JP5521497B2 - Pattern recognition apparatus, pattern recognition program, and pattern recognition method - Google Patents

Description

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

  In recent years, the e-Document Law (Electronic Document Law: Law No. 149, Law No. 149, Law No. 150, 2004) that allows documents to be stored as images has been enforced, and search needs for document images have increased. Pattern recognition technology is used in various fields.

  Pattern recognition processing for a document image can be divided into layout analysis processing for obtaining a text region from the document image and character recognition processing for converting a character image into a character code for the extracted text region. Further, it is known that when the processing time required for the pattern recognition of the document image is measured, the character recognition processing requires more processing time than the layout analysis processing. Thus, proposals have been made to reduce the time required for pattern recognition processing and to speed up pattern recognition processing.

  For example, as a method for speeding up the character recognition processing part, a simple process that sacrifices character recognition accuracy (character recognition accuracy rate) has been proposed. However, it is not preferable to reduce the character recognition accuracy in consideration of practicality such as retrieval and reuse of stored documents. Therefore, it is required to perform character recognition processing at high speed while maintaining high character recognition accuracy.

  Therefore, the following technologies relating to character recognition processing have been proposed. First, the target character code (corresponding to JIS 1st level kanji, JIS 2nd level kanji, hiragana, katakana, symbols, vertical writing characters, English letters, numbers, half-width characters, etc.) Collect a plurality of corresponding character images (learning sample patterns). The pattern recognition apparatus converts each collected character image into a feature vector, creates an average vector for each character code, records the character code and the average vector as a pair in a recording unit, and generates a dictionary. The same is done for the other character codes, and a plurality of character codes and feature vector information corresponding to the character codes are recorded in the recording unit.

Here, terms frequently used in pattern recognition are simply defined. A recognition target is called a pattern, and a set made up of all patterns is called a pattern space.
A set of one or more feature amounts obtained by extracting features from 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 feature vector value, 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 that 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. In particular, a set of patterns that can be regarded as the same type is referred to as a category pattern set, and a set of feature vectors that can be regarded as the same type is referred to as a category feature set.

  Determining to which category (category pattern set or category feature set) an unknown pattern input from a pattern recognition apparatus or the like or a feature vector obtained from the unknown pattern belongs is called pattern recognition. In particular, when it is estimated that the input pattern or feature vector may belong to a certain category in the category set, the category set is called a candidate category set.

  When performing character recognition processing, if an unknown character image is input, a feature vector is generated from the input character image, and matching is performed with a plurality of feature vectors held in a dictionary (such as a memory). . Matching is performed by calculating the distance between the feature vector of the unknown character and the feature vector of the dictionary (matching process). As a result, distance values corresponding to the number of character codes in the dictionary are obtained, and among them, the character code having the minimum distance value is output as the recognition result of the unknown character.

  However, in the character recognition processing described above, much processing time is spent on matching processing between feature vectors of unknown characters and feature vectors of all character codes in the dictionary. Further, in a language with many types of character codes such as Japanese, the matching processing time is further increased.

  Therefore, in order to speed up the matching process, a two-stage 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 that is a recognition target from a set of feature vectors in the dictionary and narrowing down a target when performing detailed classification. If the detailed classification is performed only for the narrowed-down character codes, the conventional matching processing time can be greatly reduced.

  For example, as shown in FIG. 1, a dictionary used in the large classification is generated by collecting character images belonging to one character type (category) or template (one group when one category is divided into a plurality of groups). Ask for. Next, all the sets are converted into n-dimensional (n: 1 or more integer) feature vectors to create a feature vector set of learning sample patterns of one category. Next, paying attention to one axis among n axes (axis corresponding to elements constituting the feature vector), projecting all the feature vector sets of the learning sample pattern of one category on this axis. The range of values of elements belonging to this category is obtained on the axis. Here, the range A of the axis to be projected in FIG. 1 is a range A of −127 to 128 (256 bits) by quantizing the feature quantities indicating the elements constituting all the feature vectors. Next, the minimum value min and the maximum value max in the range A on the axis are obtained, and the range B represented by the minimum value and the maximum value on the axis is expanded using a predetermined margin value margin. To range C. Assume that this category exists between the two expanded values (min-margin, max + margin). Then, the above-described enlarged range C is calculated for all n dimensions in the n dimension, and the range C is recorded in the storage unit, thereby generating a dictionary for one character type. For other character types, dictionaries are generated in the same manner as described above to generate dictionaries for all categories or templates. Hereinafter, the dictionary generated as described above is referred to as a planar dictionary.

  The plane dictionary can be represented by an array plane as shown in FIG. In the example of FIG. 2, the horizontal is a plane in which the distribution of each axis is shown in the range of −127 to 128 numerical values (256 bits), and the vertical is the number of categories. An identification number (1, 2, 3,...) Is assigned to each category for each category. FIG. 2 is an example showing an expanded range C of three types of categories in the n-dimensional space. Circles (A1, A2) indicate category 1, triangles (A3, A4) indicate category 2, and squares (A5, A6) indicate category 3.

  When attention is paid to the templates A1 and A2, the coordinate position lowered from A1 becomes the minimum value min, and the coordinate position lowered from A2 becomes the maximum value max. In FIG. 2, the margin value margin is indicated as M. The coordinate position shifted to the left and right by the margin value M from the coordinate position of the minimum value min and maximum value max is the expanded range of category 1. In order to record this range in the plane dictionary, bit “1” is set in the target range in 256 bits of −127 to 128 of category 1, and bit “0” is set in the other ranges to set the plane dictionary. To record. The same processing is performed for the other categories to complete the planar dictionary.

  In the large classification process, in the pattern recognition process, the unknown character image is first converted into an n-dimensional feature vector and projected onto each axis. Next, referring to the coordinate position of each of the n axes with respect to the unknown character image and the coordinate position on the axis held in the plane dictionary, it is determined whether or not there is a bit “1” at the same coordinate position, A category close to an unknown character image is obtained. The same processing is performed for the other axes, and categories existing in all axes are obtained, and the categories are narrowed down to obtain a large classification result (candidate category set).

Japanese Patent Laid-Open No. 10-289320

“High-speed and high-precision large classification method by projective estimation of feature regions” Fujimoto, Kamada, Kurokawa, IEICE Technical Report, Pattern Recognition / Media Understanding (PRMU), IEICE Tech. 97 no. 558, PRMU 97-220, pp. 25-32. February 19, 1998

  However, in the above method, the distribution corresponding to the feature amount of the feature vector of the learning sample pattern on the axis of one category or template is expressed by the minimum value and the maximum value. For this reason, even if a feature amount does not exist on the axis of one category or template, it is recorded in the planar dictionary as if there is a feature amount. Therefore, a template that should be excluded from the target of detailed classification in the large classification process is included in the result of the large classification, the narrowing-down ability at the time of pattern recognition is reduced, and high-precision large classification cannot be performed. When it is created using the enlarged range C, the situation shown in FIG. 3 occurs. FIG. 3 shows a two-dimensional feature space. Also, FIG. 3 shows distribution ranges (distribution range Aa of “A”, distribution range Ba of “B”, recorded in the dictionary of character types “A”, “B”, and “C” created using the expanded range C. The relationship between the distribution range Ca of “C” and the actual distribution range (range Ar of “A”, range Br of “B”, range Cr of “C”) is shown. In FIG. 3, when the unknown character “X” is input, the unknown character “X” actually has to be narrowed down to the category “C”, but in FIG. The letter “X” belongs. Therefore, the unknown character “X” is not actually close to the distribution of A, but the categories of the characters “C” and “A” are selected as category candidates close to “X”.

  Accordingly, an object of the present invention is to provide a pattern recognition apparatus, a pattern recognition program, and a pattern recognition method that perform pattern recognition with high accuracy.

  A pattern recognition apparatus which is one embodiment includes a recording unit and a classification unit. The recording unit has a candidate table. The candidate table projects, for each of a plurality of categories that classify the sample pattern for learning, an element at the same position of a reference feature vector obtained for each of a plurality of feature vectors included in the same category on an axis in a preset range. Each of the elements on the axis is generated by associating a preset margin amount with a category.

