JP2019096118A - Pattern recognition program, device and method - Google Patents

Abstract

To provide a pattern recognition program, device and method for executing pattern recognition on an input image, in which the exactness of writing in a feature space a range where there exists a learning sample is improved.SOLUTION: In a feature space, an axis group (axes 1, 2, 3, 4) matched to the distribution shape of categories A, B, C, etc., for categorizing the pattern of a learning sample is set. Secondly, a dictionary in which are correlated a category range and categories obtained by projecting feature points in the feature space that correspond to a feature vector extracted from a category learning sample to each axis of the set axis group, is created for each category. Then, a category range is extracted that includes a position obtained by projecting feature points in the feature space that correspond to a feature vector extracted form a given unknown pattern to each axis of the set axis group, and a category set corresponding to the extracted category range is outputted, whereby a category to which the given pattern belongs is, for example, broadly classified.SELECTED DRAWING: Figure 6

Description

  The present invention relates to a pattern recognition program, apparatus, and method for executing pattern recognition on an input image.

  An example of pattern recognition for an input image is character recognition that recognizes the type (character code) of each character on a document image. In such pattern recognition, it is required to reduce the time required for the recognition process to speed up while maintaining the recognition accuracy with high accuracy.

  In pattern recognition, an object to be recognized is called a pattern, and a set of all patterns is called a pattern space. A set of one or more feature quantities obtained by performing feature extraction on a pattern is called a feature vector, and the number of elements of the feature vector is called a feature vector dimension. A set of values (feature quantities) of each element of the feature vector is called a value of the feature vector, and a set formed by all the feature vector values is called a feature space. The dimension of the feature space is equal to the dimension of the feature vector which is an element of the feature space. A set of patterns or feature vectors that can be regarded as the same type is called a category. Determining which category a feature vector obtained from an unknown pattern or an unknown pattern input for pattern recognition belongs to is called pattern recognition.

  For example, in pattern recognition processing in which character recognition is performed on an unknown character image, a feature vector is generated from the input character image, and matching is performed with a plurality of feature vectors stored in a dictionary on the memory . In the matching process, distances (for example, Euclidean distances) are calculated between feature vectors of the input character image and feature vectors in the dictionary. The process of performing this calculation is called matching process. As a result, distance values corresponding to the number of feature vectors in the dictionary are obtained, and among them, the character code associated with the feature vector that is the minimum distance value is the recognition result of the character for the input character image (pattern recognition Output as a result).

  In the pattern recognition process of such an example, a large amount of processing time is required for the matching process between the feature vector of the unknown input image and all the feature vectors in the dictionary. Therefore, in order to speed up the matching process, a two-step process of large classification and detailed classification is used. The large classification is a process of extracting a feature vector close to a feature vector generated from an input character image to be recognized from a set of feature vectors in a dictionary, and narrowing down objects when performing detailed classification. If detailed classification is performed only on narrowed character codes, it becomes possible to significantly reduce the conventional matching processing time.

  For example, the following prior art is known as a large classification technique for pattern recognition (for example, Patent Document 1). In this conventional technique, a feature vector obtained from each character image for learning is projected onto an axis (orthogonal axis) of each dimension of the feature space, and a character category (type of character = character code) is formed on each axis. A dictionary recording the range of existence is created. At the time of the large classification, the feature vector of the unknown input image is projected on each axis, and the logical AND with the character category set obtained on each axis is taken, and the result is taken as the large classification result.

  For example, the following prior art is known as another major classification technique for pattern recognition (for example, Non-Patent Documents 1 and 2). This conventional technology is a large classification technology using LSH (Locality Sensitive Hashing), which is one of the Hash methods. LSH is a technology that performs major classification by Hashing using randomly generated axes, not orthogonal axes.

JP, 2011-100245, A

P. Indyk and R. Motwani, “Approximate nearest neighbors: towards removing the curse of dimensionality.”, STOC, pp. 604-613, 1998. A. Gionis, P. Indyk and R. Motwani, "Similarity Search in High Dimensions via Hashing," "Proceedings of the 25th International Conference on Very Large Database (VLDB), pp. 518-529, 1999.

  The above-mentioned large classification technology using orthogonal axes is a large classification technology with better performance in terms of classification accuracy rate, narrowing ability, and processing speed as compared with other methods. However, since this prior art uses a common orthogonal coordinate system in all categories, an area without a learning sample on the feature space is treated as if there is a learning sample on the dictionary.

  On the other hand, in the prior art using LSH, since the random multiple axes are used, the distribution of learning samples on the feature space is accurately captured, and the narrowing accuracy tends to be high. However, since the method of determining the number of axes is random, there is no guarantee of accuracy, and many axes are used, resulting in long processing time.

  Therefore, one aspect of the present invention is to improve the accuracy in describing a range in which a learning sample is present on a feature space.

  The pattern recognition program according to an example of the aspect causes the computer to execute a process of setting an axis group that matches the distribution shape of each category that classifies the pattern of the learning sample in the feature space. Next, the pattern recognition program is provided for each category based on the category range obtained by projecting feature points on the feature space corresponding to the feature vector extracted from the learning sample of the category on each axis of the axis group. The computer is caused to execute a process of classifying the identified pattern.

  It is possible to improve the accuracy in describing the range in which the learning sample is present on the feature space.

It is a figure explaining the relationship of the feature vector and the feature space in character recognition from an image, and a feature vector. It is explanatory drawing of the major classification technique for the pattern recognition by the present applicant based on patent document 1. FIG. It is explanatory drawing of the expression format of the dictionary using a hash table, and a major classification system. It is explanatory drawing of the problem of the major classification technique for the pattern recognition by the present applicant based on patent document 1. FIG. FIG. 1 is a block diagram of a first embodiment of a pattern recognition apparatus. It is explanatory drawing of 1st Embodiment. It is a flowchart which shows the process example of a feature extraction part and an eigenvector extraction part. It is explanatory drawing of hierarchical clustering. It is a flowchart which shows the process example of a hierarchical clustering part. It is a flowchart which shows the process example of a cluster selection part, an orthogonal axis determination part, and an axis | shaft output determination part. It is a flowchart which shows the example of a detailed process of an orthogonal axis determination part. It is a block diagram of 2nd Embodiment of a pattern recognition apparatus. It is explanatory drawing (the 1) of 2nd Embodiment. It is explanatory drawing (the 2) of 2nd Embodiment. It is explanatory drawing (the 3) of 2nd Embodiment. It is a flowchart which shows the process example of 2nd Embodiment. It is a figure showing the example of hardware constitutions of the computer which can realize the pattern recognition device of an embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. First, before describing the details of the present embodiment, the feature vectors and feature spaces that are the premise thereof will be described.

