JP2002329188A - Data analyzer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a data analyzer capable of forming clusters so as to attain practical classification without requiring manual operation even when the number of clusters is not previously known or there is no previous knowledge of data to be classified and allowed to be utilized for classification processing. SOLUTION: The data analyzer generates a degradation map by degrading the dimensions of a weighting vector of a lattice space map formed by SOM learning. Then the analyzer sets up a classification boundary in the lattice space map based on the degradation map and classifies the data to be classified in accordance with the classification boundary.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data analyzer for classifying and recognizing data such as images and tastes, and more particularly to an improvement in a classification procedure.

[0002]

2. Description of the Related Art In recent years, a new information processing technology that imitates information processing in the human brain has been developed. Neural networks and hierarchical module networks are known as systems that imitate the human brain, and research is being conducted in many fields in fields such as visual information processing, that is, image recognition. In this research, the process of judging the similarity of a series of image parts input from the image data acquisition unit corresponding to the retina based on its physical and image characteristics is called cluster classification or the like.

[0003] The cerebral neocortex, which controls information processing (object recognition, inference, learning, etc.) for the complex system of the brain, comprises 1000 to 2000 neurons. A column having a hierarchical structure is arranged in the visual cortex receiving information. This column is considered to be the basic unit of visual information processing in the brain, and it is known that each column responds to a specific visual pattern. Further, columns that react to the similar pattern are arranged close to each other. In other words, the neocortex has selectivity for similar input stimuli, and a specific region of the neocortex that realizes selectivity is called a "region".

[0004] More specifically, information transmission paths in the visual cortex, which is a territory responsible for information processing of the visual system, include V1 (primary visual cortex), which controls spatiotemporal analysis,
V2 (second visual cortex), V3 (tertiary visual cortex) and MT
(Dorsal path) via (Middle Temporal)
way) and V1, V2, V4 (fourth visual cortex) for performing shape analysis, PIT (Posterior Inferior Temporal), and ventral pathway through AIT (Anterior Inferior Temporal). It is known from anatomical studies that visual information is guided to the parietal association area through each path and integrated with the motor system. More specifically, in the flow of the ventral path, the initial vision such as the retina and the V1 and V2 of the visual cortex has functions such as determination of a receptive field (an area in a visual field covered by a column), color and line segment analysis. , And transmits that information to the upper layer. This function of initial vision corresponds to so-called pre-processing in image processing.

[0005] Input information passed through these low-order fields is V4
, And are converted into combination information of basic patterns (textures) based on the receptive field assigned to each column. This combination information is subjected to feature analysis as a partial image in a PIT having a receptive field wider than V4, and analyzed and recognized as an object by an AIT having a wider receptive field.

That is, in the process of realizing information processing imitating the brain using a computer system, it is necessary to use at least a first system for performing spatio-temporal analysis and a second system for shape analysis. . Conventionally, for these purposes, it is necessary to realize the following three functions as an apparatus for data analysis imitating the brain.

[0007] 1. 1. Realization of a column function model as a basic unit 2. Realization of a hierarchical territory model using columns as basic units Extraction of learning data and recognition data to column function model

[0008]

However, in the course of research for realizing each of these three functions, some functions can be realized by adjusting the recognition target to predetermined image data (such as face data). Although it is possible,
When trying to realize each function without limiting the object, it is not easy to model each function.

[Related Art] Therefore, as a simple method for providing a function equivalent to the above function, there is a method of classifying and organizing input image data on a grid space map. For this classification and organization, for example, self-organizing feature mapping (hereinafter abbreviated as SOM) (T. Kohonen. Self-organizin
g formation of topologically correct feature maps.
BiologicalCybernetics, 1982).

That is, from image data to be processed, partial image data for each recognition object (if it is face image data,
Then, the definition of the grid space map is changed for each of the partial image data. In a data processing device using this SOM, a plurality of feature values are calculated based on the partial image data, and a weight vector between a feature value set including the plurality of feature values and a grid arranged in a grid space is stored in a grid. Define as a spatial map. Here, a weight is defined for each of the feature amounts included in the feature amount set, and the sum of the weights of these components is referred to as a weight vector.

Initially, the weight vector is initialized in the following manner. That is, as shown in the above-mentioned Kohonen document, one input vector I to be referred to in the learning step is randomly selected from one input vector group to be learned (corresponding to a feature amount set here). , The weight vector of each lattice is initialized. Also, Kohone
According to n, an initial value may be set to each weight vector by a random number.