  The classification unit obtains a reference feature vector of a given pattern, classifies each element of the reference feature vector by using the candidate table, obtains a candidate category set, and outputs the classified candidate category set.

  The disclosed embodiment relating to pattern recognition has the effect of performing pattern recognition accuracy with high accuracy.

It is a figure which shows the outline | summary of the conventional candidate table (dictionary for large classification). It is a figure which shows the outline | summary of the conventional candidate table (dictionary for large classification). It is a figure which shows the relationship between the distribution range of a character type recorded in a candidate table (large classification dictionary), and an actual distribution range. It is a block diagram which shows an example of a structure of the pattern recognition apparatus in Example 1. FIG. It is a figure which shows an example of a feature vector. FIG. 10 is a flowchart illustrating an example of an operation for creating a candidate table (large classification dictionary) in the first embodiment. It is a figure which shows the relationship between a category and a template. It is a figure which shows an example of the data structure when the feature vector for every sample pattern for learning is matched and recorded on the category. A is a diagram showing an element projected on the axis, and B is a diagram showing that a margin is provided for the element projected on the axis. It is a figure which shows an example of the data structure at the time of associating the feature vector with the category, and having recorded on the axis | shaft and the position projected on the axis | shaft for every element of the feature vector. It is a figure which shows the margin width | variety (bit "1") of each axis | shaft. It is a figure which shows the outline | summary of the candidate table (large classification dictionary) in Example 1. FIG. It is a figure which shows the outline | summary of the large classification process in Example 1. FIG. It is a figure which shows an example of an operation | movement when performing a large classification process using the candidate table (large classification dictionary) in Example 1. FIG. It is a figure which shows the outline | summary of the large classification process in Example 1. FIG. It is a block diagram which shows an example of a structure of the pattern recognition apparatus in Example 2. FIG. FIG. 10 is a flowchart illustrating an example of an operation for determining a margin amount in the third embodiment. In Example 3, it is a figure which shows an example of the data structure at the time of recording by matching a category and margin amount. FIG. 10 is a flowchart illustrating an example of an operation for determining a margin amount in the fourth embodiment. It is a figure which shows an example of the outline | summary of the area on an axis | shaft in Example 4. FIG. In Example 4, it is a figure which shows an example of the data structure at the time of recording by matching a category, an area, and margin amount. FIG. 10 is a flowchart illustrating an example of an operation for determining a margin amount in the fifth embodiment. In Example 5, it is a figure which shows an example of the data structure at the time of recording by matching a font and margin amount. In Example 5, it is a figure which shows an example of the data structure at the time of recording by matching a category, a font, and margin amount. It is a figure which shows the structure when an Example is implement | achieved as a computer program.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
Example 1
The pattern recognition apparatus according to the first embodiment includes a recording unit and a classification unit (large classification unit). The recording unit has a candidate table. The candidate table is generated as follows. For each of a plurality of categories that classify the sample pattern for learning, an element at the same position of a reference feature vector obtained for each of a plurality of feature vectors included in the same category is projected onto an axis in a preset range. Then, for each element on the axis, a predetermined margin amount and a category are generated in association with each other. The reference feature vector is a feature vector calculated from the feature vector.

  A classification unit (major classification unit) obtains a reference feature vector of a given pattern, classifies each element of the reference feature vector using a candidate table, obtains a candidate category set, Output.

  FIG. 4 is a diagram showing the configuration of an embodiment of the pattern recognition apparatus. The pattern recognition apparatus 1 shown in FIG. 4 includes a feature extraction unit 2, a large classification unit 3 (classification unit), a detailed classification unit 4, a table creation unit 5, and a recording unit 7.

The feature extraction unit 2 acquires a pattern including a peripheral blank area, and makes the pattern a normalized image by scaling. Next, the feature extraction unit 2 extracts pattern feature amounts from the normalized character pattern, and arranges the extracted feature amounts to generate a feature vector. For example, when a character image is acquired as a pattern, the feature extraction unit 2 generates a feature vector by arranging the feature amounts (x1 to xn) of the input character image (Equation 1). N in Formula 1 is an integer of 1 or more.
X = (x1, x2, x3, x4,... Xn) Equation 1

FIG. 5 shows an example when variable division contour direction feature extraction is used. The feature extraction unit 2 obtains the character image “A”, converts the character image into a normalized image of 48 × 48, and divides the character region in a variable division so that the number of contour points is constant. Are extracted as variable division contour direction features. The number of divisions of the normalized image is set to 12 × 6 or 6 × 12, and the feature vector is obtained by counting the amount of contour directions in four directions of vertical, horizontal, right diagonally upper, and right diagonally lower in each divided region. In the case of FIG. 5, the character image “A” is divided into 72, and the amount of contour direction in four directions is obtained for each division (vertical, horizontal, diagonally right upward, diagonally down right) and 288 (72 X4) feature quantities (xA11 to xA1288) are arranged, and a feature vector is generated as shown in Equation 2.
XA1 = (xA11, xA12, xA13, xA14,
xA15, xA16, xA17, xA18,
.... Formula 2
xA1285, xA1286, xA1287, xA1288)

  For feature extraction, methods such as variable division contour direction feature extraction, extraction using a weighted direction index histogram, and variable division contour direction feature extraction may be used.

  The feature extraction unit 2 obtains a reference feature vector from the calculated feature vector. For example, the reference feature vector may be obtained by excluding some elements from the elements constituting the feature vector, or the feature vector may be used after dimension compression described later. It is not always necessary to generate a reference feature vector from a feature vector, and a feature vector may be used as a reference feature vector. Hereinafter, the reference feature vector will be described as a feature vector.

  Next, the feature extraction unit 2 sets the value of each element of the feature vector of the learning sample pattern to a value within a predetermined range. Here, the predetermined range is a range (feature space) including a minimum value and a maximum value that can be taken by the feature amount indicating the element of the feature vector in all the learning sample patterns, and the range is quantized. It is expressed in a numerical range. When the numerical range is represented by m bits (m: integer), it is preferable to determine the numerical range based on the storage capacity of the memory. For example, when the numerical range is represented by −127 ≦ m ≦ 128 (256 bits), the feature amount of the learning sample pattern is quantized to a value within the numerical range of −127 to 128.

  Further, when generating a candidate table or the like used for large classification using the learning sample pattern, the feature extraction unit 2 transfers the quantized feature vector of the learning sample pattern to the table creation unit 5. When pattern recognition processing is performed on an unknown input pattern, the feature extraction unit 2 transfers the quantized feature vector of the unknown input pattern to the large classification unit 3. Here, the unknown input pattern is a pattern recognition target pattern, such as a character image read by a scanner or the like.

  The large classifying unit 3 extracts a feature vector close to a feature vector generated from the input pattern from a set of feature vectors of a candidate table (large classification dictionary) described later recorded in the recording unit 7 (extracts a candidate category set) ).

  The detailed classification unit 4 calculates the distance between the reference feature vector of the candidate category set classified by the large classification unit 3 and a feature vector generated from an input pattern (such as an input character image). After that, the detailed classification unit 4 selects the smallest distance value among the calculated distance values, extracts the category corresponding to the selected distance value, and sets it as the pattern recognition result.

  The table creation unit 5 includes a margin determination unit 6 and generates a large classification dictionary, which will be described later. The table creation unit 5 acquires the feature vector of the learning sample pattern quantized from the feature extraction unit 2 and records the quantized feature vector in association with the category recorded in advance in the recording unit 7. For example, the category is a character code such as JIS level 1 kanji, JIS level 2 kanji, hiragana, katakana, symbols, vertical writing characters, letters, numbers, half-width characters, and the like. In addition, the table creation unit 5 records the margin amount determined by the margin determination unit 6 in the recording unit 7 in association with each element of the feature vector of the quantized learning sample pattern. Here, the margin amount is quantized so that it can be classified even when the normal learning sample pattern is deteriorated (character image deteriorated (fading, blurring, dirt, etc.) due to facsimile or copying). This is a value for giving a width to each element of the feature vector.

