discriminator. . Then, an ensemble classifier is created using the selected weak classifiers. Therefore, it is possible to evaluate the ensemble discriminator using the actual photographing data having no class information. Therefore, even if actual photographing data having no class information is used, a combination of weak classifiers useful for identifying a classification target can be selected, and a high-performance ensemble classifier can be created.

  In the present embodiment, the weak discriminator diversity index calculation unit 110 calculates the weak discriminator diversity index by comparing the discrimination results of the weak discriminator pairs, and the weak discriminator combination selection unit 120 outputs The case of deriving the combination of weak classifiers based on the sum of the diversity index between weak classifiers has been described. However, this is not necessary. For example, the following may be performed. That is, the weak inter-classifier diversity index calculating unit 110 compares the classification results of the L (L> 2) weak classifiers and calculates the weak classifier diversity index of the L weak classifier pairs. . Then, the weak discriminator combination selection unit 120 selects a weak discriminator combination that maximizes the diversity indicator between weak discriminators of the L weak discriminator pairs. An existing method can be used for calculating the diversity index between weak classifiers of the pair of L weak classifiers. For example, by regarding the classification result of each weak classifier as each element of the set and calculating the similarity of the L sets, it is possible to calculate the diversity index between weak classifiers.

  Further, in the present embodiment, an example has been described in which the data to be identified in the classifier is image data. However, data to be identified in the classifier is not limited to image data. For example, production result data may be used. Also, the case where the classifier identifies the class to which the identification target belongs in the input data has been described as an example. However, the discriminator is not limited to such as long as it identifies a discrimination target in input data. For example, an identifier for identifying the presence or absence of an identification target may be employed.

(Second embodiment)
Next, a second embodiment will be described. In the present embodiment, an ensemble classifier that uses a weak classifier that performs a classification process on input sample data and outputs an identification result of an identification target in the sample data based on a result of the classification process The case of using is described as an example. As described above, the present embodiment and the first embodiment mainly differ in the processing due to the difference in the method of identifying the identification target by the weak classifier. Therefore, in the description of the present embodiment, the same portions as those in the first embodiment are denoted by the same reference numerals as those in FIGS. 1 to 5B, and the detailed description is omitted. Specifically, the configuration of the information processing apparatus according to the present embodiment is the same as that of the first embodiment. Further, the processing of the weak discriminator combination selector 120 and the ensemble discriminator generator 130 in the information processing apparatus 100 shown in FIG. 1 is the same as in the first embodiment. Therefore, these detailed explanations are omitted.

  There is a decision tree as an example of a weak classifier that performs a classification process on input sample data and outputs an identification result of an identification target in the sample data based on a result of the classification process. FIG. 6 is a schematic diagram illustrating an example of the flow of identification using a decision tree. First, in the classification process, a decision tree to which data is input applies the data to a branch function set for each node of the decision tree, and based on the result, the data is divided into one of the terminal nodes (leaves). Classify into. In the example shown in FIG. 6, the input tree i is classified into leaf a by the decision tree A.

  Class information is stored in each leaf of the decision tree by a learning process in advance. After the classification process is performed, as the identification process, the decision tree outputs the class information stored in the leaf to which the input data is classified as the identification result of the identification target in the data. In the example shown in FIG. 6, class information {class P, class Q} stored in leaf a as a classification destination of input data i is output as a result of identification of input data i by decision tree A. Since the decision tree itself can be realized by a known technique, a detailed description thereof is omitted here.

  In the first embodiment, an example has been described in which the diversity index between weak classifiers is calculated by focusing on the classification result when sample data is input to each weak classifier. On the other hand, in the present embodiment, the inter-weak classifier diversity index is calculated by focusing on the classification result when sample data is input to each weak classifier.

An example of the process of the weak discriminator diversity index calculating unit 110 will be described in detail with reference to the flowchart shown in FIG.
In step S701, the inter-weak classifier diversity index calculation unit 110 inputs all sample data included in the sample data group 12 to all weak classifiers included in the weak classifier set 11. Accordingly, each weak classifier included in the weak classifier set 11 classifies the sample data input thereto.