Next, learning of a weight vector is performed. In this learning process, a feature amount set for learning is generated,
A predetermined measure (for example, a Euclidean measure) between the learning feature amount set and the weight vector of each lattice in the lattice space is calculated. Then, among the grids, a grid having a maximum relationship (minimum measure) is found. Then, the weight vector is corrected for each of the lattices existing in the neighborhood of the lattice in the lattice space so that the measure with the learning feature amount set becomes small. By repeating the learning while performing the correction of the weight vector, the lattice having the minimum measure is concentrated on a specific region for the feature amount set including the feature amounts similar to each other, and the classification of the image data is performed. Applicable state.

However, when classifying data using the grid space map after the learning, a boundary line serving as a reference for classification is formed in the grid space, and the characteristic of data given as a classification target is obtained. It is appropriate to classify the data based on which boundary line the grid having the minimum measure for the quantity set belongs to (the area on the grid space separated by this boundary line is hereinafter referred to as a cluster). It is. That is, a method for determining the boundary of a cluster is required.

In order to determine a cluster boundary, a method has been considered in which each component of a weight vector for each lattice is referred to. Specifically, if it is face image data,
As features, amounts of energy, entropy, correlation, etc. are defined.For example, by referring to a weight component corresponding to energy, a group consisting of a grid having a predetermined threshold or less is defined as a group. It is divided into two groups: a group of lattices larger than the threshold. Then, for a group of lattices larger than the threshold value, clusters are formed on the lattice space by repeating the same processing with reference to the weight component corresponding to the entropy.

According to this method, it is necessary to consider in advance what kind of feature is to be used for the image data to be classified. In addition, as a general method of forming a cluster from a plurality of sample data, maximum likelihood estimation, K-mean
s method, LBG (Linde-Buzo-Gray) method, MDS (Multi-Dimensio
nal Scaling) method is known. However, these methods require an artificial operation such as planning the number of clusters from the beginning and adjusting an appropriate threshold value setting.

Further, since the number of elements of the feature quantity set obtained from the data to be analyzed is relatively small and the map analysis after learning is easy, the above simple cluster forming method can be applied. It is known that if the number of elements of a feature amount set is increased by expanding the classification target, simply applying the above-described cluster formation method cannot determine a boundary with high accuracy.

As described above, when the number of elements in the feature set increases, the boundary becomes complicated and complicated. U-Matrix (Unified Distance Mat)
rix Methods) and potential methods. Here, the U-Matrix method is described in A. Ultsch et.a
l. "Knowledge Extraction from Artificial Neural Ne
tworks and Applications ", Proc. Transputer Anwender
Treffen / World Transputer Congress TAT / WTC 93 Aac
hen, Springer 1993, for a detailed description of the potential method, see D. Coomans, DLMassart; Anal.
Ta., 5-3, 225-239 (1981) has a detailed description.

First, in the U-Matrix method, a distance between an arbitrary grid in a grid space map and a grid adjacent thereto is obtained, and the value is three-dimensionally re-mapped between the grids. Here, as the inter-grating distance, the sum of absolute values of the differences for each component of the weight vector prototyped to two mutually adjacent gratings, the mean square of the differences, and the like are used. In short, the basic idea of U-Matrix is to reduce the distance between grids that minimizes the measure for each similar feature set, and to three-dimensionally belong to a “valley”. I do. Conversely, for each different feature set,
The distance between the grids that minimizes the measure is increased so as to be relatively “mountain”. Then, determine the boundary along this mountain. In other words, the U-Matrix compensates for the fact that the SOM performs mapping based on the similarity of the vector amount (feature amount set or the like) and the distance in the input space is not preserved. The U-Matrix method can determine a relatively appropriate boundary without the need for human intervention when the contrast between peaks and valleys is clear.However, when the boundaries change slowly or the height of the peaks or valleys for each cluster If the depths are different, human intervention is required, which is beyond the scope of the related art.

Further, it can be said that the potential method is a boundary determination method based on so-called unsupervised learning.
In this method, using a predetermined function (potential function), each input data of all input data groups (groups of feature quantity sets) is approximated by a combination of function groups, and a probability density function as a result is estimated. How to As a result, a portion (trough) of the probability density function with less overlap is used as a boundary. Here, a Gaussian type is often used as the potential function. More specifically, when there is an input data group consisting of N input vectors, assuming that each of the input data groups has a K-dimensional size, the average potential that the l-th input data receives from other input data (where the l-th input is The contribution ratio to the entire input set) Ψl is expressed by the following (1),
(2) Defined by equation.

[0020]

(Equation 1)

Note that x _kl means the k-th component of the l-th input. Α affects the number of clusters classified by the smoothing parameter. Therefore, in the potential method, optimization of the distribution function assuming the distribution shape,
Optimization of various parameters is required for each input vector set. In short, knowledge of the characteristics of data to be classified is required in advance, and artificial adjustment is indispensable.