FIG. 1 is a diagram for explaining the relationship between a feature vector and a feature space in character recognition from an image. In pattern recognition, a dictionary or an image to be recognized is represented by a feature vector which is a vector in which the numerical values of feature amounts extracted from the image are arranged. Since one character image is represented by a feature vector, one feature vector is represented as one point in the feature space in consideration of a feature space having the number of dimension axes of the vector. For example, when character recognition is performed as an example of pattern recognition, when character image 101 illustrated in FIG. 1 (here, the character “A” is displayed) is input, it is illustrated in FIG. It is expressed by a feature vector 102 = (v ₁ , v ₂ ,..., V n) consisting of n element values.

  105 and 106 in FIG. 1 are diagrams showing a method of extracting feature vectors from a character image. In this example, the feature vector 102 using the “direction index histogram” is extracted according to the following procedure.

Procedure 1: Extract each pixel constituting an outline from the character image 105.
Procedure 2: Classify contour pixels in one of four directions of "|""/""-""\".
As a pixel classification method, in any direction of eight neighboring pixels of the pixel of interest
Classification is based on the presence of contour pixel values.

Procedure 3: Divide the character image 105 into n × n (3 × 3 in the example of FIG. 1) regions,
The number of pixels in each direction in each divided area is obtained.
The number of pixels in each direction shown as 106 is called a "direction index histogram".
Procedure 4: Arrange the numerical values of the direction index in the order of the divided areas (areas 1 to 9),
The feature vector 102 is used. In the example of FIGS. 1105 and 106:
9 regions x 4 directions = feature vectors 102 with 36 direction indices
_{= (V 1, v 2,} ···, v36) is obtained.

  When the number of dimensions of the feature vector 102 is n, one character in the character image 101 illustrated in FIG. 1 corresponds to one point 104 in the n-dimensional feature space 103 illustrated in FIG. Can be represented.

  Next, the major classification technology for pattern recognition, which the applicant filed earlier, which is the basic technology of the present embodiment, will be described. FIG. 2 is an explanatory view of a large classification technique for pattern recognition by the applicant according to Patent Document 1; Orthogonal axes are determined from sample patterns of the entire image for learning. In FIG. 2, two-dimensional orthogonal axes of the X axis and the Y axis are determined to simplify the description. Next, the feature vector of each sample pattern for learning of each category (for example, character type in the case of character recognition) is projected on an axis (for example, X axis and Y axis) of each dimension of the feature space determined by the orthogonal axis. A dictionary is created in which the range of existence of character categories (type of characters = character code) is recorded on each axis. Then, at the time of major classification of pattern recognition, the feature vector of the unknown input image is projected on each axis, and the logical AND with the character category set obtained on each axis is taken, and the result is taken as the major classification result.

The specific processing of the above-mentioned large classification processing will be described below. First, the elements of the feature vector of one learning sample pattern of one category (for example, one character type) (for example, any of v ₁ , v ₂ ,..., V n in FIG. 1) In this case, _one of v ₁ and v ₂ ) is extracted because it is two-dimensional. Then, the position on the axis corresponding to the extracted element is calculated. For example, elements v _1, the value on the X axis of the element (position) is calculated. Similarly, elements v _2, the value on the Y-axis of the element (position) is calculated.

  Next, the same processing is performed on all the other learning sample patterns in which the shape of the pattern is finely deformed for one category (for example, one character type). For example, in the case of character recognition, even with the same character "A", there are variously modified "A". Each of these is given as a learning sample pattern. As a result, the range on each axis is determined for one category. A predetermined margin is added to both ends of each range. For example, in the example of FIG. 2, it is assumed that the feature points in the feature space of category A are at three positions indicated by "A". At this time, for the feature vector of each feature point “A”, the x element value of the feature vector is projected to the position on the X axis, and the y element value is projected to the position on the Y axis. A range 201Ax on the X axis and a range 201Ay on the Y axis are calculated for each feature vector. By adding a margin for each axis to these results, a range 202Ax on the X axis and a range 202Ay on the Y axis are calculated for the category A.

  The same processing as described above is performed for other categories. In the example of FIG. 2, it is assumed that the feature points in the feature space of another category B are at three positions indicated by “B”. At this time, although not shown, a range 201Bx on the X-axis and a range 201By on the Y-axis are calculated for the feature vectors of the feature points "B". By adding a margin for each axis to these results, although not shown, a range 202Bx on the X axis and a range 202By on the Y axis are calculated for the category B.

  Furthermore, similarly, in FIG. 2, it is assumed that feature points in the feature space of category C are at three positions indicated by “C”. At this time, although not shown, a range 201Cx on the X axis and a range 201Cy on the Y axis are calculated for the feature vectors of the feature points "C". By adding a margin for each axis to these results, a range 202Cx on the X axis and a range 202Cy on the Y axis are calculated for the category C, although not particularly shown.

  By repeating the above operation, a dictionary having the value of the range on each orthogonal axis in the feature space as a component is created for each category. In this case, the range of feature points of each category in the feature space is defined as an intersection partial region of the range on each orthogonal axis of the category. In the example of FIG. 2, the range of the feature point of category A in the feature space defined by the X axis and the Y axis is such that the range 202Ax on the X axis intersects with the range 202Ay on the Y axis for the category A It is defined as four areas surrounded by solid squares. In addition, the range of the feature points of category B in the feature space is surrounded by a small broken-lined square in which the range 202Bx on the X axis and the range 202By on the Y axis (both not shown) intersect for the category B. Defined as a single region. Furthermore, the range of the feature points of category C in the feature space is surrounded by a rough dashed square where the range 202Cx on the X axis and the range 202Cy on the Y axis (both not shown) intersect for the category C. Defined as a single region.

  The process in the case where the unknown pattern (for example, input character) x is roughly classified using the dictionary created as described above is as follows. It is determined to which category of the X axis and the Y axis the respective element of the feature vector of the unknown pattern x, the x element value and the y element value in the example of FIG. In the example of FIG. 2, the feature point “X” of the unknown pattern x overlaps the range of the category C on the X axis and is close to the range of the category A. In addition, in the Y axis, each range of the categories A and C overlaps. From this, the unknown pattern x is roughly classified into categories A and C.