Next, in step S702, the inter-weak classifier diversity index calculation unit 110 selects one unselected sample data from the sample data group 12.
Next, in step S703, the inter-weak discriminator diversity index calculating unit 110 uses the two weak discriminators whose classification result variation degree has not been calculated based on the classification result of the sample data in each weak discriminator performed in step S701. Calculate the classification result variation degree of the container pair. Here, the classification result variation degree is an index indicating the degree to which the classification results of the two weak classifiers are different. For comparison of the classification results between the two weak classifiers, the classification results for all the sample data performed in step S701 are used.

The leaves that are the classification results of the weak classifiers are collected, and each sample data classified into each leaf is regarded as a set element. When a certain sample data is input to two discriminators, the weak discriminator diversity index calculation unit 110 uses the similarity between sets to which the sample data belongs in each weak discriminator as the inter-set similarity. calculate. Then, the weak discriminator diversity index calculation unit 110 derives a value obtained by subtracting the calculated inter-set similarity from 1 (= 1−inter-set similarity) as the classification result variation degree. The similarity between the two sets is determined by the Jacquard coefficient J and the dice as in the first embodiment, as in the case of calculating the discrimination result variation when the weak discriminator outputs a plurality of class information as discrimination results in the first embodiment. It can be calculated using the coefficient D.
The processing from step S704 to step S706 is the same as the processing from step S204 to step S206 in the first embodiment, and a detailed description thereof will be omitted. In step S706, the weak discriminator diversity index calculation unit 110 sets the weak discriminator diversity index of each weak discriminator pair using the classification result variation degree instead of using the classification result variation degree. (Derived).

  FIG. 8 is a schematic diagram illustrating an example of a process of calculating the classification result variation degree when the sample data 1 is input to the weak classifiers A and B. The sample data 1 is classified into the classification result a by the weak classifier A, and is classified into the classification result b by the weak classifier B. A set of classification results a including all the sample data classified by the weak discriminator A into the classification result a similarly to the sample data 1, and a classification including all the sample data classified by the weak discriminator B into the classification result b. Compare with the set of results b. The classification result variation degree is calculated based on the result of this comparison.

  In the case illustrated in FIG. 8, the similarity between sets of the set of the classification results a and the set of the classification results b is calculated. The set of the classification results a is a set having {sample data 1, sample data 3, sample data 4, and sample data 7} as elements. The set of the classification results b is a set having {sample data 1, sample data 2, sample data 3, sample data 5, and sample data 8} as elements. In this case, the common element between the set of classification results a and the set of classification results b is {sample data 1, sample data 3}. Therefore, the similarity between sets when the Jacquard coefficient J is used is 2/7 from the equation (1), and the similarity between sets when the dice coefficient D is used is 4/9 ( = 2 x 2/9). Similarly, the inter-weak discriminator diversity index calculating unit 110 determines, for all the sample data included in the sample data group 12, a set of classification results in which each sample data is classified by the weak discriminator A and a weak discrimination. And a set of classification results classified by the classifier B. Then, the weak discriminator diversity index calculating unit 110 calculates the weak discriminator diversity indices of the weak discriminators A and B by averaging the calculated classification result variation degrees.

  As described above, in this embodiment, even when a weak classifier that performs classification processing on input sample data such as a decision tree and outputs a classification result based on the result of the classification processing is used, The same effect as the effect described in the first embodiment can be obtained. That is, the weakness classifier diversity index can be calculated using only the result of the classification processing of the sample data without using the class information learned by each weak classifier.

  The inter-weak discriminator diversity index calculation unit 110 may further calculate the inter-weak discriminator diversity index based on the degree of variation in the identification results described in the first embodiment. In this case, for example, the weak discriminator combination selection unit 120 configures an ensemble discriminator by combining the weak discriminator diversity index based on the classification result variation degree and the weak discriminator diversity index based on the classification result variation degree. You may select a weak classifier to perform. Also, in the present embodiment, the modified example described in the first embodiment can be adopted.

