JP2016062249A - Identification dictionary learning system, recognition dictionary learning method and recognition dictionary learning program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an identification dictionary learning device that can shorten identification time while suppressing a decrease of identification accuracy.SOLUTION: An identification dictionary learning device includes: learning means for learning multiple classifiers for classifying a data value belonging to any of multiple classes in any of multiple classes by using the data value; index calculation means 113 for calculating each optimization index of the multiple classifiers, based on multiple optimization index coefficients each representing the size of contribution to the optimization index representing identification accuracy of classification of combination between the class where a learning sample in the wrong classification belongs to and the classified class; determination means 114 for selecting the classifier among the classifiers representing that the identification accuracy of the optimization index is the highest as a weak classifier, and adding the weak classifier to the identification dictionary being aggregation of the weak classifiers stored in identification dictionary memory means 122; and coefficient update means 116 for performing the update of the optimization index coefficient based on the rule according to the addition.SELECTED DRAWING: Figure 11

Description

  The present invention relates to a technique for learning a recognition dictionary.

  One of the statistical pattern identification techniques often used in object detection is a technique called Boosting. Boosting is a technique for improving the accuracy of identification by combining a large number of simple classifiers. In Boosting, a simple classifier is called a weak classifier. A high-accuracy classifier made by combining weak classifiers is called a strong classifier. Examples of techniques based on Boosting are described in Non-Patent Documents 1 to 4, for example.

  Non-Patent Document 1 describes an example in which AdaBoost, which is one of Boosting, is used for object detection. Non-Patent Document 2 describes a boosting method called RealAdaBoost that is an improvement of AdaBoost. Non-Patent Document 3 describes GentleAdaBoost, which is one of the improved methods of AdaBoost. The techniques described in Non-Patent Document 1, Non-Patent Document 2, and Non-Patent Document 3 can be applied to, for example, detection of a specific object such as a face or a car in an image. For example, whether or not the partial region of the image is an image of the specific object can be identified using the technique described in Non-Patent Document 1, Non-Patent Document 2, or Non-Patent Document 3. .

  Non-Patent Document 4 describes an example of a technique for identifying an object having a large change in appearance due to a difference in direction, for example. In the technique described in Non-Patent Document 4, for one object that is an identification target, a plurality of categories (that is, classes) according to appearance variations and a category of a target other than the object that is the identification target And are generated.

  Patent Document 1 describes a discriminator generating device that generates a discriminator that performs multi-class discrimination, in which an object is classified into one of a plurality of classes. The discriminator generation device of Patent Document 1 generates a discriminator by combining a plurality of weak discriminators. The discriminator generation device of Patent Literature 1 performs learning to share only the feature amount with weak discriminators between a plurality of classes.

  Patent Document 2 describes a learning model generation device that optimizes parameters of a learning model. The learning model generation device of Patent Literature 2 generates a classifier for each recognition target.

JP 2011-181016 A JP 2012-038244 A

P. Viola and M. Jones, “Robust real-time object detection”, IJCV, 57 (2), pp. 137-154, 2004 R. Schapire, Y. Singer, “Improved boosting algorithms using confidence-rated predictions”, Proc. Of 11th conference on Computational Learning Theory, 1998 J. Friedman, T. Hastie, R. Tibshirani, “Additive logistic regression: a statistical view of boosting”, Annals of statistics, 28 (2), pp. 337-374, 2000 A. Torralba, K. P. Murphy, W. T. Freeman, “Sharing visual features for multiclass and multiview object detection”, PAMI2006

  In the techniques described in Non-Patent Documents 1 to 3, when the variation of the appearance of the object to be identified is large, the variation of the learning sample obtained from the object to be identified is large. That is, the variation of the learning sample classified into the class of the object to be identified is large. Therefore, in the techniques described in Non-Patent Documents 1 to 3, high discrimination accuracy cannot be obtained by a discriminator generated as a result of learning.

  When assigning a plurality of classes to an object to be identified that has a large variation in appearance, the object can be identified if a plurality of classes assigned to the object to be identified and a class assigned to a target other than the object to be identified can be identified. That is, the combinations of classes that need to be identified for object identification are limited. However, in the techniques described in Non-Patent Document 4, Patent Document 1 and Patent Document 2, there is no distinction between a class generated according to a variation in the appearance of an object to be detected and a target class that is not an object to be detected. Learning is performed. Therefore, in the identification based on the identification dictionary learned by the techniques of Non-Patent Document 4, Patent Document 1, and Patent Document 2, even if the combinations of classes that need to be identified are limited as described above, the identification dictionary It is necessary to complete all calculations associated with. For example, if the calculation is interrupted halfway, the identification accuracy may be reduced. That is, in the techniques described in Non-Patent Document 4, Patent Document 1, and Patent Document 2, when the variation in the appearance of the identification target is large, the identification time is reduced by, for example, terminating the identification process while suppressing a decrease in identification accuracy. It is not possible to obtain an identification dictionary that can shorten the above. Note that the identification dictionary is a set of data values such as identifiers and parameters representing the classifiers obtained as a result of learning, for example.

  One of the objects of the present invention is an identification dictionary that can shorten the identification time while suppressing a decrease in identification accuracy compared with a case where an object with a large appearance variation is a target of identification compared to a case where the appearance variation is small. Is to provide an identification dictionary learning apparatus capable of generating

  An identification dictionary learning device according to an aspect of the present invention performs class classification for classifying a data value belonging to any of a plurality of classes into any of the plurality of classes, including a plurality of classifiers, a plurality of the data An optimization index representing the classification accuracy of the class classification of a combination of a learning means that learns using a learning sample that is a value, the class to which the learning sample belongs in the incorrect class classification, and the classified class Index calculating means for calculating the optimization index of each of the plurality of discriminators based on a plurality of optimization index coefficients each representing the magnitude of contribution to each of the discriminators learned, the optimal The classification indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is a set of the weak classifiers stored in the identification dictionary storage unit Comprising a determining means for adding in the book, in response to the weak classifier is added, and a coefficient updating means for performing, based updating of the optimization metrics coefficient rules.

  An identification dictionary learning method according to an aspect of the present invention includes: classifying a data value belonging to any of a plurality of classes into any of the plurality of classes; Contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class learned using the learning sample that is a value The optimization index of each of the plurality of classifiers is calculated based on a plurality of optimization index coefficients each representing the size of the classifier, and among the learned classifiers, the optimization index has an identification accuracy. The classifier representing the highest is selected as a weak classifier, and the weak classifier is added to an identification dictionary that is a set of weak classifiers stored in an identification dictionary storage unit, and the weak classifier But In response to the pressure, it performed on the basis of the updating of the optimization metrics coefficient rules.

  An identification dictionary learning program according to an aspect of the present invention performs classification of a computer that classifies a data value belonging to any of a plurality of classes into any of the plurality of classes. Represents the classification accuracy of the class classification of the combination of the learning means that learns using the learning sample that is the data value, the class to which the learning sample belongs in the incorrect class classification, and the classified class. Based on a plurality of optimization index coefficients each representing the magnitude of contribution to the optimization index, index calculation means for calculating the optimization index of each of the plurality of classifiers, and among the learned classifiers The optimization indicator selects the discriminator indicating that the discrimination accuracy is the highest as the weak discriminator, and the weak discriminator is stored in the identification dictionary storage unit, the weak discriminator A decision means for adding to an identification dictionary which is a set of separate devices; and a coefficient updating means for updating the optimization index coefficient based on a rule in accordance with the addition of the weak classifier. .

  According to the present invention, it is possible to create an identification dictionary that can shorten the identification time while suppressing a decrease in identification accuracy as compared with a case where an object having a large variation in appearance is a target of identification compared to a case where the variation in appearance is small. effective.

FIG. 1 is a block diagram showing an example of the configuration of an identification dictionary learning system 100 according to the first embodiment of the present invention. FIG. 2 is a diagram schematically illustrating an example of weights stored in the weight storage unit 123. FIG. 3 is a diagram schematically illustrating an example of a label stored in the sample storage unit 124. FIG. 4A is a first diagram illustrating an example of a histogram. FIG. 4B is a second diagram illustrating an example of a histogram. FIG. 5A is a first diagram illustrating an example of an optimization index table. FIG. 5B is a second diagram illustrating an example of the optimization index table. FIG. 6 is a diagram schematically illustrating a change in the optimization index coefficient associated with the update. FIG. 7 is a flowchart showing an example of operation of the identification dictionary learning device 110 according to the first exemplary embodiment of the present invention. FIG. 8 is a diagram schematically illustrating an example of identification targets classified into hierarchical multiclasses. FIG. 9 is a block diagram showing a configuration of an object detection system 300 according to the second embodiment of the present invention. FIG. 10 is a flowchart illustrating an example of the operation of the object detection device 310 according to the second embodiment of this invention. FIG. 11 is a block diagram showing a configuration of an identification dictionary learning device 100A according to the third exemplary embodiment of the present invention. FIG. 12 is a diagram illustrating an example of a hardware configuration of a computer 1000 that can implement the apparatus according to each embodiment of the present invention.

  In the following, first, a technique for detecting an object in an image will be described for comparison with the present invention.

  A technique called template matching has long been used as a technique for detecting a specific type of object in an image and determining the position of an area where the image of the object exists. Template matching is a search that identifies a region similar to a template image by repeating matching calculations between a search window set in the image and a template image prepared in advance many times while changing the search window. It is a technique. The template image is an image of only the detected object, for example. At this time, when matching is simply performed using the pixel value of the template image and the pixel value in the image, the matching accuracy is low. In order to increase the accuracy of matching, a technique of converting pixel values into numerical sequences called feature values that are more easily matched by calculating gradient information with secondary pixels, second-order differential information, and the like is often used. . Furthermore, in order to improve accuracy, a method of optimizing (that is, learning) a parameter group used for identification processing by using a statistical pattern identification technique and calculating a matching using the dictionary is often used. . The parameter group used for the identification process is generally called a “dictionary”. The dictionary in this case is capable of classifying two classes: an object class (positive class) which is a class of a specific object to be detected and a non-object class (negative class) which is a class other than the specific object to be detected. Then, it is optimized for a certain index. This index depends on the statistical pattern identification technique used.

  One of the statistical pattern identification techniques often used in object detection is a technique called Boosting. Boosting is one of the techniques for constructing a highly accurate classifier (strong classifier) by combining a large number of simple classifiers (weak classifiers). In a classification process using a commonly used method called AdaBoost, the value obtained by sequentially executing a number of weak classifiers and adding the output scores of the individual weak classifiers with weights is the output of the strong classifiers. Derived as a score. In the learning process by AdaBoost, one weak classifier is learned one by one, and one weak classifier is constructed by adding the learned weak classifier to the weak classifier used for the strong classifier. When learning a weak discriminator, learning that emphasizes the learning sample is performed by assigning a large weight to learning data (learning sample) that is easily misidentified by an already added weak discriminator.

