JP2861816B2 - Pattern recognition dictionary generation device and pattern recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition dictionary generation device and a pattern recognition device, and more particularly to a pattern recognition dictionary generation device for generating a low-capacity dictionary and a pattern recognition device using the dictionary.

[0002]

2. Description of the Related Art In a dictionary in a pattern recognition apparatus, a vector having a fixed number of elements is represented for each category. In the case of character recognition as an example, a category in this case represents one character or individual types when there are a plurality of types in one character. Generally, one sentence includes a plurality of patterns. As in the case of each category, each category is also composed of a plurality of patterns. Characteristic portions are represented as a plurality of numerical values from a plurality of input patterns, and these are arranged in a single row, thereby forming each category. Is generated. Thus the number of elements
The optical dictionary, which is expressed without reduction, is called an uncompressed dictionary.
Huh. In many cases, the number of elements of a vector for each category reaches several hundreds, so that the size of the memory for storing the uncompressed dictionary is a problem, and it is required to reduce the capacity of the dictionary memory.

Conventionally, key information in such a pattern recognition dictionary is required.
As a method for generating a pattern recognition dictionary in which prime numbers are compressed,
There has been proposed a method of calculating the average value and the variance of the values of the elements at the same position of the feature vector obtained from different learning patterns in the same category (for example, Japanese Patent Laid-Open No. 16389/1990).
No. 0). This means that an element whose calculated average value is equal to or more than a predetermined threshold value and whose variance value is equal to or less than a certain threshold value is selected, or in which the average value is equal to or less than a certain threshold value and the variance value is equal to or less than a certain threshold value , And among the elements obtained in this way, the variance of the average value of each element in all the categories, that is, the variance between categories, is obtained. Among the elements selected in the above, the element whose category variance is equal to or more than a certain threshold is selected. In this way, the number of the element selected for each category and its value are stored as a recognition dictionary of the category, thereby generating a compressed recognition dictionary. The number of elements in such a recognition dictionary is reduced.
The created dictionary is called a compression dictionary.

[0004]

Therefore, in such a conventional pattern recognition dictionary generation method, when pattern recognition is performed based on an actually generated pattern recognition dictionary, only the elements of the selected vector are used. In order to check the similarity between the dictionary and the input feature vector, if there is a large difference in the non-selected elements, the difference between the dictionary and the input pattern cannot be detected, Was output.
In addition, in order to describe the selected element among the elements of the recognition dictionary, it is necessary to memorize not only the value of the element but also the number of the element. To reduce the data capacity of the dictionary, the same storage capacity is required, and less than half of the elements must be selected, and the similarity of more than half of the elements must be ignored, causing a reduction in recognition performance. There was a problem that it becomes. The present invention is intended to solve such a problem, and a pattern recognition dictionary generating apparatus that generates a pattern recognition dictionary with a reduced dictionary capacity without deteriorating recognition performance, and a pattern recognition using the dictionary It is intended to provide a device.

[0005]

In order to achieve the above object, a pattern recognition dictionary generating apparatus according to the present invention comprises:
A dictionary reading unit that reads a feature vector corresponding to an arbitrary category from the uncompressed dictionary, and divides each element value constituting each feature vector into a set of a plurality of element sets, and reads each of the elements read by the dictionary reading unit. an element group selection unit that sequentially selects each element pair from the element values, elements set selected by the selection unit element sets of elements based on predetermined criteria
A compression determining unit that determines whether or not to perform compression by reducing the number; and when the compression determining unit determines that compression is to be performed, a code that indicates the compression by replacing each element value of the element set with one representative value. And a non-compression processing unit that outputs, as non-compressed data, each element value constituting the element set and a code indicating non-compression when the compression determination unit determines that compression is not to be performed. When, along with constituting the compression dictionary corresponding to any category and sequentially stores the compressed and non-compressed data corresponding to each element group output from the compression processing unit and the non-compression processing unit, any catheter
Generation of compression dictionaries corresponding to Gori for all categories
And a compression dictionary storage unit for executing and executing the compression dictionary corresponding to all the categories and outputting the compressed dictionary.

[0006] Further, all the element sets are composed of the same number of elements. In addition, the compression determination unit determines a maximum element of each element value that constitutes one element set.
Compare the difference between the value and the minimum element value with a predetermined threshold,
If the difference is larger than the threshold, reduce the number of elements in the element set
It is determined that the compression is not performed, and when the difference is equal to or smaller than the threshold value, it is determined that the compression by reducing the number of elements of the element set is performed . Further, the non-compression processing unit converts the representative value of each element value constituting an arbitrary element set and each difference value indicating the difference between each element value and the representative value as uncompressed data together with a sign indicating non-compression. This is to output. The code indicating compression is described by replacing the least significant bit of the representative value, and the code indicating non-compression is replaced by the least significant bit of a predetermined element value among the element values constituting the element set. It is to describe.

Further, the pattern recognition apparatus according to the present invention comprises a feature extracting unit for extracting a feature from an input pattern and outputting an input feature vector including a plurality of element values indicating each feature, and an input feature vector. An element set selection unit that divides each element value into a set of a plurality of element sets, and sequentially selects each element set as an input element set ;
It is composed of multiple feature vectors, and
The element set that constitutes a vector is compressed by reducing the number of elements.
From which it is compression dictionary, a compression dictionary buffer for storing read the dictionary feature vectors corresponding to any category, elements of sets dictionary feature vector dictionary elements from set corresponding to the input element sets selected by the element group selection unit And a compression determining unit that determines whether or not the dictionary element set is compressed by reducing the number of elements based on the codes included in the selected dictionary element set. When it is determined that compression has been performed, a compression unit distance calculation unit that calculates a distance between the two element sets from an element value that forms the input element set and a representative value that forms the dictionary element set, and a compression determination unit When it is determined that the dictionary element set is not compressed, an uncompressed unit distance meter that calculates the distance between the two element sets from the element values forming the input element set and the element values forming the dictionary element set And the distance between the corresponding element sets calculated by the compression section distance calculation section and the non-compression section distance calculation section for all the element sets constituting an arbitrary category. The distance between the input pattern output from the distance addition unit and the category is compared for all categories, and the category corresponding to the minimum value is determined as the recognition result for the input pattern. And a recognition determining unit for outputting.

