JP2021068004A - Identification assistance data generation technique and identification information extraction technique - Google Patents

Abstract

To certainly extract unique data from a target even in a situation where an ID or the like is not assigned to the target.SOLUTION: A first phase includes the steps of: obtaining an input vector representing each of a plurality of targets; and dividing a space including the input vector such that the input vector representing each of the plurality of targets is included in different sub-spaces and generating a plurality of hyper-surfaces used to generate a unique vector that identifies each of the plurality of targets. A second phase includes the steps of: obtaining a vector representing a specific target included in the plurality of targets; and evaluating the vector representing the specific target by dividing the space including the vector such that each of the plurality of targets is included in the different sub-spaces and using the plurality of hyper-surfaces held in advance, to generate a vector that identifies the specific target.SELECTED DRAWING: Figure 1

Description

The present invention relates to a technique for extracting unique identification information from a plurality of objects.

Extract a fixed digital value (discrete value such as 0 or 1. A vector including a series of multiple digital values. Also called a fixed value from its characteristics) unique to each object from an individual or other object. In general, 0 or 1 is determined by reading out some object-specific analog value and comparing it with a threshold set for the analog value.

For example, it is assumed that a threshold value of 0.5 is set for an analog value of 0 or more and 1 or less. In this case, for a specific target for which an analog value near the threshold value of 0.5 is measured, if the noise is positive, an analog value having a threshold value of 0.5 or more is measured and "1" is extracted. If is negative, an analog value less than the threshold value of 0.5 is measured and "0" is extracted. In such cases, a unique fixed digital value cannot be obtained. Similarly, for a plurality of objects, a conversion biased to only one of "0" or "1" should be avoided because a "unique" fixed digital value cannot be obtained.

Techniques for obtaining a unique fixed digital value in such a situation include the following.
(1) A technique for determining a fixed value by reading many times and determining a fixed value by a method such as majority voting (for example, Patent Documents 1 and 2).
(2) A technique in which individual auxiliary data (helper data) is calculated and held in advance for each read pattern for which a fixed value is to be extracted, and that information is used when correcting errors. Helper data includes the exclusive OR of random codewords in the extraction pattern and the conversion of the extraction pattern into the syndrome used for error correction. These methods are also called Fuzzy Extractor. be called. (For example, Non-Patent Document 1)
(3) A technique for avoiding the use of noisy axes and values, and axes and values in which the values read throughout the object are biased (for example, Patent Document 3).

The technique (1) is effective for correcting noise temporarily generated during reading, but cannot correct noise permanently added due to pattern deformation, scratches, deterioration, or the like. Further, even for noise that temporarily occurs during reading, if the average of the reading values is not farther from the threshold value than the dispersion of noise, it is not an effective countermeasure.

The technique (2) corrects with helper data, but since this correction value differs for each read target, the read target can be specified in advance, such as when the ID of the read target is known in advance. It is an effective method for verifying whether or not the object is genuine. However, it cannot be used when extracting a fixed value such as an ID or an encryption key from a state in which the auxiliary data corresponding to the target is unknown. Although it is possible to try all the auxiliary data, there is a problem that the time required for identification becomes longer in proportion to the number of target candidates.

The technique (3) has a problem that when determining an axis to be used for each read target, it is necessary to assign an ID to the target or perform an 100% search or the like at the time of extracting a fixed value. When deciding which axes are not used for all targets, not for each read target, as the number of targets increases, the ratio of axes that cause an error for all targets or axes that cause an error for all targets decreases. For some targets, the error is small but the axis is not used, or for some targets, the error is large but the axis is left. There is a problem that the error cannot be removed.

In addition, there are various technologies such as a method of matching with a template and a method of using a digital watermark, but these are situations in which a large number of targets do not have an ID, etc., or targets for embedding a digital watermark. It is not always effective to reliably extract a fixed value from the object itself in a situation where it cannot be changed or processed.

Japanese Unexamined Patent Publication No. 2018-175762 Japanese Unexamined Patent Publication No. 11-288465 Japanese Patent No. 6305579

Y. Dodis, R. Ostrovsky, L. Reyzin, A. Smith. Fuzzy Extractors: How to Generate Strong Keys from Biometrics and Other Noisy Data. SIAM J. Comput. 38 (1): 97-139 (2008)

Therefore, an object of the present invention is, as one aspect, to provide a novel technique for reliably extracting unique data from an object.

The information processing method according to the first aspect of the present invention includes (A) a step of acquiring an input vector representing each of a plurality of objects, and (B) an input representing each of a plurality of objects in a space including the input vector. It includes a step of dividing the vector so that it is contained in different subspaces and generating a plurality of hypersurfaces used to generate a unique vector that identifies each of the plurality of objects.

The information processing method according to the second aspect of the present invention includes (C) a step of acquiring a vector representing a specific object included in the plurality of objects, and (D) each of the plurality of objects in a space including the above vector. A step of generating a vector that identifies a specific object by evaluating a vector that represents a specific object using a plurality of hypersurfaces that are divided and held in advance so that they are contained in different subspaces. Including.

According to one aspect, it is possible to reliably extract unique data from the target.

FIG. 1 is a diagram showing an outline of the system according to the first embodiment. FIG. 2 is a diagram showing a processing flow of the information processing apparatus according to the first embodiment. FIG. 3 is a diagram showing an example of a two-dimensional space. FIG. 4 is a diagram showing an example of a hypersurface surface. FIG. 5 is a diagram showing a processing flow of the hypersurface generation process. FIG. 6 is a diagram showing an example of a hypersurface surface. FIG. 7 is a diagram showing a processing flow of the frequency bias adjustment processing. FIG. 8 is a diagram showing an example of a hypersurface surface. FIG. 9 is a diagram showing a processing flow of the similarity adjustment processing. FIG. 10 is a diagram showing an example of a hypersurface surface. FIG. 11 is a diagram showing a processing flow of the distance adjustment process. FIG. 12 is a diagram showing a processing flow of the hypersurface generation process. FIG. 13 is a diagram showing a processing flow of adaptive distance adjustment processing. FIG. 14 is a diagram for explaining the setting of the target value. FIG. 15 is a diagram showing a processing flow of the similarity re-determination process. FIG. 16 is a diagram showing the positions of objects evaluated using an appropriate hypersurface surface. FIG. 17 is a diagram showing a processing flow of the detection device according to the first embodiment. FIG. 18 is a diagram showing an outline of the system according to the second to fourth embodiments. FIG. 19 is a diagram showing a processing flow of error correction data generation processing in the second embodiment. FIG. 20 is a diagram showing a processing flow of the detection device according to the second embodiment. FIG. 21 is a diagram showing a processing flow of error correction data generation processing in the third embodiment. FIG. 22 is a diagram showing a processing flow of the detection device according to the third embodiment. FIG. 23 is a diagram showing a processing flow of error correction data generation processing in the fourth embodiment. FIG. 24 is a diagram showing a processing flow of the detection device according to the fourth embodiment. FIG. 25 is a block configuration diagram of a computer device.

[Embodiment 1]
FIG. 1 shows an outline of the system according to the first embodiment. For example, an information processing device 100 installed at a production place, a collection place, a registration place, or the like and a plurality of detection devices 200 installed on a distribution channel, for example, have come to be connected via a network such as the Internet. There is.

As shown in FIG. 1, in a scene where a plurality of objects (A to D, etc. in FIG. 1) are produced, collected, registered, etc., the information processing apparatus 100 performs measurement or the like on the plurality of objects. , As a conversion function for generating unique data (hereinafter referred to as a fixed value vector) that identifies each of a plurality of objects, a space representing measurement data or the like is included in a different subspace for each of the plurality of objects. A hypersurface surface to be divided is generated, and the parameters of the hypersurface surface are distributed to the detection device 200.

The detection device 200, for example, measures a specific target on the distribution channel of the target, and then evaluates the measurement data or the like using the parameters of the hypersurface surface distributed from the information processing device 100. Generates a unique fixed value vector that identifies a particular object. Once the fixed value vector is obtained, attribute data such as data related to production or the like can be separately extracted by using the fixed value vector as an ID.

