JP2014063494A - Classification device, classification method, and electronic facility - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a classification device capable of solving the problem that a kind value of a measurement sample obtained by the conventional learning method by means of an image is not accurate, a classification method, and an electronic facility.SOLUTION: The classification device of this invention includes a classification unit for classifying a target sample, a determination unit for respectively determining training samples correlating to each aggregation of the target sample, a deletion unit for deleting a kind value of the training sample whose kind value is not accurate, and a calculation unit for acquiring, with regarding the target sample as a measurement sample, the kind value of the measurement sample according to similarity between each measurement sample and each training sample, and similarity between two measurement samples. The classification method executes each step for achieving a function of the classification device. The electronic facility includes the classification device. A technology of this invention is applied to an information processing area.

Description

  The present invention relates to information processing technology, and more particularly to a classification device, a classification method, and electronic equipment.

  As a kind of explanation method that can effectively show the relationship between information, the explanation method according to the drawings is widely used in various fields at present. For example, it is widely used for web page classification, image search, video concept search and the like. The web page classification, drawing search, image search, video concept search, and the like can be regarded as a kind of classification method in a broad sense. The diagram described in the present invention is a weighted graph showing the relationship between data, and is not a drawing generally used.

  The conventional learning method using images usually uses the similarity between training samples under ideal conditions, solves the problem by an ideal analytical expression method or iteration method, and each training sample belongs. The type value that can effectively reflect the type to be calculated is calculated. In order to spread the study results to the measurement specimens, an ideal cost function must be formed based on the limiting conditions by flexible limitation.

  However, in the learning method based on the conventional image, the type value of the training sample hardly changes in the process of disseminating the study results to the measurement sample, so that the training sample with the incorrect type value affects the calculation of the type value of the measurement sample. Can be given. That is, this can make the type value of the measurement sample obtained by calculation inaccurate. In addition, the conventional learning method based on the image to spread the study results to the measurement samples processes each measurement sample in order, but does not consider the relationship between the measurement samples at all. Specimen type value can be inaccurate.

  Hereinafter, the outline of the present invention will be briefly described so that engineers can understand the outline of the present invention. The outline of the present invention to be described later cannot show all the contents of the present invention. This summary is not intended to present an emphasis or critical part of the invention or to limit the scope of the invention. The purpose of the summary of the present invention is to submit a summary and to prepare a foreword for detailed description to be described later.

  An object of the present invention is to provide a classification device, a classification method, and electronic equipment that can solve the problem that the type value of a measurement sample obtained by a conventional image learning method is inaccurate.

  In an embodiment of the present invention, a kind of classification device is provided. The classifier classifies target samples and arranges a classification unit arranged to form at least one set of the target samples and a training sample correlated with each set of target samples, respectively. And each training sample in the training sample has a maximum unit value and a minimum value among the training sample type values correlated with the set by a definite unit having a type value and each set of the target samples. If the difference between the type value is obtained and the difference is greater than a first predetermined threshold, a deletion unit arranged to delete the type value of the training sample correlated with the set, and the target sample A measurement sample, based on the similarity between each measurement sample and each remaining training sample and the similarity between the two measurement samples, and using the type value of the remaining training sample To, by a method for optimizing seek solutions, including a calculation unit arranged to acquire the type value of the measurement specimen.

  In another embodiment of the present invention, a kind of classification method is provided. The classification method includes classifying target samples to form at least one set of target samples, determining each training sample correlated with each set of target samples, and each training in the training samples. Each sample has a type value, and each set of the target samples obtains the difference between the maximum type value and the minimum type value among the type values of the training sample correlated with the set. And when the difference is greater than a first predetermined threshold, the step of deleting the type value of the training sample correlated with the set, the target sample as a measurement sample, and between each measurement sample and the remaining training sample Based on the similarity and the similarity between the two measurement samples, and using the remaining training sample type value, the solution sample type value is calculated and optimized. Including the steps to win.

  In another embodiment of the present invention, a kind of electronic equipment is further provided. The electronic equipment includes the classification device described above.

  The following effects of the invention can be achieved by the above-described classification device, classification method, and electronic equipment of the present invention. That is, by selecting a training sample with an incorrect type value and deleting the type value of such a training sample, the training sample shows the actual distribution status of the data more accurately, and the type value of the measurement sample is displayed. The accuracy of the kind value of the training sample used in the calculation process can be ensured. In addition, the accuracy of the type value of the acquired measurement sample can be improved by adding the similarity relationship between the measurement samples in the calculation process.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the best embodiment of the present invention will be described in detail with reference to the drawings. Thereby, the effect of the invention of the embodiment of the present invention can be understood in more detail.