The compression unit distance calculation unit calculates the distance between the representative value obtained from the input element set and the representative value forming the dictionary element set, and multiplies this distance by the number of elements forming the input element set. The calculated distance is output as the distance between the two elements. Also, the compression unit distance calculation unit calculates the distance between each element value forming the input element set and the representative value forming the dictionary element set, and outputs the sum of these distances as the distance between both elements. It was done.

[0009]

Therefore, the feature vector corresponding to an arbitrary category read from the non-compressed dictionary is divided into a plurality of element sets by the element set selection section, and thereafter, the compression judgment section sets each of the element sets on a predetermined basis. Whether compression / non-compression is determined based on compression or not, and if compression is determined, the compression processing unit generates compressed data including a representative value of each element value constituting the element set and a code indicating compression, and determines whether compression is not performed. If it is determined, the non-compression processing unit generates non-compressed data including each element value and a code indicating non-compression, the compressed and non-compressed data are stored in the compression dictionary storage unit, and a compression dictionary is generated. .

The input feature vector generated from the input pattern by the feature extraction unit is divided into a plurality of element sets by the element selection unit, each of which is sequentially selected as an input element set, and an arbitrary one read from the compression dictionary. The one corresponding to the input element set is read from the dictionary feature vector corresponding to the category as a dictionary element set, and compression / non-compression is determined for the dictionary element set by the compression determination unit.
When the compression is determined, the distance between the two element sets is calculated from the element value of the input element set and the representative value of the dictionary element set by the compression unit distance calculation unit. The distance between the two element sets is calculated by the sum of the distances between the respective elements calculated from the respective element values of the input element set and the representative value of the dictionary element set by the uncompressed section distance calculation section, and the distance addition section calculates The distances between elements are totaled for each category and output as the distance between the input pattern and each category, and the recognition determination unit outputs the category corresponding to the minimum value of the distances to all the categories as the recognition result for the input pattern. You.

[0011]

Next, the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a pattern recognition dictionary generation device according to a first embodiment of the present invention.
1 is a dictionary vector storage unit that stores a dictionary vector having a predetermined number of elements for each category, and 10 is based on each dictionary vector stored in the dictionary vector storage unit 1 and a threshold value input from the terminal 2. This is a compression dictionary creation unit that creates a compression dictionary.

In the compressed dictionary creating unit 10, reference numeral 11 denotes a dictionary reading unit for reading dictionary vectors of an arbitrary category from the dictionary vector storage unit 1, and reference numeral 12 sequentially reads an arbitrary set of elements from the dictionary vectors read by the dictionary reading unit 11. The element set selection unit 13 to be selected compares the evaluation value calculated from the element set selected by the element set selection unit 12 with the threshold value stored in the threshold value storage unit 14 to determine whether compression is necessary. 15 is a compression processing section that compresses the element set by calculating a representative value of the element set output from the compression determination section 13, and 16 is the element set output from the compression determination section 13 as it is. A non-compression processing unit 17 for outputting the dictionary vector storage unit 1 uses all the element sets output from the compression processing unit 15 and the non-compression processing unit 16 as a compression dictionary.
Is a compression dictionary storage unit that outputs the compressed dictionary data.

Next, the operation of the present invention will be described with reference to the drawings. First, a method of generating a dictionary vector stored in the dictionary vector storage unit 1 in advance will be described. This dictionary vector is a source of a compression dictionary to be created, and is generated based on a learning data set. In the learning data set, a plurality of patterns are stored for each category, and pattern data of all categories is stored. From each of these patterns, a feature vector having a predetermined number of elements is obtained, and for each category, a representative feature vector of the category is obtained from a plurality of corresponding patterns. For example, an average vector of feature vectors obtained from patterns belonging to one category can be determined as a representative feature vector. This representative feature vector becomes a dictionary vector.

FIG. 8 is a flowchart showing a procedure for generating a feature vector from each pattern. In the following, a method for generating a feature vector will be described using a binary character pattern as an example. First, in step 81, an arbitrary pattern is input, and in step 82, the contour of the input character line is traced one pixel at a time, thereby examining the direction of the contour. In step 83, when the direction to which each pixel belongs is the vertical direction, a predetermined set value is written to the same coordinates of the bitmap indicating the direction. By executing steps 82 and 83 for all the contour points in this way, a bitmap representing the vertical direction among the contours of the input character is generated.

Next, at step 84, the bit map generated as described above is converted into a 16 × 4 bit map as shown in FIG.
The number of pixels into which a predetermined set value existing in each mesh is written is counted, and this value is used as a value of one element of the feature vector. A total of 64 elements of 16 × 4 in one direction are obtained. Further, by rotating this input pattern by 45 degrees and executing steps 82 to 84, the characteristics of 64 elements in the diagonal 45 degree direction are extracted, and similarly, steps 82 to 85 are further performed. By executing the process twice, 64 features representing the horizontal direction and 64 features representing the 135-degree diagonal direction are extracted, and a feature vector of 256 elements is generated in total.