FIG. 3 is an explanatory view of the expression form of the above-mentioned dictionary using the hash table and the large classification method. For each index value from i = 1 to the maximum value corresponding to each value on axis K, for each category A, B, C, etc., for each axis K (K = 1 to n) (n is the number of dimensions) A hash table dictionary 301 consisting of a dictionary bit string bit1 _AK , bit1 _BK , bit1 _CK, etc. in which the range of each category exists (= 1) or not (= 0) is set is created. When an unknown pattern x is input, each dictionary bit string bit1 _AK , bit1 _BK , bit1 _CK, etc. corresponding to the position of the k-th element value of the feature vector on the axis K The matching result bit string _bit2 x K for the axis K of the unknown pattern x is obtained. The final matching result bit string bit3 _x (not shown) is calculated by calculating the AND product for each bit position for the matching result bit string _{bit2 xK} for every axis K (K = 1 to n) obtained in this manner. Is obtained. The category corresponding to the bit position for which the value 1 is set in this bit3 _x is taken as the major classification result.

  The above-described large classification method is a technology that enables high-speed large classification processing, but in practice there may be cases where a category range in which no feature point exists is designated. In the example of FIG. 2, among the category ranges on the feature space indicated by four solid frames for the category A, the feature points of the category A are actually in the range of the solid line frame 203A at the upper left and the solid line frame 204A at the lower right. Although it does not exist, it is determined to be a range of category A. For example, in FIG. 4 (viewing is the same as in FIG. 2), the lower right one of the four regions surrounded by the solid line frame corresponding to the category A in the feature space is the feature point “X” of the unknown pattern x. It is assumed that it exists in 204A. The positions when projecting each element value of the feature vector of the unknown pattern x onto the X axis and the Y axis are both included in the range on the X axis of Category B and the range on the Y axis. Therefore, the unknown pattern x is first classified into category B. Along with this, the position when each element value of the feature vector of the unknown pattern x is projected onto the X axis and the Y axis is also included in the range 202Ax on the X axis of category A and the range 202Ay on the Y axis. Therefore, the unknown pattern x is also roughly classified into category A. However, in fact, as shown in FIG. 4, the feature point “X” of the unknown pattern x is not included in the area where the feature point “A” of category A actually exists, and the major classification of the correct answer The result is category B only. Therefore, in this large classification method, since category A is additionally selected, the narrowing ability of the large classification processing is lowered.

  In the above-described large classification method, a method may be considered in which only the "region without a learning sample" is deleted. However, since such an area can not be deleted by the dictionary description as described in FIG. 3, the check process is performed in the large classification process. For this reason, there is a problem that the large classification processing is delayed.

  The first and second embodiments of the pattern recognition apparatus described below are techniques for solving the above-mentioned problems. The method of each embodiment is a method of creating a dictionary that accurately reflects the feature distribution over the processing time, and speeding up the large classification processing without reducing the accuracy. Specifically, in accordance with the distribution of the features of each category of the learning sample, an axis that can identify the distribution shape of each category with as few axes as possible is set. The axes are based on orthogonal axis sets, and are composed of a plurality of sets that are not orthogonal to one another. This enables the range in which the learning sample is present on the feature space to be described more faithfully in the dictionary. Since unknown patterns (for example, unknown input characters) are plotted in a narrower category range than the large classification method described in FIGS. 2 to 4, it is possible to narrow down the large classification result category more accurately.

  FIG. 5 is a block diagram of a first embodiment of a pattern recognition apparatus. The first embodiment includes a feature extraction unit 501, an eigenvalue / eigenvector extraction unit 502, a hierarchical clustering unit 503, a cluster selection unit 504, an orthogonal axis determination unit 505, and an axis output determination unit 506. Each of the units 501 to 506 corresponds to a setting unit that sets an axis group that matches the distribution shape of each category that classifies the pattern of the learning sample in the feature space. The pattern recognition apparatus further includes a classification unit (not shown).

  The feature extraction unit 501 extracts feature amounts from character images 510 of learning samples each consisting of a plurality of samples prepared for each of N categories from category 1 to category N. Thereby, the feature extraction unit 501 outputs a plurality of sets of feature vectors 511 for each of N categories from category 1 to category N.

  The eigenvector extraction unit 502 calculates, for each category, eigenvectors from a plurality of sets of feature vectors 511 corresponding to the category, and causes all of the resulting eigenvectors to be included in the first axis set 512 as axis candidates.

  The hierarchical clustering unit 503 clusters the eigenvectors included in the first axis set 512, and outputs a cluster set 513 in which similar eigenvectors are classified into the same cluster.

  The cluster selection unit 504 selects an unselected cluster having the largest number of component eigenvectors in the cluster set 513 (hereinafter referred to as “selected cluster”).

  The orthogonal axis determination unit 505 calculates, from the cluster set 513, a third axis set 515 that includes representative vectors of the selected cluster and in which representative vectors substantially orthogonal to each other are stored.

  The axis output determination unit 506 adds each representative vector in the third axis set 515 to the second axis set 514 (axis candidate group). The axis output determination unit 506 determines the cluster selection unit 504 and the orthogonal axes until the number of representative vectors included in the second axis set 514 exceeds a predetermined threshold or until there is no area (see FIG. 4) in which no learning sample exists. The processing by the unit 505 is repeatedly executed. The axis output determination unit 506 outputs an output axis set 516 (axis group) including each representative vector included in the second axis set 514 as each output axis when the above condition is determined.

  After that, by projecting each feature point determined by a plurality of sets of feature vectors 511 on each axis in the output axis set 516 for each of category 1 to category N, the range on each axis of the category is a margin The dictionary determined in The expression format of the dictionary data is a hash data format similar to the case of FIG. 3 described above. Then, using such a dictionary, a category range including positions obtained by projecting feature points on the feature space corresponding to feature vectors extracted from a given pattern on each axis in the output axis set 516. Are extracted, and a set of categories corresponding to the range is output. The part that executes these processes is, for each category, based on the category range obtained by projecting feature points on the feature space corresponding to the feature vector extracted from the learning sample of the category on each axis of the axis group. It corresponds to a part of a classification unit that classifies a given pattern.