The information processing device 100 includes a measurement unit 110, a first data storage unit 120, a parameter generation unit 130, a second data storage unit 140, a registration unit 150, a distribution unit 160, and a target DB 170.

The measurement unit 110 performs a predetermined measurement on the target and stores the measurement data in the first data storage unit 120. The measurement unit 110 may perform not only the measurement but also additional processing such as feature extraction on the measurement data, and store the processing result in the first data storage unit 120. The parameter generation unit 130 uses the data stored in the first data storage unit 120 to generate a hypersurface surface that satisfies the requirements, and stores the parameters of the hypersurface surface in the second data storage unit 140.

The registration unit 150 evaluates the measurement data and the like stored in the first data storage unit 120 by using the parameters of the hypersurface stored in the second data storage unit 140, and sets a fixed value vector for each object. At the same time as generating, the attribute data of each target (for example, production place, production time, etc.) is associated and stored in the target DB 170. When the parameter generation unit 130 generates the fixed value vector of each object and stores it in the second data storage unit 140, the fixed value vector of each object stored in the second data storage unit 140 is used. It may be read. The distribution unit 160 distributes the parameters of the hypersurface surface stored in the second data storage unit 140 to the detection device 200.

The detection device 200 includes a measurement unit 210, a measurement data storage unit 220, an extraction unit 230, a post-processing unit 240, a setting unit 250, and a parameter storage unit 260.

The setting unit 250 stores the parameters of the hypersurface surface distributed from the information processing device 100 in the parameter storage unit 260. When the hypersurface parameter is received offline, it is similarly stored in the parameter storage unit 260.

The measurement unit 210 performs the same processing as the measurement unit 110 in the information processing device 100, and stores the processing result in the measurement data storage unit 220. The extraction unit 230 evaluates the measurement data and the like stored in the measurement data storage unit 220 using the parameters of the hypersurface stored in the parameter storage unit 260, and generates a unique fixed value vector that identifies a specific target. Generate. The post-processing unit 240 uses a fixed value vector as an ID to perform processing such as acquiring attribute data associated with the ID.

Next, the contents of the processing of the information processing apparatus 100 will be described in detail with reference to FIGS. 2 to 16.

First, the measurement unit 110 performs a predetermined measurement on each target i, generates an L1 dimensional vector Xi based on the measurement result, and stores it in the first data storage unit 120 (FIG. 2: step S1). ..

Here, it is assumed that there are L3 objects in total. Further, the measurement unit 110 may perform a process such as a predetermined feature extraction on the measurement data, not just the measurement. Then, the L1 dimensional vector Xi = (x ₁ , x ₂ , ..., X _L1 ) is generated for the i-th object among the three L objects.

In order to make the explanation easy to understand, an example of L3 = 4 and L1 = 2 is shown in FIG. Since L1 = 2, the vector Xi is represented in a two-dimensional space, that is, on a plane as shown in FIG. Here, the vector X ₁ = (x ₁ , x ₂ ) = (0.5, 1) of the object 1, the vector X ₂ = (x ₁ , x ₂ = (1, 0.5)) of the object 2, and the object 3 It is assumed that the vector X ₃ = (x ₁ , x ₂ ) = (0, 0.5) of the object 4 and the vector X ₄ = (x ₁ , x ₂ ) = (0.5, 0) of the object 4.

In this example, in the plane of FIG. 3, 0.5 is set as a threshold value, and if it is more than that, "1" is assigned, and if it is less than that, "0" is assigned (hereinafter, this is used). The function to be realized shall be expressed as round ()). In the plane of FIG. 3, such a threshold line is represented by a chain double-dashed line, and the objects 1 to 4 are respectively placed on the threshold line. Therefore, if there is noise in the measurement result even a little, "0" and "1" are exchanged, and a detection error occurs.

In the present embodiment, a hypersurface surface for dividing the L1 dimensional space into n subspaces is generated. When the value of each axis is classified into two types of 0 or 1, the relationship between n and the number L2 of the unique hypersurface is represented by ^{n = 2 L2.} L2 = log ₂ L3 or more, and n ≧ L3. The larger L2, the higher the noise immunity.

In the example of FIG. 3, for example, _{it aims to generate hypersurfaces y 1} and y ₂ (straight lines in the case of a plane) as shown in FIG. In the example of FIG. 4, L2 = 2 and n = 4, and the hypersurfaces y ₁ and y ₂ represented by the two dotted lines are generated. Then, each object is included in four subspaces A to D divided by these hypersurfaces, which are different from each other.

Hereinafter, the function corresponding to such a hypersurface is represented by f (), and the output range of f () is set to 0 or more and 1 or less. Also, apply the function round (), which rounds the threshold of binarization, quantization, or discretization to 0.5, that is, the first decimal place of the output of f (), to f (), and round (f ()). However, it can be extended to other cases as well.

Returning to the description of the process of FIG. 2, the parameter generation unit 130 next sets the number of hypersurfaces L2 and the like that divide the space according to the number of objects L3 (step S3). Since there is a relationship as described above, L2 and the like are determined accordingly. Further, the parameter generation unit 130 initializes the counter j for the hypersurface to 1 (step S5).

Then, the parameter generation unit 130 executes the hypersurface generation process (step S7). This hypersurface generation process will be described in detail later. The generated hypersurface parameters are stored in the second data storage unit 140.

After that, the parameter generation unit 130 increments the counter j by 1 and determines whether or not j is L2 or less (step S11). If this condition is satisfied, the hypersurface is still generated, so the process returns to step S7.

On the other hand, when j exceeds L2, the registration unit 150 evaluates each vector Xi on the hypersurface related to the generated parameter, generates a fixed value vector Zi of the target i, and generates the target DB 160 together with the attribute data. Register in (step S13). Regarding the fixed value vector Zi, the j-th value of Zi is obtained each time the j-th hypersurface is determined, or the j-th value of Zi determined at the same time when the j-th hypersurface is determined is targeted. It may be registered in the corresponding place in DB160. This is because the attribute data is later extracted from the fixed value vector Zi.

Here, using the set F () = (f ₁ (), f ₂ (), ..., f _{L2 ()) of the functions f 1 to} f _{L 2} representing the L2 hypersurface, the vector Yi = ( Calculate y _{i, 1} , y _{i, 2} , ..., y _{i, L2} ) = F (Xi) = (f ₁ (Xi), f ₂ (Xi), ..., f _L2 (Xi)) Then, by _{applying round () to each element y i, j} , it is discretized and a fixed value vector Zi is generated.

Further, the distribution unit 160 reads the generated hypersurface parameter from the second data storage unit 140 and distributes it to the predetermined detection device 200 (step S15). It may be distributed offline instead of being distributed, or it may be transmitted to the manufacturing department of the detection device 200 and recorded in the parameter storage unit 260 at the manufacturing stage of the detection device 200.

By performing such processing, the detection device 200 can reliably extract the fixed value vector without attaching an ID or the like to the target.

Next, the hypersurface generation process will be described in detail with reference to FIGS. 5 to 16.

First, the parameter generation unit 130 randomly generates a function fj () representing the j-th hypersurface that maps the entire vector Xi (FIG. 5: step S21).

Then, the parameter generation unit 130 executes the frequency bias adjustment process for _{f j () (step S23).} This frequency bias adjustment process will be described with reference to FIGS. 6 and 7.

_{It is assumed that the dotted lines y 1} = 0.5 and y ₂ = 0.5 representing the hypersurface as shown in FIG. 6 are generated with respect to FIG. 2, unlike FIG. Here, the dotted line y _j = 0.5 _{representing the hypersurface satisfies y j} = f _j (x ₁ , x ₂ ) = 0.5 (x ₁ , x ₂ ), and f ₁ (x ₁ , x 2). x ₂ ) maps the point on the upper right of the dotted line y ₁ = 0.5 to a value of 0.5 or more, which is classified as "1", and the point on the lower left of the _{dotted line y 1 = 0.5.} , Map to a value less than 0.5, classified as "0". Similarly, f ₂ (x ₁ , x ₂ ) maps the point above the dotted line y ₂ = 0.5 to a value less than 0.5, which is classified as "0", and the dotted line y ₂ = 0. The point at the lower right of .5 is mapped to a value of 0.5 or more, which is classified as "1".