The contents of the present invention can be understood in more detail from the contents described in the drawings of the present invention. In the drawings of the present invention, similar or similar parts are indicated by similar or similar reference numerals. The drawings described and the brief description thereof are included in and constitute a part of this specification to more fully describe the best embodiments of the invention, and to illustrate the principles and advantages of the invention. Will be explained in more detail.
It is a frame diagram which shows the classification device based on one Example of this invention. It is a simplified diagram showing the basic principle of processing a measurement specimen by a conventional classification technique. It is a simplified diagram showing the basic principle of processing a measurement specimen by a conventional classification technique. It is a simplified diagram showing the basic principle of processing a measurement specimen by a conventional classification technique. It is a simplified diagram showing the basic principle of processing a measurement sample by the classification device according to an embodiment of the present invention. FIG. 2 is a simplified diagram illustrating a structure of a calculation unit described in FIG. 1. It is a frame diagram which shows the classification device based on the other Example of this invention. It is a frame diagram which shows the classification method based on the other Example of this invention. It is a hardware layout drawing of information processing equipment for realizing the function of the classification device and the classification method concerning other examples of the present invention. What is desired to be described here is that the parts in the drawings of the present invention are drawn to show the parts simply / clearly, but are not drawn based on the ratio of the sizes of the parts. For example, in order to describe embodiments of the present invention in more detail, the size of a part in the drawing may be shown larger than the size of other parts.

  Hereinafter, exemplary embodiments of the present invention will be described with reference to the accompanying drawings. In order to explain the technology of the present invention in detail / understandability, not all features of the actual embodiment of the present invention are described in this specification, but when implementing any one embodiment, You must show all the features you need. Thereby, a person skilled in the art can realize the specific object of the present invention. For example, it must be shown that the limiting conditions apply to the system or method and that the limiting conditions vary with different embodiments. As is well known, creating the classification device or classification method of the present invention is very complex and time consuming, but the technicians in this technical field can implement the classification device or classification of the present invention according to the embodiments of the present invention described later. The method can be easily conceived.

  Also, to avoid lengthening the specification of the invention due to insignificant details, only the device structure and / or processing steps closely related to the technical solution of the invention are shown in the drawings of the invention. That is, description of details that are not highly relevant to the technical solution of the present invention is omitted.

  In an embodiment of the present invention, a kind of classification device is provided. The classifier is arranged to classify target samples and determine a training unit correlated to each set of target samples and a classification unit arranged to form at least one set of target samples. And each training sample in the training sample has a maximum unit value and a minimum type in the training sample type value correlated with the set by a definite unit having a type value and each set of the target samples. A deletion unit arranged to delete a training sample type value correlating to the set if the difference between the value is obtained and the difference is greater than a first predetermined threshold, and measuring the target sample A sample, based on the similarity between each measurement sample and each remaining training sample and the similarity between the two measurement samples, and using the type value of the remaining training sample , By a method for optimizing seek solutions, including a calculation unit arranged to acquire the type value of the measurement specimen.

  Hereinafter, a classification apparatus according to an embodiment of the present invention will be described in detail with reference to FIG.

  As shown in FIG. 1, the classification device 100 according to the embodiment of the present invention includes a classification unit 110, a determination unit 120, a deletion unit 130, and a calculation unit 140.

  In the classification device 100, the classification unit 110 classifies each target sample to form at least one set of target samples. The target sample of the present embodiment can be an image, a moving image, a text, or a web page.

  In the classification device according to a specific embodiment of the present invention, the classification unit 110 classifies the target samples, and between each target sample in a kind of set and at least one other target sample in the set. Is made larger than the fourth predetermined threshold.

For example, the classification method described above can be realized by a classification method. Specific steps are as follows. (A1) At first, a plurality of sets are formed so that each set includes only one target sample. (A2) Calculate the similarity between the two sets. Similarity between two sets means that the maximum similarity between the two sets is not exceeded. That is,

  In Equation 1, A and B denote any two sets, a and b denote any target sample in the two sets, and Sim (A, B) denotes set A and set B, The similarity between is shown. (A3) The similarity of the target samples in the two sets is calculated to determine whether the similarity is greater than a fourth predetermined threshold. If greater than the fourth predetermined threshold, after merging the two sets having the maximum similarity (that is, the two sets having the maximum similarity among a plurality of sets) into one set, the step ( Perform a2). In the plurality of sets, the above steps are repeated until the similarity between any two sets is smaller than a fourth predetermined threshold.

  In the present embodiment, the fourth predetermined threshold can be set by experience or by a test method, but will not be described again here.

  As shown in FIG. 1, the determination unit 120 determines each training sample correlated to each set of target samples. The training sample is an object that is the same as the target sample type. For example, the training sample and the target sample are one of the following subjects. That is, it is a kind of image, video, text, web page, and the like.

  In the classification apparatus according to the specific embodiment of the present invention, the “training samples correlated with each set of target samples” can be determined by various determination methods.

  In one type of realization, for each set of target samples, the deterministic unit 120 includes training samples that correlate training samples with similarity greater than a second predetermined threshold between at least one target sample of the set. To do.

  In an example in the implementation method, if it is assumed that the target sample set M is any one set obtained by classifying the target samples by the classification unit 110, in the set M, the deterministic unit 120 includes: A similarity between each training sample and each target sample of set M is obtained. Next, training samples whose similarity to at least one target sample in the set M is larger than a second predetermined threshold are set as training samples correlated with the set M.