  The learning sample is, for example, data used for learning by the learning unit 112 among data values obtained by observing a known identification target. The identification data is, for example, a data value obtained by observing an identification target for identifying a class to which the identification data belongs. In the following description, the identification data is also simply referred to as “sample”. The identification target belongs to one of predetermined classes. In the following description, the “class to which the learning sample belongs” refers to the class to which the identification target from which the learning sample is obtained by observation belongs. Similarly, “the class to which the identification data (sample) belongs” refers to the class to which the identification target from which the identification data (sample) is obtained by observation belongs. For example, when the identification target from which the learning sample A is obtained by observation belongs to the class B, it is described as “the learning sample A belongs to the class B”.

  As described above, Non-Patent Document 1 discloses an example in which Boosting is effectively used for object detection. The technique disclosed in Non-Patent Document 1 is a widely known technique. As a feature quantity, a Haar wavelet that can be calculated at high speed by using a RectangleFilter is adopted. As a boosting weak classifier, a process (Decision-Stamp) that simply performs threshold processing on one of feature values is employed. Therefore, in the weak classifier learning process, one feature value is selected, and one appropriate threshold is determined for the selected feature value. AdaBoost is used to generate the strong classifier. Non-Patent Document 2 discloses a boosting method called RealAdaBoost, in which AdaBoost is improved. In this method, weak classification based on the appearance probability of the positive class feature quantity and the appearance probability of the negative class feature quantity is performed instead of simple Decision-Stamp. Therefore, if there is a large amount of learning data, highly accurate object detection can be realized. Non-Patent Document 3 discloses one of the methods improved by AdaBoost called GentleAdaBoost. This technique is known to be able to learn stably by correcting the definition of an index (loss) to be optimized in the learning process, and to obtain an accuracy equal to or higher than that of RealAdaBoost. With the technique disclosed in Non-Patent Document 1 described above, it is possible to detect a specific object (for example, a face or a car) from an image with high accuracy. Further, by combining the technique disclosed in Non-Patent Document 1 with the technique disclosed in Non-Patent Document 2 or Non-Patent Document 3, a specific object (eg, face or car) can be accurately identified from an image. It is possible to detect. However, in these cases, a high detection accuracy cannot be obtained when the appearance variation of the object to be detected is large.

  When detecting an object having a large appearance variation on an image, a method of preparing one identification dictionary for each type of appearance variation and performing a detection process for each type of appearance variation is often used. However, this method has a problem that the processing time increases by the number of types of appearance. On the other hand, Non-Patent Document 4 discloses a framework for learning so that appearance variations can be classified into different categories (classes) and feature quantities and weak classifiers can be shared among a plurality of classes. ing. In this method, since weak classifiers can be shared, the processing time does not simply increase to the number of classes, and relatively high-speed processing can be realized. However, since the number of weak classifiers tends to increase as compared with the case of one appearance, the degree of speedup is not large.

  There are many multi-class identification methods in the field of statistical pattern identification. In multi-class identification that detects an identification target such as an object having a large variation in appearance, for example, the class configuration of the identification target is special in that it is a hierarchical multi-class configuration. The identification object is an object for which the class to which the identification object belongs is determined. The hierarchical multi-class configuration is, for example, a multi-class configuration in which the entire identification target is classified into a positive class and a negative class, and the positive class is further divided into a plurality of subclasses (FIG. 8). reference). In the following description, each positive class (that is, the above-described subclass) is also referred to as a “positive class”, a “positive class”, or an “object class”. The positive class is also referred to as “negative class”, “negative class”, or “non-object class”.

  FIG. 8 is a diagram schematically illustrating an example of identification objects classified into hierarchical multi-classes. In FIG. 8, identification objects included in the same class are represented by the same graphic. The position of the figure schematically represents the position in the feature space of the feature quantity extracted from the image to be identified represented by the figure. The feature space represents a feature amount space. In the example illustrated in FIG. 8, the identification target is classified into one of a positive class that is a class of a detected object and a negative class that is a target class that is not a detected object. In the example shown in FIG. 8, the positive class is further divided into three subclasses (positive class). The three positive classes are named “Pos1”, “Pos2”, and “Pos3”. When the object to be detected is a face, for example, a front face is assigned to Pos1, an oblique face is assigned to Pos2, and a side face is assigned to Pos3. Assigned. In addition, when the detection target is a face, a human front face may be assigned to Pos1, and a doll front face may be assigned to Pos2. Two types of accuracy can be considered to evaluate such a method. One is accuracy (hereinafter referred to as first accuracy) that can detect an object regardless of appearance. The first accuracy relates to the ability to discriminate between the entire positive class and the negative class. The other is accuracy (hereinafter referred to as second accuracy) that can correctly classify the appearance of the object. The second accuracy relates to the ability to discriminate between positive classes. In the object detection, it is essential that the first accuracy described above is high. The second accuracy is also important depending on the application. Note that the boundary between the positive class and the negative class shown in FIG. 8 conceptually indicates the boundary between the object and the non-object.

  In Non-Patent Document 4, learning is performed so that each positive class and negative class can be classified, and the score of each positive class is output. As a result, hierarchical class classification is possible. Further, by setting a part of the label of the positive class as a label that is neither positive nor negative during learning, it is possible to give priority to the discrimination between the entire positive class and the negative class. However, the priority is fixed and cannot be adjusted arbitrarily.

  Next, for the convenience of the following description, a multi-class identification processing framework using the identification dictionary created according to the present embodiment will be described first. Information on a number of learned weak classifiers is stored in the identification dictionary. By executing these weak classifiers one by one in order, an overall discrimination result is obtained. As described above, the process combining the weak classifiers is called a strong classifier in the field of machine learning / pattern recognition. This framework is a variant of the framework called Boosting. Each weak classifier supports one or more positive classes. The weak classifier outputs two types of scores for the input data, that is, a positive class score supported by the weak classifier and a positive class score not supported by the weak classifier. It may be designed as follows. In that case, the score value calculated by one weak classifier is the same value for the positive class supported by the weak classifier, and another same value for the remaining positive class. is there. In other words, each weak classifier outputs the same value as the score of the positive class supported by the weak classifier with respect to the input data. Each weak classifier further outputs another identical value as the score of a class other than the positive class supported by the weak classifier for the input data. The score that the weak classifier outputs as the score of the positive class supported by the weak classifier is, for example, the identification target represented by the input data belongs to one of the positive class supported by the weak classifier Represents the degree of possibility. The score output by the weak classifier as a score of a class other than the positive class supported by the weak classifier is, for example, any of the positive classes supported by the weak classifier represented by the input data. The degree of possibility that it does not belong to.

Then, the cumulative sum for each positive class of the score output from each weak classifier is calculated. The calculated cumulative sum is the output (strong discrimination score) of the strong discriminator. That is, a strong discrimination score is obtained for each positive class. Therefore, the input data can be classified into a positive class that takes the maximum value of the strong discrimination score. In the following, for the sake of brevity the positive type classes each weak classifier supports consists of one or more positive system class, a set of positive system classes, referred to as a single set S _n ( n = 1, ..., number of sets).

The output of the strong classifier is calculated using the equation shown in Equation 1. In the formula shown in Equation 1, c represents a number assigned to the positive class. H ^c and h _m ^c represent the output value of the strong discriminator and the output value of the m-th weak discriminator, respectively, indicating the likelihood of a positive class c. N _M is the number of weak classifiers, m is a weak classifier ID. The strong classifier outputs an output value (H ^c ) individually for each positive class c. Each weak classifier m also outputs an output value (h _m ^c ) individually for each positive class c. Output for all positive classes (c = 1, ..., N _C ). N _C is the number of positive classes (number of positive classes). That is, several positive class numerical values are output.

Note that when performing the classification process using such a strong classifier, it is not always necessary to calculate the output values of all weak classifiers. Intermediate output obtained by calculating the output value of only a part of weak classifiers (for example, weak classifiers whose weak classifier number m is 1 to N _M '(N _M '<N _M )) The value may be retrieved. For example, in the process of sequentially calculating the output value of the weak classifier, if the input to the classifier can be determined to be a negative class by thresholding the extracted intermediate output value, the remaining calculations can be omitted. In particular, when performing object detection, the probability that the input to the discriminator is a negative class (non-object class) is very high. Therefore, if it can be determined that the input to the discriminator is a negative class at a stage where a few weak discriminators are being calculated, the calculation cost becomes very low, and thus high-speed object detection is possible.

<First Embodiment>
Next, the identification dictionary learning system 100 which is the 1st Embodiment of this invention is demonstrated with reference to an accompanying drawing.

  FIG. 1 is a block diagram showing an example of the configuration of an identification dictionary learning system 100 according to the first embodiment of the present invention.

  Referring to FIG. 1, an identification dictionary learning system 100 according to the present embodiment includes an identification dictionary learning device 110 and a storage device 120.

  The identification dictionary learning device 110 includes an initialization unit 111, a learning unit 112, an index calculation unit 113, a determination unit 114, a weight update unit 115, and a coefficient update unit 116. The identification dictionary learning device 110 may further include a coefficient storage unit 121 and a weight storage unit 123. The identification dictionary learning device 110 may further include a sample storage unit 124.

  The storage device 120 includes an identification dictionary storage unit 122. The storage device 120 instead of the identification dictionary learning device 110 may further include a coefficient storage unit 121 and a weight storage unit 123. The storage device 120 instead of the identification dictionary learning device 110 may further include a sample storage unit 124. The identification dictionary learning device 110 may include a storage device 120.

  The following data is stored in each storage unit described above.

  The coefficient storage unit 121 stores a numerical value group of optimization index coefficients used during the operation of the identification dictionary learning device 110 of the present embodiment. The optimization index coefficient is a numerical value of 0.0 or more. The coefficient storage unit 121 may store a numerical value group of optimization index coefficients as an optimization index coefficient table in the form of a table. In this optimization index coefficient table, optimization index coefficients applied to a combination of a class to which a learning sample belongs and a positive class into which the learning sample is classified are stored. One optimization index coefficient is stored for each combination. The learning sample belongs to one of a plurality of positive class and negative class. Therefore, the number of optimization index coefficients in the table is “total number of classes × number of positive classes”. The total number of classes is the sum of the number of positive classes and the number of negative classes. The number of negative classes is always 1. The value of the optimization index coefficient stored in this optimization index coefficient table changes according to the repetition of the learning process. That is, the value of the optimization index coefficient changes according to the number of repetitions t of the learning process.

  The weight storage unit 123 stores a numerical value representing the weight of learning data (learning sample) used during the operation of the identification dictionary learning device 110 of the present embodiment. The weight stored in the weight storage unit 123 is also expressed as “sample weight”. This weight is set for each positive learning class for each learning sample. The weight storage unit 123 stores the weights of several positive classes instead of one weight for one learning sample. For this reason, the number of weights stored in the weight storage unit 123 is “the number of all learning samples × the number of positive classes”. This weight represents the likelihood of classification failure for each positive class. The more the learning sample that should be learned with greater importance, the greater the weight set for the learning sample. An example of sample weight is shown in FIG.