Further, based on these studies, Japanese Patent Application Laid-Open No. Hei 7-234854, "Cluster Classification Device", Japanese Patent Application Laid-Open No. 8-36557, "Cluster Classification Device", "Data density on self-organizing feature map" Unsupervised Cluster Classification Using Histogram ", IEICE Transactions
D-II Vol. J79-DII No.7 pp. 1280-1290 In July 1996, etc., as a development of the U-Matrix method,
A technique is disclosed that uses the distribution or the degree of integration of the number of grids that responds to an input (a measure becomes minimum for the input).
According to this technique, attention is paid to the tendency of SOM that a grid having a minimum measure is concentrated in a cluster, and it is expected that a sufficient effect can be obtained when the centroid distance between clusters is large. When the clusters are complicatedly entangled on the map and the boundary has a complicated shape, a sufficient effect cannot be obtained.

After all, in the range of the related art, when there is no guarantee of independence of data having a large number of elements or individual elements used for sensitivity classification such as image classification and wine taste, Web access log analysis and the like. Makes cluster classification difficult. That is, in this state, it is indispensable to perform the parameter adjustment artificially, and the application is limited.

The present invention has been made in view of the above-mentioned circumstances. Even when the number of clusters is not known in advance, or when there is no prior knowledge about the data to be classified, it is possible to use the data without any manual operation. It is an object of the present invention to provide a data analysis device capable of forming a cluster so as to achieve a general classification and contributing to a classification process.

[0025]

SUMMARY OF THE INVENTION The present invention for solving the above-mentioned problems of the prior art is based on n (n is 2 or more) feature data corresponding to each input data. Means for calculating a set of feature amounts, and means for generating a grid space map that defines a mapping coefficient from each component of the set of feature amounts to each grid forming a grid space. Learning is performed using input data, and as a result, the grid space map defined so that feature sets similar to each other are intensively mapped to a predetermined region in the grid space is held, and the grid In a data analysis device that classifies data using a spatial map, m is obtained by combining mapping coefficients for each grid corresponding to each component of the feature amount set under predetermined combining conditions. From the amount of synthesis, and the map degeneracy means for generating degenerate mapping to the grating on the grating space,
Boundary setting means for setting a classification boundary of data on the grid space map based on the degenerate mapping.

Here, when generating a synthetic feature, generation is performed with reference to a grid space map formed by learning. That is, the norm (Norm) of the weight vector (corresponding to the mapping coefficient) for each lattice, which forms the lattice space map, is defined, and a set of synthesis amounts for maintaining this norm distribution is generated.

It is preferable that the map degeneration means determines a synthesis condition of a synthesis amount based on a principal component analysis on a mapping coefficient. Further, the number m of the combined amount is determined by comparing the contribution ratio of each component during the principal component analysis and the change of the contribution ratio, or is determined according to the reproducibility of the original grid space map based on the combined amount. Preferably.

Further, the number m of the synthesis amount is a number m1 determined by comparing the contribution ratio of each component in the principal component analysis and the change of the contribution ratio, and the reproducibility of the original lattice space map based on the synthesis amount. It is also preferable to set the value to the smaller one of the number m2 determined accordingly.

Further, the selection of the feature amount subset may be performed on the input data for learning. That is, the present invention for solving the above-mentioned problem of the conventional example comprises means for calculating a feature amount set including n (n is 2 or more) feature amount data based on a plurality of input data; Means for generating a grid space map that defines a mapping from the set to each of the grids forming the grid space, wherein learning is performed in advance by using a plurality of learning input data, and as a result, mutually similar In a data analysis device that holds the grid space map defined so that a feature amount set is intensively mapped to a predetermined region in the grid space and classifies data using the grid space map, M (m is 1 or more and less than n) obtained by combining each component of the feature quantity set based on the input data for
And a means for learning and forming a degenerate lattice space map using the set of learning synthetic features, and a degenerate lattice space held as a result of the learning formation. And a boundary setting means for setting a classification boundary on the map.

Further, according to the present invention for solving the problems of the above-mentioned conventional example, a feature amount set consisting of n (n is 2 or more) feature amount data corresponding to each of the plurality of input data is provided. And a step of generating a grid space map that defines a mapping coefficient from each component of the feature amount set to each grid forming a grid space, using a plurality of learning input data in advance. The learning is performed, and as a result, data sets are classified using the grid space map defined so that feature sets similar to each other are intensively mapped to a predetermined area in the grid space. In the analysis method, m (m is 1 or more and less than n) synthesized amounts obtained by synthesizing mapping coefficients for each grid corresponding to each component of the feature amount set under predetermined synthesis conditions are used. A map degenerating step of generating a degenerate mapping to each grid in the grid space; and a boundary setting step of setting a data classification boundary on the grid space map based on the degenerate mapping. .