In another example in the implementation method, for each training sample, the determination unit 120 may not calculate the similarity between the training sample and each training sample in the set M. Taking the training sample S _B as an example, when the similarity between any target sample S _A and the training sample S _{B in} the set M obtained by the determination unit 120 is greater than a second predetermined threshold, the determination is made Unit 120 does not calculate the similarity between the other remainder target samples in set M and training sample S _B, and determines that training sample S _B is one training sample that is directly correlated to set M. Also good.

  The similarity between the training sample and the target sample can be calculated by selecting a calculation method to be applied according to a specific situation. For example, a conventional calculation method for calculating image similarity, text similarity, and the like can be employed. That is, when the training sample and the target sample are images, texts, and other types of objects, the similarity between the training sample and the target sample can be calculated for each type of object.

  In the present embodiment, the second predetermined threshold value can be set by experience or by a test method, but will not be described again here. In the present embodiment, the second predetermined threshold value can be set to the same value as the fourth predetermined threshold value.

  In another implementation, for each set of target samples, the deterministic unit 120 sets training samples whose similarity to at least one target sample in the set is greater than a second predetermined threshold as correlation training samples. Or “a training sample whose similarity to at least one target sample in the set is greater than a second predetermined threshold” and a training sample that correlates training samples that are greater than a third predetermined threshold to the set; To do.

  Taking a set M of any one of the target samples as an example and in a manner similar to that described above, the deterministic unit 120 has a similarity with at least one target sample in the set M greater than a second predetermined threshold. Let the training samples be training samples correlated to the set M. In order to simplify the explanation, in the following, “a training sample whose similarity to at least one target sample in the set M is larger than a second predetermined threshold value” is abbreviated as a first training sample. In addition to the first training sample, the determination unit 120 includes training samples that correlate training samples that have a similarity with any one of the training samples in the first training sample that is greater than a third predetermined threshold; can do. Similarly, in order to simplify the explanation, “a training sample whose similarity to any one training sample in the first training sample is larger than a third predetermined threshold value” is abbreviated as a second training sample in the following text. Accordingly, the “training sample correlated with the set M” in the above-described embodiment includes the first training sample and the second training sample. In the present embodiment, the third predetermined threshold can be set by experience or by a test method, but will not be described again here.

  Any training sample in the first training sample described above also has a kind value. The type value is a kind of measure indicating the degree of matching between the type of the sample to which it belongs and a predetermined type. That is, the type value is a value indicating the degree to which the type of the sample to which it belongs matches a predetermined type. In general, explain that the larger the type value, the more likely the sample belonging to the type value matches the specified type. The smaller the type value, the more samples that belong to the type value. Explain that there is a low possibility of matching a predetermined type. For example, the type value is usually placed between −1 and 1.

  In the present embodiment, the training samples can be classified into the following two types of training samples according to the difference in the method of acquiring the type value.

  A kind of training sample is a general standard sample. Such a standard sample usually has a predetermined type value. The kind value can be pre-ordered artificially, for example. However, since there is a limit to human power, the number of such standard samples that usually have a predetermined kind of value also has a certain limit.

  Another type of training specimen is an instrument standard specimen. The type value of the instrument standard sample can be obtained from the type value of the standard sample in a training process (a general classification device is a process in which training is performed using a training sample before use).

  Therefore, the training sample of the classification device according to the embodiment of the present invention can include the two types of samples described above. That is, the above-mentioned standard sample and instrument standard sample can be included. Since the kind values of the two kinds of training samples described above can be obtained by a method similar to the method described above, they will not be described again here. In this case, all of the “training samples correlated with each set of target samples” determined by the determination unit 120 may be instrument standard samples.

  As described above, the training unit correlated with each set of target samples can be determined by the processing of the determination unit 120.

  As can be estimated by general estimation, the difference in the type value of similar samples cannot be large. Thus, any two training samples (especially where the two training samples are instrumental sample samples) are each similar to any two target samples, and the two target samples are also similar to each other (eg, In the case of belonging to any one set described above), the difference in the type value of the two training samples cannot be large. Otherwise, for example, if the difference between two training sample type values that are respectively similar to two target samples in any one set is greater than a first predetermined threshold, the two training sample type values are correct. Explain that there is no. Therefore, a more accurate result can be obtained by such an estimation calculation method that removes the kind value.

  For each set of target samples, the deletion unit 130 first calculates the difference between the maximum type value and the minimum type value in the type value of “training sample correlated to the set”. Next, when the calculated difference is larger than the first predetermined threshold, the type value of such “training sample correlated with the set” is deleted.

  For example, as described above, in one set M of target samples, a training sample correlated with the set M is determined by the determination unit 120. Next, the deletion unit 130 selects the largest kind value and the smallest kind value among the kind values of all the training samples correlated with the set M, and calculates the difference between the largest kind value and the smallest kind value (said difference). Is greater than or equal to 0). If the difference is greater than the first predetermined threshold, the type values of all training samples correlated with these sets M are deleted.