In such a case, the number of objects is two on each side of the _{dotted line y 1 = 0.5, but the dotted line y 2} = 0.5 is in the upper left area (the area of round (y ₂ ) = 0). While four objects are included, the number is 0 in the lower right area (area of round (y _{2) = 1).} That is, there is a bias in the distribution of objects in the subspace divided by the hypersurface.

In such a case, for example, the parameters of the hypersurface (straight line y ₂ = f ₂ (Xi) = f ₂ (x ₁ , x ₂ ) = ax ₁ + b ₂ = c parameters a, b and c) are + or c, respectively. By changing in the − direction, for example, in the example of FIG. 6, if the dotted line y ₂ = 0.5 is moved in the upper left direction in parallel, the bias of the distribution can be corrected.

More specifically, as the target 2 and 4 close to the dotted line y ₂ = 0.5 belong to the lower right of the dotted line y ₂ = 0.5, the target 1 and 3 belong to the upper left of the dotted line y ₂ = 0.5 If you move the dotted line y ₂ = 0.5 to, the distribution will be equalized.

Since the objects 2 and 4 are X ₂ = (1,0.5) and X ₄ = (0.5,0), y _2,2 = f ₂ (X ₂ ) = a + 0.5b + c. A and b are weighted from b so that round (y _{2, 2} ) changes from 0 to 1, and all of a, b, and c are changed in the direction of increasing. Similarly, since y _2,4 = f ₂ (X ₄ ) = 0.5a + c, weight a and c are weighted from a so that round (y _{2,4) changes from 0 to 1.} Change to increase.

With such an example in mind, the frequency bias adjustment process shown in FIG. 7 is executed.

First, the parameter generation unit 130 initializes the local counter counter to 0 (step S51).

Further, the parameter generation unit 130 _{applies f j} () to each vector Xi ^{to generate a Y T} _j vector (step S53). The Y ^T _j vector shown below represents the vector in the jth row of the transposed matrix of Y, and Y is _{2 in the L3 row and L2 column with y i, j} = f _j (Xi) as the element in the i row and jth column. It is a dimensional matrix.
Y ^T _j = (y _{1, j} , y _{2, j} , ..., y _{L3, j} ) = (f _j (X ₁ ), f _j (X ₂ ), ..., f _j (X _L3 )) )

After that, the parameter generation unit 130 ^{divides each element of the Y T} _j vector into either 0 or 1 according to the threshold value, counts the frequencies of 0 and 1, and calculates the frequency difference bias (step S55). ..

In the example of FIG. 6, since the Y ^T ₂ vector is (0, 0, 0, 0), bias = | 4-0 | = 4. Here, with respect to the scalar value w, | w | represents the absolute value of w.

Then, the parameter generation unit 130 determines whether or not the frequency difference bias is smaller than the predetermined threshold value bias_threshold2 (step S57). If the distribution is completely uniform, the frequency difference bias = 0, so bias_threshold2 = 1. If this requirement is met, the process returns to the caller.

On the other hand, if the frequency difference bias is greater than or equal to the predetermined threshold value bias_threshold2, the parameter generation unit 130 determines whether or not the frequency difference bias is larger than another predetermined threshold value bias_threshold1 (step S59). .. If the frequency difference bias is too large, the hypersurface of this time cannot be tolerated. Therefore, when the requirement of step S59 is satisfied, the parameter generation unit 130 discards the hypersurface of this time (step S67) and calls the caller. Return to the processing of.

On the other hand, if the frequency difference bias is less than the predetermined threshold value bias_threshold1, the parameter generation unit 130 increments the local counter counter by 1 (step S61), and the value of the counter sets the predetermined threshold value bias_threshold3. It is determined whether or not the value has been exceeded (step S63). bias_threshold If the requirements of step S57 cannot be satisfied even after changing the parameters three times, the hypersurface surface this time is not appropriate. Therefore, when the requirement of step S63 is satisfied, the process proceeds to step S67.

On the other hand, if the value of counter is equal to or less than the predetermined threshold value bias_threshold3, the parameter generation unit 130 _{changes each parameter representing f j} () in a direction in which the frequency difference bias becomes smaller (step S65). After that, the process returns to step S53.

In this way, the frequency bias when the space is divided by the added hypersurface can be appropriately corrected.

Returning to the description of the process of FIG. 5, after step S23, the parameter generation unit 130 determines whether or not the hypersurface surface of this time has been discarded (step S25). If it is discarded, the process returns to step S21 and resumes from the generation of the hypersurface. On the other hand, if it is not discarded, the parameter generation unit 130 executes the similarity adjustment process (step S27). This similarity adjustment process will be described with reference to FIGS. 8 and 9.

_{It is assumed that the dotted lines y 1} = 0.5 and y ₂ = 0.5 representing the hypersurface as shown in FIG. 8 are generated with respect to FIG. 2, unlike FIG. Here, the dotted line y _j = 0.5 _{representing the hypersurface satisfies y j} = f _j (x ₁ , x ₂ ) = 0.5 (x ₁ , x ₂ ), and f ₁ (x ₁ , x 2). x ₂ ) maps the point on the upper right of the dotted line y ₁ = 0.5 to a value of 0.5 or more, which is classified as "1", and the point on the lower left of the _{dotted line y 1 = 0.5.} , Map to a value less than 0.5, classified as "0". Similarly, f ₂ (x ₁ , x ₂ ) maps the points to the left of the dotted line y ₂ = 0.5 to a value less than 0.5, which is classified as "0", and the dotted line y ₂ = 0. The points to the right of .5 are mapped to values of 0.5 or more, which are classified as "1".

In this case, when y ₁ and y ₂ are judged independently, the number of targets is two in each of the two regions divided by _{y 1} , and similarly, the number of targets is two in the two regions divided by _{y 2.} Since there are two each, there is no problem if it is judged by the frequency difference bias. However, the objects are not included in each of the subspaces obtained by dividing the entire space by the two hypersurfaces, but the objects 1 and 2 are included in the same subspace, and the objects 3 and 4 are in the same subspace. It has come to be included. That is, a biased distribution is formed.

To solve this problem, the similarity between hypersurfaces should be considered. Here, a Y ^T _j vector calculated for this hypersurface already similarity Similarity between Y ^T _j2 vectors for other hypersurface is set as the _second j, as the similarity between the hypersurface It is defined as follows.
Similarity (Y ^T _j , Y ^T _j2 ) = 2 · max (HD (Y ^T _j , Y ^T _j2 ), L3-HD (Y ^T _j , Y ^T _j2 )) / L3-1

max (A, B) is a function that outputs the larger value of A and B. HD (Y ^T _j , Y ^T _j2 ) is the ^{number of elements of the Y T} _j vector and the Y ^T _j2 vector whose values do not match when round () is applied. In the example of FIG. 8, if y ₁ is a hypersurface already set, the Y ^T _j2 vector becomes (1,1,0,0) when round () is applied, and if y ₂ is the hypersurface this time, Y ^T _j When round () is applied to the vector, it becomes (1,1,0,0), so HD (Y ^T _j , Y ^T _j2 ) = 0. Therefore, the Similarity is 2.max (0,4-0) / 4-1 = 1.

Regarding the similarity, the function as described above may be used for the element value of each vector being 0 or 1, and when it is a real number, the cosine similarity may be adopted. .. The cosine similarity is the absolute value of V1 · V2 / (| V1 || V2 |) for the vectors V1 and V2. Note that V1 and V2 represent the inner product of V1 and V2, and | V | represents the square root of the sum of squares of the values of each element of the vector.

The similarity adjustment process using such similarity will be described with reference to FIG.