Further, according to the present invention for solving the above-mentioned problems of the conventional example, a computer calculates a feature amount set consisting of n (n is 2 or more) feature amount data based on a plurality of input data. And generating a grid space map that defines a mapping from the feature set to the grid space, and performing similarity as learning results performed in advance using a plurality of learning input data. A data analysis program for classifying data by using the grid space map defined so that the feature set to be mapped is concentrated on a predetermined area in the grid space. A map degeneration procedure for generating a degeneration mapping obtained by degeneration under the conditions of
Setting a data classification boundary on the grid space map based on the degenerate mapping.

[0032]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [First Embodiment] A first embodiment of the present invention will be described with reference to the drawings. still,
In the following description, for the sake of explanation, the case of image data is exemplified as data to be classified, but in actual use, not only image data, but also other sensory organ data such as taste data, Furthermore, measurement data such as experimental results may be used.

The data analyzer 1 according to the present embodiment
As shown in FIG. 1, the CPU 11, the RAM 12, and the RO
M13, a hard disk 14, an image input interface 15, a display 16, and an external storage unit 17, each of which is connected to a bus. That is, the data analysis device 1 of the present embodiment is realized by software using a general personal computer. This software is generally a CD-ROM
It is distributed in a state of being stored on a recording medium such as a DVD or a DVD-ROM, or downloaded via a network (a connection interface to the network is not shown). When distributed by the recording medium, the data is read out from the external storage unit 17 and stored in the hard disk 14 by a predetermined installation process. Also, when downloaded via a network, it is similarly installed on the hard disk 14.

The CPU 11 operates in accordance with the program stored in the hard disk 14 and basically operates under the control of an operating system such as Windows (registered trademark) to implement the data analysis apparatus 1 of the present embodiment. Execute analysis software. Specifically, the data analysis program according to the present embodiment includes an image processing unit 21, a selection processing unit 22, and a classification processing unit 23, as shown in FIG. Unit 31 and Gabor filter processing unit 32
And the selection processing unit 22 includes the image extraction processing unit 3
5 and a selection signal generation processing unit 36, and the classification processing unit 2
3 includes a feature amount generation unit 41, a learning processing unit 42, a boundary determination processing unit 43, and a classification execution unit 44. Here, it is assumed that each of these units is realized as a software module, but it may be configured by a logic circuit in hardware. This CPU
The processing in 11 will be described later in detail.

The RAM 12 is used as a work memory of the CPU 11, and is used to store various parameters and data during the processing of the CPU 11. ROM
Reference numeral 13 mainly stores programs that are required when the data analyzer 1 is started, such as processing for reading an operating system. Since the contents of the start-up program are widely known, description thereof will be omitted.

The hard disk 14 has an operating system main body and various programs installed therein. In the present embodiment, a data analysis program is installed on the hard disk 14 as described above. Although the case where the data is stored in the hard disk is illustrated here, for example, an SRAM (Static Random Access Memory)
Alternatively, the CPU 11 may be installed in a non-volatile memory such as an EEPROM, or, as shown in FIG. 1, may be installed in another computer connected via a network interface (not shown) even though it is not included in the same housing as the CPU 11. It may be installed.

An image input device such as a scanner is connected to the image input interface 15, and receives image data input from the image input device and outputs the image data to the CPU 11.
The display 16 responds to an instruction from the CPU 11
Display an image.

[Details of Processing] Here, the data analysis processing of the CPU 11 will be described with reference to the functional blocks of FIG. The data analysis process of the CPU 11 can be divided into a learning process and an actual classification process, and the same functional block is used in any process.

First, an outline of the processing of each block will be described.
Is converted into grayscale data, subjected to a predetermined conversion so that the features of the image to be learned and classified are clarified, and output to the selection processing unit 22. The selection processing unit 22 divides and extracts partial image data to be learned from the image data input from the image processing unit 21, and outputs the data to the classification processing unit 23. Classification processing unit 23
In the learning process, the SOM is performed based on the input partial image data to learn and form the lattice space mapping. In the classification process, the learned lattice space mapping is performed without changing the lattice space mapping. Is performed, the input partial image data is classified. When the process shifts from the learning process to the classification process, a boundary determination process is performed, and a boundary is set in the lattice space mapping.

Next, the operation of each unit will be described in more detail. The whitening processing unit 31 of the image processing unit 21 performs Fourier transform on the input grayscale image data to divide the grayscale image data into frequency components, removes the low frequency components, and determines the magnitude of the component for each frequency. After equalization, inverse Fourier transform is performed, and the result is output to the Gabor filter processing unit 32. This process simulates the information processing performed in the ganglion cells of the human retina. In the frequency components that make up the image, the contribution of each frequency component is made uniform, low frequencies are cut off, To reduce the effect of This clarifies the detailed features of the image.

The Gabor filter processing section 32 performs the Gabor filter processing according to the following equation (3), and outputs the result to the selection processing section 22.