  In particular, in one priority example, when all of the “training samples correlated with the set” determined by the determination unit 120 are device standard samples, the deletion unit 130 deletes the type values of these device standard samples.

  In the present embodiment, the first predetermined threshold value can be set by experience or by a test method, but will not be described again here.

  The calculation unit 140 uses the type value of the remaining training sample (that is, other training samples excluding the training sample from which the type value has been deleted) to obtain a solution and optimize the type of the target sample. The type of the target sample is determined based on the type value obtained by acquiring and calculating the value.

  In the classification device according to the embodiment of the present invention, the calculation unit 140 can set the target sample as the measurement sample. Next, find a solution based on the similarity between each measurement sample and each remaining training sample and the similarity between the two measurement samples and using the type value of the remaining training sample. The type value of the measurement specimen is obtained by an optimization method.

Traditional image-based learning methods usually process each measurement sample in order and consider only the relationship between each measurement sample and the training sample, and use the relationship between the measurement sample and the training sample completely. Absent. 2A to 2C are simplified diagrams showing the basic principle adopted when a conventional image-based learning method processes a measurement sample. 2A to 2C and FIG. 2D described later, “◯” indicates a training sample (ie, S _B 1 to S _B 5), and “Δ” indicates a measurement sample (ie, S _A 1 to S _A 3). The connecting lines in the drawings show the similarity between objects to be connected.

  As shown in FIGS. 2A-2C, the conventional method only considers the similarity between the measurement sample and the training sample when processing the measurement sample. That is, the measurement samples are classified by processing each measurement sample. In FIG. 2A-2C, the connection line between a measurement sample and each training sample was not shown, but only the connection line between the some training sample and measurement sample before high similarity was shown.

FIG. 2D is a simplified diagram illustrating the principle by which the calculation unit 140 processes the measurement specimen. Comparing FIG. 2D with FIGS. 2A to 2C described above, in FIG. 2D, when the calculation unit 140 processes the measurement sample, the similarity between the measurement sample and the training sample is considered, and the measurement sample and It can be seen that the similarity between the measurement specimens (that is, the connecting line between S _A 1 and S _A 2 and between S _A 2 and S _A 3 in FIG. 2D) is also considered. it can.

  The connection lines (similarities between training samples) between the training samples shown in FIGS. 2A to 2D are not used when calculating the type value of the measurement sample. This is used when acquiring the type value of the instrument standard sample.

  Hereinafter, various arrangement examples of the calculation unit 140 will be described with reference to FIG.

  As shown in FIG. 3, the calculation unit 140 of the present embodiment may include a cost function generation module 310 and a cost function calculation module 320.

  The cost function generation module 310 generates a cost function that can reflect a predetermined restriction condition. The cost function calculation module 320 acquires the type value of the measurement sample by detecting the problem of minimizing the cost function.

  The predetermined restriction condition can include the following two conditions. That is, the higher the similarity between the measurement sample and the training sample, the closer the type value of the measurement sample and the training sample, and the higher the similarity between the two measurement samples, the two measurements The specimen type value is approaching.

The cost function generated by the cost function generation module 310 of the classification device according to the embodiment of the present invention according to the predetermined restriction condition described above may include the following functions or modifications.

は、Ｍ個の測定標本中で第ｉ個目の測定標本の種類バリューであり、

は、前記Ｎ個の訓練標本中で第ｊ個目の訓練標本の種類バリューであり、

は、第ｉ個目の測定標本と第ｊ個目の訓練標本との間の類似性であり、

は、第ｋ個目の測定標本と第ｌ個目の測定標本との間の類似性である。 In the function, M is the number of measurement samples, N is the number of remaining training samples,

Is the type value of the i-th measurement sample among the M measurement samples,

Is the type value of the jth training sample in the N training samples,

Is the similarity between the i th measurement sample and the j th training sample,

Is the similarity between the kth measurement sample and the lth measurement sample.

は、測定標本と訓練標本との間の類似性を示し、

は、測定標本と測定標本との間の類似性を示す。 In the above embodiment,

Indicates the similarity between the measurement sample and the training sample,

Indicates the similarity between the measurement sample and the measurement sample.

と

のような２つの函数を含むコスト函数は、以下のような

の形式により示すことができる。 For example,

When

The cost function including two functions such as

It can be indicated by the form.

は、線形組合係数であり、

は、すべての測定標本の種類バリューで構成されるベクトルである。

は、経験により設定するか、或いは試験方法により設定することができるが、ここでは再び説明しない。 In the function,

Is the linear combination coefficient,

Is a vector composed of all measurement sample type values.

Can be set by experience or by test method, but will not be described again here.

により、すべての測定標本の種類バリューを算出することができる。前記最適化方法

は、一種の簡単な二回最適化方法である。本技術領域の技術者は、周知する常識と公開されている資料により、最適化の方法を選び出すことができるので、ここでは再び説明しない。 Next, find the solution and optimize

Thus, it is possible to calculate the type value of all the measurement samples. The optimization method

Is a kind of simple two-time optimization method. An engineer in this technical field can select an optimization method based on common knowledge and publicly available materials, and will not be described again here.

