JP5343579B2 - Pattern recognition dictionary creation device and program - Google Patents

Description

  The present invention relates to a pattern recognition dictionary creation device and a program.

  In recent years, pattern recognition typified by character recognition has been performed by devices having a small memory capacity such as mobile phones, PDAs (Personal Digital Assistants), and car navigation devices. In an apparatus having a small memory capacity, the capacity of a dictionary used for pattern recognition and the accuracy of pattern recognition become problems. That is, there is a need for a pattern recognition dictionary that can significantly reduce the dictionary capacity without significantly reducing recognition accuracy.

For example, in the case of character recognition, first, a feature vector obtained based on a standard character pattern of each character is registered in a recognition dictionary (the feature vector will be described later). Then, the method of calculating the intervector distance between the feature vector obtained based on the input character pattern and each feature vector registered in the recognition dictionary, and making the character with the closest intervector distance as the recognition result is a normal method. It is a technique.
Further, as a distance calculation method, there are a primary identification method and a secondary identification method. The primary identification method is a method for obtaining the Euclidean distance between the feature vector of each character registered in the recognition dictionary and the feature vector obtained based on the inputted character pattern. In the secondary identification method, eigenvalues and eigenvectors representing the pixel distribution of the character pattern are obtained using techniques such as Mahalanobis distance and pseudo Bayes distance identification, and these values are registered in the recognition dictionary. The eigenvalue and eigenvector obtained from the input character pattern are compared with the eigenvalue and eigenvector registered in the recognition dictionary to recognize the input character pattern (see, for example, Patent Document 1).

JP 2003-67743 A

  In the primary identification method, since the feature vector of each character code has only to be registered as a dictionary, the dictionary capacity is small, but there is a problem that the accuracy of pattern recognition is low. The secondary identification method has high pattern recognition accuracy, but requires a large dictionary capacity because eigenvalues and eigenvectors representing pixel distributions of character patterns must be registered in the dictionary. Therefore, in a device with a small memory capacity, the primary identification method is used at the expense of pattern recognition accuracy.

  The present invention has been made in view of the above circumstances, and provides a pattern recognition dictionary creation device and program capable of improving the recognition accuracy of an image pattern even in a device having a small storage capacity for storing a dictionary. With the goal.

A pattern recognition dictionary creation device disclosed in the present specification obtains a feature amount based on an image pattern of an input image, generates a feature vector having the feature amount as an element, and feature vectors of each image And extracting image having high similarity, grouping the images having high similarity as a group, and information on pixel distribution of the image based on the feature vector generated by the feature vector generating unit. When creating a pattern recognition dictionary by storing the feature data generating means for generating a plurality of feature data, identification information for identifying each image, and feature data of each image in the storage means , without increasing the data amount of the characteristic data of an image included, by reducing the data amount of the characteristic data of the image not included in the group, the group The pattern recognition dictionary is created so that the amount of image feature data included in the group is larger than the amount of image feature data included in the image, and the pattern recognition dictionary is stored in the memory Recognizing dictionary creating means stored in the means.

  According to the pattern recognition dictionary creation device disclosed in this specification, a recognition dictionary that improves the recognition accuracy of an image pattern can be created even with a device having a small storage capacity for storing a pattern recognition dictionary.