  FIG. 2 is a diagram schematically illustrating an example of weights stored in the weight storage unit 123. In the example shown in FIG. 2, there are three positive classes. These positive classes are represented by “Pos1”, “Pos2”, and “Pos3”. The weight storage unit 123 stores the weights set for the learning samples to which the sample numbers 1 to 8 are assigned. The second line in FIG. 2 represents the class to which the learning sample belongs.

  The identification dictionary storage unit 122 stores an identification dictionary that is a parameter of a strong classifier obtained by executing the present invention. The parameters of the strong classifier obtained by the identification dictionary learning device 110 are stored in the identification dictionary storage unit 122 as an identification dictionary. A strong classifier is a set of weak classifiers. Therefore, the parameters of the strong classifier are substantially equivalent to the parameters of a plurality of weak classifiers. The identification dictionary storage unit 122 stores, for example, the following as parameters for one weak classifier.

(1) Information specifying the type of feature quantity selected as the feature quantity used for the weak classifier;
(2) a look-up table that can be used to calculate an output value of a weak classifier to be described later, based on a feature value of a learning sample
(3) information weak classifiers to identify the set S _n of positive system classes to support.

  The feature amount is a value representing the feature of the learning sample. The learning sample is a data value obtained by observing the identification target. The learning sample may include a feature amount. The learning sample may be a feature amount. In the following description, the feature amount is also expressed as “feature value” or “feature value”. The feature amount is, for example, color, pixel value, gradient (primary differentiation of pixel value), secondary differentiation of pixel value, or the like. The information that specifies the type of feature quantity is, for example, a data value that specifies which type of feature quantity has been selected as the feature quantity used in the weak classifier. For example, when a plurality of types of feature quantities are extracted from the learning sample and a serial number is assigned to each of the extracted feature quantity types, the data value specifying the type of feature quantity is the serial number or the like. May be. The data value that specifies the type of feature quantity may be a parameter used to extract the feature quantity. The lookup table described above is a reference table that can be used to calculate the output value of the weak classifier based on the feature amount of the learning sample. The identification dictionary storage unit 122 stores, for example, the parameters used to calculate the output value based on the feature amount according to the type of feature amount and the set of positive classes as the above-described lookup table. If you do. A detailed description of the lookup table will be described later. Instead of the lookup table, the identification dictionary storage unit 122 looks up the parameters used to calculate the output value based on the feature amount according to the type of feature amount and the set of positive classes. You may memorize | store in the form which is not a table.

  The positive class supported by the weak classifier is represented by a set Sn of positive class, for example. The information specifying the positive class supported by the weak classifier is represented by, for example, a data value specifying the set Sn of the positive class. The information specifying the positive class supported by the weak classifier may be, for example, the set number n assigned to the set Sn of the positive class.

As described above, in the present embodiment, each weak classifier is based on input data, the score for the positive class supported by the weak classifier, and the class other than the positive class supported by the weak classifier. Calculate the score for and. In the present embodiment, the calculation formula for calculating the score for the positive class supported by the weak classifier is different from the calculation formula for calculating the score for a class other than the positive class supported by the weak classifier. In this embodiment, the weak classifier calculates a score for a class based on a different calculation formula depending on whether or not the class is included in Sn representing a positive class supported by the weak identifier. . Weak classifier is a part of the strong classifier which operates on the basis of the obtained identification dictionary by identifying dictionary learning device 110 according to this embodiment, when calculating the score, for example, S _n of the weak identifier is supported Information for identifying the information may be read from the identification dictionary storage unit 122. Then, the weak classifier is read from the identification dictionary storage unit 122, a calculation formula which is determined depending on whether the weak discriminator includes class Sn specified by the information specifying the S _n to support The score may be calculated for each class in accordance with the calculation method based on it. Each time the determination unit 114 repeatedly determines weak classifiers, the parameters related to the weak classifiers determined by the determination unit 114 are sequentially stored in the identification dictionary storage unit 122. When the operation of determining the weak classifier and storing the parameters by the identification dictionary learning apparatus 110 is completed, the identification dictionary in which all the parameters are recorded is completed.

  The sample storage unit 124 stores learning samples. The learning sample is, for example, a set of feature amounts extracted from known detection target images. The sample storage unit 124 may further store a label described later. For example, the administrator of the identification dictionary learning device 110 may store the learning sample in the sample storage unit 124 in advance. For example, the administrator of the identification dictionary learning device 110 may store the label in the sample storage unit 124 in advance.

  Next, in order to facilitate understanding of details of the present embodiment, label setting during learning will be described first.

  In the present embodiment, one learning sample belongs to one class. However, labels for the number of all positive classes are set for one learning sample. The label is used during learning, as will be described later. The value of the label is +1 or −1. The label is uniquely determined based on the information of the class to which the learning sample belongs. The label may be determined in advance for each of the learning samples. The label may be given in advance to each learning sample. And the label provided to each learning sample should just be stored in the sample memory | storage part 124 beforehand, for example.

FIG. 3 is a diagram schematically illustrating an example of a label stored in the sample storage unit 124. As in the example shown in FIG. 3, for example, +1 is set as the label of the class to which the learning sample belongs in the learning sample. Moreover, -1 is set to the learning sample as a label of a class to which the learning sample does not belong. In the example illustrated in FIG. 8, for example, the learning samples whose sample numbers are 1 and 2 belong to the class “Pos1” that is a positive class.
In the example shown in FIG. 3, the learning samples whose sample numbers are 1 and 2 belong to the class “Pos1”, which is one of the positive classes. For learning samples whose sample numbers are 1 and 2, a label of +1 is set as the label of the class “Pos1” to which these learning samples belong. The learning samples belonging to other positive classes (“Pos2” and “Pos3”) and the negative class (“Neg”) have a label of −1 as a label of the class “Pos1” to which those learning samples do not belong. Is set.

  The initialization unit 111 sets initial values of optimization index coefficients held in the optimization index coefficient table storage unit 121. As described above, one optimization index coefficient is set for the combination of the class to which the learning sample belongs and the positive class into which the learning sample is classified. That is, one optimization index coefficient is set for a combination of two classes. The combination of the two classes is a combination of two different positive classes, a combination of a negative class and a positive class, or a combination of the same positive classes. In the following description, an optimization index coefficient set for a combination of two different positive class is also referred to as “coefficient between different positive classes”. A combination of two positive classes is also referred to as “same set”. The optimization index coefficient set for the combination of the negative class and the positive class is also referred to as “a coefficient between the positive class and the negative class”. A combination of a positive class and a negative class is also referred to as “different set”. The optimization index coefficient set for the combination of the same positive class is also referred to as “coefficient between the same positive class”. The coefficient between different positive classes is the optimization index of the combination of the positive class to which the learning sample belongs and the positive class to which the learning sample is classified in the case of class classification error that classifies the learning sample into any class Represents the size of the contribution to The coefficient between different positive classes is set to a value smaller than the coefficient between the positive class and the negative class. For example, the initialization unit 111 may set 0.0 as the initial value of the coefficient between different positive classes. For example, the initialization unit 111 may set 1.0 as the coefficient between the positive class and the negative class. The initialization unit 111 may set, for example, 1.0 as the initial value of the coefficient between the same positive classes.

  The learning unit 112 supports one positive class set Sn as a weak classifier candidate, and learns (that is, creates) a classifier having one feature amount as an input. The learning unit 112 learns a classifier as a weak classifier candidate for each combination of the positive class set Sn and the feature amount. In the following description, learning a classifier as a weak classifier candidate is also referred to as “learning a weak classifier”. The learning unit 112 repeats the learning of the discriminator the same number of times as the number of combinations of the positive class set Sn and the feature amount, so that the same number as the number of combinations of the positive class set Sn and the feature amount is obtained. A weak classifier candidate is created. As described above, weak classifier learning is to create a classifier as a candidate for a weak classifier. Specifically, the learning of the weak classifier is, for example, obtaining a parameter of the classifier used for calculating an output value based on the feature amount of the input learning sample. As the weak classifier of this embodiment, various classifiers that output a value for each positive class can be employed. The weak classifier of the present embodiment may be a classifier that calculates an output value according to a calculation formula shown in the following formula 2, for example. In the following, learning of weak classifier candidates based on the calculation formula shown in Equation 2 will be described.

In the formula shown in Equation 2, c (c = 1,..., N _C ) is a positive class number, and j (j = 1,..., N _J ) is a value that the feature quantity can take. This is the number of the section (bin) into which the range is divided, and h _j ^c is the output value of the weak classifier. N _C is the number of positive classes and N _J is the number of bins. The number of positive system class, determined in accordance with the set S _n of positive system class. The bin configuration may be determined in advance according to the type of feature amount. The number N _J of bins may be different depending on the type of feature amount. If the input feature quantity is included in the bin numbered j, the weak classifier outputs h _j ^c as an output value for the positive class c. The function δ (·) is a function whose value is 1 when the mathematical expression given as an argument is true, and whose value is 0 when the mathematical expression given as the argument is false. The symbol i is the number of the training sample, z _i ^c is a label for the positive class c assigned to the training sample i, and x _i is a feature of the training sample i that is input to the weak classifier Represents an amount. D _i ^c is a weight related to class c of the i-th learning sample. X _j represents the section (bin) in which the range of values that can be taken by the feature amount is divided, the number being j. As described above, S _n represents a set of classes whose number is n, and ε is a small constant value. In the equation 2 and the equation 3 described later, the symbol “·” between the weight and the function δ (•) and between the two functions δ (•) is an operator representing multiplication.

The learning unit 112 calculates h _j ^c according to the equation shown in Equation 2 using the given learning sample (that is, for example, all the learning samples stored in the sample storage unit 124). In that case, the learning unit 112, the output value h _j ^c about positive system classes included in the set S _n of positive system class is derived in accordance with the number 2 of the upper equation. As a result, h _j ^c is obtained for each bin. Output value h _j ^c of the learning unit 112 is calculated, it is common in positive systems classes included in the set S _n of positive system class. Thus, for positive-based classes included in the set S _n of positive-based classes, N _J-number of h _j ^c is obtained. Output value h _j ^c about positive system classes included in the set S _n of positive system class does not depend on the class, in the following description, it is referred to as "h _j". In the following description, the output value h _j ^c is also expressed as “score” or “score h _j ^c ”. Also, the learning unit 112, according to the equations shown in the lower part of the number 2, for each of the positive type classes that are not included in the set S _n of positive system classes, calculates a score h _j ^c. Score h _j ^c Positive system classes that are not included in the set S _n is common in each bin. Therefore, the positive type each class that is not included in the set S _n, 1 single score h _j ^c is obtained. As a score h _j ^c Positive system classes that are not included in the set S _n, score h _j ^c in the same number as the number of positive system classes that are not included in the set S _n is obtained. Output value h _j ^c about positive system classes that are not included in the set S _n does not depend on the characteristics of the section (not dependent on the bottle), in the following description, are referred to as "h ^c".