  By using each of the axes in the output shaft set 516 created in the first embodiment described above, the correct solution is obtained in the roughly classified set as compared with the conventional large classification method described in FIGS. 2 to 4 It is possible to further improve the narrowing ability without changing the accuracy of entering. As described in FIG. 4, in the conventional large classification method, useless category regions without learning samples are generated on the feature space depending on the category. As a result, in FIG. 4, the unknown pattern x is roughly classified into the category A in addition to the category B. On the other hand, in the first embodiment, the use of the axes 1 to 4 obtained from the output shaft set 516, for example, as shown in FIG. Can. Therefore, since the useless category area is reduced, it is possible to prevent the selection of a category which is not close to the unknown pattern input at the time of the large classification. As a result, it is possible to shorten the processing time without changing the accuracy of the large classification. In the example of FIG. 6, for example, for category A, an area shown as 601A is defined from the range of category A of axis 1 and the range of category A of axis 2 and the range of category A of axis 3 and category A of axis 4 From the range, an area shown as 602A is determined. As a result, the range of the category A on the feature space defined by the axes 1 to 4 is determined as a region 600A where the regions 601A and 602A overlap. In the range 600A of the category A, there is no region in which no feature point of the category A exists. Similarly, for Category B and Category C, an area indicated by a fine broken line and an area indicated by a coarse broken line, from which an area where there is no feature point in each category has been removed, are respectively taken as the ranges of Category B and C. It becomes settled. As a result of projecting the feature point “X” in the feature space of the unknown pattern x on the axis 1, axis 2, axis 3 and axis 4 by using the dictionary created from such a category range, Major classification can be performed such that the AND product correctly overlaps with only the category "B".

  FIG. 7 is a flowchart showing an example of processing of the feature extraction unit 501 and the eigenvector extraction unit 502 of FIG. First, a control unit (not shown) sets the first axis set 512 to an empty set (step S701).

  Next, in FIG. 5, the control unit selects one unprocessed category (hereinafter referred to as “selected category”) among the categories 1 to N (step S702).

  Next, the feature extraction unit 501 inputs a set of character images 510 of FIG. 5 for the selected category (step S703). Furthermore, the feature extraction unit 501 extracts a set of feature vectors 511 from the set of character images 510 of the selected category input in step S703 (step S704). As the feature vector 511, for example, a vector using the “direction index histogram” described with reference to 105 and 106 in FIG. 1 is employed.

  Subsequently, the eigenvector extraction unit 502 extracts a set of eigenvectors from the set of feature vectors 511 of the selected category calculated in step S704 (step S705). The eigenvectors are calculated, for example, as follows.

  Suppose that m character images are collected for one category (one type of character such as “A” for character recognition), and they are all converted into n-dimensional feature vectors.

  At this time, the k-th vector is defined by the following equation (1).

  The average vector of this category is expressed by the following equation (2).

  At this time, the i-th element of the average vector is defined by the following equation (3).

The covariance matrix Σ is calculated from each vector (k-th vector) expressed by the equation (1) and each element of the mean vector expressed by the equations (2) and (3). The element σ _ij of the i-th row and the j-th column of the covariance matrix Σ is calculated by the following equation (4).

  Next, with the n × n covariance matrix Σ as an object, a unique equation represented by the following equation (5) is created. Here, E is an n × n identity matrix.

Next, the eigenvalue λ is determined by solving the equation (5). λ has n solutions.
Furthermore, one n-dimensional eigenvector x is obtained by solving the equation shown in the following equation (6) using one λ value in the solution of the above-mentioned eigenvalues. By repeating this calculation for each eigenvalue solution, pairs of n eigenvalues and eigenvectors are obtained.

  The eigenvector extraction unit 502 adds the set of eigenvectors extracted in this manner to the first axis set 512 (step S706).

  The control unit determines whether or not the selection has been completed for all the categories of category 1 to category N (step S 707). If the determination in step S707 is NO, the control unit returns to the process in step S702, and executes the series of processes in steps S702 to S706 for the next selected category.

  If the determination in step S707 is YES, the control unit outputs the first axis set 512 (step S708). Thereafter, the control unit ends the above-described repetitive processing by the feature extraction unit 501 and the eigenvector extraction unit 502 illustrated in the flowchart of FIG. 7.

  As described above, the feature vector 511 and the eigenvector are calculated for each selected category by the processing of the feature extraction unit 501 and the eigenvector extraction unit 502 of FIG. 5 illustrated in the flowchart of FIG. The axes of the feature space for the category are obtained in a first set of axes 512.

  Next, a processing example of the hierarchical clustering unit 503 in FIG. 5 will be described. In hierarchical clustering processing, with regard to the eigenvectors obtained as the first axis set 512, those close to the center of gravity are collected at the center of gravity and grouped into a cluster, and if the distance between the center of gravity and the center of gravity is separated, another cluster is obtained. Clustering is performed. Hierarchical clustering is a method of grouping (clusters) in order from the most similar combination. For example, in agglomerative, which is a method of hierarchical clustering processing, when data of N objects is given, an initial state in which N clusters containing only one object are initially It is generated. From this initial state, the distance (dissimilarity) d (C1, C2) between clusters is calculated, and the two closest clusters of this distance are sequentially grouped into one cluster. A hierarchical structure is acquired by repeating this process until all objects are put together into one cluster.

  FIG. 8 is an explanatory diagram of hierarchical clustering processing. FIG. 8 shows an example in which each eigenvector of each category A, B, C exists on a two-dimensional feature space, for the sake of simplicity of explanation. In the feature space of FIG. 8, the axis vector (eigenvector) of category A in the first axis set 512 is the first, second, third, and fourth feature points (portion indicated by the arrow from the symbol “A”) It is indicated by. Further, the axis vector of category B is indicated by the fifth, sixth, seventh, and eighth feature points (portions indicated by arrows from the symbol “B”). Furthermore, the axis vector of category C is indicated by 9, 10, 11, and 12 feature points (portions indicated by arrows from symbol "C"). In the above-described hierarchical clustering process, the initial state starts from twelve clusters including one each of all one numbers from one to twelve. And, finally, the category A No. 2 and the category C No. 9 and No. 11 are combined into one cluster. Also, the category A No. 1 and the category C No. 10 and No. 12 are combined into one cluster. Furthermore, Category B's 5, 8 and 7 are combined into one cluster. Category A No. 4, Category A No. 3 and Category B No. 6 each form a single cluster. Of the axis vectors (eigenvectors) obtained from each category to the first axis set 512, similar axes are represented by one axis (= cluster) by the above hierarchical clustering processing.

  FIG. 9 is a flowchart showing a process example of the hierarchical clustering unit 503 of FIG. 5 based on the above-described hierarchical clustering process. First, the hierarchical clustering unit 503 regards each eigenvector in the first axis set 512 output by the eigenvector extraction unit 502 as a representative vector of one cluster to create a cluster set 513 and calculates the distance between all clusters (Step S901). As the distance, for example, the Euclidean distance between representative vectors of two clusters is calculated.

  Next, the hierarchical clustering unit 503 obtains two clusters for which the smallest distance Dmin is calculated among the distances calculated in step S901 (step S902).