First, the parameter generation unit 130 applies _{each of the set hypersurfaces f j2} () to each vector X _i ^{to generate a Y T} _j2 vector for each of _{j 2} (step S71). Further, the parameter generation unit 130 initializes the local counter counter to 0 (step S73).

The parameter generating unit 130, for all j ₂ (<j), to calculate the similarity Similarity between Y ^T _j vector and Y ^T _j2 vector, the similarity Similarity is less than the threshold sim_threshold1 a predetermined It is determined whether or not it is (step S75). If the requirement of step S75 is satisfied, there is no problem, so the process returns to the process of the caller.

On the other hand, if the requirement of step S75 is not satisfied, that is, if the similarity of even one of the hypersurfaces already set exceeds the threshold value sim_threshold 1, the parameter generation unit 130 will perform all j ₂ (<j). calculated for similarity similarity between Y ^T _j vector and Y ^T _j2 vectors, it determines whether it exceeds the threshold sim_threshold2 predetermined (step S77). This is because if the similarity is too high with any of the set hypersurfaces, the hypersurface is discarded without adjustment.

Therefore, when the condition of step S77 is satisfied, the parameter generation unit 130 discards the hypersurface surface this time (step S87). Then, the process returns to the process of the caller.

On the other hand, if the condition of step S77 is not satisfied, the parameter generation unit 130 increments the local counter counter by 1, and determines whether or not the value of this counter exceeds the predetermined threshold value sim_threshold3. (Step S81). If the requirements of step S75 cannot be satisfied even after changing the parameters three times, the hypersurface this time is not appropriate. Therefore, if the requirements of step S81 are satisfied, the process proceeds to step S87.

On the other hand, when the requirement of step S81 is not satisfied, the parameter generation unit 130 _{changes the parameter representing f j} () in a direction in which the similarity similarity becomes smaller (step S83).

More specifically, ^{among the elements (objects) of the Y T} _j vector and the Y ^T _j2 vector, the elements (objects) that match when round () is applied and are close to the threshold value (= 0.5) ( In order from the target), for example, when round () is applied to the element value, the elements i (target i) are selected by the number in which the similarity similarity becomes smaller than the threshold value sim_threshold1 due to the inversion of the values. Then, when f _j () and round () are applied to the selected element i, the parameters of the hypersurface are adjusted so that the values are inverted.

In the example of FIG. 8, assuming that the hypersurface this time is y ₂ , the objects 1 and 4 are selected because they are close _{to y 2 , and X 1} = (0.5, 1) and X ₄ = (0.5, At 0), round (y _{1, 2} ) = round (f ₂ (X ₁ )) = 1, and round (y _{4, 2} ) = round (f ₂ (X ₄ )) = 0, so each is 0. Adjust the parameters of the hypersurface so that it is inverted to 1 and 1. Then, round (Y ^T ₂ vector) becomes (0,1,0,1). Since round (Y ^T _j2 vector) is (1,1,0,0), HD (Y ^T _j , Y ^T _j2 ) = 2, and the similarity is 2.max (2,4-2). ) / 4-1 = 0, so the similarity is reduced.

When X ₁ of the _{object 1 is mapped with f 2} (x ₁ , x ₂ ), y ₁ , 2 = f ₂ (x ₁ , x ₂ ) = ax ₁ + bx ₂ + c = 0.5a + b + c. B and c are weighted and changed in the direction of decreasing all of a, b and c. Further, when X ₂ of the object 4 is mapped by _{f 2} (x ₁ , x _{2 ), y 4} , 2 = f ₂ (x ₁ , x ₂ ) = ax ₁ + bx ₂ + c = 0.5a + c, so a More weight is given to c to change a and c in the direction of increasing.

Then, the parameter generation unit 130 applies the _{function f j} () related to the changed parameter to each X _i ^{to generate a new Y T} _j vector (step S85). Then, the process returns to step S75.

By performing the above processing, the hypersurface is adjusted so that the individual objects are included in any one different subspace.

Returning to the description of the process of FIG. 5, after step S27, the parameter generation unit 130 determines whether or not the hypersurface surface of this time has been discarded (step S29). If it is discarded, the process returns to step S21. On the other hand, if it is not discarded, the parameter generation unit 130 executes the distance adjustment process (step S31). The distance adjustment process will be described with reference to FIGS. 10 and 11.

In the distance adjustment process, if the target mapping destination by the hypersurface exists near the hypersurface treated as the threshold, the result of round () may change due to noise. The parameters of the hypersurface are changed so as to increase the distance to the target mapping destination.

For example, it is assumed that the dotted lines y ₁ = 0.5 and y ₂ = 0.5 representing the hypersurface as shown in FIG. 10 are generated with respect to FIG. 2, unlike FIG. One of the objects is included in any of the subspaces obtained by dividing the space by such a hypersurface, but the objects 1 and 4 are very close to the _{dotted line y 2 = 0.5.} In such a case, it is _{preferable to rotate the dotted line y 2} = 0.5 as shown by the arrow because the objects 1 and 4 are separated from the _{dotted line y 2 = 0.5.}

FIG. 11 shows a processing flow of the distance adjustment processing.

First, the parameter generation unit 130 initializes the local counter counter to 0 (step S91). Then, the parameter generation unit 130 _{changes the parameter representing the hypersurface function f j} () so that ^{the difference between each element of the Y T} _j vector and the threshold value becomes large (step S93).

In the example of FIG. 10, because y _{1, 2} for the object 1 is smaller than 0.5, modify the parameters in the direction of y _{1, 2} is smaller than y _{4, 2} 0.5 for subject 4 Since it is large, the parameters are modified _{so that y 4, 2 becomes larger.}

More specifically, when X ₁ of the _{object 1 is mapped by f 2} (), y ₁ , 2 = f ₂ (X ₁ ) = ax ₁ + bx ₂ + c = 0.5 a + b + c, so b and c from a. Is weighted and all of a, b and c are changed in the direction of decreasing. Further, when X ₄ of the _{object 4 is mapped by f 2} (), y ₄ , 2 = f ₂ (X ₄ ) = ax ₁ + bx ₂ + c = 0.5 a + c. And c are changed in the direction of increasing.

If the parameters are modified, the mapping destination of the other object may approach the threshold value (hypersurface). In such a case, the mapping destination of the other object also deviates from the threshold value (hypersurface). Change the parameters so that.

Then, the parameter generation unit 130 _{applies the changed function f j} () to each X _i to generate a new Y ^T _j vector (step S95).

Then, the parameter generation unit 130 determines whether or not the difference between all or a certain number of elements and the threshold value is separated by a predetermined value or more ^{in the Y T} _{j vector (step S97).} It is for determining whether or not | y _{i, j} −0.5 | is equal to or greater than a predetermined value. Whether it is all elements or more than a certain number of elements shall be set in advance.

If the requirements of step S97 are satisfied, it means that a preferable hypersurface surface has been generated. Therefore, the parameter generation unit 130 sets the function f _j () of the hypersurface surface this time as satisfying the requirements (step S101). .. Then, the process returns to the process of the caller.

On the other hand, if the requirement of step S97 is not satisfied, the parameter generation unit 130 increments the local counter counter by 1 (step S99) and determines whether or not the value of the counter exceeds the threshold value dis_threshold (step S103). ). If the value of counter is equal to or less than the threshold value dis_threshold, the process returns to step S93. On the other hand, when the value of counter exceeds the threshold value dis_threshold, the process returns to the caller's processing.

By performing such processing, it is possible to adjust so that the mapping destinations of all objects or a certain number or more of objects are sufficiently separated from the threshold value (hypersurface surface). Even if this requirement cannot be satisfied, the hypersurface surface of this time may be utilized by performing the subsequent processing.

Returning to the description of the process of FIG. 5, after step S31, the parameter generation unit 130 determines whether or not the requirement satisfaction is set for the hypersurface surface this time (step S33). If the requirement satisfaction is set, the process returns to the caller's process (FIG. 2) via the terminal C in order to generate the next hypersurface. On the other hand, if the requirement satisfaction is not set, the process proceeds to the process shown in FIG. 12 via the terminal A.