(Third embodiment)
Next, a third embodiment will be described. In the present embodiment, at least one of the amount of memory used for data identification, the processing speed at the time of data identification, and the accuracy of the identification result is set, and the weak classifier configuring the ensemble classifier according to the set values is set. Change the number of As described above, the present embodiment and the first embodiment mainly differ in the configuration and processing when generating an ensemble discriminator. Therefore, in the description of the present embodiment, the same parts as those of the first embodiment are denoted by the same reference numerals as those shown in FIGS. 1 to 5 and detailed description is omitted.

  Ensemble classifiers generally improve in performance as the number of weak classifiers used therein increases, and eventually the performance saturates. Therefore, when the ensemble classifier does not achieve the desired accuracy when performing the classification process, a method of performing the classification process using an ensemble classifier configured using a larger number of weak classifiers is effective. . On the other hand, as the number of weak classifiers constituting the ensemble classifier increases, the capacity of the ensemble classifier and the time required for the classification process also increase. Therefore, when it is desired to reduce the load on the execution memory and to shorten the processing time in the identification processing, the identification may be performed using an ensemble identifier configured using a small number of weak identifiers.

FIG. 9 is a diagram illustrating an example of a basic configuration of an ensemble discriminator construction system. In the present embodiment, the ensemble classifier construction system includes an information processing device 200 that calculates a combination of weak classifiers, and an object classification device 300 that performs a classification process.
The information processing device 200 includes a weak discriminator diversity index calculation unit 210 and a weak discriminator combination selection unit 220.

The inter-weak classifier diversity index calculating unit 210 receives the weak classifier set 21 and the sample data group 22 as inputs and outputs the weak classifier-to-weak classifier diversity similarly to the inter-weak classifier diversity index calculating unit 110 in the first embodiment. Is calculated (an index indicating the diversity between weak classifiers).
The weak discriminator combination selection unit 220 is based on the weak discriminator diversity index calculated by the weak discriminator diversity index calculation unit 210, similarly to the weak discriminator combination selection unit 220 in the first embodiment. , A predetermined number of weak classifiers are selected. However, as a point different from the first embodiment, the weak discriminator combination selection section 220 selects a plurality of combinations as a combination of a plurality of weak discriminators. That is, there are a plurality of the predetermined number of candidates. The weak discriminator combination selection unit 220 outputs a combination of the weak discriminators selected in each of the predetermined number of cases as a selected weak discriminator combination list 23.

The object identification device 300 includes a target value setting unit 310, an ensemble identification unit generation unit 320, and an identification unit 330.
The target value setting section 310 sets a target value input by the user.
The ensemble classifier generation unit 320 generates an ensemble classifier based on the selected weak classifier combination list 23 output from the information processing device 200 and the target value set by the target value setting unit 310.
The identification unit 330 identifies the identification target data 31 using the ensemble identification unit generated by the ensemble identification unit generation unit 320, and outputs the result as the identification result 32.

An example of the processing of the ensemble classifier construction system (the information processing device 200 and the object identification device 300) will be described with reference to the flowchart of FIG.
In step S1001, the weak discriminator diversity index calculation unit 210 calculates a weak discriminator diversity index for all the weak discriminator pairs of the weak discriminator group included in the weak discriminator set 21. The specific processing is the same as the processing described with reference to the flowchart of FIG. 2 in the first embodiment or FIG. 7 in the second embodiment, and thus detailed description thereof will be omitted.

  Next, in step S1002, the weak discriminator combination selection unit 220 sets a plurality of patterns of the number of weak discriminators to be selected. Assuming that the number of weak classifiers prepared in advance is N, the weak classifier combination selector 220 may specify all numbers from 2 to N−1, or a predetermined number to reduce the amount of calculation. The number may be specified every other time.

  Next, in step S1003, the weak discriminator combination selection unit 220 selects an appropriate combination of weak discriminators in the number specified in step S1003 based on the inter-weak discriminator diversity index calculated in step S1001. . The process of calculating the combination of the weak classifiers is the same as the process described with reference to the flowchart of FIG. 4 in the first embodiment, and thus a detailed description thereof will be omitted.