[0042]

(Equation 2) here,

(Equation 3) Is a position vector indicating a point of interest in the image data,

(Equation 4) Is the vector of each pixel. Σ represents the variance. In this way, by applying the Gabor function obtained by modulating the harmonic function with the Gaussian function, it imitates the state equivalent to the human visual field where the point of interest of the image is more detailed than the surrounding area . In other words, the luminance variance in the peripheral portion of the image data to be processed is reduced, and the influence on the processing of the background portion is reduced.

The image extraction processing unit 35 of the selection processing unit 22
According to the selection signal input from the selection signal generation processing unit 36, partial image data is extracted from the image data input from the image processing unit 21 and output to the classification processing unit 23. The selection signal generation processing unit 36 outputs a signal for dividing the image data input from the image processing unit 21 into predetermined image blocks. Here, the image block may be obtained by dividing the image data into, for example, square blocks of 8 × 8 pixels, and is divided so as to overlap, for example, half the value of 4 pixels between adjacent blocks. I do.

It is preferable that the selection processing unit 22 extracts only significant blocks from the partial image data divided by the following equation (4) and outputs the extracted blocks to the classification processing unit 23. This equation (4) corresponds to an entropy calculation, and has a small value for partial image data corresponding to a background portion. For a face, a target portion (characteristic portion) such as an outline, an eye, a nose, and a mouth Is a large value. Therefore, by extracting only the one having a large IE in the equation (4) and outputting the extracted IE to the classification processing unit 23, a characteristic portion of the image data can be extracted.

[0045]

(Equation 5) Here, P _L is one of the pixels included in the partial image data.
This is the number of times a specific luminance level L appears.

The feature value generation unit 41 of the classification processing unit 23 calculates and outputs a plurality of feature values for each of the plurality of partial image data input from the selection processing unit 22 in accordance with a predetermined process. . A set of the plurality of feature amounts corresponds to a feature amount set. When each feature value is a scalar or the like, this feature value set is a vector of feature values. Here, a case will be described in which the feature amounts are arranged as vectors (feature amount vectors) in this manner.

The learning processing section 42 has two operation modes, a learning mode and a classification mode, and performs different processing in each operation mode. Specifically, when in the learning mode, the learning processing unit 42 manages a weight vector having the same dimension as the feature vector for each of the nodes (lattices) arranged in an M × M ′ lattice. Here, the two-dimensional arrangement in M × M ′ is merely for convenience, and is n-dimensional (n>
You may arrange in 2).

That is, in the learning mode, the learning processing unit 42 forms a lattice space map on the M × M ′ lattice space by SOM (self-organizing map) using the input feature amount vector as learning data. I do. That is, the distance between the input feature amount vector and the weight vector assigned to each lattice is calculated by a predetermined measure. Here, the Euclidean measure is used for simplicity. Then, a lattice (best matching node) c at which this distance is minimum is detected. Then, for the plurality of grids near the best matching node, the weight vectors are updated using the input feature amount vectors. Specifically, this update is performed by the following equation (5).

[0049]

(Equation 6)

The α (t), the A sigma (t), is a monotonically decreasing function of time, || r _c -r _j || includes a grating is the maximum matching node, the Euclidean distance between the lattice _j Represent. Also, t
Is “time”, and is incremented each time a feature amount vector is input (see the left side of Expression (5)).

By repeating the processing using the equation (5), a lattice space map is formed, and the most matching nodes for the feature amount vectors similar to each other form a continuous area. That is, in this lattice space, a non-linear projection from a feature vector, which is a multi-dimensional input signal, to a two-dimensional map is formed while maintaining the phase, and the feature part of the data is organized by updating the weights. As a result of the learning, a grid that responds to similar data is present in close proximity.

When the setting is changed from the learning mode to the classification mode, the learning processing unit 42 outputs an instruction to the boundary determination processing unit 43 to determine the boundary. In the classification mode, the learning processing unit 42 searches the input feature amount vector for the best matching node using the grid space map as a learning result, and outputs the best matching node to the classification execution unit 44. I do. At this time, the learning processing unit 42
Do not perform self-organization using equation (5).

The boundary determination processing unit 43 classifies each lattice into clusters according to the instruction to set the boundary input from the learning processing unit 42. The operation of the boundary determination processing unit 43 will be described later in detail.

The classification execution unit 44 operates when the learning processing unit 42 is in the classification mode, receives data to be classified (such as partial image data), holds the data, and receives In response to the input of the information of the matching node, the classification target data held is displayed on the display 16 together with information on which cluster the cluster most closely matched by the boundary determination processing unit 43 belongs to.
Etc. and output it as a classification result.