  Next, the hierarchical clustering unit 503 determines whether the minimum distance Dmin is smaller than a predetermined threshold (step S903). If the determination in step S903 is YES, the hierarchical clustering unit 503 integrates the two clusters calculated in step S902 into one new cluster (step S904).

  Next, the hierarchical clustering unit 503 sets the number of eigenvectors belonging to the new cluster generated in step S904 as the value of the total number of eigenvectors respectively belonging to the two clusters which generated the new cluster (step S905). Also, the hierarchical clustering unit 503 deletes from the cluster set 513 the two original clusters that generated the new cluster generated in step S904 (step S906). Furthermore, the hierarchical clustering unit 503 adds the new cluster generated in step S904 to the cluster set 513 (step S907).

  Subsequently, the hierarchical clustering unit 503 calculates the representative vector of the new cluster by calculating, for example, the average or the center of gravity of the representative vectors of the two original clusters which generated the new cluster generated in step S904. Then, the hierarchical clustering unit 503 calculates the distance between all clusters in the cluster set 513 again as in the case of step S901 (step S908). Thereafter, the hierarchical clustering unit 503 returns to the process of step S903 and repeats clustering.

  As a result of the above repetitive operation, when the hierarchical clustering unit 503 determines that the distance between the clusters in the cluster set 513 is sufficiently long and the minimum distance Dmin becomes smaller than the predetermined threshold in step S903, step S909 is performed. Go to the process of In step S909, the hierarchical clustering unit 503 ends the hierarchical clustering process by repeating the series of processes in steps S902 to S908, and outputs the created cluster set 513. Thereafter, the hierarchical clustering unit 503 ends the process illustrated in the flowchart of FIG.

  FIG. 10 is a flowchart showing an example of processing of the cluster selection unit 504, the orthogonal axis determination unit 505, and the axis output determination unit 506 in FIG. First, the cluster selection unit 504 sets the second axis set 514 as an empty set (step S1001).

  Next, the cluster selection unit 504 selects a cluster having the largest number of eigenvectors of components other than already selected in the cluster set 513 output from the hierarchical clustering unit 503 (hereinafter, this cluster is “selected cluster”) And written). The cluster selection unit 504 makes the selected cluster already selected in the cluster set 513 (the above, step S1002).

  Next, the orthogonal axis determination unit 505 calculates, from the cluster set 513, a third axis set 515 in which the representative vectors of the selected cluster selected in step S1002 and including the representative vectors substantially orthogonal to each other are stored (step S1003). ). The details of this process will be described later using the flowchart of FIG.

  Next, the axis output determination unit 506 adds each representative vector in the third axis set 515 to the second axis set 514, and selects a cluster of each representative vector in the third axis set 515 by the cluster set 513. (Step S1004).

  Next, the axis output determination unit 506 determines whether the number of representative vectors in the second axis set 514 has become larger than a predetermined threshold (step S1005).

  When the determination in step 905 is YES, the axis output determination unit 506 outputs an output axis set 516 in which the axes are stored, with each representative vector in the second axis set 514 as an output axis used in the dictionary. (Step S1010). Thereafter, the axis output determination unit 506 ends the processing of the flowchart illustrated in FIG.

  When the determination in step S1005 is NO, the axis output determining unit 506 determines each of the categories 1 to N obtained from the learning sample, each determined by a plurality of sets of feature vectors 511 (see FIG. 5) in the category. Feature points are projected onto all the axes in the second set of axes 514. Thereby, the axis output determination unit 506 calculates the range on each axis in the second axis set 514 of each category by adding a margin, as in the case of FIG. 2 (step S1006).

  After that, the axis output determination unit 506 repeatedly executes the series of processes of steps S1007 to S1009. First, the axis output determination unit 506 selects one of categories 1 to N in FIG. 5 (step S1007).

  Next, the axis output determination unit 506 determines that there is no feature point of the learning sample within the existence range of the category selected in step S1007 on the feature space composed of all the axes in the second axis set 514 (see FIG. It is determined whether or not there is 4) (step S1008).

  If the determination in step S1008 is NO, the axis output determining unit 506 determines whether all categories of categories 1 to N have been selected (step S1009).

  If the determination in step S1009 is NO, the axis output determination unit 506 returns to the process in step S1007, selects the next category, and repeats the determination in step S1008.

  When the determination in step S1009 is YES, the axis output determination unit 506 outputs an output axis set 516 in which the axis data is stored as axis data used in the dictionary as each representative vector in the second axis set 514. (Step S1010). Thereafter, the axis output determination unit 506 ends the processing of the flowchart illustrated in FIG.

  The axis output determination unit 506 repeats the process from step S1007 to step S1009 to find that there is no feature point of the learning sample in the existing range of the selected category on the feature space of all axes in the second axis set 514 (step If it is determined that the determination in S1008 is YES), the process returns to step S1002. Thus, the axis output determination unit 506 repeatedly executes the processing by the cluster selection unit 504 and the orthogonal axis determination unit 505. A series of operations from step S1002 to S1009 are performed until the number of representative vectors in the second axis set 514 exceeds the threshold (YES in the determination in step S1005) or until there is no area without learning samples (see FIG. 4) The determination in S1009 is YES), and is repeatedly executed. Here, the selected cluster first selected by the cluster selection unit 504 and the selected cluster next selected after step S1008 are not necessarily orthogonal to each other. Thus, ultimately, the axes included in the output axis set 516 may include non-orthogonal axes.

  FIG. 11 is a flowchart showing a detailed processing example of the orthogonal axis determination unit 505 in step S1003 of FIG. First, the orthogonal axis determination unit 505 sets the number of dimensions of the feature vector 511 to the variable d in FIG. The orthogonal axis determination unit 505 also receives the cluster set 513 output by the hierarchical clustering unit 503 and the representative vector of the selected cluster selected by the cluster selection unit 504 in step S1002 in FIG. 10 (the above, step S1101).

  Next, the orthogonal axis determination unit 505 stores only the representative vector of the selected cluster in the third axis set 515, and sets the value “1” to a variable n indicating the number of representative vectors of the third axis set 515 (step S1102).

  Further, the orthogonal axis determination unit 505 resets the intra-cluster eigenvector number maximum value Max to 0, and sets the cluster selection flag T to a value “−1” indicating a non-selected state (step S1103).

  After that, the orthogonal axis determination unit 505 selects one cluster (hereinafter referred to as “cluster C”) from the cluster set 513 other than the clusters selected in step S1002 and S1004 in FIG. 10 or step S1110 in FIG. Description) is selected (step S1104).