  In step S1004, the target value setting unit 310 sets a target value in the identification processing using the ensemble classifier. Here, one of the memory amount (capacity) used for the identification processing and the processing speed required for the identification processing (the execution speed of the identification processing) is set as the target value. The target value setting unit 310 can set, for example, a memory amount specified by the user. Further, the target value setting unit 310 may set the amount of memory to be used for the identification processing based on the maximum amount of memory that can be used for identification. Further, the target value setting unit 310 may set the processing speed specified by the user, or may calculate the processing speed from a preset overall processing time. Steps S1001 to S1003 and step S1004 are independent processes. Steps S1001 to 1003 and step S1004 may be processed in parallel, or one may be processed first and the other may be processed later.

Upon completion of the processes in steps S1001 to S1004, the process proceeds to step S1005. In step S1005, the ensemble classifier generation unit 320 sets the number of weak classifiers that satisfy the target value set in step S1004.
The amount of memory used for the classification process can be calculated from the size of the weak classifier that forms the ensemble classifier. Accordingly, the ensemble classifier generation unit 320 calculates the amount of memory to be used when performing the classification process using the combination of the weak classifiers selected in step S1003, and selects a number such that the calculated amount of memory satisfies the target value. can do.

  Further, in the classification processing in the ensemble classifier, the classification processing is performed by the weak classifier, and the classification results are totaled to output the final classification result. The execution speed of the ensemble classifier at the time of the classification process is proportional to the number of weak classifiers. For this reason, the ensemble classifier generation unit 320 can calculate the number of weak classifiers that satisfy the required speed (target value) by measuring the execution speed of the weak classifier alone at the time of the identification processing in advance. It is.

Next, in step S1006, the ensemble classifier generation unit 320 uses the weak classifiers of the combinations of the weak classifiers set in step S1005 among the combinations of weak classifiers selected in step S1003. Generate a classifier.
Next, in step S1007, the identification unit 330 performs an identification process on the input identification target data 31 by using the ensemble identifier generated in step S1006.
Next, in step S1008, the identification unit 330 determines whether the result (identification result) of the identification processing performed in step S1007 satisfies a desired performance (used memory amount or execution speed).

  When the log of the maximum memory usage and the execution speed is included in the identification result in addition to the identification result of the object to be identified, for example, the determination in step S1008 can be performed as follows. That is, the user refers to the log and instructs whether or not the desired performance is satisfied, and the identification unit 330 determines whether or not the identification result satisfies the desired performance based on the content of the instruction. Can be.

  As a result of the determination in step S1008, when the identification result satisfies the desired performance, the identification unit 330 outputs the identification result 32, and ends the processing according to the flowchart in FIG.

  On the other hand, if the classification result does not satisfy the desired performance, the process returns to step S1005, and the ensemble classifier generation unit 320 sets a smaller number of weak classifiers.

As described above, in the present embodiment, the set target value, that is, the amount of memory used in the identification processing, or the number of weak classifiers that satisfy the processing speed required in the identification processing is derived and derived. A combination pattern of weak classifiers including the number of weak classifiers is selected. Therefore, by dynamically setting the number of weak discriminators constituting the ensemble discriminator according to the set target value, the following effects can be obtained in addition to the effects described in the first embodiment. That is, it is possible to generate a higher-performance ensemble discriminator while satisfying the predetermined condition, and it is possible to realize a memory saving at the time of the identification process and a high-speed identification process.
Note that the object identification device 300 of the present embodiment can be applied to the second embodiment. Further, the ensemble classifier construction system (the information processing device 200 and the object classification device 300) may be configured by one device.

  It should be noted that each of the above-described embodiments is merely an example of a concrete example for carrying out the present invention, and the technical scope of the present invention should not be interpreted in a limited manner. That is, the present invention can be implemented in various forms without departing from the technical idea or the main features.

(Other Examples)
The present invention is also realized by executing the following processing. That is, first, software (computer program) that realizes the functions of the above embodiments is supplied to a system or an apparatus via a network or various storage media. Then, a computer (or CPU, MPU, or the like) of the system or apparatus reads and executes the computer program.