As described above, the combination of the set S _n and types of features, using the learning samples to calculate the h _j ^c, above, as a candidate of the weak classifier, learned classifier (i.e. Generation). The learned discriminator is represented by information for specifying the type of feature quantity, information for specifying the set Sn, and the above-described calculated h _j ^c .

Learning unit 112, the type of any of the feature amount obtained in the learning samples, for each of combinations of a set S _n of positive-based classes, as described above, for example according to the formula shown in Equation 2, a h _j ^c calculate. That is, the learning unit 112, while sequentially selecting the combinations of the set S _n type and positive type class of the feature amount, the calculation of the score h _j ^c, all possible combinations are selected for the combination selected Repeat until. As described later, the index calculation unit 113, and any kind of feature amount obtained in the learning samples, for each of combinations of a set S _n of positive-based classes, using a score h _j ^c, Calculate optimization metrics. Furthermore, as will be described later, determination unit 114, the calculated optimized index is the smallest, selecting a combination of said quantity and the set S _n. Then, determination unit 114, information specifying the type of the selected feature quantity, and information specifying a set S _n selected, h _j ^c calculated for the combination of the set S _n and the type of the selected characteristic quantity Are stored in the identification dictionary storage unit 122 as parameters representing a weak classifier. At that time, determination unit 114, for positive system classes that are included in the set S _n, the N _J pieces, a score h _j ^c associated with the number j of the bin, respectively, for example as part of the above-mentioned look-up table And stored in the identification dictionary storage unit 122. Determination unit 114, In addition, for each of the positive type classes that are not included in the set S _n, a score h _j ^c associated with the number c of the positive type class, for example, as part of the above-mentioned look-up table, identification dictionary What is necessary is just to store in the memory | storage part 122. FIG.

For reference, the weak classification in the identification process using the result of learning of the identification dictionary by the identification dictionary learning device 110 of the present embodiment will be described below. When deriving the score of the positive class included in the class set Sn, the device operating as a weak classifier is derived based on a sample obtained by observing the identification target (that is, the above-described identification data). Identify the bin that contains. The apparatus specifies the score calculated based on the upper expression shown in Formula 2 that is associated with the specified bin number j by referring to the reference table (lookup table) described above, for example. . The score is a score of each positive class included in the class set Sn, which is a result of the weak identification of the sample. The device, a score positive system classes that are not included in the S _n, for example, a score calculated based on the lower of the formula shown in Equation 2, may be specified in the above reference table. As described above, in the identification process, by referring to the reference table stored in the identification dictionary storage unit 122 at the time of learning, the score of each positive class can be obtained as a result of one weak identification.

Below, the example of the specific process which calculates a score based on the formula shown in Formula 2 is demonstrated in detail. Deriving a score for positive system classes that are included in the set S _n, the learning unit 112, the frequency of occurrence of the feature amount, by counting with a sample weight, and a histogram of the histogram and label -1 label +1 create. "Histogram of the label +1" is a histogram representing the feature amount of the learning samples labeled relates positive system class is +1 in the set S _n, the weighted frequency of occurrence of each section of the feature. The histogram with label +1 is also expressed as “weighted histogram with label +1”. "Histogram Labels -1" is a histogram representing the feature amount of the learning samples labeled relates positive system class is -1 in the set S _n, the weighted frequency of occurrence of each section of the feature. The histogram with label-1 is also referred to as “weighted histogram with label-1”. The weighted appearance frequency is the sum of the weights assigned to the learning samples in which the feature amount is included in the feature amount section for each feature amount section. When creating the histogram, the learning unit 112 reads the weights assigned to the learning samples from the weight storage unit 123. The learning unit 112 makes the setting of the feature amount sections the same in these two histograms. When the learning unit 112 creates these two histograms, the learning unit 112 divides the feature amount into the same range. Further, the learning unit 112 uses the logarithm of the ratio of the appearance frequencies counted in the bins having the same feature amount range in these two histograms as an output value when the feature amount to be identified is included in the bin. calculate. The learning unit 112 calculates output values (that is, scores) of all feature amount sections (for example, bin numbers j are 1 to N _J ). Deriving a score positive system classes that are not included in the set S _n, the learning unit 112 calculates the weighted number of learning samples labeled is +1, the weighted number of learning samples label is -1 . The learning unit 112 further calculates a logarithm of the ratio of the weighted numbers calculated for each of the two types of labels as an output value. The score of the positive class not included in the set Sn does not depend on the feature amount interval.

4A and 4B are diagrams illustrating examples of the above-described histogram. FIG. 4A is a weighted histogram with label +1. FIG. 4B is a weighted histogram for label-1. A normal histogram is calculated by adding 1 for each sample, whereas a weighted histogram is a histogram calculated by adding the weight values assigned to that sample for each sample. is there. The vertical axes of the histograms shown in FIGS. 4A and 4B represent weighted appearance frequencies. The horizontal axes of the histograms shown in FIGS. 4A and 4B represent feature values, that is, feature amounts. For example, when the feature value is an integer value from 0 to 255 and the bin width of the histogram is 5, the feature value interval is 0 to 4, 5 to 9, 245 to 249, 250 or more. Set to The number j of the feature amount section is given in ascending order according to the size of the feature amount, for example. In the above example, the feature of the section from the smaller size of the feature _{quantity, X 1, X 2, X} 3, ..., X j, ..., are represented as X _J. The bin number is set to the same number as the section of the feature amount represented by the bin.

The learning unit 112 may calculate the output value h _j by calculating the logarithm of the ratio of the appearance frequency values of bins having the same number j in the histograms as shown in FIGS. 4A and 4B, for example. The output value h _j ^c calculated according to Equation 2 is larger as the appearance frequency of the learning sample whose label is +1 is higher and the appearance frequency of the learning sample whose label is −1 is lower. In other words, if the characteristic quantity of a sample is included in the section j, the larger the value of the output values h _j ^c about positive system classes that are included in the set S _n, the sample is either included in the set Sn It can be said that there is a high probability of belonging to a positive class.

  As a supplement, the meaning of Equation 2 will be described in more detail. The upper equation of Equation 2 calculates the ratio between the probability that the learning sample belongs to the label + 1 class and the probability that the learning sample belongs to the class of the label-1 under the condition that the feature amount of the learning sample is obtained. It corresponds to the formula. Then, the value obtained by calculating the upper expression of Equation 2 is higher for the learning sample that seems to belong to the label + 1 class. Note that “label + 1 class” represents a positive class in which the label assigned to the learning sample is +1. “Label-1 class” represents a class in which the label assigned to the learning sample is −1. In the following, the reason why the upper equation of Equation 2 can be interpreted as a probability ratio will be described. By dividing the weighted appearance frequency in the bin of the histogram with the label +1 described above by the sum of the weights of the learning samples, the sample with the label +1 takes the feature value included in the section to which the bin is assigned. A likelihood is obtained. Similarly, by dividing the weighted appearance frequency in the bin of the histogram of the label-1 described above by the sum of the weights of the learning samples, the sample with the label of -1 is included in the section to which the bin is assigned. The likelihood that a characteristic value can be obtained is obtained. Assuming that the prior probabilities are uniform for all labels, the following holds from the Bayesian formula of statistics: By dividing the likelihood obtained for the bin assigned to the section containing the feature value of the learning sample in the histogram of label +1 by the sum of the likelihoods for that bin, the learning sample is in the class of label +1. Probability that seems to belong to. Similarly, the learning sample is obtained by dividing the likelihood obtained for the bin assigned to the section including the feature amount of the learning sample in the histogram of label-1 by the sum of the likelihoods for the bin. The probability that it belongs to the class of label-1 is obtained. Further, when the ratio of these probabilities is calculated, the above-described division by the sum of likelihoods in the denominator side and the numerator side is canceled in the equation representing the ratio of these probabilities. Therefore, the expression expressing the ratio of these probabilities is expressed by Equation 2.

  In practice, when the number and variation of learning samples are not sufficient, the weighted appearance frequency may become numerically zero. By interpolating or smoothing from adjacent bin values, the risk of zero values can be reduced.

The weak classifier of the present embodiment may be a classifier that calculates an output value according to the calculation formula shown in Formula 3 instead of Formula 2 described above. Equation 2 corresponds to calculating the logarithm of the probability ratio, while Equation 3 corresponds to calculating the difference in probability. Similar to the output value h _j ^c shown in Equation 2, the output value h _j ^c shown in _Equation ³ also has a higher appearance frequency of the sample whose label is +1 and a lower appearance frequency of the sample whose label is −1. large. When the feature amount of the sample is included in the section j, if the value of the output value h _j ^c is large, it can be said that the probability that the label relating to the class c of the sample is +1 is high.

The index calculation unit 113 calculates an optimization index related to the weak classifier candidate based on the output value of the weak classifier candidate. The index calculation unit 113 calculates an optimization index for each candidate weak classifier. As described above, a feature quantity to be used as input for one combination of the set S _n of positive system classes supported, the candidate one weak classifier is learned. That is, the candidate weak classifiers comprises a feature quantity to be used as an input, the combination of the set S _n of positive system classes supported, is uniquely associated. The optimization index is an index representing the degree of performance of the weak classifier candidate obtained by learning. Equation 4 is an expression representing an example of the optimization index. The index calculation unit 113 calculates the optimization index L according to Equation 4, for example. The optimization index corresponds to an index called loss, loss of cross entropy, or the like in the field of pattern recognition. When the optimization index L is calculated by Equation 4, the smaller the optimization index L, the higher the identification accuracy.

In Equation 4, as described above, c is a positive class number and i is a learning sample number. N _C is the number of positive classes, and N _x is the number of learning samples. The value h _i ^c in Equation 4 represents the output value h _j ^c by the weak classifier candidate for the bin j assigned to the section including the feature quantity of the sample i. R _i ^c is an optimization index coefficient. The optimization index coefficient represents, for example, the magnitude of the influence that “a weak classifier candidate classifies the learning sample i into class c” has on the optimization index of the weak classifier candidate. The optimization index coefficient R _i ^c is stored in the coefficient storage unit 121. In this embodiment, the optimization index coefficient is determined by the class to which the learning sample i belongs and the positive class c into which the learning sample i is classified. In the following description, the class to which the learning sample belongs is denoted as “correct class”. Furthermore, the class into which the learning sample is classified is denoted as “classification class” or “classification result class”. Coefficient storage unit 121, for each combination of the correct class and classification class, it is sufficient to store the optimization metrics coefficients R _i ^c associated with the combination of the correct class and classification class. Coefficient storage unit 121, for example in the form of optimization metrics coefficient table, it is sufficient to store the optimization metrics coefficients R _i ^c.