  Next, the orthogonal axis determination unit 505 calculates the inner product value of the representative vector of cluster C and all the representative vectors in the third axis set 515, and determines whether the inner product value is smaller than a predetermined threshold value, ie, 2 It is determined whether two representative vectors are substantially orthogonal (step S1105).

  If the determination in step S1105 is YES, the orthogonal axis determination unit 505 determines whether the number of eigenvectors that are components of cluster C is larger than the intra-cluster eigenvector number maximum value Max (step S1106).

  If the number of eigenvectors of cluster C is larger than the maximum value Max so far (if the determination in step S1106 is YES), the orthogonal axis determination unit 505 executes the process of step S1107. In step S1107, the orthogonal axis determination unit 505 sets the cluster number of the cluster C in the cluster selection flag T, and sets the number of eigenvectors constituting the cluster C in the intra-cluster eigenvector number maximum value Max. If the number of eigenvectors of the cluster C is not larger than the maximum value Max so far (if the determination in step S1106 is NO), the orthogonal axis determination unit 505 skips the process of step S1107.

  After that, the orthogonal axis determination unit 505 checks whether all unselected clusters other than the clusters selected in step S1002 and S1004 in FIG. 10 or step S1110 in FIG. Is determined (step S1108).

  If the determination in step S1108 is NO, the orthogonal axis determination unit 505 returns to the process of step S1104 and repeatedly executes the examination of the next unselected cluster.

  When the determination in step S1108 is YES, the orthogonal axis determination unit 505 determines whether the value of the cluster selection flag T is not “−1”, that is, a representative vector substantially orthogonal to all representative vectors in the third axis set 515. Is determined (step S1109).

  If the determination in step S1109 is YES, the orthogonal axis determination unit 505 adds the representative vector of the cluster of the number set in the cluster selection flag T to the third axis set 515 in the cluster set 513. Further, the orthogonal axis determination unit 505 makes the cluster having the above-described number already selected in the cluster set 513 (step S1110).

  In this manner, one cluster that is substantially orthogonal to all representative vectors in the third axis set 515 and has the largest number of constituent eigenvectors is selected and added to the third axis set 515 anew. It will be

  Further, the orthogonal axis determination unit 505 increments the value of the variable n indicating the number of representative vectors in the third axis set 515 by 1 (step S1111).

  Thereafter, the orthogonal axis determination unit 505 resets the intra-cluster eigenvector number maximum value Max to 0, and sets the cluster selection flag T to a value “−1” indicating a non-selected state (step S1112).

  The orthogonal axis determination unit 505 determines whether the value of the variable n indicating the number of representative vectors in the third axis set 515 is smaller than the value of the variable d indicating the number of dimensions of the feature vector 511 (step S1113). ).

  If the determination in step S1113 is YES, the orthogonal axis determination unit 505 returns to the process of step S1104 and is substantially orthogonal to all the representative vectors in the third axis set 515, and the number of constituent eigenvectors is next Selection of a large number of clusters and addition to the third axis set 515 are repeatedly attempted.

  The orthogonal axis determination unit 505 outputs the third axis set 515 (step S1114) when the following condition is satisfied in the above-described repetitive processing (step S1114), and the step S1003 in FIG. End the processing of. This end condition is a case where it is determined that a representative vector substantially orthogonal to all the representative vectors in the third axis set 515 has not been obtained (the determination in step S1109 is NO). Alternatively, it is determined that the number n of representative vectors in the third axis set 515 has reached the number d of dimensions of the feature vector 511 (the determination in step S1113 is NO).

  In the processing of the flowcharts of FIGS. 10 and 11, a cluster including as many eigenvectors as possible is selected, and a cluster including as many eigenvectors as possible even when representative vectors including the representative vectors are selected and substantially orthogonal to each other are selected. Is selected. By this processing, it is possible to obtain an output axis set 516 composed of representative vectors that can encompass the range of as many categories as possible without waste. By using each of the axes in the output axis set 516 created in this way, it is possible to remove useless category areas without learning samples on the feature space, and in the large classification, the processing time is not changed. It is possible to shorten and improve the narrowing ability.

  FIG. 12 is a block diagram of a second embodiment of the pattern recognition apparatus. The second embodiment includes a feature extraction unit 501, a common axis extraction unit 1201, each category range extraction unit 1202, an axis addition category determination unit 1203, an axis addition unit 1204, and an axis output determination unit 1205.

  The feature extraction unit 501 is the same as the feature extraction unit 501 of FIG. 5, and the feature amounts are extracted from character images 510 of learning samples each consisting of a plurality of samples prepared for each of N categories from category 1 to category N. Extract. Thereby, the feature extraction unit 501 outputs a plurality of sets of feature vectors 511 for each of N categories from category 1 to category N. The character image 510 and the feature vector 511 are the same as the character image 510 and the feature vector 511 of FIG. 5, respectively.

  The common axis extraction unit 1201 obtains a specified number of eigenvectors as common axes from the feature vectors 511 of all the learning samples of all the categories from category 1 to N, and creates a common axis set 1210.

  The category range extraction unit 1202 projects the feature points in the feature space corresponding to the feature vectors 511 of all the learning samples in the category on each common axis in the common axis set 1210 for each of the categories 1 to N. Find the value on each common axis and add a fixed margin. Thus, the category range extraction unit 1202 creates a category range on each common axis for each category.

  The axis additional category determination unit 1203 specifies the categories of the learning sample of the specified category in the range of the category in the feature space obtained from the range of the specified category of each common axis while specifying the categories 1 to N one by one. It is determined whether there is an area having no feature point. The axis additional category determination unit 1203 determines a category having a region having no feature point of the learning sample as a category to which an axis is to be added.

  The axis addition unit 1204 calculates an eigenvalue and an eigenvector only from the feature vector 511 of the learning sample of the category determined by the axis addition category determination unit 1203. The axis adding unit 1204 adds one by one to the additional axis set 1211 (initial value is the common axis set 1210) using the eigenvectors as new additional axes in descending order of eigenvalues. The axis addition unit 1204 adds the above eigenvectors one by one until the range of the above-described determined category in the feature space defined by the additional axis set 1211 has no feature point of the learning sample of that category. Repeat the process.

  The axis output determination unit 1205 repeats the processing by the above-described additional axis category determination unit 1203 and the additional axis unit 1204. The axis output determination unit 1205 outputs an output axis set 1212 having each axis in the additional axis set 1211 as an output axis when the processing for all the categories is completed.