It is a figure which shows the structure of a pattern recognition dictionary creation apparatus. It is a figure which shows the structure of a dictionary creation part. It is a figure which shows the character prepared in order to demonstrate grouping, and its average vector. (A) is a figure which shows a mode that the character was arrange | positioned in the vector space based on the average vector of each character, (B) is a figure which shows the result of the 1st hierarchical clustering. (A) is a figure which shows the result of the 2nd hierarchical clustering, (B) is a figure which shows the result of the 3rd hierarchical clustering. (A) is a figure which shows the result of the 4th hierarchical clustering, (B) is a figure which shows the result of the 5th hierarchical clustering. It is a figure which shows the result of the 6th hierarchical clustering. It is a figure which shows an example of a structure of an index table. It is a figure which shows an example of a structure of a code book table. (A) is a figure which shows a mode that the subvector which partially matches among the element strings of an eigenvector is extracted, (B) is an example which assigns an index number to the extracted subvector and registers it in the codebook table (C) is a diagram showing a state of extracting sub-vectors that partially match the element sequence of the eigen vector while allowing an error, and (D) shows an index number for the extracted sub-vector. It is a figure which shows the example given and registered to a codebook table. It is a flowchart which shows the process sequence of a dictionary preparation part. It is a figure for demonstrating the other method of grouping.

  Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

  As shown in FIG. 1, the pattern recognition dictionary creation device 1 of this embodiment includes a dictionary creation unit 10, a scanner device 21, an operation unit 22, a storage device 23, and a display device 24.

  The scanner device 21 reads an image such as characters printed on paper as pattern image data. The read pattern image data is stored in the storage device 23 under the control of a CPU (Central Processing Unit) 11.

  The operation unit 22 is an operation input reception unit that receives a user instruction, and receives an instruction to start the operation of the scanner device 21, an instruction to display the dictionary created by the dictionary creation unit 10 on the display device 24, and the like.

  The storage device 23 stores pattern image data read by the scanner device 21, a pattern recognition dictionary created by the dictionary creation unit 10, and the like.

  The display device 24 displays the data of the pattern recognition dictionary created by the dictionary creation unit 10 under the control of the CPU 11.

  As shown in FIG. 1, the dictionary creation unit 10 includes a CPU 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, an input / output interface 14, a graphic interface 15, and a network interface 16. ing.

The ROM 12 stores a program for controlling the CPU 11. The CPU 11 reads a program recorded in the ROM 12 and performs a calculation according to the read program. A functional block of the dictionary creation unit 10 realized by cooperation of hardware such as the CPU 11 and a program stored in the ROM 12 will be described later with reference to FIG. Further, the RAM 13 records data being calculated by the CPU 11 and data after the calculation. For example, the pattern image data read by the scanner device 21 and stored in the storage device 23 is read by the control of the CPU 11 and stored in the RAM 13.
The program does not necessarily have to be stored in the ROM 12. For example, a computer-readable flexible disk (FD), DVD (Digital Versatile Disc), DVD-RAM, CD-ROM (Compact Disc Read Only Memory) ), CD-R (Recordable) / RW (ReWritable), magneto-optical disk, portable storage medium such as an IC card, or storage medium such as HDD provided in a computer, public line, Internet, LAN, WAN, etc. The program may be stored in another computer (or server) connected to the computer via the computer, and the computer may read and execute the program from this. Alternatively, the program may be stored in a portable storage medium or storage medium from another computer (or server) via a public line, the Internet, a LAN, a WAN, or the like, and the computer may read the program from there and execute it.

  The input / output interface 14 is a signal input / output unit. The pattern image data read by the scanner device 21 is input to the input / output interface 14. The input / output interface 14 outputs the input pattern image data to the storage device 23 under the control of the CPU 11. The input / output interface 14 outputs the pattern recognition dictionary read from the storage device 23 under the control of the CPU 11 to, for example, the external storage device 30.

  The graphic interface 15 is an interface for displaying an image processed by the CPU 11 on the display device 24, and converts graphic data into a waveform electrical signal for display on the display device 24.

  The network interface 16 is connected to the network. For example, the pattern recognition dictionary read from the storage device 23 is transferred to another device via the network under the control of the CPU 11.

Next, functional blocks of the dictionary creation unit 10 realized by cooperation of hardware such as the CPU 11 and a program stored in the ROM 12 will be described with reference to FIG.
The function unit 110 realized by the program-controlled CPU 11 includes a feature vector generation unit 111, an average vector calculation unit 112, a principal component analysis unit 113, a grouping unit 114, and a dictionary generation / registration unit 115. ing. The dictionary generation / registration unit 115 includes a grouped character information registration unit 116, an ungrouped character information registration unit 117, and a table storage unit 118. Hereinafter, each block will be described.