The present invention is not limited to the feature extraction method described above, but may be any method that generates a feature vector having a predetermined number of elements.
By applying such a feature vector generation method to all learning patterns belonging to one category and calculating an average vector of the generated feature vectors, a dictionary vector of the category is generated, and these procedures are applied to all categories. By executing the above, an uncompressed pattern recognition dictionary that is a source of the compressed pattern recognition dictionary is generated.

FIG. 4 is a block diagram showing a dictionary vector having a predetermined number of elements. In FIG. 4, reference numeral 40 denotes a category code for identifying each category, 41 to 42, 45 to 4
7, are elements included in each element set. Now, assuming that the number of elements of this dictionary vector is n, a dictionary for one category is composed of a category code 40 and a dictionary vector including n elements. The dictionary vectors of all these categories are generated in advance based on the above-described procedure, and are stored in the dictionary vector storage unit 1 of FIG. 1 as uncompressed pattern recognition dictionaries.

Next, the operation of the compression dictionary creating section 10 will be described. First, the dictionary reading unit 11 selects an arbitrary category, reads a dictionary vector corresponding to the category from the dictionary vector storage unit 1, and stores the dictionary vector. The element set selecting unit 12 sequentially reads out the element sets from the dictionary vectors stored in the dictionary reading unit 11 and transmits them to the compression determining unit 13. FIG. 9 is a block diagram illustrating the element set selection unit 12. In FIG. 9, reference numeral 90 denotes a dictionary buffer in the dictionary reading unit 11, in which each element of the dictionary vector corresponding to an arbitrary category read from the dictionary vector storage unit 1 is temporarily stored.

The element number storage unit 91 stores a head number and an end number in advance as storage location information of a plurality of elements constituting each element set (0 to K) stored in the dictionary buffer 90. The element set reading unit 92 reads a plurality of elements from the dictionary buffer 90 based on the start number and the end number stored in the element set number storage unit 91, and uses these elements as an element set. The data is stored in the set storage unit 93 and output to the terminal 94. For example, if the head number corresponding to the k-th element set is k1 and the end number is km, the element set reading unit 92
The m elements from the k1st element dk1 to the kmth element dkm of the dictionary buffer 90 are read.

In the above description, the case where the number m of elements included in each element set is variable has been described.
May be fixed. FIG. 10 is a block diagram showing another example of the element set selection unit 12, and particularly shows a case where the number of elements included in each element set is fixed. In FIG. 10, reference numeral 101 denotes an element set number calculation unit. In this case,
Since the number m of elements included in each element set is fixed,
An element set number calculation unit 101 is provided instead of the above-described element set number storage unit 91, and the head number of the k-th element set is set to k
× m + 1 and the end number is k × m + m
Is obtained by As described above, by setting the number of elements constituting all the element sets to be the same, the element set selection processing is simplified, and the processing can be performed at higher speed.

In FIG. 4, the element 41 indicates the first element d11 of the first element set.
Reference numeral 5 denotes the first element d21 of the second element set. Also, FIG. 7 described above shows a case where each direction is divided into 64 elements by employing the directional feature extraction method of four directions, and two elements surrounded by a thick line are used as an element set. That is, element 1 and element 2 constitute a first element set, and element 3 and element 4 constitute a second element set. Note that, in this embodiment, a case where two elements are made into one element set has been described as an example, but an element set may be made up of elements other than two.

Next, in FIG.
An evaluation value is obtained by using m elements belonging to an arbitrary element set stored in the element set storage unit 93 (see FIGS. 9 and 10) of the element set selection unit 12. By comparing with a threshold value stored in the threshold value storage unit 14, it is determined whether or not to compress these elements.
In general, when each element value is replaced with a predetermined representative value for the purpose of compressing a dictionary, it is expected that pattern recognition performance will decrease due to an error between each element value and the representative value. Therefore,
A predetermined evaluation value indicating an error between each element value and the representative value, that is, a variation of each element value, is calculated, and compression / non-compression is determined based on the evaluation value, whereby the generated pattern recognition dictionary is recognized. It is possible to suppress a decrease in performance.

In particular, the compression determining section 13 selects a maximum value and a minimum value from each of the elements constituting an arbitrary set of elements and evaluates the difference between the maximum value and the minimum value. This evaluation value generation circuit calculates an evaluation value for each element set. In addition, as another evaluation value calculation method, another method, for example, a variance value of each element in an element set may be calculated.

Here, when the calculated evaluation value is equal to or smaller than the threshold value, it is determined that the variation of each element constituting the element set is relatively small, and it is determined that compression is performed. In response, the m elements constituting the element set are output to the compression processing unit 15, and a representative value is obtained from each element by the compression processing unit 15, and the representative value and a code indicating compression are assigned to the element set. Is stored in the compression dictionary storage unit 17 as the information of. The average value of each element is used as the representative value of the element set, but the maximum value or the minimum value of each element may be selected and used.

On the other hand, when the calculated evaluation value is larger than the threshold value, it is determined that the variation of each element constituting the element set is relatively large, and it is determined that compression is not performed. In response to this, m elements constituting the element set are transmitted to the non-compression processing unit 16, and the non-compression processing unit 16 compresses each element value with a code indicating non-compression without being compressed.
7 is stored. As described above, as a compression / non-compression determination method, the difference between the maximum value and the minimum value of each element value constituting an arbitrary element set is compared with a predetermined threshold value, and the difference is determined by the threshold value. If the difference is smaller than the threshold, it is determined that compression is not performed. If the difference is smaller than the threshold value, compression is determined. Is simplified and the determination process is sped up.