  5A and 5B are diagrams illustrating examples of the above-described optimization index coefficient table. FIG. 5A shows an example of initial values of the optimization index coefficient table. The table shown in FIG. 5B is an example of an optimized index coefficient table updated as described later. FIG. 5B shows an example of the maximum value of each optimization index coefficient included in the optimization index coefficient table.

When calculating the optimization index L, the index calculation unit 113 calculates the optimization index coefficient R _i ^c in the optimization index coefficient table based on, for example, the optimization index coefficient table, the class to which the learning sample belongs, and the class ^c. Should be specified. The index calculation unit 113 may read the specified optimization index coefficient from the coefficient storage unit 121 and calculate the optimization index using the read optimization index coefficient.

Index calculation unit 113, for example, a combination of a set S _n of feature amounts and positive system classes, may be selected one combination. Then, the index calculation unit 113 may calculate the optimization index L for the selected combination. Index calculation unit 113, until all combinations of the set S _n of feature amounts and positive system class is selected, the selection of the combination, and may be repeated to calculate the optimization index L. As described above, the optimization index L is calculated for each of the weak classifier candidates.

The determination unit 114 identifies a combination of the positive class set _Sn and the feature quantity that has the smallest optimization index value L calculated by the conversion index calculation unit 113. Determination unit 114, an input feature amount included in the specified combination, the candidate weak classifiers that supports a set S _n of positive system classes included in the identified combination is determined as a weak classifier. In the following description, a weak classifier candidate (classifier) determined as a weak classifier is also referred to as a “determined weak classifier”. The determination unit 114 adds the determined weak classifier to an identification dictionary representing a set of a plurality of weak classifiers constituting the strong classifier. That is, the determination unit 114 stores the parameters of the weak classifier in the identification dictionary storage unit 122. As described above, the parameters of the weak classifier include, as described above, information specifying the type of feature quantity, parameters used to calculate an output value based on the feature quantity, and a set of positive classes. Information specifying S _n . As described above, the determination unit 114 may store the parameters used to calculate the output value based on the feature amount in the identification dictionary storage unit 122 as a lookup table. When the determination unit 114 stores the parameters of the weak classifier in the identification dictionary storage unit 122, the determination unit 114 does not overwrite the parameters of the weak classifier stored in the identification dictionary storage unit 122, and the parameters of the weak classifier are identified. It is added in the storage unit 122.

The weight update unit 115 updates the weight assigned to the learning sample stored in the weight storage unit 123 using the output value h _i ^c output from the weak classifier determined by the determination unit 114. For example, the weight update unit 114 may update the weight given to each learning sample according to Equation 5. Specifically, when updating the weight of the learning sample i, the weight updating unit 115 calculates the output value h of the weak classifier determined by the determination unit 114, which is calculated based on the feature amount of the learning sample i, for example. _j ^c is calculated. The weight update unit 115 gives the calculated h _j ^c , the weight D _i ^c of the learning sample i stored in the weight storage unit 123, and the learning sample i stored in the sample storage unit 124, for example. The value of the left side of Equation 5 is calculated based on the label z _i ^c that has been set. Weight updating unit 115 is stored in the weight storage unit 123, a weighting D _i ^c of the learning sample i, was calculated is overwritten by the value of the left-hand side of Equation 5. The symbol e in Equation 4 is the base of the natural logarithm (ie, the Napier number).

  The coefficient updating unit 116 updates the optimization index coefficient held in the coefficient storage unit 121 so that, for example, the following condition is satisfied. First, the first condition is that the coefficient between different positive classes does not exceed the coefficient between the positive class and the negative class. The “coefficient between different positive classes” is an optimization index coefficient in which the associated “correct class” and “classification class” are two different positive classes. “Coefficient between positive class and negative class” is an optimization index coefficient associated with a correct class that is a negative class and a classified class that is a positive class. Furthermore, the second condition is that the coefficient between different positive classes increases monotonously with each update by the coefficient updating unit 116. The coefficient updating unit 116 may perform updating so that the coefficient between different positive classes increases monotonously (that is, monotonously non-decrease) in a broad sense. The coefficient updating unit 116 may increase the coefficient between different positive classes according to, for example, a predetermined rule (that is, rule). The user of the identification dictionary learning apparatus 110 may be able to set an amount (a magnitude of the update) by which the coefficient between different positive classes increases in one update.

  FIG. 6 is a diagram schematically illustrating a change in the optimization index coefficient associated with the update. The horizontal axis of the graph shown in FIG. 6 represents the number of updates (that is, the number of learning repetitions t). The vertical axis of the graph shown in FIG. 6 represents the value of the optimization index coefficient. The broken line in the graph shown in FIG. 6 represents the transition of the optimization index coefficient between the positive class and the negative class. In the graph shown in FIG. 6, a curve drawn by a solid line represents the transition of the optimization index coefficient between different positive classes. The initial value of the optimization index coefficient between different positive classes may be 0, for example. As shown in FIG. 6, the coefficient updating unit 116 updates the optimization index coefficient between the different positive classes so that the optimization index coefficient between the different positive classes gradually approaches the limit value, as shown in FIG. Also good. The coefficient updating unit 116 may update the optimization index coefficient between the different positive classes so that the optimization index coefficient between the different positive classes reaches the maximum value by repeating learning. The example shown in FIG. 6 represents an example of transition of optimization index coefficients between different positive classes when the coefficient updating unit 116 updates the coefficients between positive classes so as to monotonically increase according to the sigmoid function. The sigmoid function is a function represented by the following equation 6, for example.

  The coefficient updating unit 116 may increase t by a predetermined amount according to repetition of learning (that is, the number of updates). The coefficient updating unit 116 may calculate an optimization index coefficient using the t. For example, the user may set the degree of increase by increasing or decreasing a and b in Equation 6 independently, for example. The optimization index coefficient between different positive classes may be the same regardless of the combination of two different positive classes. The optimization index coefficient between different positive classes may be different depending on the combination of two different positive classes. For example, when each positive class is assigned to a different posture of an object (for example, a face), the larger the difference in the angle of the posture of the object between two different positive classes, the larger the optimization index coefficient. It only has to be assigned. In that case, learning can be performed so that a large erroneous classification of the face posture is unlikely to occur. For example, three curves drawn by solid lines in FIG. 6 represent an example of transition of coefficients according to the angle difference of the face posture.

  The update setting unit 117 receives, from the terminal device 140, for example, a rule (that is, a rule) that increases the coefficient between different positive classes. The update setting unit 117 may receive a parameter representing a rule that increases the coefficient between different positive classes from the terminal device 140. The parameter representing the rule that increases the coefficient between different positive classes is, for example, at least one of a and b, which are the parameters of the sigmoid function described above. The user of the identification dictionary learning device 110 may input parameters representing rules that increase the coefficient between different positive classes by operating the terminal device 140.

  Next, the operation of the identification dictionary learning device 110 of this embodiment will be described in detail with reference to the drawings.

  FIG. 7 is a flowchart illustrating an example of the operation of the identification dictionary learning device 110.

In the following description, as described above, N _C represents the number of positive classes, and N _f represents the number of features. In each embodiment of the present invention, the feature amount is not limited to a specific feature amount.

First, the initialization unit 111 initializes an optimization index coefficient table (step S111). The initialization unit 111 may set, for example, 0 as an initial value of the optimization index coefficient associated with a different combination of positive classes as in the example illustrated in FIG. 5A. The initialization unit 111 may set a sufficiently small value (that is, a value that is substantially 0) as an initial value of an optimization index coefficient associated with a combination of different positive classes. For example, the initialization unit 111 may set 1 as an initial value of another optimization index coefficient. Occurs at a later stage, for the set S _n of all positive-based classes available, those optimization index for later good are searched. Information specifying all S _n may also be given in advance to the learning unit 112. Learning unit 112, learning samples stored in the sample storage unit 124 based on the belonging class may identify all S _n.

Then, the learning unit 112, from a set of positive type class, not selected, selects one set S _n of positive type class (step S112-1). The learning unit 112 further selects one type of feature value that has not been selected for the set Sn selected in step S112-1 from all of the predetermined feature value types (step S112-2). . Then, as described above, the learning unit 112 learns weak classifier candidates (step S112-3).

  Next, the optimization index calculation unit 113 calculates the optimization index L related to the weak classifier candidate learned in step S112-3 (step S113-1).

For example, the learning unit 112 determines whether all feature amounts have been selected (step S113-2). If all the feature values have not been selected (No in step S113-2), the operations from step S112-2 to step S113-2 are repeated. If all feature amounts have been selected (Yes in step S113-2), for example, the learning unit 112 determines whether all sets Sn have been selected (step S113-3). If there is a set S _n which are not selected (in step S113-3 No), the operation from step S112-1 to step S113-3 is repeated. Operation from step S112-1 to step S113-3, until all of the set S _n is selected (Yes in step S113-3) are repeated.

In the above operation, the operation from step S112-2 to S113-2 is repeated as many times as the number N _f of the feature. Repeating operations in steps S112-2 to S113-2 is repeated more times as the number of sets S _n positive system class. Thus, a feature quantity, a positive type of weak classifier candidates associated with each combination of the set S _n classes, optimization index value L is calculated.

Further, the number of sets S _n is the number of combinations of any number that is selected so as not to overlap the N _C-number. Therefore, if positive system the number of classes is large, the number of sets S _n is a huge number. Thus, if the repeated operation of all the set S _n from step S112-1 to step S113-3, the number becomes huge number, especially repeats the operations of steps S112-3 and step S113-1. Therefore, instead of the above-described operation is repeated for all S _n, as follows, may be limited to search range of candidate weak classifiers. For example, first, the learning unit 112 may select one feature amount. Then, the learning unit 112 for the set S _n of one of the positive type classes, may be performed to learn the candidate weak classifiers. The index calculation unit 113 may calculate an optimization index for the weak classifier candidates learned by the learning unit 112. Learning unit 112 in the set S _n of one of the positive type classes, the most optimized index L candidates associated weak classifiers is good (small) may specify a set S _n. Learning unit 112, a class included in the specified set S _n, it may be determined as the classes that make up the set S _n. Learning unit 112, similarly, for a set S _n made definite class and another one positive system class by learning the candidate weak classifiers, described above, and specific set Sn, the set S _n What is necessary is just to repeat the determination of the class which comprises. In that case, the repeat count is approximately proportional to the number N _C of positive system class. Therefore, the identification dictionary learning device 110 can find an approximate solution (that is, a weak classifier candidate having an approximate minimum optimization index) with a small number of iterations.