  13, 14 and 15 are explanatory diagrams of the second embodiment. In the second embodiment described above, the common axis extraction unit 1201 and the category range extraction unit 1202 are similar to the conventional large classification method described in FIG. 2 to FIG. Each category range of is extracted. In the category range on the feature space obtained from the category range on each common axis, the axis additional category determination unit 1203 determines whether there is a region where there is no feature point obtained from the feature vectors 511 of all the learning samples of the category. Determine For example, in the example of FIG. 13, the range of category A on the feature space obtained from the category range of common axis 1 and common axis 2 is the area indicated by 1301A and 1302A. Among these, in the area 1301A, a feature point “A” obtained from the feature vector 511 of the category A of the learning sample is present. However, as in the case of FIG. 4, there is no feature point “A” of category A in the area 1302A. The axis additional category determination unit 1203 determines that there is an area 1302A in which there is no feature point of category A for category A. Based on the determination result, the axis addition unit 1204 adds additional axes 1 and 2 determined as eigenvectors from the feature vector 511 of only the category A of the learning sample, in descending order of eigenvalues, until the region 1302A disappears. As a result, for example, for the category A, the areas shown as 1301 A and 1302 A are determined from the range of the category A of the common axis 1 and the range of the category A of the common axis 2. Further, an area indicated as 1303 A is determined from the range of the category A of the additional axis 1 and the range of the category A of the additional axis 2. As a result, the range of the category A on the feature space defined by the common axes 1 and 2 and the additional axes 1 and 2 is determined as an area 1300A where the areas 1301A and 1303A overlap. In the range 1300A of this category A, there is no region in which no feature point of category A exists.

  As illustrated in FIG. 14, similarly for Category B and Category C, a region indicated by a fine broken line and a rough region in which regions where there is no feature point in each category have been removed by additional axes 1 and 2 Areas indicated by broken lines are determined as the ranges of categories B and C, respectively. In addition, as for the category B and the category C, an additional axis may be further added when it is not possible to remove an area in which there is no feature point of each category by the additional axes 1 and 2 alone. As a result of projecting the feature point “X” in the feature space of the unknown pattern x on the common axis 1, the common axis 2, the additional axis 1, and the additional axis 4 by using the dictionary created from such a category range, Major classification can be performed such that the AND product of each projection result correctly overlaps with only the category "B".

  As shown in FIG. 15, for each of the categories A, B and C, for example, the axes 1 and 2 of A, the axes 1 and 2 of B, and the axes 1 and 2 of C, respectively. There is an optimal axis along the distribution. Here, since the category A and the category C show close distributions, axes 1 and 2 obtained from the categories A and C can be adopted.

  FIG. 16 shows a process example of the common axis extraction unit 1201, the category range extraction unit 1202, the axis addition category determination unit 1203, the axis addition unit 1204, and the axis output determination unit 1205 in the second embodiment shown in FIG. It is a flowchart.

  First, the common axis extraction unit 1201 obtains a defined number of eigenvectors as common axes from the feature vectors 511 of all the learning samples of all the categories 1 to N extracted by the feature extraction unit 501, and generates a common axis set 1210. (Step S1601).

  Next, the category range extraction unit 1202 projects the feature points in the feature space corresponding to the feature vectors 511 of all the learning samples in the category on each common axis in the common axis set 1210 for each of the categories 1 to N. Calculate the value on each common axis and add a fixed margin. Thus, the category range extraction unit 1202 creates a category range on each common axis for each category. The category range extraction unit 1202 sets the common axis set 1210 as the initial value of the additional axis set 1211 (the above, step S1602).

  Next, the axis output determination unit 1205 designates one unprocessed category among the categories 1 to N (hereinafter, this category is referred to as “designated category”) (step S1604).

  Next, in the range of the category in the feature space obtained by the axis additional category determination unit 1203 from the range of the specified category of each axis in the additional axis set 1211, there is a region where there are no feature points of the learning sample of the specified category. It is determined whether there is any (step S1605). If the determination in step S1605 is NO, the process proceeds to step S1610 by the axis output determination unit 1205.

  If the determination in step S1605 is YES, the axis adding unit 1204 calculates eigenvalues and eigenvectors only from the feature vector 511 of the learning sample of the specified category, and sorts the eigenvectors in descending order of the corresponding eigenvalues (step S1606).

  The axis adding unit 1204 selects one eigenvector in the sorted order, and adds an additional axis defined by the eigenvector to the additional axis set 1211 (step S1607). The additional axes added here are not necessarily orthogonal to the common axis and the additional axes already added. Thus, ultimately, the axes included in the output shaft set 1212 may include non-orthogonal axes.

  The axis adding unit 1204 determines whether or not there is a region having no feature point of the learning sample of the specified category in the range of the category in the feature space obtained from the range of the specified category of each axis in the additional axis set 1211 It determines (step S1608).

  If the determination in step S1608 is YES, the axis adding unit 1204 determines whether the number of axes in the additional axis set 1211 is larger than a predetermined threshold (step S1609).

  If the determination in step S1609 is NO, the axis adding unit 1204 returns to the process of step S1607, selects the next eigenvector in the rearranged order, and adds an additional axis defined by the eigenvector to the additional axis set 1211 to add. Then, the axis adding unit 1204 repeats the determination in step S1608 for the feature space determined from the new additional axis set 1211.

  As a result of the series of processes from step S1607 to S1609 being repeated, if there is no area having no feature point of the learning sample of the specified category in the feature space and the reflection of step S1608 becomes NO, the axis output determination unit 1205 The process proceeds to step S1610. Also, when the number of axes of the additional axis set 1211 exceeds the predetermined threshold and the determination in step S1609 is YES, the process proceeds to step S1610 by the axis output determination unit 1205.

  In step S1610, the axis output determination unit 1205 determines whether the processing of all the categories 1 to N has been completed (step S1610).

  If the determination in step S1610 is NO, the axis output determination unit 1205 returns to the process of step S1604, specifies one unprocessed new category, and executes the series of processes from step S1605 to S1609 for the specified category. Do.

  If the determination in step S1610 is YES, the axis output determination unit 1205 outputs an output axis set 1212 including each axis obtained in the additional axis set 1211 as an output axis used in the dictionary. Thereafter, the axis output determination unit 1205 ends the process illustrated in the flowchart of FIG.

  By using each axis in the output axis set 1212 created by the processing of the flowchart in FIG. 16, it is possible to remove useless category areas without learning samples in the feature space, and the accuracy is not changed at the time of major classification. Processing time can be shortened, and improvement in narrowing capability can be realized.