In the case of non-compression, each element value is not stored in the compression dictionary storage unit 17 as it is, but the average value of each element value in the element set is calculated as a representative value. The difference value from each element may be stored together with the representative value. In this case, one representative value, m difference values, and a code indicating non-compression are stored for each element set. Is done. As a result, the number of bits required to represent each element value is reduced, and the dictionary capacity can be further reduced even in the case of non-compression.

Further, an element having a predetermined number is deleted from the element set, and m-1 element values or a difference value between each element value and a representative value and a total of m values consisting of one representative value are obtained. , And a code indicating non-compression. By repeatedly performing the series of compression dictionary creation processing for an arbitrary element set as described above for all element sets, a new dictionary in which all element sets are compressed / uncompressed is created, and a compressed dictionary storage is performed. By storing the contents stored in the unit 17 in the dictionary vector storage unit 1, a new compressed pattern recognition dictionary is completed.

FIG. 5 is an explanatory diagram showing a compressed pattern recognition dictionary. Compared to the uncompressed pattern recognition dictionary initially stored in the dictionary vector storage unit 1 shown in FIG. 4, the first element set, that is, from element d11 to element d1m, Are compressed to d1 'which is a representative value 52 generated from
Is stored with it. Further, the second element set is determined to be uncompressed. In this case, S2 is stored as a code 53 indicating non-compression, and the first element 45 d21 is used. Up to d2m which is the m-th element 47 is stored as it is.

As described above, the apparatus for generating a pattern recognition dictionary according to the present invention reads out a feature vector corresponding to an arbitrary category from an uncompressed dictionary, divides it into a plurality of element sets, and sets It is determined whether or not to compress the element set, and in the case of compression, each element value constituting the element set is replaced with one representative value to generate compressed data, and all the categories are obtained from the compressed and uncompressed data. Is generated, it is possible to generate a compressed dictionary without reducing the elements of the pattern recognition dictionary more than necessary, and without reducing the recognition performance drastically. Can be generated.

FIG. 6 is an explanatory diagram showing another example of the compressed pattern recognition dictionary, particularly showing a case where each element value is expressed by a representative value and its difference when the pattern is not compressed. In FIG. 6, when the element set is compressed, FIG.
Are stored in the same manner as the reference numeral 51 and the representative value 52. On the other hand, in the case of non-compression, as in the example of the second element set, the code 53 indicating non-compression is stored as S2 and the average value of m elements from the original element 45 to the element 47 is stored. Is stored in d2 'as a representative value 61, and the element difference value 65
Indicates a value obtained by subtracting the representative value 61 from the first element 45 of the element set.

Similarly, the difference between the m-1 elements and the representative value is stored, and the value of the element difference value 66 becomes the difference value between the m-1 element and the representative value, and the mth element The value for element 47 may or may not be stored. When the value relating to the m-th element is deleted, a representative value is added to the first element difference value 65 of the element set in the uncompressed portion distance calculation unit 26 of the pattern recognition device shown in FIG. 6
By adding 1, the original element value 45 is restored, and similarly, the m-1st element values are restored. Further, since the representative value 61 is the average value of the original element values, the m-th element value 47 is obtained by subtracting all m-1 element values already restored from m times the representative value 61. Can be restored. In the above description, the representative value 61 is described as being arranged at the head of the value of the element set. However, the place where the representative value 61 is arranged is not fixed, but can be set flexibly as necessary. is there.

Further, in the above description, the compression / non-compression is indicated by adding reference numerals 51 and 53 (see FIG. 5) to the element sets of each category stored in the generated pattern recognition dictionary. Although the case has been described above, each element set may be incorporated into a predetermined element value. FIG. 11 is an explanatory diagram showing another example of the compressed pattern recognition dictionary. In particular, a code indicating compression / non-compression for each element set is incorporated in the least significant bit of a predetermined element value.
Although it is assumed that it is represented by 8 bits, it is not actually limited to 8 bits, and it can be easily analogized even if it is represented by another length.

The code indicating compression or non-compression is placed at the least significant bit of the 8 bits constituting the element value. Here, these codes are assigned to the first element value or the representative value. Assuming that the code indicating compression is 1 and the code indicating non-compression is 0, the code 51 indicating compression in FIG.
2, the 8-bit value 1 shown in FIG.
11 is stored. Similarly, the code 53 indicating uncompressed in FIG. 5 is placed at the least significant bit of the first element value 45 of the corresponding element set and stored as the 8-bit value 115 in FIG. The values of the second and subsequent element values 46 and 47 are stored as they are.

As described above, the code indicating compression is described by replacing the least significant bit of the representative value with the least significant bit, and the code indicating non-compression is the lowest one of the predetermined element values of the element values constituting the element set. Since the description is made by replacing the lower one bit, it is possible to suppress an increase in the dictionary capacity due to the code added to each element set.

Next, as a second embodiment of the present invention, a pattern recognition apparatus that performs pattern recognition using a dictionary generated by the above-described pattern recognition dictionary generation apparatus will be described with reference to FIG. FIG. 2 is a block diagram showing a pattern recognition device. In FIG. 2, reference numeral 20 denotes a pattern recognition device that recognizes a pattern input from a terminal 3 based on a compressed pattern recognition dictionary stored in a dictionary vector storage unit 1. , A recognition unit that outputs information indicating a predetermined category from the terminal 4 as a recognition result.

In the recognizing unit 20, a feature extracting unit 21 extracts a feature vector from a pattern input from the terminal 3 based on a feature extracting method (see FIG. 8), and a temporary storing unit 22 stores the extracted feature vector. A vector buffer 23 is an element set selection unit that selects a predetermined element set from the feature vectors in the vector buffer 22. Reference numeral 24 denotes a compression dictionary buffer for reading out a dictionary vector corresponding to an arbitrary category from the dictionary vector storage unit 1 and temporarily storing the dictionary vector. Reference numeral 25 denotes a compression dictionary buffer 24 for storing an element set corresponding to the element set selected by the element set selection unit 23. And a compression determination unit that outputs the read element set to the non-compression unit distance calculation unit 26 or the compression unit distance calculation unit 27 according to the code indicating compression / non-compression included in the data.

Reference numeral 26 denotes a non-compression section distance calculation section for calculating the distance between the element set in the non-compression state input from the compression determination section 25 and the element set from the element set selection section. The compression unit distance calculation unit 28 calculates the distance between the element set in the compressed state and the element set from the element set selection unit, and the distance 28 output as a calculation result from the non-compression distance calculation unit 26 or the compression unit distance calculation unit 27. Is added to all the element sets for each category. 29 is a distance addition unit that calculates the pattern input from the terminal 3 and the dictionary vector of each category read from the dictionary vector storage unit 1 calculated by the distance addition unit 28. The recognition determining unit outputs the category having the smallest distance among the distances to the terminal 4 as a recognition result.

Next, the operation of the pattern recognition device will be described. It is assumed that the pattern recognition dictionary generated by the above-described pattern recognition dictionary generation device is stored in the dictionary vector storage unit 1. First, the feature extraction unit 21 extracts a feature vector including a predetermined number n of elements from the pattern input from the terminal 3 by using the same feature extraction method as that used when the pattern recognition dictionary is generated (see FIG. 8). , Are stored in the vector buffer 22.

Next, the distance between the read pattern and all dictionary vectors corresponding to each category in the compressed pattern recognition dictionary is calculated, and recognition judgment is performed based on the distance between each category. . Now, when calculating the distance to the L-th category, the dictionary vector of the L-th category is read from the dictionary vector storage unit 1, and
It is stored in the compression dictionary buffer 24. Element set selector 23
Has the same configuration as that shown in FIG. 9 or FIG. 10 described above, sequentially reads out the element sets from the input feature vectors stored in the vector buffer 22, and sets the uncompressed portion distance calculating section 26 and Output to the compression section distance calculation section 27.

In response, an element set corresponding to the input element set read by the element set selection unit 23 is read from the compression dictionary buffer 24 to the compression determination unit 25 as a dictionary element set. The compression determination unit 25 refers to the reference numerals 51 and 53 stored in the pattern recognition dictionaries shown in FIGS. 5 and 6 to determine whether each element constituting the element set is compressed or not. The compression unit distance calculation unit 26 or the compression unit distance calculation unit 2
7 is output to these dictionary element sets.

The non-compression unit distance calculation unit 26 calculates the individual element values constituting the input element set selected by the element set selection unit 23 and the dictionary element set correspondingly output from the compression determination unit 25. The sum of the absolute values of the differences from the individual element values that make up is calculated, and this calculation result is output to the distance addition unit 28 as the distance between one corresponding element set. In addition,
Although the distance is calculated by the sum of the absolute values of the differences, the distance may be expressed by another calculation method, for example, the sum of the squares of the differences. When the dictionary element set is described by the representative value of the element set and the difference between each element and the representative value, the representative value is added to the stored difference value, and the element value of the original dictionary is added. Is restored, the same distance calculation as described above is performed. In this case, when one of the difference values between the element value and the representative value is deleted, the difference value is deleted by subtracting the sum of the restored element values excluding the element from the number of elements of the representative value. Restore the value of an element.

On the other hand, in the compression section distance calculation section 27, the individual element values constituting the input element set selected by the element set selection section 23 and the dictionary element set output from the compression determination section 25 corresponding thereto are set. Is calculated from the representative values of the individual element sets that constitute. FIG. 12 is a block diagram illustrating the compression unit distance calculation unit 27. Reference numeral 121 denotes a representative value calculation unit that calculates a representative value from each element value stored in the element set storage unit 93 of the element selection unit 23. Is an element number storage unit for storing the number of each element value stored in the element set storage unit 93, and 124 is a compression determination unit 25 via a terminal 123.
The dictionary representative value storage unit 125 stores the representative value of the element set input from the dictionary representative value calculation unit 121 and calculates the distance from the input representative value calculated by the representative value calculation unit 121 and the dictionary representative value from the dictionary representative value storage unit. The distance calculation unit 126 multiplies the distance calculated by the distance calculation unit 125 by the number of elements in the number-of-elements storage unit 122 and outputs the result to the distance addition unit 28 via the terminal 127 as the corresponding number of elements. It is a multiplication unit.

First, the representative value calculation unit 121 reads each element constituting the input element set from the element set storage unit 93 in which an arbitrary input element set selected from the input feature vector by the element set selection unit 23 is stored. Is read out, the average value is calculated and used as an input representative value, and the number-of-elements storage unit 122 stores the number of elements. On the other hand, the dictionary representative value from the compression determination unit 25 is input through the terminal 123 and stored in the dictionary representative value storage unit 124. The distance calculation unit 125 stores the dictionary representative value and the representative value calculation in the dictionary representative value storage unit 124. Part 12
The absolute value of the difference from the input representative value of 1 is obtained as a distance and output to the multiplication unit 126.

Next, the multiplication unit 126 multiplies the value from the distance calculation unit 125 by the value of the number-of-elements storage unit 122, and uses the result as the distance of the element set via the distance addition unit 2 via the terminal 127.
8 is output. Although the distance calculating unit 125 sets the absolute value of the difference between the two representative values to be input, the distance may be calculated by another distance calculating method, for example, the square of the difference between the two representative values.

In the representative value calculation unit 121, the average value of each input element value is obtained as the representative value. However, the representative value calculation unit 121 calculates the element value based on another representative value calculation method, for example, the maximum value or the minimum value of each element value. A set of representative values may be calculated. in this way,
When the dictionary element set is compressed, the distance between the representative value obtained from the input element set and the representative value forming the dictionary element set is calculated, and this distance is multiplied by the number of elements forming the input element set. Since the distance is output as the distance between the two elements, the distance calculation is simplified and the distance calculation process is sped up.

In the above description, the input representative value is calculated from the input element set, and the distance between the two element sets is calculated based on the input representative value and the dictionary representative value. Alternatively, a distance for each element may be calculated from each element of the input element set and the dictionary representative value, and the distance of the element set may be calculated by adding these distances. FIG. 13 is a block diagram showing another example of the compression unit distance calculation unit 27. In particular, the distance of an element set is calculated by calculating the distance for each element. In FIG. 13, the representative value calculation unit 121 and the multiplication unit 12 shown in FIG.
6, the element value storage unit 131 and the distance addition unit 1
It is the same as FIG. 12 except that the 306 is provided.

First, the element value storage section 131 reads out and stores each element constituting the input element set stored in the element set storage section 93 one by one, and the distance calculation section 125 stores the element number storage section 122 The absolute value of the difference between each element value in the element value storage unit 131 and the dictionary representative value stored in the dictionary representative value storage unit 124 is obtained as a distance in accordance with the number of elements stored in Output to distance adding section 126. The distance adding unit 126 adds the distances corresponding to all the elements output from the distance calculating unit 125 to obtain the total sum, and outputs the sum to the distance adding unit 28 via the terminal 127 as the element set distance.

Although the distance calculator 125 sets the absolute value of the difference between the two representative values to be input, the distance is calculated using another distance calculation method, for example, the square of the difference between the two representative values. Is also good. In this way, when the dictionary element set is compressed, the distance between each element value forming the input element set and the representative value forming the dictionary element set is calculated, and the sum of these distances is calculated between the two elements. Since the distance is output, the distance can be calculated by simple arithmetic processing, and the configuration for calculating the distance is simplified.

As described above, the compression section distance calculation section 26
And the non-compression unit distance calculation unit 27 calculates the distance between arbitrary element sets and outputs the calculated distance to the distance addition unit 28, and sequentially calculates the distances between all element sets constituting the corresponding category. The sum is calculated by the adder 28 and output to the recognition determiner 29 as the distance between the feature vector of the pattern input from the terminal 3 and the dictionary vector corresponding to an arbitrary category in the dictionary vector storage 1. Further, when the distances to all the categories are calculated and input, the recognition determination unit 29 selects the smallest one of the distances corresponding to each category, and determines the corresponding category by using the input pattern. Is output to the terminal 4 as a recognition result for.

As described above, the pattern recognition apparatus according to the present invention reads a dictionary feature vector corresponding to an arbitrary category from the compression dictionary generated by the above-described pattern recognition dictionary generation apparatus, selects a dictionary element set, and selects this dictionary element set. When the dictionary element set is compressed, the distance between the two element sets is calculated from the representative value and the element value of the input element set selected from the input feature vector. Calculate the distance between the two element sets based on the sum of the distances between the element values, and sum up these for all the element sets, calculate the distance between the input pattern and any category, and calculate the distance corresponding to each category. Since the recognition determination for the input pattern is performed based on the input pattern, it is possible to perform pattern recognition having a predetermined recognition performance even with a dictionary whose capacity is compressed. To become.

Next, as a third embodiment of the present invention, a pattern recognition dictionary generating apparatus will be described with reference to FIG. FIG. 3 is a block diagram of a pattern recognition dictionary generation device according to another embodiment of the present invention. In particular, the pattern recognition dictionary generation unit 10 (see FIG. 1) and the recognition unit 20 (see FIG. 2) are provided. ing. 3, the same or equivalent parts as those described above are denoted by the same reference numerals, 32 is a pattern data storage unit that stores a data set of patterns used for dictionary generation, and 33 is a procedure shown in FIG. Generates an original uncompressed pattern recognition dictionary based on the dictionary vector storage unit 1
This is a dictionary creation unit that outputs to the dictionary.

Reference numeral 10 denotes a compression dictionary creating unit for creating a pattern recognition dictionary stored in the dictionary vector storage unit 1 and a pattern recognition dictionary compressed based on a threshold value controlled from the terminal 2; The pattern read from the pattern data storage unit 32 via the pattern recognition unit 10 in the dictionary vector storage unit 1 is determined based on the pattern recognition dictionary newly created, and the result is output to the terminal 4. And 31 accumulate the recognition results from the recognizing unit 20, respectively, calculate a recognition performance value based on the number of patterns for which recognition determination has been performed and the number of correct patterns of the recognition result, and perform pattern recognition based on the recognition performance values. The control unit evaluates the recognition dictionary.

First, the dictionary generation unit 33 generates an original uncompressed pattern recognition dictionary from a pattern data set stored in the pattern data storage unit 32 based on the procedure shown in FIG. Output to the dictionary vector storage unit 1. Next, the control unit 31 activates the recognizing unit 20, and in response, the recognizing unit 20 stores the pattern read from the pattern data storage unit 32 via the terminal 3 in the dictionary vector storage unit 1. The recognition determination is performed using the uncompressed pattern recognition dictionary, and the result is output to the control unit 31 via the terminal 4. The control unit 31 calculates and stores a recognition performance value by using a ratio of the number of patterns for which the recognition unit 20 has performed the recognition determination and the number of correct patterns of the recognition result.

Next, the control section 31 sets a predetermined threshold value in the threshold value storage section 14 (see FIG. 1) in the compression dictionary Is started, and a new dictionary compressed based on the uncompressed pattern recognition dictionary stored in the dictionary vector storage unit 1 is created. After a new dictionary is created in the dictionary vector storage unit 1, the control unit 31 activates the recognition unit 20 again to execute the same recognition determination as described above based on the new compressed pattern recognition dictionary. A recognition performance value is calculated from the result.

Here, the control unit 31 compares the recognition performance value for the new compressed pattern recognition dictionary with the recognition performance value for the already stored uncompressed pattern recognition dictionary and determines that the reduction in the recognition performance is within a predetermined value. In the case of a decrease, the threshold value set in the threshold value storage unit 14 in the compression dictionary generation unit 10 via the terminal 2 is corrected, and then the compression dictionary generation unit 10 and the recognition unit 20 are again adjusted. Start up,
Compare the recognition performance values. On the other hand, when the recognition performance has decreased by a predetermined value or more, the corresponding pattern recognition dictionary 1
The pattern recognition dictionary obtained last time is determined as the final pattern recognition dictionary, and a series of dictionary creation processing ends.
This satisfies the predetermined recognition performance,
It is possible to generate the most compact low-capacity dictionary.

[0056]

As described above, the pattern recognition dictionary generation device according to the present invention reads out a feature vector corresponding to an arbitrary category from an uncompressed dictionary, divides it into a plurality of element sets, and based on a predetermined criterion. Key of each element set
It is determined whether or not to perform compression by reducing the prime number. In the case of performing compression, each element value forming the element set is replaced with one representative value to generate compressed data. Since the compression dictionary corresponding to the category is generated, it is possible to generate a compressed dictionary without reducing the elements of the pattern recognition dictionary more than necessary. A capacity pattern recognition dictionary can be generated.

[0057] Further, since all elements set to be configured with the same number of elements, in compressing the number of elements, it is possible to select an element sets by separating each predetermined number of elements, the elements The set selection processing is simplified, and processing can be performed at higher speed. Further, as a compression / non-compression determination method, a difference between a maximum value and a minimum value of each element value constituting an arbitrary element set is compared with a predetermined threshold value, and the difference is larger than the threshold value. In this case, compression is determined not to be performed, and compression is determined to be performed when the difference is equal to or smaller than the threshold value. Therefore, it is possible to determine accurately with a simple calculation, and the configuration for the determination process is simplified. And the speed of the determination process is increased.

When it is determined that the data is not compressed, a representative value of each element value forming an arbitrary element set and a difference value indicating a difference between each element value and the representative value are converted to a non-compressed value. Since the data is output as non-compressed data together with the indicated code, the number of bits required to represent each element value is reduced, and the dictionary capacity can be further reduced even in the case of non-compression. Also, as a code indicating compression, the least significant one of the representative values is used.
Bits are described by replacing them with bits, and a code indicating non-compression is represented by the least significant one of predetermined element values among the element values constituting the element set.
Since the description is made by replacing the bits, it is possible to suppress an increase in the dictionary capacity due to the code added to each element set.

Further, the pattern recognition apparatus according to the present invention provides a compression dictionary generated by the above-described pattern recognition dictionary generation apparatus , that is, a plurality of feature bases corresponding to each category.
And arbitrary feature vectors
Element set is compressed by reducing the number of elements
From the dictionary, select the dictionary element pair reads the dictionary feature vectors corresponding to any category, the dictionary element set is its
If the compression is performed by reducing the number of elements, the distance between the two element sets is calculated from the representative value and the element value of the input element set selected from the input feature vector. Calculates the distance between the two element sets by the sum of the distances between the corresponding element values, sums up these for all the element sets, calculates the distance between the input pattern and any category, and Since the recognition determination for the input pattern is performed based on the distance, it is possible to perform pattern recognition having a predetermined recognition performance even with a dictionary whose capacity is compressed.

When the dictionary element set is compressed, the distance between the representative value obtained from the input element set and the representative value forming the dictionary element set is calculated, and the distance which constitutes the input element set is calculated. Since the distance multiplied by the number is output as the distance between the two elements, the distance calculation is simplified, and the distance calculation process is sped up. Further, when the dictionary element set is compressed, the distance between each element value forming the input element set and the representative value forming the dictionary element set is calculated, and the sum of these distances is calculated as the distance between both elements. Since the output is performed, the distance can be calculated by simple arithmetic processing, and the configuration for calculating the distance is simplified.

[Brief description of the drawings]

FIG. 1 is a block diagram of an apparatus for creating a pattern recognition dictionary according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a pattern recognition device according to a second embodiment of the present invention.

FIG. 3 is a block diagram of a pattern recognition dictionary creation device according to a third embodiment of the present invention.

FIG. 4 is an explanatory diagram showing an uncompressed pattern recognition dictionary.

FIG. 5 is an explanatory diagram showing a compressed pattern recognition dictionary.

FIG. 6 is an explanatory diagram showing another example of a compressed pattern recognition dictionary.

FIG. 7 is an explanatory diagram showing element division when generating a feature vector.

FIG. 8 is a flowchart illustrating a feature vector generation procedure.

FIG. 9 is a block diagram illustrating an element set selection unit.

FIG. 10 is a block diagram showing another example of the element set selection unit.

FIG. 11 is an explanatory diagram showing another example of a compressed pattern recognition dictionary.

FIG. 12 is a block diagram illustrating a compression unit distance calculation unit.

FIG. 13 is a block diagram illustrating another example of a compression unit distance calculation unit.

[Explanation of symbols]

1 ... Dictionary vector storage unit, 10 ... Compression dictionary creation unit, 11
... Dictionary reading unit, 12 ... Element set selection unit, 13 ... Compression determination unit, 14 ... Threshold storage unit, 15 ... Compression processing unit, 16
... non-compression processing unit, 17 ... compression dictionary storage unit, 20 ... recognition unit, 21 ... feature extraction unit, 22 ... vector buffer, 23
... An element set selection unit, 24 a compression dictionary buffer, 25 a compression determination unit, 26 a non-compression unit distance calculation unit, 27 a compression unit distance calculation unit, 28 a distance addition unit, 29 a recognition determination unit, 31.
Control unit 32 pattern data storage unit 33 dictionary creation unit 90 dictionary buffer 91 element set number storage unit 9
2 ... Element set reading unit, 93 ... Element set storage unit, 101 ...
Element set number calculation unit, 121: representative value calculation unit, 122: element number storage unit, 124: dictionary representative value storage unit, 125: distance calculation unit, 126: multiplication unit, 131: element value storage unit.

Claims (8)

(57) [Claims] 1. A feature vector comprising a plurality of feature vectors corresponding to each category, each feature vector comprising a plurality of element values indicating features extracted from a predetermined learning pattern set , and Compression by reduction
A pattern recognition dictionary generation device that generates a compressed pattern recognition dictionary for recognizing an input pattern as one of categories based on an uncompressed dictionary expressed without being performed. A dictionary reading unit for reading a feature vector corresponding to an arbitrary category; each element value constituting each feature vector is divided into a set of a plurality of element sets, and each of the element values read by the dictionary reading unit. An element set selecting unit that sequentially selects each element set from a compression determining unit that determines whether to perform compression by reducing the number of elements of the element set selected by the element set selecting unit based on a predetermined criterion. When it is determined by the compression determination unit to compress, each element value constituting the element set is replaced with one representative value,
A compression processing unit that outputs as compressed data together with a code indicating compression, and when the compression determination unit determines that compression is not to be performed,
A non-compression processing unit that outputs each element value constituting the element set and a code indicating non-compression as uncompressed data; and the compression corresponding to each element set output from the compression processing unit and the non-compression processing unit. And a compressed dictionary corresponding to an arbitrary category by sequentially storing the compressed dictionary and uncompressed data, and generating a compressed dictionary corresponding to the arbitrary category.
And a compression dictionary storage unit configured to execute and perform a compression dictionary for all categories and output the compression dictionaries for all categories. 2. The pattern recognition dictionary generation device according to claim 1, wherein all of the element sets have the same number of elements. 3. The pattern recognition dictionary generation device according to claim 1, wherein the compression determining unit is configured to determine a maximum element value and a maximum element value of an element value forming one element set.
Comparing the difference with the small element value with a predetermined threshold, and if the difference is greater than the threshold, the number of elements of the element set
It is determined that compression by reduction is not performed, and when the difference is equal to or less than the threshold, the number of elements in the element set is reduced.
A pattern recognition dictionary generation device, which determines to perform compression . 4. The pattern recognition dictionary generation device according to claim 1, wherein the non-compression processing unit calculates a representative value of each element value constituting an arbitrary element set and a difference between each of the element values and the representative value. A pattern recognition dictionary generation device, which outputs each difference value shown together with a code indicating non-compression as the uncompressed data. 5. The pattern recognition dictionary generation device according to claim 1, wherein the code indicating the compression is described by replacing the least significant bit of the representative value with the least significant bit, and the code indicating the non-compression corresponds to the element set. The least significant one of the predetermined element values among the constituent element values
A pattern recognition dictionary generation device characterized by being described by replacing bits. 6. A pattern recognition apparatus for recognizing an input pattern based on a predetermined pattern recognition dictionary, wherein a feature is extracted from the input pattern, and an input feature vector including a plurality of element values indicating each feature is extracted. a feature extraction unit for outputting, with the respective element values constituting the input feature vector is divided into a plurality of elements of sets of the set, the element set selection unit for sequentially selecting each element sets as an input element sets, corresponding to each category From multiple feature vectors
The element set that constitutes an arbitrary feature vector is
A compression dictionary buffer for reading and storing a dictionary feature vector corresponding to an arbitrary category from a compression dictionary compressed by prime reduction, and an element set corresponding to an input element set selected by the element set selection unit; A compression determining unit that selects a dictionary element set from the feature vector and determines whether or not the dictionary element set is compressed by reducing the number of elements from codes included in the selected dictionary element set; When the determination unit determines that the dictionary element set is compressed, the compression for calculating the distance between the two element sets from the element values forming the input element set and the representative values forming the dictionary element set A unit distance calculation unit, and when the compression determination unit determines that the dictionary element set is not compressed, an element value forming the input element set and the dictionary element set An uncompressed part distance calculator for calculating the distance between the two element sets from the constituent element values, and a distance between the corresponding element sets calculated by the compressed part distance calculator and the non-compressed part distance calculator, A distance adding unit that outputs the distance between the input pattern and the category by summing up all the element sets that constitute the category of the category; and the input pattern and the category that are output from the distance adding unit. A recognition determining unit that compares distances between all categories and outputs a category corresponding to the minimum value as a recognition result for the input pattern. 7. The pattern recognition device according to claim 6, wherein the compression unit distance calculation unit calculates a distance between a representative value obtained from the input element set and a representative value constituting the dictionary element set. A pattern recognition apparatus, wherein a distance obtained by multiplying a distance by the number of elements constituting the input element set is output as a distance between the two elements. 8. The pattern recognition device according to claim 6, wherein the compression unit distance calculation unit calculates a distance between each element value forming the input element set and a representative value forming the dictionary element set. And a sum of the distances is output as a distance between the two elements.