  Next, the determination unit 114 determines the weak classifier candidate having the smallest optimization index value L obtained in the operation of step S114 as the weak classifier among the weak classifier candidates (step S114). . The determination unit 114 adds the determined weak classifier to the identification dictionary (that is, the determined set of weak classifiers) stored in the identification dictionary storage unit 122. The selection of the weak classifier candidate having the smallest optimization index value L is to select the combination Sn of the feature quantity type and the positive class having the smallest optimization index value L. The determined weak classifier outputs the same output value for each positive class included in the set Sn supported by the weak classifier based on the selected type of feature quantity. In other words, it can also be interpreted that the weak classifier (and the feature amount) are shared between the classes belonging to the positive class combination Sn. The determination unit 114 adds the adopted weak classifier parameters to the identification dictionary (that is, the determined set of weak classifier parameters) stored in the identification dictionary storage unit 122. Specifically, the determination unit 114 includes information for specifying the feature amount of the learning sample, a reference table used for obtaining the output value of the weak classifier, and a set of positive classes supported by the weak classifier. And added to the identification dictionary stored in the identification dictionary storage unit 122.

  Next, the weight update unit 115 updates the weights assigned to the learning samples by class (step S115).

  Next, the coefficient updating unit 116 updates the optimization index coefficient stored in the coefficient storage unit 121, for example, in the form of an optimization index coefficient table (step S116).

  Next, the learning unit 112 performs end determination for determining whether to end learning (step S117). The condition for determining that the learning unit 112 ends the learning process in the end determination in step S117 may be, for example, that the number of times of accommodation determination exceeds a predetermined number. The condition that the learning unit 112 determines to end the learning process is that the strong classifier optimization index L based on the determined weak classifier is the strong classifier optimization index to which the last determined weak classifier is not added. The amount of decrease from may be less than a predetermined threshold. The condition for determining that the learning unit 112 finishes the learning process is that the identification accuracy of the strong classifier for the evaluation data set, which is a set of evaluation samples prepared in advance and does not include learning data, It may be that a predetermined standard is reached. The learning unit 112 determines the accuracy of identification based on the absolute value or the identification error rate of the optimization index L obtained by causing the strong classifier based on the determined weak classifier to identify the evaluation data set. Also good. When the end of the learning process is not determined (No in step S117-1), the identification dictionary learning device 110 further repeats the operations from step S111 to S117-1. When the end of the learning process is determined (Yes in step S117-1), the operation of the identification dictionary learning device 110 illustrated in FIG. 7 ends.

  Next, the effect of this embodiment will be described.

  In this embodiment, it is possible to create an identification dictionary that can shorten the identification time while suppressing a decrease in identification accuracy as compared with a case where an object having a large appearance variation is a target of identification compared to a case where the appearance variation is small. There is an effect.

  The reason is that the coefficient updating unit 116 updates the optimization index coefficient as the weak classifier is added. The optimization index coefficient represents the magnitude of contribution to the optimization index of the combination of the class to which the learning sample belongs and the class into which the learning sample is classified. At the time of learning, the weak identifier with the highest identification accuracy is selected based on the optimization index. Therefore, the optimization index coefficient represents the degree to which the improvement of the accuracy of identification between the two classes is emphasized during learning in the combination of two classes to which the optimization index coefficient is given. In the process of sequentially selecting weak classifiers, the coefficient updating unit 116 can update the optimization index coefficients given to the two classes in accordance with the degree of importance placed on the accuracy of identification between the two classes. . Therefore, by updating the optimization index coefficient by the coefficient updating unit 116, for example, in the early stage, the accuracy of discriminating between the learning sample belonging to the negative class and the other learning sample is higher than that of other weak classifier candidates. High classifier candidates can be determined to be weak classifiers. Similarly, by updating the optimization index coefficient by the coefficient updating unit 116, at a later stage, weak discriminator candidates whose discrimination accuracy between different positive class is higher than other weak classifier candidates are weak classifiers. Can be determined. In the object detection process using the identification dictionary obtained as described above by the identification dictionary learning device 110 of the present embodiment, the following effects can be obtained by using the weak classifiers in the determined order. . In the early stage of identification processing, samples belonging to the negative class and other samples can be accurately identified. That is, it is possible to suppress a decrease in identification accuracy of an object having a large appearance variation compared to a case where the appearance variation is small. Furthermore, if the sample is identified as belonging to the negative class, the remaining identification processing can be omitted. Therefore, compared with the case where the remaining identification processing is not omitted, the processing for detecting the object (that is, the processing for determining whether or not the sample is a sample obtained by observing the object) is accelerated. Can do. That is, the time required for the identification process can be reduced as compared with the case where the identification process is not omitted. In the case described above, the identification accuracy of samples belonging to different positive classes is further improved as the identification process proceeds.

  With the techniques described in each of the above-mentioned patent documents and non-patent documents, the effect of the above-described embodiment cannot be obtained. With these techniques, it is difficult to set a priority level for which accuracy is to be learned and how much importance is assigned to the classification between the positive class and the negative class and the classification between the positive classes. Or in those techniques, the priority is fixed. As described above, for example, by setting the priority of the classification accuracy by the optimization index coefficient, the effect of the above-described embodiment can be obtained. When the priority of classification accuracy cannot be set, it is difficult to obtain the effect of the above-described embodiment.

  Below, the effect of this embodiment is demonstrated in more detail.

  The identification dictionary learning device 110 according to the present embodiment includes an initialization unit 111 that can set the priority of classification between a negative class and a positive class. The identification dictionary learning device 110 of the present embodiment further includes a coefficient updating unit 116 that changes the optimization index coefficient as the value of the learning repetition number m increases. The identification dictionary learning device 110 of the present embodiment further includes a learning unit 112 that can learn a strong classifier according to the updated optimization index coefficient, an index calculation unit 113, a determination unit 114, and a weight update unit 115. . Therefore, by using the identification dictionary learned by the identification dictionary learning device 110, a high detection rate and high-speed object detection process can be realized.

  Hereinafter, the reason why this effect is obtained will be described.

  In the present embodiment, by using the optimization index coefficient, it is possible to freely design the accuracy priority for the identification between the positive class and the negative class and the identification between each positive class. When the coefficient between different positive classes is set to a value smaller than the coefficient between positive class and negative class, the optimization index (loss) when misclassifying a positive class sample into another positive class Get smaller. In this case, learning is performed with an emphasis on discrimination between positive class and negative class.

  For example, by repeating the update so as to monotonously increase the coefficient between different positive classes, learning with an emphasis on identification between positive classes is performed as the number of updates increases. An apparatus that performs identification processing using the learning result may calculate the output value of the weak classifier in the order determined in the learning. Then, each time the device or the like calculates the output value of the weak classifier, the device or the like may update the output value of the strong classifier by adding the calculated output value of the weak classifier to the output value of the strong classifier. . Each time the device or the like updates the output value of the strong classifier, the device or the like may perform identification processing for identifying which class the sample belongs to based on the updated output value of the strong classifier.

  When learning and identification processing are performed in this way, at the beginning of the identification processing (for example, when m in Formula 1 is small), identification between the entire positive class and the negative class is prioritized. Therefore, when the identification process for identifying which class the sample belongs to is performed based on the output value of the strong classifier at the stage where m is small, the identification accuracy of the sample belonging to the negative class is increased. In this case, for example, the above-described apparatus that performs the identification process may determine whether or not the sample belongs to the negative class based on a predetermined criterion every time the output value of the strong classifier is updated. When the above-described apparatus determines that the sample to be identified belongs to the negative class, the remaining identification process may be omitted. By performing learning as described above, high accuracy can be obtained in the identification process for determining whether or not a sample belongs to the negative class even when m is small. Therefore, by designing the apparatus for performing the identification process as described above, the detection process can be greatly speeded up while suppressing a decrease in the detection rate. In addition, if the identification process is executed up to a stage where m is large, the accuracy of identification between the positive classes is also increased.

  In the present embodiment, for each combination of two positive classes, the priority of identification between positive classes can also be set. Therefore, it is possible to adjust the discrimination accuracy between the positive classes finally obtained. For example, if the face direction (front direction, diagonal direction, horizontal direction) is assigned as a positive class, the original face direction is a continuous change, so the boundary between the front direction and the diagonal direction is ambiguous. It is. Therefore, by adjusting the optimization index coefficient so that the priority between the front direction class and the diagonal direction class is lowered, a discriminator suitable for rough classification of the face direction can be realized. For example, when the discriminator is used to detect a person's looking aside while driving, determining that the face direction is diagonal when the face direction is the front direction is that the face direction is horizontal. In the case of the direction, the error is not so large as to determine that the face direction is the front direction. Therefore, there is an advantage in being able to adjust as described above.

<Second Embodiment>
Next, a second embodiment of the present invention will be described in detail with reference to the drawings. The object detection system according to the present embodiment is an example of an object detection system using an identification dictionary learned by the identification dictionary learning device 110 according to the first embodiment of the present invention.

  FIG. 9 is a block diagram showing the configuration of the object detection system 300 according to this embodiment.

  Referring to FIG. 9, the object detection system 300 according to the present embodiment includes an object detection device 310 and a storage device 120. The object detection system 300 may further include an identification dictionary learning device 110. The identification dictionary learning device 110 and the storage device 120 of the present embodiment are the same as the identification dictionary learning device 110 and the storage device 120 of the first embodiment. Therefore, the detailed description regarding the identification dictionary learning device 110 and the storage device 120 is omitted. Note that the identification dictionary storage unit 122 included in the storage device 120 stores the identification dictionary learned by the identification dictionary learning device 110. In the present embodiment, the object detection system 300 is also referred to as an identification system 300.

  The object detection device 310 includes a setting unit 311, an extraction unit 312, a weak identification unit 313, an identification score calculation unit 314, and a determination unit 315. In FIG. 9, the identification dictionary learning device 110 and the object detection device 310 are depicted as separate devices. However, the same device may operate as the identification dictionary learning device 110 and the object detection device 310. Further, the same device may include the storage device 120.

  An image obtained by photographing an object is input to the setting unit 311. The setting unit 311 receives the image. Then, the setting unit 311 sets a search window in a part of the received image including a part that is not set as the search window. The setting unit 311 may repeat setting of the search window while changing the part to be set as the search window until all parts of the received image are set as the search window at least once. The partial image included in the search window is the above-described sample.

  The extraction unit 312 extracts a feature amount from the partial image set in the search window. The extraction unit 312 may extract all feature amounts used by at least one of the weak classifiers. The feature amount used by each weak classifier is recorded in the identification dictionary stored in the identification dictionary storage unit 122. For example, the extraction unit 312 may select weak classifiers in order from the weak classifier having the determined order. Then, when the feature amount used by the selected weak classifier has not been extracted, the extraction unit 312 may extract the feature amount.

  The weak classifier 313 uses the extracted feature value to calculate the output value of the weak classifier recorded in the identification dictionary. The weak classifier 313 may calculate the output value of each weak classifier recorded in the identification dictionary using the extracted feature quantity. The weak classifier 313 may calculate the output value of the weak classifier selected by the extraction unit 312 among the weak classifiers recorded in the identification dictionary using the extracted feature amount.

  The score calculation unit 314 may calculate the output value of the strong classifier based on the calculated output value of each weak classifier. The score calculation unit 314 may calculate the output value (that is, the identification score) of the strong classifier based on the calculated output value of each weak classifier.