  In the embodiment of the present invention, the calculation unit 140 can also use the training sample and the target sample from which the type value is deleted as the measurement sample, and obtain the type value of the measurement sample by the above-described method, The classification result is determined by the sample type value. In this case, the variable formed in the calculation process not only includes the type value of the target sample, but also includes the new type value of the training samples from which the type value has been deleted. After obtaining a new type value for all measurement samples, such a target sample type value can determine the final classification result (the purpose of the classification is to acquire the target sample type). For example, assuming that the type value is between −1 and 1, when the first, second and third type values of each target sample obtained by calculation are 1, 0.8 and −1, respectively Indicates that the first target sample and the predetermined type are perfectly matched, indicates that the second second target sample and the predetermined type are relatively consistent, and indicates the third target sample. And that the given type does not match exactly. The predetermined type may be, for example, a type of training sample whose type value is 1 in the training sample.

  In actual applications, the measurement specimens are usually acquired in large quantities, but in the conventional method, the similarity between each measurement specimen and the training specimen is considered without considering the similarity between the measurement specimens acquired in large quantities. Consider. In the embodiment of the present invention, by further detecting the similarity between the measurement specimen and the measurement specimen, higher classification accuracy and classification result can be obtained.

  Hereinafter, a classification device according to another embodiment of the present invention will be described in detail with reference to FIG.

  As shown in FIG. 4, the classification apparatus 400 not only includes a classification unit 410, a determination unit 420, a deletion unit 430, and a calculation unit 440, but further includes a sample update unit 450. The classification unit 410, the determination unit 420, the deletion unit 430, and the calculation unit 440 of the classification device 400 shown in FIG. 4 have the same structure and function as each unit in the device 100 described in FIG. 1 and / or FIG. And a similar technical effect can be achieved, and will not be described again here.

  The training sample of the present embodiment includes the above-mentioned standard sample and instrument standard sample. As shown in FIG. 4, in the classification device 400, the sample update unit 450 sets the measurement sample that has acquired a new type value as a device standard sample in the next classification process. The measurement sample becomes a part of the instrument standard sample to be used in the next classification process, and the original instrument standard sample is still used as the instrument standard sample.

  As described above, the classification device according to the embodiment of the present invention selects a training sample whose type value is not correct, and deletes the type value of the training sample so that the training sample has an actual distribution state of data. And the accuracy of the type value of the training sample used in the process of calculating the type value of the measurement sample can be ensured. In addition, the classification apparatus according to the embodiment of the present invention can improve the accuracy of the type value of the acquired measurement sample by adding the similarity relationship between the measurement samples in the calculation process.

  In an embodiment of the present invention, a kind of classification method is further provided. Hereinafter, the classification method according to the present embodiment will be described with reference to FIG.

  As shown in FIG. 5, the processing flow 500 of the classification method according to the embodiment of the present invention starts from step S510 and first performs step S520.

  In step S520, the target samples are classified to form at least one set of target samples. Next, step S530 is performed. The processing performed in step S520 is the same as the processing performed by the classification unit 110 described with reference to FIG. 1, and a similar technical effect can be obtained, and therefore will not be described again here.

  In step S530, each training sample that correlates to each set of target samples is determined. Each training sample in the training sample has a kind value. Next, step S540 is performed. The process performed in step S530 is the same as the process performed by the confirmation unit 120 described with reference to FIG. 1, and a similar technical effect can be obtained, and therefore will not be described again here.

  In step S540, the difference between the maximum type value and the minimum type value is acquired from the set of target samples in the type value of the training sample correlated with the set. If the difference is larger than the first predetermined threshold, the training sample type value correlated with the set is deleted. Next, step S550 is performed. The processing performed in step S540 is similar to the processing performed by the deletion unit 130 described with reference to FIG. 1, and similar technical effects can be obtained.

  In step S550, the target sample is a measurement sample, and the type value of the remaining training sample is used based on the similarity between each measurement sample and the remaining training sample and the similarity between the two measurement samples. Then, the type value of the measurement sample is obtained by a method of obtaining and optimizing the solution. Next, step S560 is performed. The processing performed in step S550 is the same as the processing performed by the calculation unit 140 described with reference to FIG. 1 and / or FIG. 3, and a similar technical effect can be obtained.

  The processing flow 500 of the present embodiment ends with step S560.

  In step S550 according to the preferred embodiment of the present invention, the training sample from which the type value is deleted and the target sample can be used as a measurement sample, and the classification result can be obtained from the type value of the obtained target sample.

  As described above, the classification method according to the embodiment of the present invention selects a training sample whose type value is not correct, and deletes the type value of the training sample so that the training sample has an actual distribution state of data. And the accuracy of the type value of the training sample used in the process of calculating the type value of the measurement sample can be ensured. In addition, the classification method according to the embodiment of the present invention can improve the accuracy of the type value of the acquired measurement sample by adding the similarity relationship between the measurement samples in the calculation process.