  The margin determination unit 6 determines, for example, the margin amounts provided in the minimum value direction and the maximum value direction around the values on the axis corresponding to each element. Further, the margin amount may be associated with the same margin amount for all elements of all categories, or may be associated with each element individually. The margin amount is indicated by a value in a numerical range obtained by quantizing a range including the minimum value and the maximum value that can be taken by the feature amount indicating the element of the feature vector in all the above-described learning sample patterns. In addition, the margin determination unit 6 records the margin amount in the recording unit 7 in association with each element of the quantized feature vector.

  The feature extraction unit 2, the large classification unit 3, the detailed classification unit 4, and the table creation unit 5 may be realized using a CPU (Central Processing Unit). A programmable device (FPGA (Field Programmable Gate Array), PLD (Programmable Logic Device), etc.) may be used.

  The recording unit 7 records programs, tables, data, and the like. The recording unit 7 is a memory such as a ROM (Read Only Member), a RAM (Random Access Memory), or a hard disk. The recording unit 7 may record data such as parameter values and variable values, and can also be used as a work area. In the first embodiment, a candidate table (large classification dictionary) and the like are recorded.

(Operation of the table creation unit)
FIG. 6 is a flowchart showing an example of the operation of the table creation unit 5.
In step S1, the table creation unit 5 extracts one feature vector element of one learning sample pattern, and obtains the position on the axis of the axis corresponding to the extracted element. Note that the feature vector of the learning sample pattern is obtained from the learning sample pattern input to the feature extraction unit 2. Here, the sample pattern for learning is a character image such as JIS level 1 kanji, JIS level 2 kanji, hiragana, katakana, symbols, characters for vertical writing, letters, numbers, and half-width characters. The learning sample pattern is a deteriorated character image (a character image that has been deteriorated (blurred, blurred, smudged, etc.) by facsimile or copying). In this example, description is made using characters as a pattern, but is not limited to characters.

  FIG. 7 is a diagram showing the relationship between categories and templates for character images “A”, “B”,... That are learning sample patterns. In FIG. 7, the learning sample pattern of the character image “A” is included in category 1 (corresponding to the character code of the character “A”), and the learning sample pattern of the character image “B” is category 2 (the character “B”). (Corresponding to the character code). Category 1 is divided into templates A1 to A5. Category 2 is divided into templates B1 to B5.

  FIG. 8 is a diagram illustrating an example in which the characters “A”, “B”,... Shown in FIG. 7 are converted into 288-dimensional feature vectors as described above. Each learning sample pattern converted into a feature vector is recorded in the recording unit 7 in association with a category via the table creation unit 5. In the example of FIG. 8, the feature vector names “XA1”, “XA2”,... Are recorded in association with “category 1” corresponding to the character “A”. Further, the feature vector names “XB1”, “XB2”,... Are associated with “category 2” corresponding to the character “B” and recorded in the recording unit 7. Note that the feature extraction unit 2 may record the category and the feature vector directly in the recording unit 7 without using the table creation unit 5.

  Next, the table creation unit 5 extracts one feature vector element of the learning sample pattern recorded in the recording unit 7 and projects the extracted element on the axis to obtain the position of the element on the axis. In the example of FIG. 9A, each element corresponding to “x1” of the feature vector of each learning sample pattern is projected on the axis corresponding to “x1” of the feature vector of category 1 (character image “A”). doing. 9A shows the coordinates on the axis obtained by quantizing the feature quantity of each element of the feature vector in the numerical range of -127 to 128 on the axis indicated by the numerical range of -127 to 128 (256 bits). The position is determined to represent the distribution of each axis.

  For example, in the case of the feature quantity “xA11” corresponding to the element “x1” of the category 1 feature vector, the table creation unit 5 quantizes the feature quantity “xA11” to obtain the coordinate position “PA11”. Then, the table creation unit 5 records the feature quantity “xA11” and the coordinate position “PA11” in the recording unit 7 in association with each other as shown in FIG. Similarly, for each element of the feature vector of the other learning sample pattern, the table creation unit 5 records each element and each coordinate position in the recording unit 7 in association with each other.

  Next, in step S2 of FIG. 6, the table creation unit 5 acquires the margin amount calculated in advance by the margin determination unit 6, and in step 3, the table creation unit 5 performs the margin amount with the coordinate position on the axis as the center. To the left and right (minimum value direction and maximum value direction). Then, the table creation unit 5 determines the range of the category (or template) of the learning sample pattern and records the range in the recording unit 7. FIG. 9B is a diagram showing that a preset margin amount is added to the left and right at the coordinate position corresponding to the element “x1” of the feature vector of category 1 shown in FIG. 9A.

  For example, in the case of the coordinate position “PA11” corresponding to the element “x1” of the category 1 feature vector, the table creation unit 5 sets the margin calculated in advance by the margin determination unit 6 around the coordinate position “PA11”. The quantity “MA11” is set. In the example of FIG. 10, the margin amount “MA11” is recorded in the recording unit 7 in association with the coordinate position “PA11”. Similarly, for the other margin amounts, the table creation unit 5 records each coordinate position and each margin amount in the recording unit 7 in association with each other. In the case of indicating the coordinate position, for example, if the numerical value range is 256 bits, it is conceivable to set “1” to the bit corresponding to the calculated coordinate position in the range of 0 to 255 bits. For example, if the coordinate position “PA11” corresponds to the 50th bit (−78) of 0 to 255 bits, “50” is set to “PA11”. If the margin amount is “5 bits”, “5” is recorded in “MA11”. As a result, “1” is set in the 45th to 55th bits (−83 to −73). “0” is set to bits that do not correspond to the coordinate position and the margin amount. Note that the minimum value and maximum value range (−83 to −73) obtained from the coordinate position and the margin amount may be directly recorded without recording the coordinate position and the margin amount. Further, for example, a 256-bit bit string may be prepared, and “1” may be set in the corresponding range (45 bits to 55 bits).

  In step S4 of FIG. 6, the table creation unit 5 determines whether all elements (x1 to x288) of the target feature vector have been processed. If the processing has been completed for all elements, the process proceeds to step S6, and if there are still unprocessed elements, the process proceeds to step 5. In step 5, the next element is selected and the process proceeds to step S1.

  In step S <b> 6, the table creation unit 5 determines whether the target category (1, 2 to) or the feature vectors of all templates have been processed. If processing has been completed for all feature vectors, the process proceeds to step S8, and if unprocessed feature vectors still remain, the process proceeds to step 7. In step 7, the feature vector of the next category or template is selected, and the process proceeds to step S1.

In step S8, it is determined whether the table creation unit 5 has processed all the learning sample patterns (categories (1, 2 to) or templates). If the processing has been completed for all the learning sample patterns, the process proceeds to step S10. If unprocessed learning sample patterns still remain, the next learning sample pattern is selected in step 9, Move to 1.
The table creation unit 5 creates a candidate table (large classification dictionary) by performing the table creation processing in steps S1 to S8.

  FIG. 11 shows an example of the result of the table creation process shown in FIG. 6 performed for all the learning sample patterns. The coordinate position and the margin amount corresponding to each element of each category are represented by a bit string. This is an example in which a bit “1” is set at a corresponding location on the axis to be processed.

  FIG. 12 is a diagram showing an outline of the planar dictionary. Compared to the conventional planar dictionary (see FIG. 2), the planar dictionary of the first embodiment appropriately captures the distribution for each element of each feature vector for each category. Can be improved.

  FIG. 13 shows distribution ranges (distribution ranges Aa ′ and “B” of “A”) recorded in the dictionary of the character types “A”, “B”, and “C” when the planar dictionary is created by the method described in the first embodiment. ”Distribution range Ba ′,“ C ”distribution range Ca ′) and the actual distribution range (“ A ”range Ar,“ B ”range Br,“ C ”range Cr). By using the planar dictionary according to the first embodiment, the unknown character “X” is appropriately input even when an unknown character “X” is input as compared with the case where the conventional planar dictionary (see FIG. 3) is used. “X” is narrowed down to the category of “C”, and the unknown character “X” does not belong to the category of the letter “A”.