  The determination unit 315 determines whether the partial image included in the region of the image included in the search window is an object or a non-object. That is, the determination unit 315 determines whether the partial image included in the search window belongs to the negative class based on the identification score. If the determination unit 315 determines that the partial image included in the search window belongs to the negative class, the determination unit 315 determines that the identification regarding the search window is terminated.

  FIG. 10 is a flowchart showing an example of the operation of the object detection apparatus 310 of the present embodiment.

  First, the setting unit 311 sets a search window in a part of the image (step S311). The setting unit 311 sets, for example, a part of an image including a part not set as a search window, for example, having a predetermined size, as a search window. The setting unit 311 may set the search window while moving the search window so as to perform raster scan in the image, for example, in the repetition of the operations from step S311 to step S316-2 shown in FIG. The setting unit 311 may set the search window at a predetermined interval so that the search window does not overlap with the previously set search window. The setting unit 311 may set the search window at a predetermined interval so that the search window overlaps the previously set search window. In the following description, the object detection device 310 determines whether or not the partial image included in the set search window is an object image. The class of the image of the object is a positive class, and the class of an image that is not an image of the object (non-object image) is a negative class.

  The object detection device 310 identifies whether or not the image of the portion where the search window is set is a specific object through operations from step S312-1 to step S315 described later. In particular, step S312-3 and step S313 correspond to weak classification processing by one weak classifier. The operation of step S312-3 and step S313 for calculating the output value by the selected weak classifier is repeated while changing the selected weak classifier. The output value of the weak classifier obtained is sequentially added to the output value of the strong classifier in step S314. As a result, the output value (identification score) of the strong classifier is calculated. Equation 1 is a formula for calculating this identification score. Hereinafter, the above will be described in more detail.

  The extraction unit 312 selects the weak classifier that is determined in the earliest order among the unselected weak classifiers (step S312-1). The extraction unit 312 determines whether or not a feature amount that is an input of the selected weak classifier has been extracted (step S312-2). When the feature quantity that is the input of the selected weak classifier has been extracted (Yes in step S312-2), the extraction unit 312 does not have to extract the feature quantity. If the feature quantity that is the input of the selected weak classifier has not been extracted (No in step S312-2), the extraction unit 312 extracts the feature quantity (step S312-3). Information that identifies the type of feature quantity that is input to the weak classifier is stored in the identification dictionary storage unit 122.

  Next, the weak identification unit 313 calculates weak identification (step S313). In other words, the weak classifier 313 calculates the output value of the selected weak classifier. At this time, the weak classifier 313 is used to calculate the output value of the weak classifier stored in the identification dictionary storage unit 122 based on the feature quantity that is the input of the selected weak classifier. The output value of the selected weak classifier may be obtained using a lookup table. Note that the weak classifier 313 calculates an output value for each positive class as an output value of the weak classifier.

Next, the score calculation unit 314 adds the calculated output value of the selected weak classifier to the output value of the strong classifier (step S314). If all weak classifiers have been selected, in step S314, the sum (ie, the final identification score) H ^c shown in Equation 1 is calculated for each positive class c. If there is a weak classifier that has not been selected, in step S314, the intermediate result H ^c of the calculation of the identification score shown in Equation 1 is calculated for each positive class c. The intermediate result of the calculation of the identification score represents the sum of the output values of some of the weak classifiers.

Next, the determination unit 315 performs an abort determination to determine whether or not to terminate the identification process regarding the partial image included in the search window (step S315). Determination unit 315, based on the value of the calculated H ^c, partial images included in the search window is equal to or belongs to the negative class. For example, if the partial image included in the search window is likely to be an object image as the value of H ^c increases, the determination unit 315 determines that the partial image included in the search window is negative when the value of H ^c is smaller than a predetermined threshold. What is necessary is just to determine with belonging to a class. When it is determined that the partial image included in the search window belongs to the negative class, it is not necessary to continue the identification process. If the determination unit 315 determines that the partial image included in the search window belongs to the negative class, the determination unit 315 determines that the identification process related to the search window is to be terminated (ie, “canceled”) (Yes in step S316-1). In this case, if all the parts of the image are not set as search windows (No in step S316-2), the object detection device 310 performs the operation from step S311 again. That is, the processing by the object detection device 310 proceeds to processing for the next search window. If it is not determined to be “censored” (No in step S316-1), the object detection device 310 performs the operation from step S312-1 again. That is, the object detection device 310 performs calculation for the next weak classifier. As described above, H ^c is calculated for each class c. Therefore, in S315, the determination unit 315 may compare the maximum value _{c c} (H ^c ) of H ^c calculated in step S314 with a predetermined threshold value. Further, the threshold value used in S315 may be set to a value that can be determined experimentally beforehand for each class c so that the partial image clearly belongs to the negative class. In that case, the determination unit 315 may compare the maximum value _{c c} (H ^c ) of H ^c calculated in step S314 with a threshold experimentally obtained for class c. If the determination unit 315 determines that the partial image included in the search window belongs to the negative class as a result of the comparison, the determination unit 315 may determine that the identification process related to the search window is terminated.

In addition, when the determination unit 315 is not determined to be “censored” and the value of H ^c is larger than a threshold set in advance as a positive class threshold, the determination unit 315 includes a portion included in the search window. It may be determined that the image belongs to the positive class. That is, the determination unit 315 may determine that an object is detected in the set search window. When there are a plurality of positive classes, and when H ^c of a plurality of classes c exceeds the threshold of the positive class, the determination unit 315 determines that the partial image included in the search window has the largest H ^c value. What is necessary is just to determine with belonging to c. When there is no positive class in which H ^c exceeds the threshold of the positive class, the determination unit 315 may determine that the partial image included in the search window belongs to the negative class. That is, the determination unit 315 may determine that no object is detected in the search window.

<Third Embodiment>
Next, a third embodiment of the present invention will be described in detail with reference to the drawings. The third embodiment represents a concept common to the above-described embodiments.

  FIG. 11 is a block diagram showing the configuration of the identification dictionary learning device 100A of the present embodiment.

  Referring to FIG. 11, the identification dictionary learning device 100A of the present embodiment includes a plurality of classifiers that perform class classification that classifies data values belonging to any of a plurality of classes into any of the plurality of classes. The classification accuracy of the class classification of the combination of the learning unit 112 that learns using the learning sample that is the data value, the class to which the learning sample in the wrong class classification belongs, and the classified class An index calculation unit 113 that calculates the optimization index of each of the plurality of classifiers based on a plurality of optimization index coefficients each representing the magnitude of the contribution to the optimization index that represents Among the classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. A decision unit 114 that adds to the identification dictionary that is a set of weak classifiers, a coefficient update unit 116 that updates the optimization index coefficient based on a rule in accordance with the addition of the weak classifiers, Is provided.

<Other embodiments>
The identification dictionary learning system 100, the identification dictionary learning device 110, the identification dictionary learning device 100A, the storage device 120, the object detection system 300, and the object detection device 310 are respectively a computer and a program for controlling the computer, dedicated hardware, or It can be realized by a combination of a computer and a program for controlling the computer and dedicated hardware.

  FIG. 12 shows a hardware configuration of a computer 1000 that can realize the identification dictionary learning system 100, the identification dictionary learning device 110, the identification dictionary learning device 100A, the storage device 120, the object detection system 300, and the object detection device 310. It is a figure showing an example. Referring to FIG. 12, a computer 1000 includes a processor 1001, a memory 1002, a storage device 1003, and an I / O (Input / Output) interface 1004. The computer 1000 can access the recording medium 1005. The memory 1002 and the storage device 1003 are storage devices such as a RAM (Random Access Memory) and a hard disk, for example. The recording medium 1005 is, for example, a storage device such as a RAM or a hard disk, a ROM (Read Only Memory), or a portable recording medium. The storage device 1003 may be the recording medium 1005. The processor 1001 can read and write data and programs from and to the memory 1002 and the storage device 1003. The processor 1001 can access the terminal device 140, for example, via the I / O interface 1004. The processor 1001 can access the recording medium 1005. The recording medium 1005 stores a program that causes the computer 1000 to operate as the identification dictionary learning device 110, the identification dictionary learning device 100A, or the object detection device 310.

  The processor 1001 loads a program stored in the recording medium 1005 that causes the computer 1000 to operate as the identification dictionary learning device 110, the identification dictionary learning device 100A, or the object detection device 310 into the memory 1002. When the processor 1001 executes the program loaded in the memory 1002, the computer 1000 operates as the identification dictionary learning device 110, the identification dictionary learning device 100A, or the object detection device 310.

  Initialization unit 111, learning unit 112, index calculation unit 113, determination unit 114, weight update unit 115, coefficient update unit 116, update setting unit 117, setting unit 311, extraction unit 312, weak identification unit 313, score calculation unit 314 The determination unit 315 is realized by, for example, a dedicated program that can be read from a recording medium 1005 that stores the program into the memory 1002 and that can realize the function of each unit, and a processor 1001 that executes the program. Can do. The coefficient storage unit 121, the identification dictionary storage unit 122, the weight storage unit 123, and the sample storage unit 124 can be realized by a memory 1002 included in the computer 1000 or a storage device 1003 such as a hard disk device. Alternatively, the initialization unit 111, the learning unit 112, the index calculation unit 113, the determination unit 114, the weight update unit 115, the coefficient update unit 116, the update setting unit 117, the coefficient storage unit 121, the identification dictionary storage unit 122, and the weight storage unit 123. Also, part or all of the sample storage unit 124, the setting unit 311, the extraction unit 312, the weak identification unit 313, the score calculation unit 314, and the determination unit 315 may be realized by a dedicated circuit that realizes the function of each unit. it can.

  Moreover, although a part or all of said embodiment can be described also as the following additional remarks, it is not restricted to the following.

(Appendix 1)
Learning means for learning a plurality of discriminators by using a plurality of learning samples that are the data values for classifying data values belonging to any of the plurality of classes into any of the plurality of classes. When,
A plurality of optimization index coefficients each representing a magnitude of contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class An index calculation means for calculating the optimization index of each of the plurality of classifiers,
Among the learned classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. Determining means for adding to an identification dictionary which is a set of weak classifiers;
Coefficient updating means for updating the optimization index coefficient based on a rule in response to the addition of the weak classifier;
An identification dictionary learning device comprising:

(Appendix 2)
The class classification includes a plurality of positive classes to which the data value obtained from the recognition target belongs, and a negative system to which the data value not obtained from the recognition target belongs. Classifying the data values into one of the classes,
The identification dictionary learning device further includes:
The magnitude of the contribution to the optimization index of the in-kind set that is the combination of two different positive class is the optimization index of the different set that is the combination of the positive class and the negative class. Initialization is performed to set an initial value of each of the plurality of optimization index coefficients, which is a value used for deriving the optimization index, before the update is performed so as to be smaller than the magnitude of the contribution of The identification dictionary learning device according to attachment 1, further comprising:

(Appendix 3)
The identification dictionary learning device according to claim 2, wherein the initialization unit performs the initialization so that the contribution to the optimization index of each of the related sets is eliminated or substantially eliminated.