[Boundary Determination Processing] Here, the operation of the boundary determination processing unit 43 will be described. The boundary determination processing unit 43
As shown in FIG. 3, the principal component analyzer 51 and the filter 5
2 and a boundary analysis unit 53.
The principal component analysis unit 51 acquires information of a lattice space map (weight vector of each lattice), and performs principal component analysis (PCA; Principal Component Analysis) on each of the weight vectors.
is) yields a linear composition of the weight vectors,
Among the components of the composite weight vector, the components are rearranged in descending order of contribution ratio, and the result is output. Since this PCA is widely known, its detailed description is omitted. Note that the contribution rate is a general quantity defined in the principal component analysis, and specifically, the weight vector is (6)
When expressed as

(Equation 7) Assuming that a combined weight vector obtained by linear combination with respect to this is Expression (7),

(Equation 8) The set of coefficients of the linear combination is obtained as an eigenvector of a variance-covariance matrix based on the weight vector of each lattice, and a value obtained by dividing by a sum of corresponding eigenvalues corresponds to a contribution ratio.

The filter unit 52 determines a composite weight vector component having the largest contribution ratio among the composite weight vectors sorted in the order of contribution ratio, and a predetermined ratio (for example, the maximum contribution ratio) to the maximum contribution ratio. Is output up to the combined weight vector component having a contribution ratio of 10% or more.
That is, the filter unit 52 selectively outputs m (where m is equal to or more than 1 and the number of weight vector components is equal to or less than n) m contribution weights whose contribution rate is greater than a predetermined value. The component of the selected composite weight vector corresponds to the composite amount of the present invention, and the principal component analysis unit 51 and the filter unit 52 implement a map reduction unit that generates a reduction mapping. Although the contribution rate is used here, the eigenvalue itself related to the squared fraction may be used in the same manner.

The boundary analysis unit 53 uses the statistical values (intermediate value, average value, median value, etc.) of the components of the selected composite weight vector selectively output by the filter unit 52 as threshold values. , Different codes are given to those exceeding the threshold and those below the threshold. However, in the case of PCA, the average value is 0. SOM based on random learning data or high-dimensional learning data
It is preferable to use the median for. Here, as the code, for example, among the components of the composite weight vector, for the first component having the maximum contribution rate, the value of the first component is larger than the median value of the distribution of the first component in the lattice. “00” is assigned to “11” and the grid below the median (FIG. 4A). Similarly, for the second component having the second largest contribution ratio, “01” is assigned to a grid having a value of the second component larger than the median value, and “00” is assigned to a grid value having a median value or less.
(FIG. 4 (b)), and assigns “10” to the third grid having the third component larger than the median and “00” to the grid below the median (FIG. 4 (c)). . This process corresponds to a process of binarizing the grid space. Note that the filter unit 52 may set the boundary with a median value for the sum of the selected combinations of the components.

Then, the boundary analyzing unit 53 generates a 6-bit code by connecting the codes related to the first to third components for each lattice. Specifically, for the first component, "1
When “1” is assigned to the second component, “01” is assigned to the second component, and “10” is assigned to the third component.
110 "is generated. The boundary analysis unit 53 classifies the lattices having the same code after the connection into one cluster. That is, in the grid space, a boundary is set at a portion where the sign after connection to each of the adjacent grids is different.

As described above, the synthesis is performed so that the contribution ratio becomes large, and the boundary is set using the component selected according to the contribution ratio, so that the weight vector that defines the original grid space map is used. Practical classification can be realized with more accurate criteria.

The filter unit 52 may determine a component to be used for the boundary analysis among the combined weight vectors by the following method. From the composite weight vector component having the maximum contribution rate and the maximum contribution rate, the change rate of the contribution rate is determined in order, and the composite weight vector up to the component having a predetermined rate (for example, the change rate is 5% or more) is calculated. Output. This method is advantageous when the contribution or eigenvalue does not decrease monotonically. Specifically, for the magnitude of the contribution rate of the first component,
Assuming that the second component has a size of 30%, the third component has a size of 25% with respect to the first component, and the fourth component has a content of 10%, a component having a change rate of 5% or more; That is, up to the third component is used for the boundary analysis.

Further, the filter unit 52 may determine a component to be used for the boundary analysis in the composite weight vector by the following method. That is, initially, l is set to “1”, and the norm of the original weight vector (the magnitude at a predetermined measure)

(Equation 9) And l as a candidate from those having a large contribution rate
Is selected, and the norm of the vector obtained by inverting the vector of the synthesis amount consisting of l

(Equation 10) Is calculated, and based on these, how much of the original weight vector is reproduced (recall of the original weight vector) is calculated using the following equation (8).

[0062]

[Equation 11]

Then, 1 is incremented until the recall of the original weight vector becomes larger than a predetermined ratio. According to this method, the recall is high,
Classification becomes easier.