(Example 2)
FIG. 14 is a diagram illustrating an operation of the large classification unit 3 illustrated in FIG. 4.
In step S1401, the large classification unit 3 secures a bit string area temp_bit having a width corresponding to the number of all categories in a memory such as the recording unit 7, and sets “1” to all bits. In the example of FIG. 15, since there are n categories, temp_bit having an n-bit width is secured and “1” is set to all n bits. temp_bit can be expressed by Equation 3.
temp_bit: (1, 1, 1... 1) Equation 3
temp_bit is n bits

In step S1402, the large classification unit 3 obtains a feature vector of an unknown input character, projects each element of the feature vector onto each corresponding axis, and obtains a coordinate position on each axis. In the example of FIG. 15, when the coordinate position on the axis of the target of unknown input is represented by k, the coordinate position k on the axes of categories 1 to n corresponding to the coordinate position k of the target axis of unknown input is set. Get the set value. If the unknown input character “X” is included in the category, the acquired coordinate position k on the axes of categories 1 to n can be expressed by Equation 4.
Bit string of coordinate position k of target axis: (1, 0, 0...) Equation 4
The bit string of the coordinate position k of the target axis is n bits

In step S1403, the large classification unit 3 performs a logical product (bit AND) with a bit string composed of the coordinate positions of categories 1 to n corresponding to the coordinate position of the target axis of the unknown character, and temp_bit. In the example of FIG. 15, the logical product of temp_bit and the bit string of the coordinate position k of the target axis is calculated.
temp_bit AND (bit string of coordinate position k of target axis) Equation 5
Temp_bit = (1, 0, 0...) That is the result of the logical product

  In step S1404, the large classification unit 3 determines whether or not the processing in step S1403 has been performed on all axes of unknown characters. If the process of step S1403 has been completed for all axes of unknown characters, the process proceeds to step S1406, and if not, the process proceeds to step S1405. In step S1405, the large classification unit 3 selects the next axis.

  In step S1406, the large classification unit 3 selects a category (or template) in which “1” is set in the bit of the corresponding category number in the bit string calculation area temp_bit of each axis. Then, the large classification unit 3 records the selected category in the recording unit 7 as a large classification result. As described above, the accuracy of the large classification can be improved by using the candidate table (large classification dictionary) created in the first embodiment, but there is still room for reducing the processing time of the large classification. .

Therefore, a method for generating a large classification dictionary using the dimension-compressed feature vector by compressing the dimension of the feature vector using canonical discriminant analysis will be described. FIG. 16 is a diagram illustrating an example of a configuration of a pattern recognition apparatus 1601 in which a feature compression unit 1602 is provided in the feature extraction unit 2 illustrated in FIG. The feature compression unit 1602 converts the original feature vector into a compressed feature belt having a small number of dimensions. The feature compression processing for obtaining the compression feature beltle is to calculate initial coordinates of the compression feature space, orthogonalize the coordinate axes, and calculate the compression feature beltle. In the calculation process of the initial coordinates of the compressed feature space, the coordinate axis that maximizes the variance between categories and the variance ratio within the category is extracted by canonical discriminant analysis that is an existing technique. For example, in a 288-dimensional feature space, when obtaining a 16-dimensional compressed feature space, a character category variance matrix Sb = matrix B and a character category variance matrix Sw = matrix W are defined as follows (Formula 6). The matrix B and the matrix W are 288 × 288 matrices.

Next, the feature compression unit 1602 obtains an eigenmatrix and an eigenvalue matrix that satisfy Expression 7.

The eigenvector of the number of dimensions to be compressed is selected from the larger eigenvalue obtained as described above. For example, in the case of 16 dimensions, the eigenvector shown in Expression 8 becomes the initial coordinates.

Next, 16 288-dimensional vectors changed to orthogonal coordinate axes by Schmidt orthogonalization are obtained. Equation 9 shows a transformation matrix in which the initial coordinates are orthonormalized and 16 288-dimensional vectors are arranged in a matrix form.

Next, calculation of the compressed feature (projection of the feature vector) is performed by multiplying the transposed matrix of the 288-dimensional feature vector obtained from one character image by the transformation matrix, as shown in Expression 10, to obtain a 16-dimensional feature vector. Ask for.

Each element y ₁ , y ₂ ... Y ₁₆ of this 16-dimensional compressed feature vector is a value on each axis when the feature vector X represented by one point in the 288-dimensional space is projected onto 16 axes. (Coordinate position). For example, y ₁ is a value (coordinate position) on the first axis when the feature vector X is projected onto the first axis. The same is true for the other elements.

  In the second embodiment, the feature extraction unit 2 generates an original feature vector as described in the first embodiment, and the feature compression unit 1602 dimensionally compresses the original feature vector. Then, the table creation unit 5 creates a large classification plane dictionary using the dimension-compressed feature vector. Also, when performing a large classification, the feature compression unit 1602 compresses the dimension of an unknown input feature vector, and then the large classification unit 3 uses the candidate table (large classification dictionary) generated in the second embodiment to perform a large classification. Therefore, the number of vectors handled in the large classification process can be reduced. Therefore, it is possible to shorten the processing time for the large classification. In addition, since the number of vectors to be handled can be reduced by performing dimension compression, the storage area of the candidate table (large classification dictionary) of the recording unit 7 can be reduced.

(Example 3)
In the third embodiment, a margin amount on the axis is determined based on a difference (deformation degree) between learning sample patterns for each category or template, and a margin amount is determined for each category or template of the learning sample pattern. The difference between the learning sample patterns is a difference between the original learning sample pattern and a pattern obtained by degrading the original learning sample pattern (this pattern is also a learning sample pattern).

FIG. 17 is a flowchart illustrating an example of margin determination processing according to the third embodiment.
In step S1701, the feature extraction unit 2 converts a plurality of learning sample patterns deteriorated with respect to the original learning sample pattern into feature vectors, the feature vector and the category, and the features of the original learning sample pattern. The vector is associated and recorded in the recording unit 7.

  In step S1702, the margin determining unit 6 selects a category or a template. In step S1703, the margin determination unit 6 selects a learning sample pattern as a base. In step S1704, the margin determination unit 6 selects an element from a feature vector of a pattern obtained by degrading the original learning sample pattern.

  In step S1705, the difference between the element value of the original feature vector and the element value of the deteriorated feature vector is obtained on the axis corresponding to the element selected by the margin determination unit 6. For example, assuming that the coordinate position on the axis corresponding to the element of the feature vector of the original learning sample pattern is Pbase, and the coordinate position on the axis of the deteriorated pattern is Pdete, | Pbase−Pdete | Find the difference. Next, the margin determination unit 6 compares the magnitudes of Pbase and Pdete to obtain a magnitude relationship, and rightward (maximum value) around the value corresponding to the element of the feature vector of the original learning sample pattern Direction) margin or left margin (minimum value direction). When the axis is in a numerical range of 0 to 255, Pbase-Pdete is calculated, and if the calculation result is a positive value, it is set as a left margin, and the value is recorded in the variable mLtemp. If it is a negative value, the margin amount in the right direction is set, and this value is recorded in the variable mRtemp.

In step S1706, the margin determination unit 6 adds the margin amount calculated in step S1705 for each category. For example, as shown in Expression 11, the variable mRtemp is added to the variable mR if the margin amount in the right direction is added, and the variable mLtemp is added to the variable mL if the margin amount is added in the left direction.
mR ← mR + mRtemp
mL ← mL + mLtemp Formula 11
←: Indicates assignment

At that time, whenever the variable mRtemp is added to the variable mR, 1 is added to the variable mRcount as shown in Expression 12, and the number of additions is recorded. Each time the variable mLtemp is added to the variable mL, 1 is added to the variable mLcount as shown in Expression 12, and the number of additions is recorded.
mRcount ← mRcount + 1
mLcount ← mLcount + 1 Formula 12

  In step S1707, the margin determination unit 6 determines whether or not all the elements of the feature vector of the learning sample pattern whose target has been deteriorated have been processed. If all elements have been processed, the process proceeds to step S1709. If there is an unprocessed element, the process proceeds to step S1708. In step S1708, the margin determination unit 6 selects the next element, and the process proceeds to step S1705.