(Appendix 4)
The identification dictionary learning device according to any one of supplementary notes 1 to 3, wherein the coefficient updating unit performs the update so that a magnitude of contribution to the optimization index of each of the related sets increases.

(Appendix 5)
The coefficient updating means performs the updating so that the magnitude of the contribution of each of the related sets to the optimization index approaches the magnitude of the contribution of the different set to the optimization index. 4. The identification dictionary learning device according to any one of 4 above.

(Appendix 6)
An update setting means for receiving the rule in updating the optimization index of at least one of the related sets;
The coefficient update means performs the update according to the received rule.
The identification dictionary learning device according to any one of appendices 1 to 5.

(Appendix 7)
The learning means includes, for each of a plurality of different combinations of a class set including one or more classes and a type of feature amount derived based on the learning sample, the combination and the plurality of learning samples. The identification dictionary learning device according to any one of supplementary notes 1 to 6, wherein learning of the weak classifier is performed based on the identification classifier.

(Appendix 8)
Including an identification device and the identification dictionary learning device according to any one of appendices 1 to 7,
The identification dictionary learning device includes:
Further comprising the identification dictionary storage means,
The determining means adds the selected weak classifier to the identification dictionary,
The identification device includes:
Setting means for receiving an identification sample which is the data value;
Determining means for classifying the classification sample by integrating each of the classification results of the classification sample by the plurality of weak classifiers included in the identification dictionary;
Comprising
Identification system.

(Appendix 9)
A plurality of classifiers that perform classification that classifies data values belonging to any of a plurality of classes into any of the plurality of classes, and learning using a plurality of learning samples that are the data values,
A plurality of optimization index coefficients each representing a magnitude of contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class And calculating the optimization index for each of the plurality of classifiers,
Among the learned classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. Adding to an identification dictionary which is a set of weak classifiers;
In response to the addition of the weak classifier, the optimization index coefficient is updated based on a rule.
Identification dictionary learning method.

(Appendix 10)
The class classification includes a plurality of positive classes to which the data value obtained from the recognition target belongs, and a negative system to which the data value not obtained from the recognition target belongs. Classifying the data values into one of the classes,
further,
The magnitude of the contribution to the optimization index of the in-kind set that is the combination of two different positive class is the optimization index of the different set that is the combination of the positive class and the negative class. An initialization is performed to set an initial value of each of the plurality of optimization index coefficients, which is a value used for deriving the optimization index, before the update is performed so as to be smaller than the magnitude of the contribution of 9. The identification dictionary learning method according to 9.

(Appendix 11)
The identification dictionary learning method according to claim 10, wherein the initialization is performed so that the contribution of each of the related sets to the optimization index is eliminated or substantially eliminated.

(Appendix 12)
The identification dictionary learning method according to any one of appendices 9 to 12, wherein the update is performed so that the magnitude of contribution of each of the related sets to the optimization index increases.

(Appendix 13)
The update is performed such that the magnitude of the contribution to the optimization index of each of the related sets approaches the magnitude of the contribution to the optimization index of the heterogeneous set. Identification dictionary learning method.

(Appendix 14)
Receiving the rules in updating the optimization index of at least any of the cognate sets;
Performing the update according to the received rules;
The identification dictionary learning method according to any one of appendices 9 to 13.

(Appendix 15)
For each of a plurality of different combinations of a class set including one or more classes and a type of feature amount derived based on the learning sample, the weak identification is performed based on the combination and the plurality of learning samples. The identification dictionary learning method according to any one of supplementary notes 9 to 14, wherein learning of a vessel is performed.

(Appendix 16)
Computer
Learning means for learning a plurality of discriminators by using a plurality of learning samples that are the data values for classifying data values belonging to any of the plurality of classes into any of the plurality of classes. When,
A plurality of optimization index coefficients each representing a magnitude of contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class An index calculation means for calculating the optimization index of each of the plurality of classifiers,
Among the learned classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. Determining means for adding to an identification dictionary which is a set of weak classifiers;
Coefficient updating means for updating the optimization index coefficient based on a rule in response to the addition of the weak classifier;
Identification dictionary learning program to be operated.

(Appendix 17)
The class classification includes a plurality of positive classes to which the data value obtained from the recognition target belongs, and a negative system to which the data value not obtained from the recognition target belongs. Classifying the data values into one of the classes,
Computer
The magnitude of the contribution to the optimization index of the in-kind set that is the combination of two different positive class is the optimization index of the different set that is the combination of the positive class and the negative class. Initialization is performed to set an initial value of each of the plurality of optimization index coefficients, which is a value used for deriving the optimization index, before the update is performed so as to be smaller than the magnitude of the contribution of The identification dictionary learning program according to supplementary note 16, which is operated as a conversion means.

(Appendix 18)
Computer
The initialization means for performing the initialization so that the contribution to the optimization index of each of the cognate sets is eliminated or substantially eliminated;
The identification dictionary learning program according to appendix 17, which is operated as described above.

(Appendix 19)
Computer
The coefficient updating means for performing the updating so that the magnitude of contribution to the optimization index of each of the cognate sets increases;
The identification dictionary learning program according to any one of appendices 9 to 18 that is operated as described above.

(Appendix 20)
Computer
The coefficient updating means for performing the updating so that the magnitude of the contribution to the optimization index of each of the related sets approaches the magnitude of the contribution of the heterogeneous set to the optimization index;
The identification dictionary learning program according to any one of appendices 9 to 19 that is operated as described above.

(Appendix 21)
Computer
Update setting means for receiving the rule in updating the optimization index of at least one of the related sets;
The coefficient updating means for performing the updating according to the received rule;
The identification dictionary learning program according to any one of appendices 9 to 20 that is operated as described above.

(Appendix 22)
Computer
For each of a plurality of different combinations of a class set including one or more classes and a type of feature amount derived based on the learning sample, the weak identification is performed based on the combination and the plurality of learning samples. The learning means for learning the vessel;
The identification dictionary learning program according to any one of appendices 9 to 21 that is operated as described above.

  The present invention has been described above with reference to the embodiments, but the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  INDUSTRIAL APPLICABILITY According to the present invention, the present invention can be applied to use for detecting a specific object from an image. In particular, the present invention can be applied practically to detection applications relating to objects that have large variations in appearance on images, for example, human heads, faces, human bodies, automobiles, and the like.

DESCRIPTION OF SYMBOLS 100 Identification dictionary learning system 110 Identification dictionary learning apparatus 110A Identification dictionary learning apparatus 111 Initialization part 112 Learning part 113 Index calculation part 114 Determination part 115 Weight update part 116 Coefficient update part 117 Update setting part 120 Storage apparatus 121 Coefficient storage part 122 Identification Dictionary storage unit 123 Weight storage unit 124 Sample storage unit 140 Terminal device 300 Object detection system 310 Object detection device 311 Setting unit 312 Extraction unit 313 Weak identification unit 314 Score calculation unit 315 Determination unit 1000 Computer 1001 Processor 1002 Memory 1003 Storage device 1004 I / O interface 1005 Recording medium

Claims (10)

Learning means for learning a plurality of discriminators by using a plurality of learning samples that are the data values for classifying data values belonging to any of the plurality of classes into any of the plurality of classes. When,
A plurality of optimization index coefficients each representing a magnitude of contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class An index calculation means for calculating the optimization index of each of the plurality of classifiers,
Among the learned classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. Determining means for adding to an identification dictionary which is a set of weak classifiers;
Coefficient updating means for updating the optimization index coefficient based on a rule in response to the addition of the weak classifier;
An identification dictionary learning device comprising: The class classification includes a plurality of positive classes to which the data value obtained from the recognition target belongs, and a negative system to which the data value not obtained from the recognition target belongs. Classifying the data values into one of the classes,
The identification dictionary learning device further includes:
The magnitude of the contribution to the optimization index of the in-kind set that is the combination of two different positive class is the optimization index of the different set that is the combination of the positive class and the negative class. Initialization is performed to set an initial value of each of the plurality of optimization index coefficients, which is a value used for deriving the optimization index, before the update is performed so as to be smaller than the magnitude of the contribution of The identification dictionary learning device according to claim 1, further comprising: The identification dictionary learning device according to claim 2, wherein the initialization unit performs the initialization so that the contribution to the optimization index of each of the related sets is eliminated or almost eliminated. 4. The identification dictionary learning device according to claim 1, wherein the coefficient updating unit performs the updating so that a magnitude of contribution to the optimization index of each of the related sets is increased. 2. The coefficient updating unit performs the updating so that the magnitude of contribution of each of the related sets to the optimization index approaches the magnitude of contribution of the heterogeneous set to the optimization index. 5. The identification dictionary learning device according to any one of 1 to 4. An update setting means for receiving the rule in updating the optimization index of at least one of the related sets;
The coefficient update means performs the update according to the received rule.
The identification dictionary learning device according to claim 1. The learning means includes, for each of a plurality of different combinations of a class set including one or more classes and a type of feature amount derived based on the learning sample, the combination and the plurality of learning samples. The identification dictionary learning device according to claim 1, wherein learning of the weak classifier is performed based on the identification dictionary learning device. An identification device, and an identification dictionary learning device according to any one of claims 1 to 7,
The identification dictionary learning device includes:
Further comprising the identification dictionary storage means,
The determining means adds the selected weak classifier to the identification dictionary,
The identification device includes:
Setting means for receiving an identification sample which is the data value;
Determining means for classifying the classification sample by integrating each of the classification results of the classification sample by the plurality of weak classifiers included in the identification dictionary;
Comprising
Identification system. A plurality of classifiers that perform classification that classifies data values belonging to any of a plurality of classes into any of the plurality of classes, and learning using a plurality of learning samples that are the data values,
A plurality of optimization index coefficients each representing a magnitude of contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class And calculating the optimization index for each of the plurality of classifiers,
Among the learned classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. Adding to an identification dictionary which is a set of weak classifiers;
In response to the addition of the weak classifier, the optimization index coefficient is updated based on a rule.
Identification dictionary learning method. Computer
Learning means for learning a plurality of discriminators by using a plurality of learning samples that are the data values for classifying data values belonging to any of the plurality of classes into any of the plurality of classes. When,
A plurality of optimization index coefficients each representing a magnitude of contribution to an optimization index representing the classification accuracy of the class classification of a combination of the class to which the learning sample belongs in the wrong class classification and the classified class An index calculation means for calculating the optimization index of each of the plurality of classifiers,
Among the learned classifiers, the optimization indicator selects the classifier representing that the classification accuracy is the highest as a weak classifier, and the weak classifier is stored in the identification dictionary storage unit. Determining means for adding to an identification dictionary which is a set of weak classifiers;
Coefficient updating means for updating the optimization index coefficient based on a rule in response to the addition of the weak classifier;
Identification dictionary learning program to be operated.

Images