Moving on to the description of the process of FIG. 12, the parameter generation unit 130 executes the adaptive distance adjustment process (step S35). This adaptive distance adjustment process will be described with reference to FIGS. 13 and 14.

In the adaptive distance adjustment process, when there is a set of objects that are not easily separated from the hypersurface, the frequency distribution after applying round () does not change significantly, and after round () of the objects included in the set. The parameter of the hypersurface is changed so that the value of is inverted. For example, the value after round () of about m objects is inverted from 0 to 1, and the value after round () of about m objects is changed from 1 to 0. The number of flips from 0 to 1 and the number of flips from 1 to 0 do not have to be exactly the same.

FIG. 13 shows the processing flow of the adaptive distance adjustment process.

First, the parameter generation unit 130 initializes the local counter counter to 0 (step S111). Further, the parameter generation unit 130 _{specifies a set S of the target i in which | f j} (X _i ) −0.5 | is equal to or less than a certain value (step S113). In this way, a set of objects i that are not easily separated from the hypersurface this time is specified.

_{Then, the parameter generation unit 130 calculates the target values C i and j} for the target i included in the set S (step S115).

The target value is set to 1-round (f _j (X _i )). This is because, as schematically shown in FIG. 14, when f _j (X _i ) is smaller than 0.5, round (f _j (X _i )) = 0, so the target value C _{i, j} Becomes 1. That is, the _{parameters are changed so that f j} (X _i ) is 0.5 or more.

On the other hand, when f _j (X _i ) is larger than 0.5, round (f _j (X _i )) = 1, so the target values C _{i and j} become 0. That is, the _{parameters are changed so that f j} (X _i ) is less than 0.5.

The processing up to this point is a precondition processing of the adaptive distance adjustment processing, and the set S and the target values C _{i and j} are not changed even if the processing after step S117 is performed.

Then, the parameter generation unit 130 sets the parameter representing the _{hypersurface function f j} () to have a small difference between _{the element value and the target value C i, j} for the target i included in the set S in the ^{Y T} _{j vector.} And so that the difference between the element value of the other object and the threshold value becomes large (step S117). This process is basically the same as in step S93, but the direction of reducing the difference from the target value is different for some elements related to the target.

Then, the parameter generation unit 130 applies a new _{hypersurface function f j} () to each X _i ^{to generate a new Y T} _j vector (step S119).

Then, the parameter generation unit 130 determines whether or not the difference between all or a certain number of elements and the threshold value is separated by a predetermined value or more ^{in the Y T} _{j vector (step S121).} This step is the same as in step S97.

If the requirement of step S121 is satisfied, the process returns to the caller's process in order to perform the next process.

On the other hand, if the requirement of step S121 is not satisfied, the parameter generation unit 130 increments the local counter counter by 1 (step S123) and determines whether or not the value of the counter exceeds the threshold value dis_threshold (step S125). ). If the value of counter is equal to or less than the threshold value dis_threshold, the process returns to step S117. On the other hand, when the value of counter exceeds the threshold value dis_threshold, the parameter generation unit 130 discards the hypersurface surface this time (step S127). Then, it returns to the processing of the caller.

In this way, it becomes possible to satisfy the requirements for distance while maintaining the frequency distribution. However, when the adaptive distance adjustment process is performed, the state obtained by the similarity adjustment process may no longer be destroyed.

Returning to the description of the process of FIG. 12, after step S35, the parameter generation unit 130 determines whether or not the hypersurface surface of this time has been discarded (step S37). If the hypersurface surface this time is discarded, the process returns to step S21 in FIG. 5 via the terminal B.

On the other hand, when the hypersurface surface this time is not discarded, the parameter generation unit 130 executes the similarity re-determination process (step S39). The similarity re-determination process will be described with reference to FIG.

Parameter generating unit 130, for all j ₂ (<j), to calculate the similarity Similarity between Y ^T _j vector and Y ^T _j2 vector, the similarity Similarity is less than the threshold sim_threshold1 a predetermined Whether or not it is determined (step S131). This is the same as step S75. If this requirement is met, there is no problem, so the process returns to the caller's process.

On the other hand, if the requirement of step S131 is not satisfied, that is, if the similarity of even one of the hypersurfaces already set exceeds the threshold value sim_threshold1, the parameter generation unit 130 discards the hypersurface this time ( Step S133) Then, the process returns to the process of the caller.

Returning to the description of the process of FIG. 12, after step S39, the parameter generation unit 130 determines whether or not the hypersurface surface of this time has been discarded (step S41). If it is discarded, the process returns to step S21 of FIG. 5 via the terminal B.

On the other hand, if the hypersurface this time is not discarded, the hypersurface this time is adopted, so the process returns to the caller's process.

By executing the above-mentioned hypersurface generation process, L2 hypersurfaces can be appropriately generated.

The hypersurface generation process described above is an example, and various deformations are possible. In particular, although an example of processing in which hypersurfaces are added one by one is shown, a plurality of hypersurfaces may be generated at once and similarity adjustment processing may be performed on them. The frequency bias adjustment process, similarity re-judgment process, distance adjustment process, etc. are executed in series and the parameters are changed each time. The parameters may be changed at once based on the result.

In the hypersurface generation process described above, when j = 1, the similarity adjustment process and the similarity redetermination process are not performed.

An example of preferable hypersurfaces y ₁ and y ₂ is shown in FIG. 4, but more specifically, since these hypersurfaces have a threshold value of 0.5, y ₁ = 0.5 = x ₂ + x ₁ −. It is 0.5, and y ₂ = 0.5 = x _2- x ₁ +0.5. The parameters of y ₁ are a = 1, b = 1, and c = −0.5. The parameters of y ₂ are a = -1, b = 1, and c = 0.5.

Then, if the input vectors X _{1 to} X _{4 of the objects 1 to 4 are evaluated using such y 1} and y ₂ , it becomes as shown in FIG. That is, the fixed value vector Z ₁ = (1, 1) is calculated for the target 1, the fixed value vector Z ₂ = (1, 0) is calculated for the target 2, and the fixed value vector Z ₃ = is calculated for the target 3. (0,1) is calculated, and a fixed value vector Z ₄ = (0,0) is calculated for the target 4. As a result, a unique vector that identifies each of the objects can be obtained from only the _{input vectors X 1 to} X ₄ obtained from the measurement results of the objects 1 to 4.

Next, the processing content of the detection device 200 will be described with reference to FIG.

First, the measurement unit 210 makes a predetermined measurement on a specific target, generates an L1 dimensional vector X based on the measurement result, and stores it in the measurement data storage unit 220 (FIG. 17: step S201). The measuring unit 210 performs the same processing as the measuring unit 110 of the information processing apparatus 100.

The extraction unit 230 evaluates the vector X on the hypersurface related to the parameter distributed from the information processing device 100 and stored in the parameter storage unit 260, and generates a target fixed value vector (step S203).

Since the parameters of f _j () (1 ≦ j ≦ L2) are stored in the _{parameter storage unit 260, Y'= (y 1} , y ₂ , ..., y _L2 ) = (f ₁ (X). ), F ₂ (X), ..., f _L2 (X)), and round () is applied to each of _{y j to obtain the fixed value vector Z.}

The extraction unit 230 outputs this fixed value vector to the post-processing unit 240, and the post-processing unit 240 extracts the attribute data corresponding to the fixed value vector from, for example, the target DB 170 in the information processing apparatus 100 via the network 300. Then, the process related to the attribute data and the like is executed (step S205).

As described above, in the present embodiment, the hypersurface surface related to the parameter stored in the parameter storage unit 260 is the same hypersurface surface for all objects, and no ID or the like is attached to each object. , You will be able to get a fixed value vector.

[Embodiment 2]
Although the first embodiment also has sufficient noise immunity, an error correction function may be added.