Further, the number of synthesis amounts to be selected according to the contribution ratio is m1, the number of synthesis amounts to be selected according to the change in the contribution ratio is m2, and is determined according to the recall of the weight vector. It is also preferable that the number of synthesis amounts to be selected be m3 and the smaller number be the number m to be selected. The m pieces determined in this way are selected in descending order of the contribution rate.

[Operation] According to the data analyzer according to the present embodiment, the norm of the original weight vector is reproduced at a predetermined ratio for the n-dimensional weight vector for each lattice in the lattice space obtained by the SOM learning. M defined as
An amount of dimension (m is 1 or more and less than n) is generated by synthesis or the like. Then, grid classification is performed based on this amount.
As a result, even if the number of clusters is not known in advance, or if there is no prior knowledge about the data to be classified, clusters can be formed so that practical classification can be achieved without any manual operation. Can contribute.

[Second Embodiment] In the description so far, the synthesis amount used for the boundary setting has been selected by principal component analysis on the grid space map already formed by learning. Principal component analysis is performed on the feature amount vector generated based on this, and similarly, m synthesized feature amounts smaller than the dimension n of the feature amount vector are generated and selected. It is also preferable to form a map (boundary map).

Further, the boundary map may be used as a grid space map without separately generating the boundary map. In this case, for a feature amount vector obtained from data to be classified input in the classification mode, a combined feature amount defined at the time of learning is generated, and the combined feature amount and a grid serving as a boundary map are generated. Classification is performed with reference to the spatial map.

[0068]

FIG. 5 shows a data analyzer 1 according to the first embodiment of the present invention, in which wines manufactured in three wineries A to C are set as a plurality of taste feature amounts (feature amount vectors). An example (a) in the case of classification and an example (b) in the case of manual classification using the MDS method are shown.

Since the meanings of the vertical and horizontal coordinates are different in FIGS. 5A and 5B, there is not necessarily a correlation between the corresponding positions. As shown in FIG. 5A, in the result of the classification by the data analysis device 1, the wine is classified into three wineries based on the taste feature amounts, whereas FIG.
As shown in (b), according to the MDS method, the feature amount vector group was divided into four regions. In detail, winery C was classified as having different characteristics while having common parts.

On the other hand, the data analyzer 1
Are obtained, and the winery C is divided into two. FIG.
In (a), the grid is defined as a continuous upper and lower side and a continuous left and right side. (Specifically, in a 37 × 37 grid space, a 1 × 1 grid is a 37 × 1 grid. Note that the grid is adjacent to the grid of 1 row and 37 columns.

[0071]

According to the present invention, for a plurality of input data, a feature amount set consisting of n (n is 2 or more) feature amount data corresponding to each input data is calculated, and each of the feature amount sets is calculated. A grid space map that defines mapping coefficients for each grid in the grid space from the components is generated, and learning is performed in advance by using a plurality of learning input data. As a result, feature sets similar to each other are obtained. A data analyzer that holds a grid space map defined to be intensively mapped to a predetermined region in a grid space and classifies data using the grid space map, wherein each component of a feature amount set is M (m is 1 or more and n is obtained by combining mapping coefficients for each grid corresponding to
), A degenerate mapping to each grid in the grid space is generated from the combined amount, and the data analysis device sets a data classification boundary on the grid space map based on this degenerate mapping. Even if the user does not know the data to be classified or does not have prior knowledge about the data to be classified, a cluster can be formed so as to achieve a practical classification without any manual operation, which can contribute to the classification process.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of a data analysis device according to an embodiment of the present invention.

FIG. 2 is a functional block diagram of a data analysis processing program in a CPU 11.

FIG. 3 is a functional block diagram of a boundary determination processing unit 43.

FIG. 4 is an explanatory diagram illustrating a boundary determination process in the data analyzer.

FIG. 5 is an explanatory diagram showing an analysis example by the data analysis device of the present invention.

[Explanation of symbols]

1 data analyzer, 11 CPU, 12 RAM, 1
3 ROM, 14 hard disk, 15 image input interface, 16 display, 17 external storage unit, 21 image processing unit, 22 selection processing unit, 23 classification processing unit, 31 whitening processing unit, 32 Gabor filter processing unit, 35 image extraction processing Unit, 36 selection signal generation processing unit, 41 feature amount generation unit, 42 learning processing unit,
43 boundary determination processing unit, 44 classification execution unit, 51 principal component analysis unit, 52 filter unit, 53 boundary analysis unit.