  FIG. 17 is a diagram showing an example of a hardware configuration of a computer 1701 capable of realizing the pattern recognition apparatus of the first embodiment of FIG. 5 or the second embodiment of FIG. The computer 1701 shown in FIG. 17 includes a processor 1705, a memory 1706, a hard disk drive (which may be a semiconductor disk drive) 1707, and an input / output interface 1708. Furthermore, the computer 1701 has an optical drive device 1709 into which a portable recording medium 1711 such as a DVD disk is inserted. These components are connected to one another by a bus 1710. A scanner 1702 and a camera 1703 are externally connected to the input / output interface 1708, and a character image 510 (see FIG. 5 or 12) can be read. The configuration shown in FIG. 17 is an example of a computer that can realize the pattern recognition apparatus, and such a computer is not limited to this configuration.

  The memory 1706 is, for example, a semiconductor memory such as a read only memory (ROM), a random access memory (RAM), or a flash memory, and stores a program and data used for processing.

  The processor 1705 operates as each block shown in FIG. 5 or 12 by executing a program corresponding to the processing of the flowchart of FIG. 7, FIG. 9 to FIG. 11 or FIG. Do.

  For example, the hard disk drive 1707 includes the character image 510 or the feature vector 511 of FIG. 5 or 12, the first axis set 512 of FIG. 5, the cluster set 513, the second axis set 514, the third axis set 515, the output axis set 516, The common axis set 1210, the additional axis set 1211 or the output axis set 1212 shown in FIG. 12 are stored.

  The pattern recognition apparatus shown in FIG. 5 or 12 stores programs and data corresponding to the processes of the flowcharts of FIG. 7 and FIGS. 9 to 11 or 16 in the hard disk drive 1707 and loads them into the memory 1706 Can be used.

  The optical drive device 1709 drives the portable recording medium 1711 and accesses the recorded contents. An operator or a user can store the program and data described above in the portable recording medium 1711 and load it into the memory 1706 for use.

  Thus, the computer-readable recording medium for storing the program and data described above is a physical (non-transitory) recording medium such as the memory 1706, the hard disk drive 1707, or the portable recording medium 1711. .

  Note that the above-described program or data may be received from an external device via a network interface (not shown) and loaded into the memory 1706 for use.

  In addition, it is not necessary to include all the components of the pattern recognition apparatus of FIG. 5 or FIG. 12, and it is also possible to abbreviate | omit a part of components according to a use or conditions.

  While the embodiments of the disclosure and the advantages thereof have been described in detail, those skilled in the art can make various changes, additions, and omissions without departing from the scope of the present invention as set forth in the claims. .

  In addition, the present invention is not limited to the embodiment described above, and can be variously modified in the implementation stage without departing from the scope of the invention. Also, the functions executed in the above-described embodiment may be implemented in combination as appropriate. The embodiment described above includes various steps, and various inventions can be extracted by appropriate combinations of a plurality of disclosed configuration requirements. For example, if some configuration requirements are removed from all the configuration requirements shown in the embodiment, a configuration from which this configuration requirement is removed can be extracted as the invention if the effect can be obtained.

501 Feature Extraction Unit 502 Eigenvector Extraction Unit 503 Hierarchical Clustering Unit 504 Cluster Selection Unit 505 Orthogonal Axis Determination Unit 506 Axis Output Determination Unit 510 Character Image 511 Feature Vector 512 First Axis Set 513 Cluster Set 514 Second Axis Set 515 Third Axis Set 516, 1212 Output axis set 1201 Common axis extraction section 1202 Category range extraction section 1203 Axis additional category determination section 1204 Axis addition section 1205 Axis output determination section 1210 Common axis set 1211 Additional axis set

Claims (6)

Set the axis group according to the distribution shape in the feature space of each category that classifies the pattern of learning sample,
For each of the categories, given patterns are provided based on category ranges obtained by projecting feature points on the feature space corresponding to feature vectors extracted from learning samples of the categories on each axis of the axis group. Classify,
Pattern recognition program for making a computer execute processing. A dictionary in which a category range obtained by projecting feature points on the feature space corresponding to a feature vector extracted from a learning sample of the category onto each axis of the axis group is associated with the category for each category Create
The category range including the position obtained by projecting the feature point on the feature space corresponding to the feature vector extracted from the given pattern onto each axis of the axis group is extracted, and the extracted category range is extracted. Classify the category to which the given pattern belongs by outputting a set of corresponding categories,
The pattern recognition program according to claim 1, further causing the computer to execute a process. In the process of setting the axis group,
Calculating an eigenvector from the feature vector of the learning sample for each category, clustering all eigenvectors obtained as a result, and outputting a cluster set in which similar eigenvectors are classified into the same cluster;
An axis candidate as an axis candidate by calculating, from the cluster set, representative vectors that include representative vectors of the selected cluster and that are orthogonal to each other within a predetermined error while selecting a cluster with the largest number of eigenvectors in the cluster set The process of sequentially adding to the group,
The feature points on the feature space corresponding to the learning samples of each category are projected to each axis candidate included in the axis candidate group until the number of axis candidates included in the axis candidate group exceeds a predetermined threshold After repeatedly executing the process of sequentially adding the axis candidate to the axis candidate group until there is no area where the feature point does not exist in the intersection area in the feature space of the category range obtained by A process of outputting as the axis group;
The pattern recognition program according to claim 1 or 2, which executes. In the process of setting the axis group,
A process of obtaining a predetermined number of eigenvectors as a common axis from feature vectors of all the learning samples of all the categories;
A process of extracting a category range on the common axis by projecting feature points on the feature space corresponding to learning samples of the category on the common axis for each category;
After each common axis is included in the axis candidate group, there is also one feature point corresponding to a learning sample of the category in the intersection region in the feature space of the range of the category on the common axis for each category. If there is no region, a process of calculating eigenvalues and eigenvectors only from feature vectors of learning samples of the category, and adding the eigenvectors as new additional axes to the axis candidate group in descending order of the eigenvalues is the axis candidate A process of repeating until the range of the category in the feature space defined by the axis in the group has no feature point corresponding to the learning sample of the category;
The pattern recognition program according to claim 1 or 2, which executes. In the feature space, the computer sets an axis group that matches the distribution shape of each category that classifies the pattern of the learning sample,
For each of the categories, given patterns are provided based on category ranges obtained by projecting feature points on the feature space corresponding to feature vectors extracted from learning samples of the categories on each axis of the axis group. Classify,
Pattern recognition method. A setting unit configured to set an axis group according to the distribution shape of each category that classifies the pattern of the learning sample in the feature space;
For each of the categories, given patterns are provided based on category ranges obtained by projecting feature points on the feature space corresponding to feature vectors extracted from learning samples of the categories on each axis of the axis group. A classification unit to classify,
Pattern recognition apparatus comprising: