JP5391876B2 - Representative feature extraction system, method and program - Google Patents

Description

  The present invention relates to a representative feature extraction system, method, and program for extracting feature data representing a data group from a data group consisting of a plurality of data.

  A representative feature extraction technique for extracting features representative of a data group from a large number of data groups is a technique applied to a wide range of fields such as image processing, image recognition, and data mining. For example, it is possible to compress an image by extracting representative features that well represent the characteristics of the entire image or partial image from the image data, and expressing the image as a difference between the representative feature and the representative feature. In addition, pixels that represent the pixel values of the sea surface are extracted from the image taken from the sky and used as representative features, and by detecting pixels that are different from the representative features, drowning people and drifting objects are detected. However, it can also be applied to maritime rescue. Furthermore, the behavior is analyzed using representative features by dividing the series of behavior patterns into partial patterns so that each becomes a similar behavior pattern, and extracting features that represent the partial patterns for each partial pattern. Application to action mining is also possible.

  As an example of representative feature extraction, the image processing apparatus described in Patent Document 1 (Japanese Patent Application Laid-Open No. 2007-142883) groups motion vectors into groups with similar motions, and obtains an average vector for each group. Thus, a representative vector representing each group is obtained. In the image segmentation method described in Patent Document 2 (Japanese Patent Laid-Open No. 11-167634), the RGB values of each pixel of the image are histogrammed to divide the image into regions of similar colors, and for each region. The average color or the most frequent color is obtained as the representative color. In the image processing apparatus described in Patent Document 3 (Japanese Patent Application Laid-Open No. 2008-066583), a local maximum value of a histogram in a color space is selected as a representative color, and is expressed in the representative color, thereby being used in the input image. Limited color processing is performed to reduce the number of colors that are used.

  As for data hierarchization in the related art, in the predictive encoded image data management system described in Patent Document 4 (Japanese Patent Laid-Open No. 9-294277), an image data string is separated into hierarchies according to the data structure. Then, a physical feature of the image data is extracted from the separated hierarchy, and a feature vector that characterizes the image is generated. Also, the similarity of the feature vectors is calculated, and the video is structured according to the similarity. In the multi-data division management method described in Patent Document 5 (Japanese Patent Laid-Open No. 2006-338298), division data is managed in a hierarchy in order to improve the efficiency of management of small-capacity data in a memory. In the document storage / retrieval apparatus described in Patent Document 6 (Japanese Patent Laid-Open No. 7-219968), the input document data is separated into a plurality of hierarchies and stored in a page turning form based on a visual viewpoint. When displaying document data, the target page can be found at high speed.

JP 2007-142883 A JP-A-11-167634 JP 2008-0665803 A JP-A-9-294277 JP 2006-338298 A Japanese Patent Application Laid-Open No. 7-219968

  Of the representative feature extraction techniques in the related art described above, Patent Documents 1 and 2 describe a method of obtaining an average of feature amounts to obtain a representative feature. However, the technique of extracting representative features using an average has a problem that it cannot cope with a feature value having an outlier. An outlier is a value that deviates significantly from the original value due to the influence of noise or the like. For the average, the influence of the feature amount that is far from the average greatly contributes, and therefore, when there is an outlier in the feature amount, an appropriate representative value cannot be obtained based on the feature amount average. .

  Patent Documents 2 and 3 describe methods for obtaining a maximum value and a maximum value in a histogram and using them as representative features. However, the method using the histogram has a problem that it cannot cope with sparse data. If the data is sparse and few data have the same value, a histogram cannot be created. In particular, when the dimension of the feature is high, the data is relatively sparse, and it is often difficult to create a histogram when using a high-dimensional feature.

  SUMMARY OF THE INVENTION An object of the present invention is to provide a representative feature extraction system that can cope with a case where there is an outlier in a feature amount or a case where data is sparse.

  Another object of the present invention is to provide a representative feature extraction method that can cope with the case where there is an outlier in the feature amount or the case where data is sparse.

  The representative feature extraction system of the present invention is a representative feature extraction system that selects representative features from input data, and includes a dividing unit that hierarchically divides input data into a plurality of partial data groups, and a hierarchy that has not been processed. Using the lowest hierarchy as the processing hierarchy, for each partial data group in the processing hierarchy, for each data contained in the partial data group, the sum of distances from that data to other data contained in the partial data group is calculated and evaluated An evaluation value calculation means for obtaining a value, a data selection means for selecting, as representative data, the data with the smallest evaluation value among the data included in the partial data group for each partial data group in the processing hierarchy; If it is an upper hierarchy, the representative data selected by the data selection means is the representative feature for the input data, and the processing hierarchy is a hierarchy other than the highest hierarchy. Data selection means having a aggregation means for generating a partial data group is selected representative data aggregation and processing hierarchy than one layer on the hierarchy.

  The representative feature extraction method of the present invention is a representative feature extraction method for selecting a representative feature from input data. The representative feature extraction method divides input data hierarchically into a plurality of partial data groups and The evaluation value is calculated by calculating the sum of the distance from the data to the other data included in the partial data group for each data included in the partial data group in each partial data group of the processing hierarchy with the lower hierarchy as the processing hierarchy Selecting the one with the smallest evaluation value as the representative data of the partial data group, and if the processing hierarchy is the highest hierarchy, making the selected representative data the representative feature for the input data, and processing When the hierarchy is a hierarchy other than the highest hierarchy, the selected representative data is aggregated to generate a partial data group one hierarchy above the processing hierarchy and obtain an evaluation value Having a repeating and selecting the representative data.

  According to the present invention, the representative feature can be stably selected even when the feature amount has an outlier. The reason is that the data having the minimum distance to other data in each data group is hierarchically set as data representing each data group. Since the data including outliers is excluded from the selection in the lower hierarchy, the representative feature is selected from the data including no outliers in the upper hierarchy. In the present invention, the representative feature can be obtained even if the data is sparse. The reason is that the data that minimizes the total distance to other data can always be obtained regardless of the number of data.

It is a block diagram which shows the structure of the representative feature extraction system of the 1st exemplary embodiment of this invention. It is a flowchart explaining operation | movement of the system shown in FIG. It is a block diagram which shows the structure of the representative feature extraction system of the 2nd exemplary embodiment of this invention. It is a flowchart explaining operation | movement of the system shown in FIG. It is a block diagram which shows the structure of the representative feature extraction system of the 3rd exemplary embodiment of this invention. It is a flowchart explaining operation | movement of the system shown in FIG. It is a block diagram which shows the structure of the representative feature extraction system of the 4th exemplary embodiment of this invention. It is a flowchart explaining operation | movement of the system shown in FIG. It is a block diagram which shows the structure of the representative feature extraction system of another example embodiment of this invention. It is a figure which shows an example of an input image. It is a figure which shows an example of the division | segmentation method. It is a figure which shows an example of a hierarchical structure. It is a figure which shows an example of a process target. It is a figure which shows an example of a process target. It is a figure which shows an example of a clustering result. It is a figure which shows an example of the representative data for every partial image. It is a figure which shows an example of a process target. FIG. 10 is a block diagram illustrating a configuration of a representative feature extraction system according to a fourth embodiment.

  Next, embodiments for carrying out the invention will be described in detail with reference to the drawings.

  The representative feature extraction system of the first exemplary embodiment of the present invention shown in FIG. 1 includes an input unit 11 that reads a data group to be processed, a data group storage unit 12 that stores the read data group, and a data group storage. The evaluation value calculation unit 13 for obtaining an evaluation value for each data in the data group stored in the unit 12 and the evaluation value calculation unit 13 from among the data in the data group stored in the data group storage unit 12 A data selection unit 14 that selects the data having the smallest evaluation value as a representative feature of the data group, and an output unit 15 that outputs the data selected by the data selection unit 14 to the outside as the representative feature of the data group. ing. Here, the evaluation value calculation unit 13 calculates a distance representing the degree of similarity with other data in the data group for each data in the data group, and obtains an evaluation value of each data using the total of the distances. The input unit 11, the data group storage unit 12, the evaluation value calculation unit 13, the data selection unit 14, and the output unit 15 function as an input unit, a storage unit, an evaluation value calculation unit, a data selection unit, and an output unit, respectively.

  Next, the flowchart of FIG. 2 is demonstrated about operation | movement of the representative feature extraction system shown in FIG.

  First, in step A 1, the input unit 11 reads a data group to be processed and stores it in the data group storage unit 12. Next, in step A2, the evaluation value calculation unit 13 calculates a distance representing the degree of similarity of each data in the data group in the data group storage unit 12 with other data in the data group. The evaluation value of each data is obtained using the total. Thereafter, in step A3, the data selection unit 14 selects, as the representative feature of the data group, the data having the smallest evaluation value obtained in step A2 for each data in the data group in the data group storage unit 12. In step A4, the output unit 15 outputs the data selected in step A3 as a representative feature. As will be described later, for example, the evaluation value is determined so as to be smaller as the total distance to other data is smaller. Thus, the total distance is small, that is, the degree of similarity with other data in the data group. Data having a large value is selected as a representative feature.

  The representative feature extraction system of the first exemplary embodiment can also be realized by using a computer such as a personal computer or a workstation and executing a program on the computer. These computers have a camera or the like connected as an input unit 11 and a display or the like connected as an output unit 15, and generally have a central processing unit (CPU) and a hard disk device for storing programs and data. And an input device such as a keyboard and a mouse for inputting commands and the like, a reading device for reading a recording medium such as a CD-ROM, and an interface used for connection to an external network. The data group storage unit 12 is configured by securing a storage area on the main memory or the hard disk device. Then, the computer implements the representative feature extraction system of the first exemplary embodiment by causing the computer to read and execute the computer program for realizing the functions of the evaluation value calculation unit 13 and the data selection unit 14 described above. It will be done. The program is read into a computer from a recording medium such as a CD-ROM or via a network.

  Next, a representative feature extraction system according to a second exemplary embodiment of the present invention will be described. In the second exemplary embodiment, the data is hierarchically divided into a plurality of partial data groups, and each partial data group has the data having the smallest total distance from other data in the partial data group This is characterized in that data representing a data group is selected, and a representative feature is selected by hierarchically executing a process of selecting such representative data.

  The representative feature extraction system of the second exemplary embodiment shown in FIG. 3 includes an input unit 21 that reads data to be processed (input data), and a data division unit 22 that hierarchically divides the input data into a plurality of partial data groups. A partial data group storage unit 23 for storing a partial data group for each layer, an evaluation value calculation unit 24 for obtaining an evaluation value for each data in each partial data group for each layer, and each partial data group for each layer The data selection unit 25 for selecting the representative data, the data aggregating unit 26 for aggregating the data selected by the data selection unit 25, and the output that outputs to the outside data representative of the input data by the data aggregating unit 26 Part 27. Here, the evaluation value calculation unit 24 represents the degree of similarity of each data in each partial data group of the processing target hierarchy stored in the partial data group storage unit 23 with other data in the partial data group. The distance is calculated, and the evaluation value of each data is obtained using the total distance. The data selection unit 25 represents the data with the smallest evaluation value by the evaluation value calculation unit 24 from the data of each partial data group of the processing target hierarchy stored in the partial data group storage unit 23 and represents the partial data group. Select as data. If the representative data selected by the data selection unit 25 is in the highest hierarchy, the data aggregating unit 26 uses the representative data as a representative feature for the input data, and if the representative data is in a hierarchy other than the highest, A partial data group of a hierarchy one level higher than the hierarchy is generated from the representative data and stored in the partial data group storage unit 23.

  In this configuration, the input unit 21, the data division unit 22, the partial data group storage unit 23, the evaluation value calculation unit 24, the data selection unit 25, the data aggregation unit 26, and the output unit 27 are respectively input means, division means, It functions as storage means, evaluation value calculation means, data selection means, aggregation means, and output means.

  Next, the flowchart of FIG. 4 is demonstrated about operation | movement of the representative feature extraction system shown in FIG.

  First, in step B1, the input unit 21 reads data to be processed (input data), and in step B2, the data dividing unit 22 hierarchically divides the input data into a plurality of partial data groups, and a partial data group storage unit 23. Store in. Then, in step B3, the evaluation value calculation unit 24 sets the lowest hierarchy among the plurality of hierarchies that have not been processed at that time as the processing hierarchy, and sets each processing hierarchy in the partial data group storage unit 23. For each partial data group, for each data in the partial data group, a distance representing similarity to other data in the partial data group is calculated, and an evaluation value of each data is obtained using the total distance. The unprocessed hierarchy here is a hierarchy in which evaluation value calculation and data selection processing described below are not completed at that time. Thereafter, in step B4, the data selection unit 25, for each partial data group of the processing hierarchy at that time in the partial data group storage unit 23, the data with the smallest evaluation value obtained in step B3 is assigned to the partial data group. Select as representative data. In step B5, the data aggregation unit 26 determines whether the current processing hierarchy is the highest hierarchy. If it is determined in step B5 that the layer is the highest layer, the data aggregation unit 26 sets the representative data in the processing layer as the representative feature for the input data, and the output unit 27 sets the representative feature as the external feature in step B7. Output to. On the other hand, if it is determined in step B5 that it is not the highest hierarchy, in step B6, the data aggregating unit 26 aggregates representative data of each partial data group of the current processing hierarchy, so that the current processing hierarchy A partial data group belonging to a hierarchy one level above is generated and stored in the partial data group storage unit 23. Thereafter, the processing from step B3 is repeated in order to repeat the above-described processing for a layer one layer higher than the current processing layer. Therefore, the partial data group of the hierarchy one level generated by the data aggregation unit 26 is a target of the next evaluation value calculation by the evaluation value calculation unit 24.

  As described above, in the second exemplary embodiment, the data dividing unit 22 that hierarchically divides input data into a plurality of partial data groups, the partial data group storage unit 23 that stores each partial data group for each hierarchy, By providing a data aggregating unit 26 that collects representative data of each partial data group and creates a partial data group of the next hierarchy for each, so that evaluation value calculation and data selection for representative data can be performed hierarchically. Become. The evaluation value calculation unit 24 and the data selection unit 25 perform the same processing as the evaluation value calculation unit 13 and the data selection unit 14 in the first exemplary embodiment on each partial data group, thereby making each partial data group Select representative data.

  In such a second exemplary embodiment, the overall calculation amount can be reduced by performing evaluation value calculation for each hierarchical partial data group. In addition, since the data including outliers can be removed at the stage of processing the lower hierarchy, the influence of the outliers can be reduced and the representative feature can be selected stably.

  The representative feature extraction system of the second exemplary embodiment can also be realized by executing a program on a computer, as in the case of the first exemplary embodiment. In this case, the partial data group storage unit 23 is configured by securing a storage area on the main memory or hard disk device of the computer. The program executed by the computer is a program that realizes the functions of the data dividing unit 22, the evaluation value calculating unit 24, the data selecting unit 25, and the data aggregating unit 26 described above.

  Next, a representative feature extraction system according to a third exemplary embodiment of the present invention will be described. The representative feature extraction system of the third exemplary embodiment shown in FIG. 5 is an occurrence distribution storage unit 15 that stores an occurrence distribution of feature quantities assumed to be input in the system of the first exemplary embodiment shown in FIG. And the occurrence distribution of the feature quantity is given to the evaluation value calculation unit 13 from the occurrence distribution storage unit 15. The occurrence distribution storage unit 15 functions as an occurrence distribution storage unit.

  In the present exemplary embodiment, the evaluation value calculation unit 13 calculates an evaluation value using the occurrence distribution of the feature amount assumed to be input. The occurrence distribution of the feature quantity assumed to be input is a distribution of values that each data is expected to take in the data group input to the input unit 12. Specifically, as in the case of the first exemplary embodiment, for each data in the data group, a distance representing the similarity to the other data in the data group is calculated, and the total of the distances is used to calculate each distance. Find the evaluation value of the data. At this time, it is assumed that the distance value between the two data is virtually generated based on the occurrence distribution stored in the occurrence distribution storage unit 15, and the virtually generated data is currently compared. It is expressed by the probability of being between two data. In other words, assuming that a pattern is virtually generated from the occurrence distribution of the feature amount, the distance from one data A in the data group to another data B in the data group is greater than the data B when viewed from the data A. Is the probability that a close pattern will be observed. Then, the evaluation value calculation unit 13 directly uses this distance value or the sum thereof as an evaluation value. Alternatively, the evaluation value calculation unit 13 determines the evaluation value from the distance value using an appropriate function that increases the evaluation value if the distance value is large.

  Next, the flowchart of FIG. 6 is demonstrated about operation | movement of the representative feature extraction system shown in FIG.

  First, in step A 1, the input unit 11 reads a data group to be processed and stores it in the data group storage unit 12. Next, the evaluation value calculation unit 13 uses the occurrence distribution stored in the occurrence distribution storage unit 15, and in step A <b> 2 a, for each data in the data group in the data group storage unit 12, A distance value representing the degree of similarity with other data is obtained as described above, and an evaluation value of each data is obtained using the total distance. Thereafter, in step A3, the data selection unit 14 selects, as the representative feature of the data group, the data having the smallest evaluation value obtained in step A2 for each data in the data group in the data group storage unit 12. In step A4, the output unit 15 outputs the data selected in step A3 as a representative feature.

  According to the third exemplary embodiment, the occurrence distribution in the feature quantity is taken into consideration, so that the pattern virtually generated from the occurrence distribution is more than the data including outliers as seen from some data. The probability of being close to data increases and approaches 1. That is, when viewed from some data, data including outliers is farther than other data. Considering the original meaning of the occurrence distribution of the feature amount, the probability 1 is the same evaluation as when there is no feature. Therefore, the influence of outliers on the evaluation values of other data is reduced. For this reason, even if an outlier is included in the feature amount of the data, more appropriate representative feature extraction can be performed by performing the processing of the present exemplary embodiment.

  The representative feature extraction system of the third exemplary embodiment can also be realized by executing a program on a computer, as in the case of the first exemplary embodiment. In this case, the data group storage unit 12 and the occurrence distribution storage unit 16 are configured by securing a storage area on the main memory or hard disk device of the computer.

  Next, a representative feature extraction system according to a fourth exemplary embodiment of the present invention will be described. The representative feature extraction system of the fourth exemplary embodiment shown in FIG. 7 is obtained by adding a clustering unit 28 and a cluster representative extraction unit 29 to the system of the second exemplary embodiment shown in FIG. The clustering unit 28 and the cluster representative extraction unit 29 function as clustering means and cluster representative extraction means, respectively. The clustering unit 28 is provided on the output side of the data aggregation unit 26 and divides the representative feature selected as the representative feature of the partial data group into a plurality of clusters having similar properties. The cluster representative extraction unit 29 extracts representative features representing each cluster from the clusters divided by the clustering unit 28, and sends them to the output unit 29. As a result, the output unit 29 outputs the representative feature representing each cluster as the representative feature of the input data.

  Next, the operation of the representative feature extraction system shown in FIG. 7 will be described with reference to the flowchart of FIG.

  From step B1 to step B6, processing is executed in the same manner as shown in FIG. Then, after the execution of step B5, when the processing hierarchy at that time is the highest hierarchy, in step C1, the clustering unit 28 divides (clusters) the input data into clusters having similar feature values. . In step C2, the cluster representative extraction unit 29 extracts a representative feature for each cluster created by the clustering unit 28. Thereafter, in step C3, the output unit 27 outputs the representative feature extracted in step C2 to the outside.

  According to the fourth exemplary embodiment as described above, by using the clustering technique, it is possible to select the representative feature corresponding to each data group when the input data is composed of data groups having a plurality of properties. become.

  Similarly to the case of the first exemplary embodiment, the representative feature extraction system of the fourth exemplary embodiment can also be realized by executing a program on a computer. In this case, the partial data group storage unit 23 is configured by securing a storage area on the main memory or hard disk device of the computer. A program executed by the computer is a program that realizes the functions of the data dividing unit 22, the evaluation value calculating unit 24, the data selecting unit 25, the data aggregating unit 26, the clustering unit 28, and the cluster representative extracting unit 29 described above. .

  FIG. 9 is a block diagram showing a configuration of a representative feature extraction system of still another exemplary embodiment of the present invention.

  The representative feature extraction system shown in FIG. 9 is the same as the representative feature extraction system of the second exemplary embodiment shown in FIG. 3, but includes an input unit 21, a partial data group storage unit 23, and an output unit 27. This is different from that shown in FIG. That is, this configuration includes a data dividing unit 22 that hierarchically divides input data into a plurality of partial data groups, an evaluation value calculating unit 24 that obtains an evaluation value for each data in each partial data group for each hierarchy, A data selection unit 25 that selects representative data of each partial data group and a data aggregation unit 26 that aggregates the data selected by the data selection unit 25 are provided. The data dividing unit 22 gives each partial data group of the lowest hierarchy to the evaluation value calculating unit 24. The evaluation value calculation unit 24 calculates, for each data in each partial data group in the hierarchy to be processed, a distance representing the similarity to other data in the partial data group, and uses each sum of distances to calculate each distance. Find the evaluation value of the data. The data selection unit 25 selects data having the smallest evaluation value by the evaluation value calculation unit 24 as representative data of the partial data group from the data of each partial data group of the processing target hierarchy. If the representative data selected by the data selection unit 25 is in the highest hierarchy, the data aggregating unit 26 outputs the representative data as a representative feature for the input data. Then, a partial data group in a hierarchy one level higher than the representative hierarchy is generated from the representative data and output to the evaluation value calculation unit 24. In this configuration as well, representative features are extracted by the same procedure as in the second exemplary embodiment.

  The representative feature extraction system of the exemplary embodiment shown in FIG. 9 also has the above-described data dividing unit 22, evaluation value calculating unit 24, data selecting unit 25, and data on the computer in the same manner as in the exemplary embodiments described above. This can also be realized by executing a program that realizes the function of the aggregation unit 26.

  Next, the operation of the above-described exemplary embodiment will be described using specific examples.

[Example 1]
Example 1 corresponds to the second exemplary embodiment described above.

  In the first embodiment, features representing the sea surface itself are extracted from an image obtained by imaging the sea surface with a hyperspectral camera, and pixels that are not the sea surface, such as people and drifting objects, are detected. The second exemplary embodiment is applied to the marine rescue system to be searched. Specifically, the representative feature extraction system of the second exemplary embodiment is used when extracting features representing sea level data.

  A hyperspectral camera is a camera that has a function of finely measuring the spectrum of an object with high wavelength resolution. For example, for each pixel, a bandwidth of several nanometers to several tens of nanometers, and several tens to several hundreds of bands. Spectral intensity can be measured with the number of bands. The information of each pixel can be expressed as D-dimensional vector data composed of D spectral intensities, where D is the number of bands. If the number of pixels of an image is N, the information of all the images can be expressed as a collection (data string) of N data of D-dimensional vectors.

  In the present embodiment, K representative features that well represent the overall features are extracted from the N data strings, and data different in nature from these K representative features is detected as abnormal data. A pixel that is not the sea surface is detected from an image of the sea surface, and a rescuer such as a deaf person is detected. For example, as shown in FIG. 10, when most of the target image is a sea region and there is a deaf person, the feature vector of the sea surface itself that occupies most of the image is used as reference data, and the pixel is different from the reference data. Detecting deaf people can help with rescue activities.

  In the present embodiment, a hyperspectral camera is used as the input unit 21, a display is used as the output unit 27, and the data dividing unit 22, the evaluation value calculating unit 24, the data selecting unit 25, and the data aggregating unit 26 are programmed on a computer. Is implemented by software. As the partial data group storage unit 23, a computer storage device is used.

Image information is input from the hyperspectral camera as the input unit 21. The image is two-dimensional, the information of pixel i is I _i, and the set of all input data is G = {I ₁ , I ₂ ,..., I _N }. The information of each pixel has a spectral intensity as a feature quantity, and a D-dimensional vector composed of D feature quantities, I _i = (v _i ⁽¹⁾ , v _i ⁽²⁾ ,..., V _i ^(D) ) Expressed.

The data dividing unit 22 divides the input data G = {I _i } into a plurality of partial data groups. As shown in FIG. 11, the division into partial data groups is divided into a two-dimensional lattice pattern to form a hierarchical structure as shown in FIG. This is because the spatially close regions are considered to have close spectral properties, and thus the spatially close regions are set as the same partial data group.

For example, when the input image is 256 × 256 pixels (= 65536 data) and the number k of representative features to be output is 1, in each layer, the data in the 2 × 2 region one layer below is one data. In order to divide into a hierarchical structure, since 256 = 2 ⁸ ,
G = (G ₁ ⁽¹⁾ , G ₂ ⁽¹⁾ , G ₃ ⁽¹⁾ , G ₄ ⁽¹⁾ )
G _i ⁽¹⁾ = (G _{i, 1} ⁽²⁾ , G _{i, 2} ⁽²⁾ , G _{i, 3} ⁽²⁾ , G _{i, 4} ⁽²⁾ )
G _{i, j} ⁽²⁾ = (G _{i, j, 1} ⁽³⁾ , G _{i, j, 2} ⁽³⁾ , G _{i, j, 3} ⁽³⁾ , G _{i, j, 4} ⁽³⁾ )
...
^{_{G i, ..., k (7}} ) = (G i, ..., k, 1 (8), G i, ..., k, 2 (8), G i, ..., k, 3 (8), G i, … _{, K, 4} ⁽⁸⁾ )
And can be divided into 8 hierarchies, and each hierarchy can be divided into 2 × 2 partial data groups. When the division ends, these partial data groups are stored in the partial data group storage unit 23.

  Next, the evaluation value calculation unit 24 takes out the partial data group of the processing hierarchy from the partial data group storage unit 23, with the lowest hierarchy among the plurality of hierarchies currently unprocessed as the processing hierarchy, For each piece of data in each extracted partial data group, a distance from other data in the partial data group is calculated. The processing for the partial data group is executed for each hierarchy, but first, the evaluation value calculation unit 24 is the lowest hierarchy passed from the data dividing unit 22 via the partial data group storage unit 23. The evaluation value calculation process is executed for the partial data group belonging to.

As the distance between the pixel I _a and the pixel I _b , any distance reference such as the L ₂ distance represented by the expression (1) or the L ₁ distance represented by the expression (2) can be used. The L ₂ distance is based on the definition of the distance in the two-dimensional Euclidean space and is generally used as a distance.

The evaluation value calculation unit 24 obtains an evaluation value of each data as the sum of the distances from other data in the partial data group for each data in the partial data group. For example, partial data group _{{(I i, I j,} I k, I l)} If consisting of four data, a distance of I _a and I _b as D _ab, the evaluation value V _i of I _i

Similarly, V _j , V _k , and V _l are also _obtained as the sum of distances from other data.

When the evaluation value is calculated for each data in each partial data group in the processing hierarchy in this way, the data selection unit 25 next selects the data of the partial data group for each partial data group. Select the one with the smallest evaluation value. For example, a partial data group consists of four data {(I _i , I _j , I _k , I _l )}, and the evaluation value of each data is {(V _i , V _j , V _k , V _l )}. If the minimum of {(V _i , V _j , V _k , V _l )} is V _i , the data selection unit 25 sets the representative data of this partial data group as I _i . .

  When representative data is selected for each partial data group of the processing hierarchy, the data aggregating unit 26 determines whether or not the current processing hierarchy is a predetermined highest hierarchy. In the case of the highest hierarchy, the data aggregation unit 26 notifies the output unit 27 of the representative data. On the other hand, if it is not the highest level, the data aggregating unit 26 creates a partial data group for a hierarchy one level higher than the current processing hierarchy. The partial data group in the upper hierarchy is a collection of some representative data belonging to the current processing hierarchy selected by the data selection unit 25. When the partial data group of the upper hierarchy is created, the data aggregation unit 26 stores the created partial data group in the partial data group storage unit 22 and supplies it to the evaluation value calculation unit 24.

  For example, the input image is 256 × 256 pixels (= 65536 data), the number k of representative features to be output is 1, and in each layer, 2 × 2 data one layer below is one data, When the input data is divided into hierarchical structures, there are 16384 (= 128 × 128) partial data groups each consisting of 2 × 2 data in the lowest hierarchy. For the second layer from the bottom, 4096 (= 64 × 64) partial data groups are created based on 128 × 128 representative data of the lowest layer. By performing this process over 8 hierarchies, one representative data is selected from 2 × 2 partial data groups in the highest hierarchy. Since one target representative data is obtained at the highest hierarchy, the data aggregation unit 26 finally notifies the output unit 27 of this representative data.

  A display provided as the output unit 27 displays representative data.

  In the first embodiment, the evaluation value calculation unit 24 calculates the total distance from other data in the partial data group as the evaluation value of each data, and the data selection unit 25 calculates the portion for each partial data group. The data with the smallest evaluation value is selected from the data in the data group. Therefore, unlike the method of determining the representative value by obtaining the histogram, the representative data can be obtained reliably even when the feature dimension is higher than the number of data.

When the representative data is determined by obtaining the histogram, the representative data cannot be obtained correctly if there is not enough data to contain a sufficient number of data at one segment in the histogram. For example, when the number of data is small and the number of data that falls within one segment is one or two, the segment that happens to contain two data is selected as a representative and does not represent the whole. This becomes more prominent as the feature dimension increases. This is because the number of breaks increases in proportion to the power of the number of dimensions. When one feature is divided into 100, one-dimensional features can be divided into 100 pieces, but 60-dimensional features can be divided into 100 ⁶⁰ pieces (= 10 ¹²⁰ pieces).

  Compared with the method using the histogram, the method according to the first embodiment can always obtain representative data even if the dimension of the feature amount is high.

  Further, in this embodiment, the evaluation value calculation unit 24 calculates the total distance from other data in the partial data group as the evaluation value of each data, and the data selection unit 25 calculates the partial data group for each partial data group. The data having the smallest evaluation value is selected from the data in a hierarchical manner. Therefore, even if outliers are included in the data, representative features can be extracted stably.

  On the other hand, when the average value is used as a representative feature, the average value includes the influence of an outlier. Usually, when calculating an average value, the influence of the distance is increased due to the nature of the operation itself of obtaining the average, so that there is a problem that the influence of the outlier is inevitably increased.

In the case of the present embodiment, for example, as shown in FIG. 13, it is assumed that the lowermost partial data group consists of four data of {(I ₁ , I ₂ , I ₃ , I ₄ )}, and I ₄ is the influence of noise. It is assumed that the data contains outliers. _Let D _ij be the distance between each data. Since I ₄ is data including outliers, the distances D ₁₄ , D ₂₄ , and D ₃₄ are larger than the distances D ₁₂ , D ₁₃ , and D ₂₃ . Therefore, among the evaluation values (V ₁ , V ₂ , V ₃ , V ₄ ) of each data, which is the total distance from other data, V ₄ is the other evaluation values (V ₁ , V ₂ , V ₃ ). Larger than In this case, the data I ₄ is not selected as the representative data because the data having a small evaluation value is used as the representative data. Therefore, according to the present embodiment, there is a very low possibility that data including outliers will be propagated to the upper hierarchy as representative data. By performing such selection hierarchically, according to the first embodiment, data that includes an outlier due to the influence of noise or the like is not included in the data of the higher hierarchy, and can be represented stably. Features can be extracted.

  Furthermore, this embodiment has an effect that it is not affected by not only outliers due to noise or the like, but also a minimum number of data having different properties. For example, in order to search for a drowning person from an image of the sea surface, when the spectrum of the sea surface itself is extracted as a representative feature, simply taking the average, not only the sea surface but also the sea surface of a drowning person or the like to be detected The average including things other than itself will be the representative feature. According to this embodiment, by selecting data hierarchically and excluding those having different properties in the lower hierarchy, it is possible to finally select representative features without including those having different properties.

  Furthermore, this embodiment also has an effect that the representative feature can be selected at high speed.

  For example, it is assumed that the input data consists of N = a × b data. If all distances to other data are obtained for each data without hierarchization, ab (ab-1) / 2 distance calculation is required.

On the other hand, it is assumed that n = a × b data is divided into one partial data group each consisting of b data. The number of distance calculations in the partial data group composed of b data is b (b-1) / 2 times, and this is executed for each of the a partial data groups. The total is ab (b-1) / 2 times. The distance calculation in the second layer is a (a-1) / 2 times because it is a distance calculation within a data. Therefore, by hierarchizing, calculation of ab (b-1) / 2 + a (a-1) / 2 is sufficient. The difference in the number of distance calculations between the case of not hierarchizing and the case of hierarchization is a (b ² −1) (a−1). Here, since a and b are larger than 1, the calculation amount is inevitably reduced by hierarchization. Here, the calculation amount in the case of dividing into one layer is shown, but if it is divided into a plurality of layers, a larger reduction effect on the calculation amount can be obtained.

  In the first embodiment, the number K of representative features to be finally obtained is set to 1, and the whole is symbolized by one feature. However, K is set to a value larger than 1, and a plurality of representative features or a plurality of representative features are represented. It may be extracted. For example, by selecting a representative feature for each area obtained by dividing the image into 4 × 4 pixels, the image can be reduced to 1/16. If the input data is assumed to consist of multiple elements with different properties, rather than representing the whole as a single feature, obtain multiple representative features and represent each of multiple elements with different properties. Is preferable.

  In the above description, in order to obtain the partial data group, the two-dimensional space is divided into continuous square lattices, but the two-dimensional space is hierarchized using other shapes such as a rectangular lattice, a triangular lattice, and a hexagonal lattice. It is also possible to divide automatically. Alternatively, it is possible to divide spatially discontinuously, such as dividing every predetermined number of pixels.

  In this embodiment, the division is performed without gaps and without overlapping, but it is also possible to divide so that there are gaps or overlaps. Even if there is a gap and not all data is used, or there is an overlap, it is possible to obtain the overall characteristics. For example, in general, division into powers of 2 makes it easy to implement on a computer. Therefore, when the number of input data is not a power of 2, it is possible to exclude fractions. Similarly, if a certain degree of uniformity can be assumed in the input data, there is little influence even if there is an overlap.

  In the first embodiment described above, a hyperspectral camera is used as the input means, and image data is directly input from the camera to the computer. However, the camera and the computer are arranged in different locations connected via a network. It is also possible. For example, a camera may be mounted on an aircraft, data from the camera may be transmitted to a ground control base via a wireless network, and representative feature extraction may be performed by a computer on the ground control base. In addition, it can be configured to directly input the image from the camera to the computer for processing in real time, and the data is stored in the storage device, for example, a post-survey after a problem occurs, Data collection and representative feature extraction can be performed at different timings.

  As a representative feature output method, not only the representative feature is displayed on the display described in this embodiment, but the representative feature is electronically notified to other devices through a network, or a hard disk device or USB (Universal Serial Bus) is used. ) Any method can be used, such as recording a representative feature on a recording medium such as a memory.

  In this embodiment, the case where the representative feature extraction system is used for a system for detecting a deaf person or a drifting object from an image of the sea surface has been described. However, in order to detect non-sediment from an image of a sediment disaster site, The present embodiment can be similarly applied to a system that extracts a feature that represents a desert, and a system that extracts a feature that represents a desert in order to detect an oasis from a desert image. In addition, as the data from which representative features are extracted, not only spectral data obtained by the hyperspectral camera used in this embodiment but also arbitrary data can be used. For example, a behavior is expressed using a plurality of feature quantities expressing behavior such as moving speed and staying time, representative behavior is extracted from a series of behavior history, and abnormal behavior is compared with typical behavior. The present embodiment can be applied to a failure detection system that detects and detects the occurrence of a failure. Furthermore, by expressing speech speech of one word as a multidimensional vector, detecting utterances found in normal conversation as a representative feature from a series of conversations, and detecting conversations that differ from the representative feature, for criminal investigations, etc. The present embodiment can also be applied. Furthermore, this embodiment can be applied not only to selecting representative features for detecting abnormalities that detect unusual data, but also to selecting representative pixels that well represent the features of the image in order to compress the image. It is. The present embodiment can also be applied to classify data using representative features.

[Example 2]
Example 2 corresponds to the fourth exemplary embodiment described above. The second embodiment is similar to the first embodiment, but differs from the first embodiment in that a clustering unit 28 and a cluster representative extraction unit 29 are provided as shown in FIG. Also, it is assumed that the input data is composed of a plurality of elements having different properties, and the number of representative features to be finally obtained is a plurality. The number of representative data K is also 2 or more.

  For example, assuming that representative data is selected for each region obtained by dividing the image into 8 × 8 pixels, by performing the same processing as in the first embodiment, the data input to the clustering unit 28 is the input image of 64. The representative data (compressed image) is reduced to a factor of one. The clustering unit 28 divides the input image into partial images (clusters) having similar feature amounts. For example, if the image to be processed is composed of a sea area and a sand beach area as shown in FIG. 14, the clustering unit 28 performs a clustering process on the entire input image, and as shown in FIG. 15, the sea area and the sand beach area. The input image is divided into two partial images 1 and 2 corresponding to the area. For the clustering process, for example, an arbitrary method such as a k-means can be used.

  As described above, the cluster representative extraction unit 29 extracts a representative feature for each created cluster, assuming that a cluster is generated. For example, when two clusters (partial images 1 and 2) indicating a sea area and a sand beach area are created as shown in FIG. 15, as shown by the spectrum curve in FIG. The representative value of the cluster representing the value and the sandy beach is obtained.

  As a method of obtaining the representative value of the cluster, there is a method in which data having the minimum total distance to other data is used as the representative value, as in the processing in the evaluation value calculation unit 24 and the data selection unit 25. Alternatively, a representative value can be obtained as an average value of data in the cluster. In the present embodiment, representative features are selected hierarchically before clustering, and outlier data and a small number of data with different properties are excluded in advance. It is also divided into clusters with similar properties. For this reason, since only data having substantially the same property exists in one cluster, even if the average is obtained here, it is not affected by data having different outliers or properties.

  For clustering, a method of dividing into similar regions using spatial continuity may be used, or a method of dividing into similar regions using only spectral similarity without using spatial continuity. Also good. When the input image is composed of a plurality of regions that can be spatially distinguished as shown in FIG. 16, it is possible to select a representative feature that is preferable to divide into similar regions using spatial continuity. . On the other hand, when the input image is an image in which the soil is partially visible from the forest, the input image is converted into the forest by focusing only on the spectral similarity without using spatial continuity. Can be clustered on each tree and soil.

  According to the second embodiment, the input image is divided into clusters having similar feature amounts, and representative feature extraction is performed on each cluster, so that the input image has a plurality of regions having several different properties. It will be possible to cope with the case of consisting of.

  In this embodiment, since hierarchical selection is performed before clustering, an input image can be processed at high speed. It is also conceivable to perform clustering directly on input data without performing hierarchical selection. However, since a clustering algorithm such as the k-means method uses repetitive operations, it is an extremely time-consuming algorithm. Therefore, significant speedup can be realized by performing hierarchical selection in advance and reducing the number of data to be clustered.

[Example 3]
Example 3 corresponds to the third exemplary embodiment described with reference to FIG.

The occurrence distribution storage unit 16 has an occurrence distribution Q (X) = (q ⁽¹⁾ (x), q ⁽²⁾ (x),..., Q ^(D) (x)) of each data I _i , that is, each feature. The occurrence distribution q ^(j) (x) of the quantity v _i ^(j ) is stored. For example, the distribution of the feature quantity of the j-th component is a normal distribution with an average of m ^(j) and a variance of σ ^(j) , or between [min ^(j) and max ^(j) ]. Is stored in the occurrence distribution storage unit 16. It is assumed that the feature quantity v _i ^(j) occurs according to the distribution stored in the occurrence distribution storage unit 16. Based on prior knowledge about the object to be photographed, the occurrence distribution can be given in advance including parameters such as mean and variance, and only the shape of the distribution is known, and the parameters such as mean and variance are actually The occurrence distribution can also be given by obtaining from the data input to. Furthermore, when prior knowledge about the occurrence distribution cannot be obtained or when the occurrence distribution cannot be determined from the prior knowledge, a uniform distribution is assumed, and the maximum value max ^(j) and the minimum value min of the uniform distribution are assumed. It is also possible to determine ^(j) from the maximum value and the minimum value of the j-th component of all image data actually input.

The evaluation value calculation unit 13 uses a probability P (I _a that a value virtually generated using the occurrence distribution Q (x) stored in the occurrence distribution storage unit 16 is a value between the data I _a and I _b. , I _b ), the distance between these data I _a and I _b is defined. Since occurrence distribution of the j component is q ^{(j) (x),} the value v _a value of the j component of the pattern to be virtually generated from the occurrence distribution is the j-th component of data I _a ^(j) values become probability p between the value of the j component of ^{_{I b v b (j) (}} v a (j), I b (j)) is represented by the formula (4). Therefore, the value produced virtually the I _a and I _b value become probability P (I _a, I _b) during shall be occurrence distribution of each component is independently represented by the formula (5) The In a normal case, the two data are different as the distance is larger, and the two data are closer as the value is smaller. The probability defined by Equation (5) is also small, meaning that an event that takes a value between two values rarely occurs, that is, the distance between the two data is short, and the two data are similar. It represents that. On the other hand, a large probability value indicates that it is not uncommon for an event to take a value between two data values, that is, the distance between the two data is long and the two data are dissimilar. .

By taking the logarithm of this probability p as the distance D (I _i , I _R ) between I _i and I _R , it is also possible to obtain equation (6). Since the logarithmic function (log) is a monotonically increasing function, even if the logarithm is taken, the property as a distance does not change.

Here, as compared with the case of using the general distance represented by Expression (1), the effect of using the distance as defined in this embodiment will be described with reference to FIG.

For example, when a sensor corresponding to the d-th component is lost or an outlier is taken under the influence of noise, I _i and I _R are originally similar data, but the d-th data in those data is similar. It is assumed that the component difference (v _i ^(d) −v _R ^(d) ) becomes a large value. In FIG. 17, the horizontal axis indicates the nth component, and the vertical axis indicates the value of the feature amount of the nth component. The data I _S is the same target data as I _R , but noise enters the d-th component of I _S , and the feature value v _S ^(d) is an outlier. Data I _T is the data of another subject and I _R, has a totally different characteristic amount from the I _R. However, when the normal distance L ₂ is used, since v _S ^(d) is an outlier, (v _S ^(d) −v _R ^(d) ) ² becomes a very large value, except for the d-th component. The distance L ₂ (I _S , I _R ) is a large value even though the feature values of are substantially the same, and as a result, it is determined that the data I _S and I _R are not similar. On the other hand, although I _T and I _R are slightly different as a whole, since there is no component having a large difference, the distance L ₂ (I _T , I _R ) is not so large, and I _T and I _R are I _S It would be judged to be similar than I _R and.

On the other hand, in the case of the distance represented by the equation (6) used in the present embodiment, when the contribution of the d-th component is considered, when the difference between v _S ^(d) and v _R ^(d) increases due to an outlier. p (v _S ^(d) , v _R ^(d) ) becomes close to 1, and the logarithmic value log (p (v _S ^(d) , v _R ^(d) )) of the d-th component in equation (6) is 0. Close to. For this reason, the contribution to the entire d-th component including outliers is reduced. For the i-th component (i ≠ d) that does not include any other outlier, the difference between v _S ⁽ⁱ⁾ and v _R ⁽ⁱ⁾ is small, and p (v _S ⁽ⁱ⁾ , v _R ⁽ⁱ⁾ ) is The logarithmic value log (p (v _S ⁽ⁱ⁾ , v _R ⁽ⁱ⁾ )) of the i-th component of Equation (6) becomes a small value (a negative value with a large absolute value). Since the d-th component including outliers is 0 and the component not including outliers is a negative value having a large absolute value, the sum D (I _S , I _R ) is a negative value having a large absolute value. The distance defined in the embodiment is a small value. Thus data I _S and I _R are determined to be similar. On the other hand, since I _T and I _R are totally different, p (v _T ⁽ⁱ⁾ , v _R ⁽ⁱ⁾ ) becomes a value close to 1 for all components, and logarithmic value log of each component in Equation (6) (p (v _T ⁽ⁱ⁾ , v _R ⁽ⁱ⁾ )) are both close to zero. Even if a value close to 0 is added, it is also close to 0, so D (I _T , I _R ) is close to 0. Distance represented by the formula (6), since 0 is the maximum value, the distance data I _T and I _R becomes a large value.

  Due to this property of the distance shown in Equation (6), even if there is a deficiency or outlier in a part of the feature amount, it is less susceptible to that. Therefore, the method of the third embodiment has a property that a representative feature can be stably extracted even if an outlier is included in input data.

  Similarly, due to the property that the contribution of the component having a large difference with the distance represented by Equation (6) is small, the method of this embodiment stably extracts representative features even when the feature dimension is high. It has the property that it can.

On the other hand, when the normal distance L ₂ shown in the equation (1) is used, the representative feature cannot be extracted stably when the feature dimension becomes high. In general, when the normal distance L ₂ is used, it is said that the determination of data proximity becomes unstable as the dimension increases. The reason why the determination of the proximity based on the normal distance becomes unstable in the high dimension is that the normal distance scale assumes that D is a large natural number, and among the components of the D-dimensional pattern, the contribution of the component with the smaller distance Is much smaller than the contribution of components with large distances. The influence of a minute change of a component with a large distance, such as an outlier, is greater than the contribution of a component with a small distance, causing instability in proximity determination. In addition, since the probability that a component having a large distance will appear as the dimension D increases, the probability that such instability of neighborhood determination will occur in a high-dimensional pattern increases.

  On the other hand, when a pattern is virtually generated from the occurrence distribution of the feature quantity of an arbitrary pattern as in the present embodiment, the probability that the pattern is closer to normal data than the feature quantity of each data for each data Is close to 1 for components with large differences in feature quantities. The closer the value is to 1, the smaller the influence is. Therefore, the influence of the component having a large feature amount difference on the other component is reduced. For this reason, if the other components are similar, the distance between the two patterns is reduced. Thus, even if the dimension is increased, according to the method of this embodiment, the distance does not become unstable.

  Here, an effect is shown in the case where the feature can be expressed as a high-dimensional vector composed of a plurality of feature amounts, and an outlier is included in a certain feature amount of the plurality of feature amounts. Next, the effect of the present embodiment for the case where abnormal data is included in a plurality of data will be described.

As shown in FIG. 13, when a representative value is selected from four data of {(I ₁ , I ₂ , I ₃ , I ₄ )}, I ₄ is data including an outlier due to the influence of noise. And _Let D _ij be the distance between each data. Since I ₄ is data including outliers, the distances D ₁₄ , D ₂₄ , and D ₃₄ are larger than the distances D ₁₂ , D ₁₃ , and D ₂₃ .

As described above, when the normal distance L ₂ is used, since the influence of the distance is large, for example, when considering the total distance from I ₂ to other data, D ₂₄ compared to D ₁₂ and D _23. Becomes much larger and the influence of D ₂₄ becomes smaller, and the influence of D ₁₂ and D ₂₃ becomes larger. Therefore, the evaluation value is determined centering on the positional relationship between I ₁ , I ₂ , and I _{3 where} the data is dense, and representative features that express the properties of the data excluding outliers are selected. .

[Example 4]
Next, Example 4 will be described. The fourth embodiment is the same as the first embodiment, but further includes the occurrence distribution storage unit 16 shown in the third embodiment, and the evaluation value calculation unit 24 has the same procedure as the evaluation value calculation unit 13 in the third embodiment. This is different from the first embodiment in that the distance is calculated and the evaluation value is obtained. FIG. 18 shows the configuration of a representative feature extraction system according to the fourth embodiment.

  The occurrence distribution storage unit 16 stores an occurrence distribution of feature quantities assumed to be input. Similarly to the first embodiment, the input data is divided into a plurality of hierarchical partial data groups, and the data is processed hierarchically. That is, the data dividing unit 22 divides input data into a plurality of partial data groups in a hierarchical manner so that evaluation value calculation and data selection can be performed in a hierarchical manner. The evaluation value calculation unit 24 and the data selection unit 25 calculate, for each data in each partial data group for each hierarchy, a distance representing the similarity with other data in the partial data group, and the distance is calculated. To evaluate the evaluation value of each data. The distance here is a value between two data values to be compared when the pattern is virtually generated from the occurrence distribution stored in the occurrence distribution storage unit 16. Probability. The data aggregating unit 26 collects representative data for each partial data group and creates a partial data group of the next hierarchy based on the collected data, thereby enabling evaluation value calculation and data selection to be performed hierarchically.

  In the fourth embodiment, since the hierarchical representative feature selection based on the first embodiment and the evaluation value calculation based on the occurrence distribution based on the third embodiment are combined, the effects based on both the embodiments are combined. Have.

  Specifically, according to the present embodiment, by using a hierarchical structure, the influence of data including outliers or a small number of data having different properties is not propagated from a lower hierarchy to an upper hierarchy. As in the case of Example 1, the representative feature can be extracted without being affected by a small number of data having different outliers and properties. In addition to using a hierarchical structure, the probability that the pattern takes a value between two values of the current comparison target data as a virtual pattern generated from the occurrence distribution is used. Similarly, it is less likely to be affected by a small number of data with different outliers and properties.

  By calculating hierarchically and obtaining the representative feature by the sum of distances, the representative feature can be obtained even when the number of data is small, and the representative feature can be obtained at high speed, as in the first embodiment. .

By using the probability that the pattern takes a value between the two values of the current comparison target, assuming that the pattern is virtually generated from the occurrence distribution, the feature becomes high-dimensional as in the third embodiment. However, in the fourth embodiment, a clustering unit and a cluster representative extraction unit can be added as in the second embodiment. In that case, the representative feature can be stably extracted even when the input image is composed of a plurality of regions having different properties.

  The present invention can be applied to a technical field in which, for example, representative features are extracted from images, sounds, spectrum patterns, behavior patterns, etc., and abnormality detection, data compression, classification, and the like are performed.

11, 21 Input unit 12 Data group storage unit 13, 24 Evaluation value calculation unit 14, 25 Data selection unit 15, 27 Output unit 16 Occurrence distribution storage unit 22 Data division unit 23 Partial data group storage unit 26 Data aggregation unit 28 Clustering unit 29 Cluster representative extractor

Claims (16)

A representative feature extraction system that selects representative features from input data,
Dividing means for dividing the input data hierarchically into a plurality of partial data groups;
The distance from the data to other data included in the partial data group for each data included in the partial data group in each partial data group of the processing hierarchy, with the lowest hierarchy among the unprocessed hierarchies An evaluation value calculating means for calculating an evaluation value by calculating a sum of
For each partial data group of the processing hierarchy, data selection means for selecting, as representative data, the one with the smallest evaluation value among the data included in the partial data group;
If the processing hierarchy is the highest hierarchy, the representative data selected by the data selection means is the representative feature for the input data, and if the processing hierarchy is a hierarchy other than the highest hierarchy, the data selection means Aggregating means for aggregating the representative data selected by and generating a partial data group in a hierarchy one layer higher than the processing hierarchy;
An occurrence distribution storage means for storing an occurrence distribution of a feature quantity assumed to be input;
I have a,
The evaluation value calculation means obtains a distance between two data as a probability that a data value according to the pattern exists between the two data values when a pattern is virtually generated by the occurrence distribution. Representative feature extraction system. Clustering means for dividing the representative feature obtained by the aggregation means into clusters;
Cluster representative extraction means for extracting representative features for each cluster;
The representative feature extraction system of claim 1 further comprising:   The representative feature extraction system according to claim 2, wherein the cluster representative extraction unit extracts a representative feature from the cluster using a distance reference different from the distance reference used in the evaluation value calculation unit.   The representative feature extraction system according to claim 1, wherein the dividing unit divides the input data into partial data groups using spatial continuity in the input data. Before SL cluster representative extraction unit, and the representative feature an average of the data in the cluster, the representative feature extraction system according to claim 2. Input means for reading the input data;
Storage means for storing the partial data group;
An output means for outputting the representative feature;
The representative feature extraction system according to any one of claims 1 to 5 , further comprising: A representative feature extraction method for selecting a representative feature from input data,
Dividing the input data hierarchically into a plurality of partial data groups;
The distance from the data to other data included in the partial data group for each data included in the partial data group in each partial data group of the processing hierarchy, with the lowest hierarchy among the unprocessed hierarchies To calculate an evaluation value, selecting the smallest evaluation value as representative data of the partial data group,
If the processing hierarchy is the highest hierarchy, the selected representative data is the representative feature for the input data;
When the processing hierarchy is a hierarchy other than the highest hierarchy, the selected representative data is aggregated to generate a partial data group of a hierarchy one level higher than the processing hierarchy, and to obtain the evaluation value; Repeating representative data selection,
I have a,
The distance between the two data is such that when a pattern is virtually generated by an occurrence distribution prepared in advance for a feature quantity that is assumed to be input, the data value of the pattern is between the values of the two data. A representative feature extraction method expressed as a probability of existence . Clustering the representative features into clusters;
Extracting representative features for each cluster;
The representative feature extraction method according to claim 7 , further comprising: 9. The representative feature extraction method according to claim 8 , wherein a distance reference used when extracting a representative feature for each cluster is different from a distance reference used when obtaining a distance between the data. Wherein said input data using a spatial continuity in the input data is divided into the partial data group, the representative feature extraction method according to any one of claims 7 to 9. The average of the data before Symbol cluster as a representative features extracted from the cluster, the representative feature extraction method according to claim 8. The representative feature extraction method according to claim 7 , wherein a uniform distribution between a maximum value and a minimum value in the input data is used as the occurrence distribution of the feature amount. The representative feature extraction method according to any one of claims 7 to 12 , wherein spectrum data of a hyperspectral camera is converted into multidimensional data as the input data. After extracting the representative feature from said input data, representative feature further comprising detecting as abnormal data data different from the representative feature extraction method according to any one of claims 7 to 13. On the computer,
A process of hierarchically dividing the input data into a plurality of partial data groups;
The distance from the data to other data included in the partial data group for each data included in the partial data group in each partial data group of the processing hierarchy, with the lowest hierarchy among the unprocessed hierarchies A data selection process for calculating an evaluation value by calculating the sum of the above, selecting the smallest evaluation value as representative data of the partial data group,
If the processing hierarchy is the highest hierarchy, the selected representative data is the representative feature for the input data; and
When the processing hierarchy is a hierarchy other than the highest hierarchy, a process of aggregating the selected representative data to generate a partial data group of a hierarchy one level higher than the processing hierarchy and repeating the data selection process When,
Was executed,
The distance between the two data is such that when a pattern is virtually generated by an occurrence distribution prepared in advance for a feature quantity that is assumed to be input, the data value of the pattern is between the values of the two data. A program expressed as a probability of existence . In the computer,
Clustering to divide the representative features into clusters;
Processing for extracting representative features for each cluster;
The program according to claim 15 , further executing:

Images