  11: weak discriminator set, 12: sample data group, 13: ensemble discriminator, 100: information processing device, 110: inter-weak discriminator diversity index calculation unit, 120: weak discriminator combination selection unit, 130: ensemble discrimination Unit generator

Claims (12)

An information processing apparatus for performing a process for selecting a plurality of weak classifiers constituting the ensemble classifier,
Input means for inputting a candidate of a weak classifier constituting the ensemble classifier,
The same data is given to at least two of the weak discriminators among the candidates input by the input means, and information obtained when performing discrimination processing on the data in each of the weak discriminators is compared. Deriving means for, for each of a plurality of combinations of the at least two weak discriminators, deriving an index representing the diversity of the discrimination results between the at least two weak discriminators based on the result of
For each of the candidates of the number of the plurality of weak classifiers constituting the ensemble classifier , based on an index derived by the derivation unit and representing the diversity of the classification result between the at least two weak classifiers. And a processing unit for selecting a combination of the plurality of weak classifiers.   The information processing apparatus according to claim 1, wherein the processing unit further includes a generating unit configured to generate the ensemble classifier using a combination of the selected weak classifiers.   The information processing apparatus according to claim 1, wherein the information obtained when performing the identification processing is a result of identification of an identification target by the identifier. The weak classifier classifies the input data, and outputs an identification result of the identification target based on the classification result,
The information processing apparatus according to claim 1, wherein the information obtained when performing the identification processing is a result of the classification performed by the identifier.   The information processing apparatus according to claim 1, wherein a label is not assigned to the data. The deriving unit is obtained by giving the same data to at least two of the weak discriminators among the candidates input by the input unit, and performing the discriminating process on the data in each of the weak discriminators. Means for obtaining information for each of the plurality of data,
Based on the information obtained when performing the identification processing, in the at least two weak classifiers, means for deriving a degree of variation in information obtained when performing the identification processing,
Means for deriving an index indicating the diversity of identification results between the at least two weak classifiers based on a degree of variation in information obtained when performing the identification processing. The information processing device according to any one of claims 1 to 5 . The deriving unit provides the same data to at least two of the weak classifiers among the candidates input by the input unit, and performs a classification process on the data in each of the weak classifiers. 7. The information processing apparatus according to claim 6 , wherein when does not contribute to the identification of the identification target, when the data is used, the degree of variation in information obtained when performing the identification processing is not derived. . The number of the plurality of weak classifiers, the information processing apparatus according to any one of claim 1 to 7, characterized in that a predetermined number. An ensemble classifier construction system having the information processing device according to claim 1 ,
Target value setting means for setting a target value for identification processing by the ensemble identifier,
Deriving a number of the weak classifiers to achieve the target value, before hexene-option has been among the combinations of the plurality of weak classifiers, using said combination of said plurality of weak classifiers of the number obtained by the derived Generating means for generating an ensemble discriminator; 10. The system according to claim 9 , wherein the target value includes at least one of an amount of memory used when the identification by the ensemble identifier is performed and a processing speed when the identification by the ensemble identifier is performed. 4. The ensemble discriminator construction system according to 1. An information processing method for performing a process for selecting a plurality of weak classifiers constituting the ensemble classifier,
An inputting step of inputting a candidate of a weak classifier constituting the ensemble classifier,
Among the candidates input in the input step, the same data is given to at least two of the weak classifiers, and information obtained when performing the identification process on the data in each of the weak classifiers is compared. Deriving an index representing the diversity of the classification result between the at least two weak classifiers based on the result, performing a deriving step for each of a plurality of combinations of the at least two weak classifiers;
For each of the candidates for the number of the plurality of weak classifiers constituting the ensemble classifier , based on an index that is derived in the derivation step and that represents the diversity of the classification results between the at least two weak classifiers. the information processing method characterized by having a step of selecting a combination of said plurality of weak classifiers. Program for causing a computer to function as each unit of the information processing apparatus according to any one of claims 1-8.

Images