In the present embodiment, for example, as shown in FIG. 18, the information processing apparatus 100b further includes an error correction data generation unit 180 and a third data storage unit 190, and instead of the distribution unit 160, the parameters and errors of the hypersurface surface are used. It has a distribution unit 160b that distributes correction data.

Further, each detection device 200b further has an error correction data storage unit 270, and instead of the setting unit 250 and the extraction unit 230, stores the hypersurface parameters in the parameter storage unit 260 and in the error correction data storage unit 270. It has a setting unit 250b for storing error correction data and an extraction unit 230b having an error correction function.

Next, the processing content of the error correction data generation unit 180 of the information processing apparatus 100b will be described with reference to FIG. This process is performed, for example, before or after step S15 in FIG.

First, the error correction data generation unit 180 generates a fixed value vector Zi for each target i based on the generated hypersurface surface (step S301). This step is the same as a part of the process in step S13 of FIG.

That is, the vector Yi = (y _{i, 1} ) using the set F () = (f ₁ (), f ₂ (), ..., f _{L2 ()) of the functions f 1 to} f _{L 2} representing the hypersurface. , Y _{i, 2} , ..., y _{i, L2} ) = F (Xi) = (f ₁ (Xi), f ₂ (Xi), ..., f _L2 (Xi)). Further, the fixed value vector Zi is generated by discretizing each element y _{i, j of the vector Y i by applying round ().}

Then, the error correction data generation unit 180 generates redundant data for error correction for each fixed value vector Zi (step S303). Specifically, redundant data of the organization code having Zi as an information symbol is generated according to existing error correction codes such as BCH code, RS code, LDPC code, and turbo code.

For example, when the fixed value vector Zi = (1,1) is obtained for Xi = (0.5, 0.5), the redundant data Z _{r, according to a predetermined error correction code for Zi. i} = (0,0,1,0,1,0,0) is generated.

Then, the error correction data generation unit 180 performs machine learning using the input vector Xi as an input and the redundant data Z _{r and i} as teacher data to generate a trained model (step S305). For example, a trained model is generated by preparing a predetermined model based on a neural network and training with an existing algorithm. The error correction data generation unit 180 stores the parameters of the trained model in the third data storage unit 190.

The distribution unit 160b distributes the parameters of the trained model stored in the third data storage unit 190 to each detection device 200b (step S307). This data may also be distributed offline.

On the other hand, the detection device 200b executes the process as shown in FIG.

First, the measurement unit 210 makes a predetermined measurement on a specific target, generates an L1 dimensional vector X based on the measurement result, and stores it in the measurement data storage unit 220 (step S311). This step is the same as step S201 in FIG.

The extraction unit 230b is a hypersurface surface related to the parameters distributed from the information processing device 100b and stored in the parameter storage unit 260, evaluates the vector X, and generates a target fixed value vector (step S313). This step is the same as step S203 in FIG.

Further, the extraction unit 230b constructs a trained model from the parameters of the trained model stored in the error correction data storage unit 270, and generates redundant data for the vector X (step S315).

Then, the extraction unit 230b performs error correction processing on the fixed value vector and the redundant data to generate the final fixed value vector (step S317).

For example, even if X = (0.4,0.6) is obtained in step S313 and the fixed value vector Z = (1,0) is obtained for the same object as in the above example, in step S315. If redundant data (1,0,1,0,1,0,0) or data close to it is obtained, it can be corrected to Z = (1,1) by error correction processing.

In addition, after step S317, the process as in step S205 in FIG. 17 may be performed.

By distributing the above error correction data to the detection device 200b, the noise immunity to the measurement by the detection device 200b is increased.

[Embodiment 3]
Error correction data may be generated in a mode different from that of the second embodiment, and error correction processing may be performed using the data.

The processing content of the error correction data generation unit 180 according to the present embodiment will be described with reference to FIG. This process is performed, for example, before or after step S15 in FIG. However, L2 is set longer with redundancy.

First, the error correction data generation unit 180 generates a fixed value vector Zi for each target i based on the generated hypersurface surface (step S321). This step is the same as a part of the process in step S13 of FIG.

Then, the error correction data generation unit 180 performs machine learning for each element i of the fixed value vector by inputting the value of the element other than the element i and using the value of the element i as the teacher data, and the trained model. Is generated (step S323). For example, a trained model is generated by preparing a predetermined model based on a neural network and training with an existing algorithm. The error correction data generation unit 180 stores the parameters of the trained model in the third data storage unit 190.

Here, assuming that the fixed value vector Zi = (a, b, c, d, e, f, g, h, k) (a to k are all 1 or 0) is obtained, the first element "a" is obtained. Is used as the teacher data, and machine learning is performed with the remaining elements "b, c, d, e, f, g, h, k" as inputs, and the trained model g ₁ is generated. Further, machine learning is performed using the second element "b" as teacher data and the remaining elements "a, c, d, e, f, g, h, k" as inputs to obtain the trained model g ₂ . Generate. Similarly, L2 = 9th element "k" is used as teacher data, and machine learning is performed with the remaining elements "a, b, c, d, e, f, g, h" as inputs, and the learning has been completed. Generate model g _9.

Note that this is an example, and a plurality of elements may be used as teacher data, and the remaining elements of the plurality of elements may be used as inputs. For example, machine learning is performed with "a, b, c" as teacher data and the remaining elements "d, e, f, g, k" as inputs, and "d, e, f" is used as teacher data and the rest. Machine learning is performed by inputting the elements "a, b, c, g, h, k" of the above, "g, h, k" is used as the teacher data, and the remaining "a, b, c, d, e, f". It is also possible to transform it by performing machine learning with "" as an input. In short, we generate multiple trained models and use them to make complementary inferences about all the elements.

Then, the distribution unit 160b distributes the parameters of the trained model stored in the third data storage unit 190 to each detection device 200b (step S325). This data may also be distributed offline.

On the other hand, the process in the detection device 200b is shown in FIG.

First, the measurement unit 210 makes a predetermined measurement on a specific target, generates an L1 dimensional vector X based on the measurement result, and stores it in the measurement data storage unit 220 (step S311). This step is the same as step S201 in FIG.

Further, the extraction unit 230b generates a target first evaluation vector from the vector X based on the hypersurface surface related to the parameter stored in the parameter storage unit 260 (step S333).

In the present embodiment, since the parameter of f _j () (1 ≦ j ≦ L2) is stored in _{the parameter storage unit 260, Y'= (y 1} , y ₂ , ..., y _L2 ). = (F ₁ (X), f ₂ (X), ..., f _L2 (X)) is calculated and used as the initial first evaluation vector.

Further, the extraction unit 230b constructs a plurality of trained models from the parameters of the plurality of trained models stored in the error correction data storage unit 270, and applies the plurality of trained models to the first evaluation vector. Then, a second evaluation vector is generated (step S335).

For example, the first evaluation vector Y'= [0.9,0.1,0.5,0.3,0.8,0.5,0.7,0.1,0.6] is obtained. Assuming, in the above example, g ₁ (0.1, 0.5, 0.3, 0.8, 0.5, 0.7, 0.1, 0.6), g ₂ (0.9) , 0.5, 0.3, 0.8, 0.5, 0.7, 0.1, 0.6) ,. .. .. Calculate g ₉ (0.9, 0.1, 0.5, 0.3, 0.8, 0.5, 0.7, 0.1). Then, for example, the second evaluation vector Y ”= (0.8,0.1,0.5,0.3,0.8,0.5,0.7,0.1,0.6) is obtained. Be done.

Then, the extraction unit 230b generates a candidate vector from the first evaluation vector and the second evaluation vector (step S337). Here, for the same element of the first evaluation vector and the second evaluation vector, the one having a larger difference from the threshold value is adopted, or an arbitrary value between the same elements (a value determined according to a predetermined rule, for example, an experiment) is adopted. The function G ₂ (Y', Y ") that adopts) is applied.