Continued on the front page (72) Inventor Noriji Kato 430 Nakai-cho Sakaigami-gun, Kanagawa Green Tech Naka Fuji Xerox Co., Ltd. F-term (reference) 5L096 FA19 GA55 GA59

Claims (12)

[Claims] 1. A means for calculating a feature amount set consisting of n (n is 2 or more) feature amount data corresponding to each input data for a plurality of input data, and a grid from each component of the feature amount set Means for generating a grid space map that defines a mapping coefficient for each grid that forms a space, comprising: performing learning using a plurality of learning input data in advance;
As a result, the grid space map defined so that feature sets similar to each other are mapped intensively in a predetermined area in the grid space is held, and data is classified using the grid space map. In the data analysis device, a degenerate mapping to each grid in the grid space is performed from m synthesis quantities obtained by synthesizing mapping coefficients for each grid corresponding to each component of the feature quantity set under predetermined synthesis conditions. A data analysis apparatus, comprising: map reduction means for generating; and boundary setting means for setting a classification boundary of data on the grid space map based on the reduction mapping. 2. The data analysis device according to claim 1, wherein the map degeneration means determines a synthesis condition of a synthesis amount based on principal component analysis of a mapping coefficient. 3. The data analyzer according to claim 2, wherein the number m of the combined amounts is determined by comparing the contribution ratio of each component or a change in the contribution ratio during principal component analysis. Data analyzer. 4. The data analysis device according to claim 2, wherein the number m of the combined amounts is determined according to the reproducibility of the original grid space map based on the combined amount. . 5. The data analysis apparatus according to claim 2, wherein the number m of the combined amounts is a number m1 determined by comparing the contribution ratio of each component or a change in the contribution ratio during principal component analysis, and the combined amount. A data analysis apparatus characterized in that the smaller one of a number m2 determined according to the reproducibility of the original grid space map based on the above is used. 6. A means for calculating a feature quantity set consisting of n (n is 2 or more) feature quantity data based on a plurality of input data, and mapping the feature quantity set to each grid forming a grid space. Means for generating a grid space map to be defined, and performing learning using a plurality of learning input data in advance,
As a result, the grid space map defined so that feature sets similar to each other are mapped intensively in a predetermined area in the grid space is held, and data is classified using the grid space map. In the data analysis device, m components obtained by combining components of a feature amount set based on the learning input data under predetermined combining conditions (m is 1 or more and n
And a means for learning and forming a reduced grid space map by using the set of learning combined feature quantities, wherein the reduced means is held as a result of the learning formation. A boundary setting means for setting a classification boundary on a grid space map. 7. A step of calculating a feature amount set including n (n is 2 or more) feature amount data corresponding to each input data for a plurality of input data; and a grid from each component of the feature amount set. Generating a grid space map that defines mapping coefficients for each of the grids forming the space, and performing learning using a plurality of learning input data in advance,
As a result, in a data analysis method for classifying data using the grid space map defined such that feature sets similar to each other are intensively mapped to a predetermined area in the grid space, M mapping coefficients obtained by synthesizing mapping coefficients for each grid corresponding to each component of the quantity set under predetermined synthesis conditions (m
A map reduction step of generating a degenerate mapping to each grid in the grid space from a synthesis amount of 1 or more and less than n); and a boundary for setting a data classification boundary on the grid space map based on the degenerate mapping. A data analysis method, comprising: a setting step. 8. The data analysis method according to claim 7, wherein in synthesizing the synthesis amount, a synthesis condition is determined based on principal component analysis of the mapping coefficient, and the number m of the synthesis amount is determined by calculating A data analysis method characterized in that the data analysis method is determined by comparing the contribution ratio of the synthesis amount or the change in the contribution ratio during the analysis. 9. The data analysis method according to claim 7, wherein in synthesizing the synthesis amount, a synthesis condition is determined based on principal component analysis of the mapping coefficient, and the number m of the synthesis amount is determined by the synthesis amount. A data analysis method characterized in that the data analysis method is determined according to the reproducibility of the original grid space map based on the data. 10. The data analysis method according to claim 7, wherein, when synthesizing the synthesis amount, a synthesis condition is determined based on principal component analysis of the mapping coefficient, and the number m of the synthesis amount is determined by calculating Either the number m1 determined by comparing the contribution ratio of each synthesis amount or the change in the contribution ratio at the time of analysis or the number m2 determined according to the reproducibility of the original lattice space map based on the synthesis amount is smaller. A data analysis method, characterized in that the value is set to one of the values. 11. A computer which calculates a feature amount set including n (n is 2 or more) feature amount data based on a plurality of input data, and defines a mapping from the feature amount set to a lattice space. And generating a grid space map to be executed.As a learning result performed in advance using a plurality of learning input data, feature sets similar to each other are concentrated in a predetermined area in the grid space. A data analysis program for classifying data by using the grid space map defined to be mapped by using the map space reduction method for generating a reduced map obtained by reducing the grid space map under predetermined conditions. And setting a classification boundary of the data on the grid space map based on the degenerate mapping. Analysis program. 12. A computer-readable recording medium storing the data analysis program according to claim 11.