  In step S1709, the margin determination unit 6 determines whether or not processing has been performed for all feature vectors of the degraded learning sample pattern. If processing has been performed for all feature vectors, the process proceeds to step S1711. If there is an unprocessed feature vector, the process proceeds to step S1710. In step S1710, the margin determination unit 6 selects the next degraded feature vector of the learning sample pattern, and the process proceeds to step S1704.

  In step S1711, it is determined whether the margin determination unit 6 has processed all the categories. If processing has been performed for all categories, the process proceeds to step S1712, and if there is an unprocessed category, the process proceeds to step S1702. Each time the category changes, initial values (for example, “0”) are set as the count values mRcount and mLcount.

In step S1712, the margin determination unit 6 obtains average values mRave and mLave of the right and left margin amounts for each category as shown in Expression 13.
mRave ← mR / mRcount
mLave ← mL / mLcount Formula 13

  The margin amount may be adjusted by adding a fixed multiple to the average values mRave and mLave. Then, the margin determining unit 6 associates the categories with the average values mRave and mLave of the margin amounts and records them in the recording unit 7 as shown in FIG. FIG. 18 is a diagram illustrating an example of a data structure when the margin amounts (mRave, mLave) calculated by the margin determination method illustrated in the third embodiment are recorded in association with each category. In FIG. 18, “category” is recorded with an identification number (1, 2 to n) for identifying the category, and “mRave” and “mLave” are margin amounts (MAL1 to MALn / MAL) for each category determined by the above processing. MAR1 to MARn) are recorded.

  As described above, the margin determination process is performed by combining elements at the same position in each feature vector of a plurality of original patterns included in a category and patterns obtained by degrading the original pattern at the same positions in a preset range. Project to the axis corresponding to the element. Next, the difference between the coordinate position on the axis of the original pattern and the coordinate position on the axis of each deteriorated pattern is calculated. Then, the average value of the difference in the minimum value direction and the average value of the difference in the maximum value direction are calculated as the margin amount with the coordinate position on the axis of the original pattern as the center, and the margin amount is associated with each category and recorded. Record in part 7.

According to the third embodiment, the margin amount is determined based on the original learning sample pattern and a plurality of deteriorated learning sample patterns, so that the large classification is performed using the conventional large classification dictionary. , Accuracy can be improved. That is, since the planar dictionary of Example 3 appropriately captures the distribution for each element of each feature vector for each category, the accuracy of the large classification can be improved.
Note that the margin amount may be determined by the method described in the third embodiment using the dimension-compressed vector described in the second embodiment.

Example 4
In the fourth embodiment, the axis is divided by a predetermined size, each element of the feature vector of all the learning sample patterns for each category is projected onto the axis, and the margin amount based on the section and the coordinate position on the axis Ask for.

  FIG. 19 is a flowchart illustrating an example of margin determination processing according to the fourth embodiment. Further, in the fourth embodiment, as shown in FIG. 20, the margin determination unit 6 sets a section divided on the axis by a predetermined size. For example, when the axis is indicated by a numerical range of 0 to 255, 256 is divided at equal intervals and an identification number is assigned to each section. In the case of FIG. 20, the section is divided into 10 and divided into 10 equal parts. Note that the sections do not necessarily have to be equally divided.

  In step S1901, a plurality of learning sample patterns degraded by the feature extraction unit 2 with respect to the original learning sample pattern are converted into feature vectors, and the feature vector, category, section, and original learning sample pattern are converted. Are recorded in the recording unit 7 in association with each other. In step S1902, the margin determination unit 6 selects a category or a template. In step S1903, the margin determination unit 6 selects a learning sample pattern as a base. In step S1904, the margin determination unit 6 selects an element from a feature vector of a pattern obtained by degrading the original learning sample pattern. In step S1905, the margin determination unit 6 selects a section having a value corresponding to the target element selected in step S1904.

  In step S1906, as in step S1705 described in the first embodiment, the element value of the original feature vector and the element value of the deteriorated feature vector on the axis corresponding to the element selected by the margin determining unit 6 are used. The difference is calculated as a margin. For example, assuming that the coordinate position on the axis corresponding to the element of the feature vector of the original learning sample pattern is Pbase, and the coordinate position on the axis of the deteriorated pattern is Pdete, | Pbase−Pdete | Find the difference. Next, the margin determination unit 6 compares the magnitudes of Pbase and Pdete to obtain a magnitude relationship, and rightward (maximum value) around the value corresponding to the element of the feature vector of the original learning sample pattern Direction) margin or left margin (minimum value direction). When the axis is set to a numerical value range of 0 to 255, for example, for the interval 1, if Pbase-Pdette is calculated and the calculation result is a positive value, it is set as a left margin, and the value is recorded in the variable mLint1_temp. . If it is a negative value, the margin amount in the right direction is set, and this value is recorded in the variable mRint1_temp.

In step S1907, the margin determination unit 6 adds the margin amount calculated in step S1906 for each section. For example, in the case of section 1, as shown in Expression 14, if the margin amount in the right direction is added, the variable mRint1_temp is added to the variable mRint1, and if the margin amount in the left direction is added, the variable mLint1 is added. The variable mLint1_temp is added.
mRint1 ← mRint1 + mRint1_temp
mLint1 ← mLint1 + mLint1_temp Equation 14

At that time, each time the variable mRint1_temp is added to the variable mRint1, 1 is added to mRint1_count as shown in Expression 15, and the number of times of addition is recorded. Further, whenever the variable mLint1_temp is added to the variable mLint1, 1 is added to mLint1_count as shown in Expression 15, and the number of additions is recorded.
mRint1_count ← mRint1_count + 1
mLint1_count ← mLint1_count + 1 Equation 15

  In step S1908, the margin determination unit 6 determines whether or not all the elements of the feature vector of the learning sample pattern whose target has been deteriorated have been processed. If all elements have been processed, the process proceeds to step S1910. If there is an unprocessed element, the process proceeds to step S1909. In step S1909, the margin determination unit 6 selects the next element, and the process proceeds to step S1905.

  In step S1910, the margin determination unit 6 determines whether or not processing has been performed for all feature vectors of the deteriorated learning sample pattern. If processing has been performed for all feature vectors, the process proceeds to step S1912. If there is an unprocessed feature vector, the process proceeds to step S1911. In step S1911, the margin determination unit 6 selects the next degraded feature vector of the learning sample pattern, and the process proceeds to step S1904.

  In step S1912, it is determined whether the margin determining unit 6 has processed all the categories. If all categories have been processed, the process proceeds to step S1913. If there is an unprocessed category, the process proceeds to step S1902. Each time the category changes, an initial value (for example, “0”) is set as the count value. In the case of section 1, initial values are set to the values of mRint1_count and mLint1_count.

In step S1913, the margin determination unit 6 obtains an average value of the respective margin amounts in the right direction and the left direction in each section of each category. In the case of section 1, as shown in Expression 16, average values mRint1_ave and mLint1_ave of the respective right and left margin amounts in each section of each category are obtained.
mRint1_ave ← mRint1_ / mRint1_count
mLint1_ave ← mLint1_ / mLint1_count Equation 16

  The margin amount may be adjusted by adding a fixed multiple to the average values mRint1_ave and mLint1_ave. Further, as shown in FIG. 21, the margin determining unit 6 associates the category with the average value of the margin amount and records them in the recording unit 7 respectively. FIG. 21 is a diagram illustrating an example of a data structure when a margin amount calculated by the margin determination method illustrated in the fourth embodiment is recorded in association with each section of each category. 21, identification numbers (1, 2 to n) for identifying the category are recorded, and “section 1” to “section w” (w is an integer) are used to identify the sections. Identification numbers (1, 2 to w) are recorded. In addition, in “mRint1_ave” to “mRintw_ave” and “mLint1_ave” to “mLintw_ave” in FIG. 21, the margin amounts (MAL11 to MALwn / MAR11 to MARwn) of each section determined by the above processing are recorded. .

  As described above, in the fourth embodiment, the section in which the axis is divided into the preset range is set by the margin determination process. Next, the element at the same position in each feature vector of the plurality of original patterns included in the category and the pattern obtained by degrading the original pattern is used as an axis corresponding to the element at the same position in the preset range. Project. Next, the difference between the coordinate position on the axis of the original pattern and the coordinate position on the axis of each deteriorated pattern is calculated. Then, the average value of the difference in the minimum value direction and the average value of the difference in the maximum value direction are calculated as margin amounts around the coordinate position on the axis of the original pattern, and the coordinate position on the axis of the original pattern Is recorded in the recording unit 7 in association with a margin amount for each section including the.

In the conventional method, a large number of categories belong to one point on the axis, and the refining ability at the time of recognition is low, and it is not possible to narrow down to a small number of categories while maintaining high accuracy. However, according to the fourth embodiment, the margin amount is determined for each section based on the learning sample pattern based on the margin amount and a plurality of deteriorated learning sample patterns. The accuracy of large classification can be improved using a dictionary. That is, since the planar dictionary of Example 4 appropriately captures the distribution for each element of each feature vector in each section of each category, the accuracy of the large classification can be improved.
Note that the margin amount can be determined by the method described in the fourth embodiment using the dimension-compressed vector described in the second embodiment.

(Example 5)
In the fifth embodiment, paying attention to the font type of the character pattern included in each category or each set of templates, the margin amount is obtained based on the distribution on the axis for each font type in the same category.

FIG. 22 is a flowchart illustrating an example of margin determination processing according to the fifth embodiment.
In step S2201, the feature extraction unit 2 converts the character patterns of the sample patterns for learning of different font types in the same category into feature vectors, and records the feature vectors, fonts, and categories in the recording unit 7 in association with each other.

In step S2202, the margin determination unit 6 selects a target category. In step S2203, the margin determination unit 6 selects a font.
In step S2204, the margin determination unit 6 selects an element from the feature vector of the font selected in step S2203, and projects the feature vector element on the corresponding axis. For example, when there are a plurality of different font type learning sample patterns (character patterns: F11 to F1n) in the same category 1, each of the character patterns F11 to F1n is converted into a 288-dimensional feature vector. An axis indicated by a numerical value range of 0 to 255 is assigned to each of the 1st to 288th elements constituting each feature vector of the character patterns F11 to F1n. Each element corresponding to the first of the feature vectors of the character patterns F11 to F1n is projected onto the axis corresponding to the first element. For the other axes, the target elements of the feature vectors of the character patterns F11 to F1n are projected onto the target axes. That is, all the elements of the feature vectors of the character patterns F11 to F1n are all projected on 288 axes.

  In step S2205, the margin determination unit 6 determines whether or not all the elements of the feature vector for the target font have been processed. If all elements have been processed, the process proceeds to step S2207. If there is an unprocessed element, the process proceeds to step S2206. In step S2206, the margin determination unit 6 selects the next element, and proceeds to step S2204.

  In step S2207, the margin determination unit 6 determines whether or not all the feature vectors of the font selected in step S2203 have been processed. If processing has been performed for all feature vectors, the process proceeds to step S2209. If there is an unprocessed feature vector, the process proceeds to step S2208. In step S2208, the margin determination unit 6 selects the next font feature vector, and the process proceeds to step S2204.

  In step S2209, the margin determination unit 6 obtains an average value of a plurality of elements projected on each axis of the same category. For example, in the case of 288 axes on which all the elements of the feature vectors of the character patterns F11 to F1n included in the category 1 described above are projected, the average value of the elements on the axis for every 288 axes Calculate Next, the difference between the average value and the value of the element on the axis of the feature vector for each font type is obtained. For example, assuming that the coordinate position on the axis corresponding to the average value is Pfbase and the coordinate position on the axis of the character pattern for each font is Pfdete, | Pfbase-Pfdete | is calculated to obtain the difference. Next, the margin determination unit 6 compares the magnitudes of Pfbase and Pfdete to obtain a magnitude relationship, and determines whether the margin is in the right direction (maximum value direction) or the left direction (minimum value direction) around the average value. ). When the axis is in a numerical range of 0 to 255, Pfbase-Pfdete is calculated, and if the calculation result is a positive value, it is set as a left margin, and the value is recorded in the variable mLf_temp. If it is a negative value, the margin amount in the right direction is set, and this value is recorded in the variable mRf_temp.

In step S2210, the margin determination unit 6 obtains an average value of differences for each font, and sets it as a font margin value for each category. First, the margin amount calculated in step S2209 is added for each font. For example, as shown in Expression 17, the variable mRf_temp is added to the variable mRf if the margin amount in the right direction is added, and the variable mLf_temp is added to the variable mLf if the margin amount is added in the left direction.
mRf ← mRf + mRf_temp
mLf ← mLf + mLf_temp Equation 17

At that time, each time the variable mRf_temp is added to the variable mRf, 1 is added to mRf_count as shown in Expression 18, and the number of times of addition is recorded. Further, whenever the variable mLf_temp is added to the variable mLf, 1 is added to mLf_count as shown in Expression 18, and the number of additions is recorded.
mRf_count ← mRf_count + 1
mLf_count ← mLf_count + 1 Equation 18

Next, the margin determination unit 6 obtains the average values mRf_ave and mLf_ave of the right and left margin amounts for each font, as shown in Expression 19.
mRf_ave ← mRf / mRf_count
mLf_ave ← mLf / mLf_count Equation 19

  The margin amount may be adjusted by adding a fixed multiple to the average values mRf_ave and mLf_ave. Each time the font changes, initial values (for example, “0”) are set as the count values mRcount and mLcount.

  In step S2211, it is determined whether the margin determining unit 6 has processed all the categories. If all categories have been processed, the process proceeds to step S2212, and if there is an unprocessed category, the process proceeds to step S2202.

  Then, as shown in FIG. 23, the margin determining unit 6 records the average values mRf_ave and mLf_ave of the margin amounts for each font in the recording unit 7 in association with each font. In the case of the font F1 in FIG. 23, average values MARF1 and MALF1 corresponding to the font F1 in FIG. 23 are set. The average value is set for each of the other fonts.

  As described above, in the fifth embodiment, the elements at the same position of the feature vectors of the different font patterns included in the same category are projected on the axes corresponding to the elements at the same position in the preset range, The average value of the coordinate positions on the pattern axis is calculated. Next, the difference between the coordinate position on the axis of the pattern for each font and the coordinate position of the average value is calculated, and the difference is the difference between the minimum value direction and the maximum value direction centered on the coordinate position of the average value. Divide into Then, the average value of the difference in the minimum value direction and the difference in the maximum value direction of the same font are calculated as the margin amount for each category, and the margin amount is associated with the category and recorded in the recording unit.

  In the conventional method, a large number of categories belong to one point on the axis, and the refining ability at the time of recognition is low, and it is not possible to narrow down to a small number of categories while maintaining high accuracy. However, according to the fifth embodiment, since the margin amount is determined based on the font for each category, the accuracy of the large classification can be improved by using the large classification dictionary. That is, since the planar dictionary of Example 5 appropriately captures the distribution for each element of each feature vector for each category as compared to the conventional case, the accuracy of the large classification can be improved.

  Note that the margin amount can be determined by the method described in the fifth embodiment using the dimension-compressed vector described in the second embodiment.

(Modification 1)
Note that instead of the category in the third embodiment, by using a template classified by font type, it is possible to record the font and the merge amount in association with each category as shown in FIG. In FIG. 24, “category” is recorded with identification numbers (1, 2 to n) for identifying the category, and “font” is recorded with identification numbers (F1, F2 to Fn) for identifying the font. ,ing. In “mRf_ave2” and “mLf_ave2”, the margin amounts (MALF11, MALF12˜ / MARF11, MARF12˜) for each font in each category determined by the processing of the third embodiment are recorded.

(Modification 2)
In the second modified example, when pattern recognition is performed using the pattern recognition processing described in the above-described embodiment, it is found that there is a misrecognized character. The plane dictionary for classification is adjusted, and the adjustment result is recorded in the plane dictionary for large classification.

FIG. 25 is a diagram illustrating an example of a hardware configuration of a computer that can implement the apparatus according to the embodiment.
The computer hardware 2500 includes a CPU 2501, a recording unit 2502 (ROM, RAM, hard disk drive, etc.), a recording medium reader 2503, an input / output interface 2504 (input / output I / F), a communication interface 2505 (communication I / F), and the like. It has. Further, each of the above components is connected by a bus 2506.

The CPU 2501 executes the above-described pattern recognition processing (the processing shown in FIGS. 6, 13, 17, 19, 22, etc.) stored in the recording unit 2502.
The recording unit 2502 records programs executed by the CPU 2501 and data. It is also used as a work area. The recording unit 2502 has the function of the recording unit 7 described above.

  The recording medium reader 2503 controls reading / writing of data with respect to the recording medium 2503a according to the control of the CPU 2501. Then, the data written under the control of the recording medium reader 2503 is stored in the recording medium 2503a, or the data stored in the recording medium 2503a is read. The detachable recording medium 2503a includes a computer readable recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory. The magnetic recording device includes a hard disk device (HDD). Examples of the optical disc include a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only Memory), and a CD-R (Recordable) / RW (ReWritable). Magneto-optical recording media include MO (Magneto-Optical disk).

  An input / output device 2504 a (for example, a display) is connected to the input / output interface 2504, receives information input by the user, and transmits the information to the CPU 2501 via the bus 2506. Further, operation information and the like are displayed on the display screen in accordance with a command from the CPU 2501.

  A communication interface 2505 is an interface for LAN connection, Internet connection, or wireless connection with another computer as necessary. It is also connected to other devices and controls data input / output from external devices.

When the user finds that there is a misrecognized character in the pattern recognition result displayed on the screen of the monitor that is the input / output device 2504a shown in FIG. 25, in order for the user to change the misrecognized character to the correct character Select the wrong recognition character on the screen with the mouse. And a user inputs a correct answer character instead of a misrecognized character. At that time, the CPU 2501 obtains from the recording unit 2502 and compares the data related to the large classification axis related to the erroneously recognized character and the data related to the large classification axis related to the correct character to determine which axis caused the error. . Then, the CPU 2501 displays the axis and distribution in error on the monitor screen as the determination result (display as shown in FIGS. 12 and 15). The user adjusts the distribution area (margin amount) of the axis in error displayed on the screen with a mouse or the like. The CPU 2501 can reflect the adjusted result in the large classification plane dictionary of the recording unit 2502. Therefore, in addition to rewriting the misrecognized character as the correct answer character as in the past, even if it is misrecognized, the user can easily adjust the distribution of the target axis of the planar dictionary for large classification visually be able to.
(Configuration when this embodiment is realized as a computer program)
By using a computer having a hardware configuration as shown in FIG. 25, the various processing functions described above (the processing (such as a flowchart) described in the embodiment) are realized. In that case, a program describing the processing contents of the functions that the system should have is provided. By executing the program on a computer, the above processing functions are realized on the computer. The program describing the processing content can be recorded in a computer-readable recording medium 2503a.

  When distributing the program, for example, a portable recording medium such as a DVD or a CD-ROM in which the program is recorded is sold. It is also possible to store the program in a storage device of a server computer and transfer the program from the server computer to another computer via a network.

  The computer that executes the program stores, for example, the program recorded on the portable recording medium or the program transferred from the server computer in its own storage device. Then, the computer reads the program from its own storage device and executes processing according to the program. The computer can also read the program directly from the portable recording medium and execute processing according to the program. Further, each time the program is transferred from the server computer, the computer can sequentially execute processing according to the received program.

  The present invention is not limited to the above-described embodiment, and various improvements and modifications can be made without departing from the gist of the present invention. Each embodiment may be combined with each other as long as there is no contradiction in processing.

Regarding the embodiment including the above-described examples, the following additional notes are further disclosed.
(Appendix 1)
For each of a plurality of categories that classify the sample pattern for learning, the element at the same position of the reference feature vector obtained for each of the plurality of feature vectors included in the same category is projected onto the axis of the preset range, For each element, a recording unit that records a candidate table generated by associating a preset margin amount and category,
Obtaining a reference feature vector of a given pattern, using the candidate table, classifying each element of the reference feature vector to obtain a candidate category set, and outputting a classified candidate category set;
A pattern recognition apparatus comprising:
(Appendix 2)
The pattern recognition device includes a feature extraction unit that makes the learning sample pattern or the given pattern a feature vector of the same dimension;
A feature compression unit that dimensionally compresses the feature vector into the reference feature vector;
The pattern recognition apparatus according to appendix 1, further comprising:
(Appendix 3)
The margin amount is set in advance for elements at the same position in each reference feature vector of a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern. Projecting on the axis of the range, calculating the difference between the coordinate position on the axis of the original learning sample pattern and the coordinate position on the axis for each deteriorated pattern, the original learning The pattern recognition apparatus according to appendix 1 or 2, wherein the average value of the difference in the minimum value direction and the average value of the difference in the maximum value direction centered on the coordinate position on the axis of the sample pattern for use.
(Appendix 4)
The margin amount sets a section obtained by dividing the axis into a predetermined range, and a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern Are projected onto the axis of a preset range, the coordinate position on the axis of the original learning sample pattern, and the axis for each degraded pattern The difference between the coordinate position and the average value of the difference between the minimum value direction and the average value of the maximum value direction around the coordinate position on the axis of the original learning sample pattern is calculated as a margin amount. The pattern according to claim 1 or 2, wherein the margin amount is associated with each of the sections including the coordinate position on the axis of the original learning sample pattern. Recognition device.
(Appendix 5)
The margin amount is calculated by projecting an element at the same position of each feature vector of the sample pattern for learning of different fonts included in the same category onto an axis corresponding to the element at the same position in a preset range. Calculating the average value of the coordinate position on the axis of the learning sample pattern, calculating the difference between the coordinate position on the axis of the learning sample pattern for each font and the coordinate position of the average value, The difference is divided into a difference in the minimum value direction and a difference in the maximum value direction centered on the coordinate position of the average value, and the average value of the difference in the minimum value direction and the difference in the maximum value direction of the same font for each category. The pattern recognition apparatus according to appendix 1 or 2, wherein a value is calculated as a margin amount, and the margin amount is associated with each font.
(Appendix 6)
On the computer,
A process for obtaining a reference feature vector of a given pattern;
For each of a plurality of categories that classify the sample pattern for learning, the element at the same position of the reference feature vector obtained for each of the plurality of feature vectors included in the same category is projected onto the axis of the preset range, A process for obtaining a candidate category set by classifying each element of the reference feature vector using a candidate table generated by associating a predetermined margin amount with a category for each element;
A process of outputting the classified candidate category set;
A pattern recognition program characterized in that
(Appendix 7)
Computer
Find a reference feature vector for a given pattern,
For each of a plurality of categories that classify the sample pattern for learning, the element at the same position of the reference feature vector obtained for each of the plurality of feature vectors included in the same category is projected onto the axis of the preset range, For each element, using a candidate table generated by associating a preset margin amount and category, classifying each element of the reference feature vector to obtain a candidate category set,
Output the classified candidate category set,
A pattern recognition method that performs that.
(Appendix 8)
Processing the learning sample pattern or the given pattern into a feature vector of the same dimension;
Processing to dimensionally compress the feature vector into the reference feature vector;
The pattern recognition program according to appendix 6, which causes a computer to execute
(Appendix 9)
Projecting an element at the same position of each reference feature vector of a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern onto an axis in a preset range And calculating the difference between the coordinate position on the axis of the original learning sample pattern and the coordinate position on the axis for each deteriorated pattern, and Appendix 6 or 8 that causes a computer to execute a process of calculating an average value of a difference in a minimum value direction and an average value of a difference in a maximum value direction as a margin amount around a coordinate position on an axis Pattern recognition program.
(Appendix 10)
A section obtained by dividing the axis into predetermined ranges is set, and each reference feature vector of a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern Are projected onto an axis in a preset range, and the coordinate position on the axis of the original learning sample pattern and the coordinate position on the axis for each deteriorated pattern Calculating the difference, calculating the average value of the difference in the minimum value direction and the average value of the difference in the maximum value direction as a margin amount around the coordinate position on the axis of the original learning sample pattern, Causing the computer to execute a process of associating and recording the margin amount for each section including the coordinate position on the axis of the original learning sample pattern. Pattern recognition program according to Appendix 6 or 8.
(Appendix 11)
Projecting an element at the same position of each feature vector of a sample pattern for learning of different fonts included in the same category onto an axis corresponding to the element at the same position of a preset range, and learning sample patterns of the different fonts The average value of the coordinate position on the axis is calculated, the difference between the coordinate position on the axis of the learning sample pattern for each font and the coordinate position of the average value is calculated, and the difference is calculated as the average value. The difference between the minimum value direction and the maximum value direction is centered on the coordinate position, and the average value of the difference between the minimum value direction and the average value of the maximum value direction of the same font for each category is used as a margin amount. 9. The pattern recognition program according to appendix 6 or 8, wherein the computer executes a process of calculating and associating the margin amount for each font and recording it in a recording unit. Grams.
(Appendix 12)
Making the learning sample pattern or the given pattern a feature vector of the same dimension;
Dimensionally compressing the feature vector into the reference feature vector;
The pattern recognition method according to appendix 7, wherein the computer executes this.
(Appendix 13)
Projecting an element at the same position of each reference feature vector of a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern onto an axis in a preset range And calculating the difference between the coordinate position on the axis of the original learning sample pattern and the coordinate position on the axis for each deteriorated pattern, and The pattern recognition method according to appendix 7 or 12, wherein the computer executes to calculate the average value of the difference in the minimum value direction and the average value of the difference in the maximum value direction as a margin amount centering on the coordinate position on the axis.
(Appendix 14)
A section obtained by dividing the axis into predetermined ranges is set, and each reference feature vector of a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern Are projected onto an axis in a preset range, and the coordinate position on the axis of the original learning sample pattern and the coordinate position on the axis for each deteriorated pattern Calculating the difference, calculating the average value of the difference in the minimum value direction and the average value of the difference in the maximum value direction as a margin amount around the coordinate position on the axis of the original learning sample pattern, (7) The computer executes that the margin amount is associated with each section including the coordinate position on the axis of the original learning sample pattern and recorded in the recording unit. Pattern recognition method according to 2.
(Appendix 15)
Projecting an element at the same position of each feature vector of a sample pattern for learning of different fonts included in the same category onto an axis corresponding to the element at the same position of a preset range, and learning sample patterns of the different fonts The average value of the coordinate position on the axis is calculated, the difference between the coordinate position on the axis of the learning sample pattern for each font and the coordinate position of the average value is calculated, and the difference is calculated as the average value. The difference between the minimum value direction and the maximum value direction is centered on the coordinate position, and the average value of the difference between the minimum value direction and the average value of the maximum value direction of the same font for each category is used as a margin amount. The pattern recognition method according to appendix 7 or 12, wherein the computer executes calculation and records the margin amount in association with each font in the recording unit.

DESCRIPTION OF SYMBOLS 1,1601 Pattern recognition apparatus 2 Feature extraction part 3 Large classification | category part 4 Detailed classification | category part 5 Table preparation part 6 Margin determination part 7 Recording part 1602 Feature compression part 2500 Hardware 2501 CPU
2502 Recording unit 2503 Recording medium reader 2503a Recording medium 2504 Input / output interface 2504a Input / output device 2505 Communication interface 2506 Bus

Claims (7)

For each of a plurality of categories of classification of learning sample pattern is a pattern used in classifying a given pattern, determine the reference feature vector for each feature vector of the plurality of the learning samples pattern included in the same category, determined and the elements of the reference feature vectors by projecting the corresponding coordinate position of the axis represented by the predetermined range, the coordinate position of the element on the shaft which is projected on each of the elements, which is set in advance and margin amount, a recording unit for recording by generating a candidate table in association with, and the category,
Obtains a reference feature vector of said given pattern, using the candidate table, a classification unit which obtains a candidate category set by the classification for each element of the reference feature vectors, and outputs the candidate category set classified,
A pattern recognition apparatus comprising: The pattern recognition device includes a feature extraction unit that makes the learning sample pattern or the given pattern a feature vector of the same dimension;
A feature compression unit that dimensionally compresses the feature vector into the reference feature vector;
The pattern recognition apparatus according to claim 1, further comprising:   The margin amount is set in advance for elements at the same position in each reference feature vector of a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern. Projecting on the axis of the range, calculating the difference between the coordinate position on the axis of the original learning sample pattern and the coordinate position on the axis for each deteriorated pattern, the original learning The pattern recognition apparatus according to claim 1, wherein an average value of a difference in a minimum value direction and an average value of a difference in a maximum value direction are centered on a coordinate position on the axis of the sample pattern for use.   The margin amount sets a section obtained by dividing the axis into a predetermined range, and a plurality of original learning sample patterns included in the category and a pattern obtained by degrading the original learning sample pattern Are projected onto the axis of a preset range, the coordinate position on the axis of the original learning sample pattern, and the axis for each degraded pattern The difference between the coordinate position and the average value of the difference between the minimum value direction and the average value of the maximum value direction around the coordinate position on the axis of the original learning sample pattern is calculated as a margin amount. The pattern according to claim 1, wherein the margin amount is associated with each section including the coordinate position on the axis of the original learning sample pattern. Emissions recognition device.   The margin amount is calculated by projecting an element at the same position of each feature vector of the sample pattern for learning of different fonts included in the same category onto an axis corresponding to the element at the same position in a preset range. Calculating the average value of the coordinate position on the axis of the learning sample pattern, calculating the difference between the coordinate position on the axis of the learning sample pattern for each font and the coordinate position of the average value, The difference is divided into a difference in the minimum value direction and a difference in the maximum value direction centered on the coordinate position of the average value, and the average value of the difference in the minimum value direction and the difference in the maximum value direction of the same font for each category. The pattern recognition apparatus according to claim 1, wherein a value is calculated as a margin amount, and the margin amount is associated with each font. On the computer,
A process for obtaining a reference feature vector for each feature vector of the plurality of learning sample patterns included in the same category, for each of a plurality of categories that classify the learning sample pattern that is a pattern used when classifying a given pattern; ,
For each of a plurality of categories of classification of the learning sample pattern, obtains a reference feature vector for each feature vector of the plurality of the learning samples pattern included in the same category, the elements of the reference feature vectors obtained, preset is projected onto the corresponding coordinate position of the axis represented by the range, generated in association with the coordinate position of the element on the shaft which is projected on each of the elements, a preset margin amount, and the categories, the Using the candidate table, classifying each element of the reference feature vector and obtaining a candidate category set;
And outputting the candidate category set that was classified,
A pattern recognition program characterized in that Computer
For each of a plurality of categories that classify a learning sample pattern that is a pattern used when classifying a given pattern, a reference feature vector is obtained for each feature vector of the plurality of learning sample patterns included in the same category ,
For each of a plurality of categories of classification of the learning sample pattern, obtains a reference feature vector for each feature vector of the plurality of the learning samples pattern included in the same category, the elements of the reference feature vectors obtained, preset is projected onto the corresponding coordinate position of the axis represented by the range, generated in association with the coordinate position of the element on the shaft which is projected on each of the elements, a preset margin amount, and the categories, the Using the candidate table, a classification is made for each element of the reference feature vector to obtain a candidate category set,
To output the candidate category set that was classified,
A pattern recognition method that performs that.

Images