  In an embodiment of the present invention, a kind of electronic equipment is further provided. The electronic equipment includes the classification device described above.

  The electronic device according to the embodiment of the present invention is any one kind of equipment in the equipment described later. That is, it is a kind of equipment in a mobile phone, a computer, a doublet personal computer, a portable information terminal, a media player, and the like. The electronic equipment has the various functions and technical effects of the classification device described above, but will not be described again here.

  The functions of each component unit, child unit, module, and the like included in the classification device according to the embodiment of the present invention described above can be realized by software, firmware, hardware, or any combination thereof. When the function is realized by software or firmware, a program that configures the software or firmware from a storage medium or network is transferred to a device having a dedicated hardware structure (for example, the common device 600 shown in FIG. 6). You can enter and install. The device in which the various programs are installed can realize the functions of the constituent units and the child units.

  FIG. 6 is a simplified diagram showing a hardware arrangement structure diagram of a kind of information processing equipment for realizing the functions of the classification device and the classification method according to the embodiment of the present invention.

  In FIG. 6, a central processing unit (CPU) 601 performs various processes by a program stored in a read-only memory (ROM) 602 or a program read from a memory device 608 to a random access memory (RAM) 603. The RAM 603 also stores data demanded when the CPU 601 performs various processes. The CPU 601, ROM 602, and RAM 603 are connected to each other via a bus 604. An input / output interface 605 is also coupled to the bus 604.

  The following components are coupled to the input / output interface 605. For example, an input device 606 (including a keyboard, a mouse, etc.), an output device 607 (including a display device (eg, a cathode ray tube display device (CRT), a liquid crystal display device (LCD)), a speaker, etc.), a memory device 608 ( A communication device 609 (including a network interface card (for example, a LAN card, a modulation / demodulation device, etc.)). The communication device 609 performs communication via a network such as the Internet. Depending on demand, the drive device 610 can also be coupled to the input / output interface 605. A removable memory medium 611, for example, a magnetic disk, a write disk, a magneto-optical disk, a semiconductor memory device, etc. can be mounted on the drive device 610. Therefore, the program stored in the memory medium 611 can be read into the memory device 608 according to demand.

  When the above-described processing is performed by software, predetermined software can be read from a network, for example, the Internet, or a memory medium, for example, a removable memory medium 611.

  As known to those skilled in the art, the memory media is not limited to the removable memory media 611 shown in FIG. 6 where the program is stored or transmits the program to the user away from the facility. Absent. The removable memory medium 611 includes, for example, a magnetic disk (including a floppy (registered trademark) disk), a write disk (including a compact disk read-only memory (CD-ROM) and a digital versatile disk (DVD)), and a magneto-optical disk. A disk (mini disk (MD) (registered trademark)) or a semiconductor memory. The memory medium may be a ROM 602, a hard disk included in the memory device 608, or the like. A program is stored in the memory medium, and a user can use the memory medium and the program stored therein.

  In an embodiment of the present invention, a program in which an apparatus-readable instruction code is stored is provided. When the apparatus reads and executes the instruction code, the classification method according to an embodiment of the present invention is performed. A memory medium such as a device for storing the program, for example, a magnetic disk, a write disk, a magneto-optical disk, a semiconductor memory device, etc. is also disclosed in the specification of the present invention.

  In particular embodiments of the present invention described above, features described and / or disclosed in one type of implementation may be used in one or more other implementations in a similar or similar manner, or It can be used in combination with features of other embodiments, or can be used in place of features of other embodiments.

  The method according to the embodiment of the present invention is not performed in accordance with the time order described in the specification or the time order shown in the drawings, and is performed in accordance with another time order, or is performed together. Or can be implemented separately. That is, the technical scope of the present invention is not limited to the order of implementation described in this specification.

  Each step of the method of the present invention can be implemented by a computer-executable program stored in a device-readable memory medium.

  The object of the present invention can also be realized by a method described later. That is, the executable program code stored in a memory medium is directly or indirectly provided to a system or facility, and a computer or a central processing unit (CPU) in the system or facility reads and executes the program code. It is a method to do.

  In this case, the system or equipment must have the function of executing the program code. The implementation method of the present invention is not limited to the above-described program. The format of the program may be any format. For example, a target program, an interpreter executable program, or a script provided to the operating system.

  The above-described device-readable memory media includes the following devices, but is not limited to the following devices. For example, memory devices and memory units, semiconductor equipment, disk units (eg, magnetic disks, write disks, magneto-optical disks), other media capable of storing information, and the like.

  The technology of the present invention can also be realized by connecting a user's computer to a predetermined website on the Internet, downloading the computer program code of the present invention, and executing the program on the computer.

  Terms described herein, for example, related terms such as left, right, first, second, etc., are for separating one device or step from another device or step, It is not intended to limit or present any relationship or order that exists between the devices or steps. In addition, various terms that do not limit the inclusion of other things such as “include”, “comprise”, and the like are not limited to the process, method, apparatus, or facility that includes any element. , Which may further include other elements not described, or may further include elements inherent in the process, method, apparatus, or facility described above. That is, in the absence of other restrictive terms, terms such as “... include one” include not only the element in which the process, method, apparatus, or facility is described, but also other elements. Means that you can.

  According to the embodiment of the present invention described above, the present invention provides the following solutions. However, the present invention is not limited to the method described later.

(Appendix 1)
In the classifier, the classifier classifies the target samples and determines a classification unit arranged to form at least one set of the target samples and a training sample correlated with each set of target samples. And each training sample in the training sample is a maximum in the type value of the training sample correlated to the set by a deterministic unit having a type value and each set of target samples. A deletion unit arranged to acquire a difference between the type value and the minimum type value and, if the difference is greater than a first predetermined threshold, to delete the type value of the training sample correlated with the set; Based on the similarity between each measurement sample and each remaining training sample and the similarity between the two measurement samples, with the target sample as a measurement sample, and the remaining training sample Using a type value, a method of optimizing seeking solutions, including a calculation unit arranged to acquire the type value of the measurement specimen.

(Appendix 2)
The calculation unit determines that the higher the similarity between the measurement sample and the training sample, the closer the kind value of the measurement sample and the training sample is, and the similarity between the two measurement samples. The higher the performance, the closer the two sample types are closer to each other, the cost function generation module arranged to generate a cost function that can reflect that, and the minimization of the cost function. The classification apparatus according to appendix 1, further comprising: a cost function calculation module that acquires a type value of a measurement sample by detecting a problem.

と

を含み、
前記函数において、Ｍは、測定標本の数量であり、Ｎは、残った訓練標本の数量であり、

は、Ｍ個の測定標本中で第ｉ個目の測定標本の種類バリューであり、

は、前記Ｎ個の訓練標本中で第ｊ個目の訓練標本の種類バリューであり、

は、第ｉ個目の測定標本と第ｊ個目の訓練標本との間の類似性であり、

は、第ｋ個目の測定標本と第ｌ個目の測定標本との間の類似性である付記２に記載の分類装置。 (Appendix 3)
The cost function is a function as shown below, or a modified example,

When

Including
In the function, M is the number of measurement samples, N is the number of remaining training samples,

Is the type value of the i-th measurement sample among the M measurement samples,

Is the type value of the jth training sample in the N training samples,

Is the similarity between the i th measurement sample and the j th training sample,

Is the classification device according to attachment 2, which is a similarity between the k-th measurement sample and the l-th measurement sample.

(Appendix 4)
4. The training sample according to any one of appendix 1 to appendix 3, including a standard sample having a predetermined type value and a device standard sample that obtains a type value from the type value of the standard sample in a training process. Sorter.

(Appendix 5)
The classification device according to appendix 4, wherein a training sample correlated with each set of target samples determined by the determination unit is the device standard sample.

(Appendix 6)
In any one of appendix 1 to appendix 3, the calculation unit uses the training sample and the target sample from which the type value has been deleted as the measurement sample, and determines the classification result based on the obtained type value of the target sample. The classification device described.

(Appendix 7)
The classification apparatus according to appendix 4 or appendix 5, wherein the calculation unit uses the training sample and the target sample from which the type value has been deleted as the measurement sample, and determines the classification result based on the obtained type value of the target sample.

(Appendix 8)
8. The classification device according to any one of appendices 3, 5, and 7, further including a sample update unit that uses a measurement sample that has acquired a new type value as an instrument standard sample in the next classification process.

(Appendix 9)
The deterministic unit is arranged such that, with each set of target samples, a training sample whose similarity is greater than a second predetermined threshold among at least one target sample of the set is a training sample that correlates to the set. The classification device according to any one of appendix 1 to appendix 8.

(Appendix 10)
The deterministic unit determines, for each set of target samples, training samples whose similarity to at least one target sample in the set is greater than a second predetermined threshold and whose similarity is greater than a third predetermined threshold. The classification device according to appendix 9, which is arranged so as to be a training sample correlated with the set.

(Appendix 11)
The classification unit is arranged so that the similarity between each target sample in the set and at least one other target sample in the set is greater than a fourth predetermined threshold according to each type of the target sample. The classification device according to any one of appendix 1 to appendix 10, wherein:

(Appendix 12)
The classification device according to any one of Supplementary Note 1 to Supplementary Note 11, wherein the target specimen and the training specimen are a kind of an image, a moving image, a text, or a web page.

(Appendix 13)
In the classification method, a target sample is classified to form at least one set of the target samples, and each training sample correlated with each set of the target samples is determined, and each training sample in the training sample is determined. Each of which has a kind value and each set of the target samples obtains a difference between the largest kind value and the smallest kind value among the kind values of the training sample correlated with the set. And if the difference is greater than a first predetermined threshold, deleting the training sample type value correlated with the set, and using the target sample as a measurement sample, the similarity between each measurement sample and the remaining training sample The measurement sample type variable by a method of finding and optimizing the solution based on the similarity between the two measurement samples and using the type value of the remaining training sample. Including a step to win over, the.