For example, when Y'= [0.9,0.1,0.5,0.3,0.8,0.5,0.7,0.1,0.6], Y "= (0. If 8,0.1,0.5,0.3,0.8,0.5,0.7,0.1,0.6) is obtained, the only different value is the first element. Here, "0.9", which is the larger difference between 0.9 and 0.8, for example, the threshold value of 0.5, is adopted. Then, Y'''= (0.9, 0.1, 0.5, 0.3, 0.8, 0.5, 0.7, 0.1, 0.6) is obtained.

Then, the extraction unit 230b determines whether or not all the elements of the candidate vector Y'''' are separated from the threshold value by a predetermined value or more (step S339). For example, for each element, it is determined whether or not | element value − threshold value | ≧ predetermined value.

If this requirement is not satisfied, the extraction unit 230b sets the candidate vector as a new first evaluation vector (step S341). Then, the process returns to step S335 to generate a new second evaluation vector from the new first evaluation vector.

On the other hand, when the requirement of step S339 is satisfied, the extraction unit 230b performs discretization processing (round ()) on the candidate vector to generate a final fixed value vector (step S343).

After step S343, the process as in step S205 in FIG. 17 may be performed.

Even when the above error correction data is distributed to the detection device 200b, the noise immunity to the measurement by the detection device 200b is high.

If the requirements of step S339 are checked a plurality of times and the requirements are not satisfied, an error may be output.

[Embodiment 4]
Error correction data may be generated in a mode different from that of the second and third embodiments, and error correction processing may be performed using the error correction data.

The processing content of the error correction data generation unit 180 according to the present embodiment will be described with reference to FIG. 23. This process is performed, for example, before or after step S15 in FIG.

First, the error correction data generation unit 180 generates a fixed value vector Zi for each target i based on the generated hypersurface surface (step S351). This step is the same as a part of the process in step S13 of FIG.

That is, the vector Yi = (y _{i, 1} ) using the set F () = (f ₁ (), f ₂ (), ..., f _{L2 ()) of the functions f 1 to} f _{L 2} representing the hypersurface. , Y _{i, 2} , ..., y _{i, L2} ) = F (Xi) = (f ₁ (Xi), f ₂ (Xi), ..., f _L2 (Xi)). Further, the fixed value vector Zi is generated by discretizing each element y _{i, j of the vector Y i by applying round ().}

The distribution unit 160b distributes the fixed value vector Zi itself as error correction data to each detection device 200b (step S353). Regarding Zi, all the objects i (1 ≦ i ≦ L3) may be transmitted to each detection device 200b, or each detection is limited to the objects having low error correction capability in the embodiments other than the present embodiment. It may be delivered to the apparatus 200b and combined with an error correction method of another embodiment.

This data may also be distributed offline.

On the other hand, the detection device 200b executes the process as shown in FIG. 24.

First, the measurement unit 210 makes a predetermined measurement on a specific target, generates an L1 dimensional vector X based on the measurement result, and stores it in the measurement data storage unit 220 (step S361). This step is the same as step S201 in FIG.

Then, the extraction unit 230b evaluates the vector X on the hypersurface related to the parameter distributed from the information processing device 100b and stored in the parameter storage unit 260, and generates a target candidate vector (step S363). This step is the same as step S203 in FIG. 17, but is shown as a candidate vector because processing for error correction is performed below.

After that, the extraction unit 230b compares the fixed value vector stored in the error correction data storage unit 270 with the candidate vector, identifies the fixed value vector closest to the candidate vector, and outputs the fixed value vector (step S365).

Even if such processing is performed, the target fixed value vector can be surely obtained. In addition, the data size can be reduced compared to saving the Xi itself, and the data size can be reduced even for random data that has no feature in the Xi.

Although the embodiments of the present invention have been described above, the present invention is not limited thereto. For example, the processing flow is an example, and as long as the processing result does not change, the order of the steps may be changed or a plurality of steps may be executed in parallel.

At least a part of the information processing devices 100 and 100b and the detection devices 200 and 200b described above are computer devices, and as shown in FIG. 25, a memory 2501, a CPU (Central Processing Unit) 2503, and a hard disk. The drive (HDD: Hard Disk Drive) 2505, the display control unit 2507 connected to the display device 2509, the drive device 2513 for the removable disk 2511, the input device 2515, and the communication control unit 2517 for connecting to the network are connected to the bus 2519. It is connected with. The HDD may be a storage device such as a solid state drive (SSD). The operating system (OS) and the application program for performing the processing according to the embodiment of the present invention are stored in the HDD 2505 and read from the HDD 2505 to the memory 2501 when executed by the CPU 2503. .. The CPU 2503 controls the display control unit 2507, the communication control unit 2517, and the drive device 2513 according to the processing contents of the application program to perform a predetermined operation. Further, the data in the process of processing is mainly stored in the memory 2501, but may be stored in the HDD 2505. In an embodiment of the present technology, the application program for performing the above-described processing is stored and distributed on a computer-readable removable disk 2511 and installed from the drive device 2513 to the HDD 2505. It may be installed on the HDD 2505 via a network such as the Internet and a communication control unit 2517. Such a computer device realizes various functions as described above by organically collaborating with the hardware such as the CPU 2503 and the memory 2501 described above and the program such as the OS and the application program. ..

The data used by executing the processing as described above is stored in a storage device such as a memory 2501 or an HDD 2505 regardless of whether the data is in the middle of processing or is the processing result.

The above-described embodiments can be summarized as follows.

The information processing method according to the first aspect of the present embodiment includes (A) a step of acquiring an input vector representing each of a plurality of objects, and (B) each of the plurality of objects in a space including the input vector. It includes a step of dividing the input vector to be represented into different subspaces and generating a plurality of hypersurfaces used to generate a unique vector that identifies each of the plurality of objects.

By generating and holding a hypersurface surface having such properties, even in a situation where the target is not given an ID or the like or the target cannot be changed or processed, a later extraction phase (or detection) is performed. In the phase), it becomes possible to reliably extract unique data from the target.

The information processing method further includes (C) a step of outputting the parameters of the generated hypersurface surface to an apparatus that generates a unique vector that identifies the target with respect to at least one of the plurality of targets. It may be included. If the parameters of the hypersurface are output to such a device, it becomes possible to generate a unique vector that identifies the object with the device, and it is easy to extract attribute data and other post-processing based on the unique vector. Will be able to.

Further, the information processing method may further include (D) a step of generating data for error correction of the unique vector and outputting it to the apparatus. The noise in the measurement in the above device can be dealt with appropriately.

In the step of generating the plurality of hypersurfaces described above, a certain number or more of the plurality of hypersurfaces or the plurality of hypersurfaces map an input vector representing each of the plurality of objects on the hypersurface. It may be generated so that the distance between the point and the threshold value is separated by a predetermined value or more. Increased noise immunity.

Further, in the step of generating the plurality of hypersurfaces described above, the plurality of hypersurfaces may be generated so that their similarity to each other is less than the threshold value. You will be able to correct the distribution bias.

Further, in the process of generating data for error correction described above, (d1) a plurality of input vectors representing each of a plurality of objects are evaluated by using the generated plurality of super-curved surfaces. It represents a step of generating a unique vector for identifying each of the objects, (d2) a step of generating redundant data with an error correction code for each of the above unique vectors, and (d3) each of a plurality of objects. It may include a step of performing machine learning on a model that takes an input vector as an input and outputs redundant data for a unique vector about the target, and generates a trained model. It utilizes an already well-known error correction code.

Further, the process of generating data for error correction described above (d4) evaluates an input vector representing each of a plurality of objects by using a plurality of generated hypersurfaces, thereby performing a plurality of objects. The step of generating a unique vector that identifies each of the above, and (d5) some elements in the unique vector that identifies each of the plurality of objects are output, and elements other than the above-mentioned some elements in the unique vector are output. It may include a step of performing machine learning on a model of input and generating a plurality of trained models so that all the elements of the unique vector can be inferred. In this way, a trained model for complementarily reconstructing a unique vector that identifies an object can be obtained.

Further, the process of generating data for error correction described above (d6) evaluates an input vector representing each of a plurality of objects by using a plurality of generated hypersurfaces, thereby performing a plurality of objects. It may include steps to generate a unique vector that identifies each of the above. Since unique vectors for the number of targets are generated as error correction data, the data size may be large, but the processing load for generating error correction data is low.

The information processing method according to the second aspect of the present embodiment includes (A) a step of acquiring a vector representing a specific object included in a plurality of objects, and (B) a space including the above vectors. Generate a vector that identifies a specific object by evaluating a vector that represents a specific object using a plurality of hypersurfaces that are divided and held in advance so that each of them is contained in a different subspace. Including steps.

As described above, if the data of the hypersurface surface having the above-mentioned properties is obtained in advance, unique data can be reliably extracted from the target even in a situation where the target is not given an ID or the like. Then, by using the generated vector, for example, it becomes possible to acquire the associated attribute data and perform various processes.

The information processing method may further include (C) a step of performing error correction processing on the generated vector by using the data for error correction held in advance.

For example, when the data for error correction is a machine-learned trained model in which an input vector representing an object is input and redundant data with an error correction code is output for the object, as described above. The error correction process (c1) generates redundant data with an error correction code from an input vector representing a specific target by a trained model, and (c2) a vector that identifies a specific target based on the error correction code. And the generated redundant data, a step of generating a corrected vector that identifies a specific target may be included. By using such a trained model, redundant data for error correction can be restored, and an error in a vector that identifies an object can be corrected by utilizing an already well-known error correction code.

Further, when the data for error correction is a unique vector that identifies each of the plurality of objects, the error correction process described above is (c3) among the unique vectors that identify each of the plurality of objects. , May include a step of identifying a unique vector that is closest to the vector that identifies a particular object.

Further, the data for error correction has some elements in the vector for a certain object as an output, and elements other than the above-mentioned some elements in the vector for the certain object as inputs, and for a certain object. In the case of a trained model in which all the elements of the vector can be inferred, the generation step described above is (b1) from the first vector obtained by transforming the input vector by the plurality of supercurves. , The first step of generating the second vector by the trained model, and (b2) the second step of generating the candidate vector based on the element-by-element comparison between the first vector and the second vector, and ( b3) The third step of determining whether each element of the candidate vector has a difference from the threshold value in the dispersal processing for each element is equal to or more than a predetermined value, and (b4) each element of the candidate vector is element-by-element. When the difference from the threshold value in the dispersal processing of is greater than or equal to a predetermined value, the step of executing the dispersal process for each element of the candidate vector and (b5) each element of the candidate vector are When the difference from the threshold value in the dispersal processing for each element is less than a predetermined value, the candidate vector may be set as the first vector, and the steps of executing the first step to the third step may be included.

A program for causing a computer to execute the information processing method described above can be created, and the program is stored in various storage media.

Further, the information processing device that executes the information processing method as described above may be realized by one computer or a plurality of computers, and information processing may be performed by combining them. It shall be referred to as a system or simply a system.

100,100b Information processing device 110 Measuring unit 120 First data storage unit 130 Parameter generation unit 140 Second data storage unit 150 Registration unit 160, 160b Distribution unit 170 Target DB
180 Error correction data generation unit 190 Third data storage unit 200, 200b Detection device 210 Measurement unit 220 Measurement data storage unit 230, 230b Extraction unit 240 Post-processing unit 250 Setting unit 260 Parameter storage unit 270 Error correction data storage unit

Claims (15)

A step to get an input vector representing each of multiple objects,
It is used to divide the space containing the input vector so that the input vectors representing each of the plurality of objects are included in different subspaces, and to generate a unique vector that identifies each of the plurality of objects. Steps to generate multiple hypersurfaces and
A program to make a computer execute. A claim for causing a computer to further perform a step of outputting the generated parameters of the plurality of hypersurfaces to a device that generates a unique vector that identifies the target for at least one of the plurality of objects. 1 The program described. The program according to claim 2, wherein the computer further performs a step of generating data for error correction of the unique vector and outputting the data to the apparatus. In the step of generating the plurality of hypersurfaces,
The distance between the plurality of hypersurfaces or the point at which a certain number or more of the plurality of hypersurfaces map the input vector representing each of the plurality of objects on the hypersurface and the threshold value is separated by a predetermined value or more. The program according to any one of claims 1 to 3 generated in. In the step of generating the plurality of hypersurfaces,
The program according to any one of claims 1 to 4, wherein the plurality of hypersurfaces are generated so that their similarity to each other is less than a threshold value. The process of generating data for error correction is
A step of generating a unique vector that identifies each of the plurality of objects by evaluating an input vector representing each of the plurality of objects using the generated plurality of hypersurface surfaces.
For each of the unique vectors, a step of generating redundant data with an error correction code, and
A step of executing machine learning on a model that takes an input vector representing each of the plurality of objects as an input and outputs redundant data for the unique vector for the object to generate a trained model.
3. The program according to claim 3. The process of generating data for error correction is
A step of generating a unique vector that identifies each of the plurality of objects by evaluating an input vector representing each of the plurality of objects using the generated plurality of hypersurface surfaces.
Machine learning is executed on a model that outputs a part of the elements in the unique vector that identifies each of the plurality of objects and inputs elements other than the part of the unique vector. Steps to generate multiple trained models so that you can infer all the elements of a unique vector,
3. The program according to claim 3. The process of generating data for error correction is
3. The third aspect of the present invention includes a step of generating a unique vector that identifies each of the plurality of objects by evaluating an input vector representing each of the plurality of objects using the generated plurality of hypersurface surfaces. Described program. A step to get a vector representing a specific object contained in multiple objects, and
By dividing the space containing the vector so that each of the plurality of objects is included in a different subspace and using a plurality of hypersurfaces held in advance, the vector representing the specific object is evaluated. , A generation step that generates a vector that identifies the particular object,
A program to make a computer execute. The program according to claim 9, wherein the computer further executes a step of performing error correction processing on the generated vector using the data for error correction held in advance. When the data for error correction is a machine learning trained model in which an input vector representing an object is input and redundant data with an error correction code is output for the object.
The error correction process
A step of generating redundant data with the error correction code from the input vector representing the specific target by the trained model, and
A step of generating a corrected vector that identifies the specific target from the vector that identifies the specific target and the generated redundant data based on the error correction code.
10. The program according to claim 10. When the data for error correction is a unique vector that identifies each of the plurality of objects,
The error correction process
The program according to claim 10, further comprising a step of identifying a unique vector that is closest to the vector that identifies the specific object among the unique vectors that identify each of the plurality of objects. Data for error correction
To be able to infer all the elements of the vector for the object by using some elements in the vector for the object as the output and the elements other than the part of the elements in the vector for the object as the input. If there are multiple trained models generated in
The generation step
The first step of generating the second vector by the trained model from the first vector obtained by transforming the input vector by the plurality of hypersurface surfaces.
The second step of generating a candidate vector based on the element-by-element comparison between the first vector and the second vector, and
The third step of determining whether each element of the candidate vector has a difference from the threshold value in the discretization process for each element is equal to or larger than a predetermined value.
When each element of the candidate vector has a difference from the threshold value in the discretization process for each element of a predetermined value or more, the step of executing the discretization process for each element of the candidate vector. When,
When each element of the candidate vector has a difference from the threshold value in the discretization process for each element less than a predetermined value, the candidate vector is used as the first vector, and the first step to the third step are performed. Steps to perform steps and
9. The program according to claim 9. A means to obtain an input vector representing each of multiple objects,
It is used to divide the space containing the input vector so that the input vectors representing each of the plurality of objects are included in different subspaces, and to generate a unique vector that identifies each of the plurality of objects. A means to generate multiple hypersurfaces,
Information processing system with. A means to obtain a vector representing a specific object included in multiple objects, and
By dividing the space containing the vector so that each of the plurality of objects is included in a different subspace and using a plurality of hypersurfaces held in advance, the vector representing the specific object is evaluated. , A means of generating a vector that identifies the particular object, and
Information processing system with.

Images