(Appendix 14)
The step of obtaining the type value of the measurement sample by a method of obtaining and optimizing a solution is performed by measuring the measurement sample and the training sample as the similarity between the measurement sample and the training sample increases. Generating a cost function that can reflect that the type value of the two is closer and the higher the similarity between the two measured samples, the closer the type value of the two measured samples are; and The classification method according to claim 13, further comprising: obtaining a type value of the measurement sample by detecting a problem of minimizing the cost function.

(Appendix 15)
15. The classification method according to appendix 13 or appendix 14, wherein the training sample includes a standard sample having a predetermined type value and an equipment standard sample that obtains a type value from the type value of the standard sample in a training process.

(Appendix 16)
The training sample and the target sample from which the type value has been deleted are used as a measurement sample together, and the classification result is determined based on the obtained type value of the target sample, further including the step of any one of the supplementary notes 13 to 15 Classification method.

(Appendix 17)
Electronic equipment, comprising the classification device according to any one of appendix 1 to appendix 12.

(Appendix 18)
The electronic equipment according to supplementary note 17, wherein the electronic equipment is one of the following equipments: a mobile phone, a computer, a doublet personal computer, a portable information terminal, a media player, and the like.

(Appendix 19)
In the program in which the device-readable instruction code is stored, the classification method according to any one of Supplementary Note 13 to Supplementary Note 16 is performed by executing the program.

(Appendix 20)
In the computer-readable memory medium, the program described in Appendix 19 is stored in the memory medium.

DESCRIPTION OF SYMBOLS 100 Classification apparatus 110 Classification unit 120 Determination unit 130 Deletion unit 140 Calculation unit 310 Cost function generation module 320 Cost function calculation module 400 Classification apparatus 410 Classification unit 420 Determination unit 430 Deletion unit 440 Calculation unit 450 Specimen update unit 600 Common device 601 Central processing Unit (CPU)
602 Read only memory (ROM)
603 Random access memory (RAM)
604 bus 605 input / output interface 606 input device 607 output device 608 memory device 609 communication device 610 drive device 611 removable memory medium

Claims (10)

In the classification device,
A classification unit arranged to classify target samples and form at least one set of target samples;
A training unit arranged to determine each training sample correlated to each set of target samples, and each training sample in the training sample has a determination unit having a kind value;
With each set of target samples, in the type value of the training sample correlated with the set, if the difference between the maximum type value and the minimum type value is obtained and the difference is greater than a first predetermined threshold, A deletion unit arranged to delete the type value of the training sample correlated to the set; and
Using the target sample as a measurement sample, based on the similarity between each measurement sample and each remaining training sample and the similarity between two measurement samples, and using the type value of the remaining training sample And a calculation unit arranged to obtain the type value of the measurement sample by a method for obtaining and optimizing a solution. The calculation unit is
The higher the similarity between the measurement sample and the training sample, the higher the similarity between the measurement sample and the training sample and the higher the similarity between the two measurement samples. A cost function generation module arranged to generate a cost function that can reflect the fact that the two sample type values are close to each other, and
The classification apparatus according to claim 1, further comprising: a cost function calculation module that acquires a type value of the measurement sample by detecting a problem of minimization of the cost function.   The classification apparatus according to claim 1, wherein the training sample includes a standard sample having a predetermined type value and a device standard sample that obtains a type value from the type value of the standard sample in a training process.   4. The classification apparatus according to claim 3, wherein a training sample correlated with each set of target samples determined by the determination unit is the instrument standard sample.   The classification device according to claim 3 or 4, wherein the calculation unit uses the training sample and the target sample from which the type value is deleted as a measurement sample, and determines the classification result based on the obtained type value of the target sample.   6. The classification apparatus according to claim 3, further comprising a sample update unit that uses a measurement sample that has acquired a new type value as an instrument standard sample in the next classification process.   The deterministic unit is arranged such that, with each set of target samples, a training sample whose similarity is greater than a second predetermined threshold among at least one target sample of the set is a training sample that correlates to the set. The classification device according to any one of claims 1 to 6.   The deterministic unit determines, for each set of target samples, training samples whose similarity to at least one target sample in the set is greater than a second predetermined threshold and whose similarity is greater than a third predetermined threshold. The classification device according to claim 3, wherein the classification device is arranged so as to be a training sample correlated with the set. In the classification method,
Classifying target samples to form at least one set of target samples;
Determining each training sample correlated to each set of target samples, and each training sample in the training sample all having a type value;
With each set of target samples, in the type value of the training sample correlated with the set, if the difference between the maximum type value and the minimum type value is obtained and the difference is greater than a first predetermined threshold, Deleting the training sample type value correlated to the set; and
Using the target sample as a measurement sample, based on the similarity between each measurement sample and the remaining training sample, and the similarity between the two measurement samples, and using the type value of the remaining training sample, Obtaining a type value of the measurement sample by a method for obtaining and optimizing a solution. In electronic equipment,
Electronic equipment including the classification device according to any one of claims 1 to 8.

Images