First, the feature vector generation unit 111 will be described.
The pattern image data scanned by the scanner device 21 and stored in the storage device 23 is read from the storage device 23 and stored in the RAM 13 under the control of the CPU 11.
The feature vector generation unit 111 acquires pattern image data stored in the RAM 13 from the RAM 13.
In the following description, a case where a pattern recognition dictionary is created using a character image obtained by reading characters printed on paper with the scanner device 21 will be described as an example. However, the present embodiment is not limited to characters. For example, it may be a human face image or voice data obtained by inputting a voice with a microphone.
The feature vector generation unit 111 obtains a feature amount based on the character pattern of the character image, and generates a feature vector having the feature amount as an element. Specifically, for example,
`` Son Ning, Toru Tahara, Hiroki Aso, Masayuki Kimura, `` High-precision character recognition using directional element features '', IEICE Transactions (D-II), J74-D-II, 3, pp. 330-339 (1991-2) ''
The character image is converted into a D-dimensional (D is an arbitrary natural number) feature vector using a method disclosed in the literature. The storage device 23 collects a plurality of learning samples such as characters with different fonts and handwritten characters for the same character. The feature vector generation unit 111 reads the character images of the plurality of learning samples from the RAM 13 and creates feature vectors of the read character images. In this embodiment, pattern image data read by the scanner device 21 is described as an example of pattern image data. However, pattern image data created by the pattern recognition dictionary creating device 1 itself may be used. In addition, the pattern image data generated by another apparatus may be input via a network. In the formula (1), N learning samples X ¹ of the same character (N is an arbitrary natural number) are used. , X ² , X ³ ,..., X ^N feature vectors, respectively.

特徴ベクトル生成部１１１は、上述した処理を複数の文字ごとにそれぞれ行う。特徴ベクトル生成部１１１は、特徴ベクトルを生成すると、文字ごとに複数収集した学習サンプルの特徴ベクトルに文字を識別する文字コードを割り付けて記憶装置２３に保存する。また、特徴ベクトル生成部１１１は、文字を識別する文字コードと、文字ごとに複数収集した学習サンプルの特徴ベクトルとを平均ベクトル算出部１１２と、主成分分析部１１３とに出力する。

The feature vector generation unit 111 performs the above-described processing for each of a plurality of characters. When the feature vector generation unit 111 generates the feature vector, the feature vector generation unit 111 assigns a character code for identifying the character to the feature vector of the learning sample collected for each character and stores the character code in the storage device 23. Further, the feature vector generation unit 111 outputs a character code for identifying a character and a feature vector of learning samples collected for each character to the average vector calculation unit 112 and the principal component analysis unit 113.

The average vector calculation unit 112 calculates the average vector for each character using the feature vector of the same character acquired from the feature vector generation unit 111.
The average vector is obtained by the following equation (2).

また、式（２）に従って求めた平均ベクトルを式（３）のように表す。

Moreover, the average vector calculated | required according to Formula (2) is represented like Formula (3).

平均ベクトル算出部１１２は、式（２）に示すように各特徴ベクトルの同一次元のベクトル要素同士を加算し、これを学習サンプル数Ｎで除算して平均ベクトルを求める。
例えば、特徴ベクトルの１次元のベクトル要素であれば、上述した式（１）に示すｘ^１ _１，ｘ^２ _１，ｘ^３ _１，・・・，ｘ^Ｎ _１の値を加算して、これをＮで除算した値が、式（３）に示す平均ベクトルの１次元のベクトル要素となる。
平均ベクトル算出部１１２は、文字ごとに平均ベクトルを算出し、算出した平均ベクトルを文字コードに対応付けて記憶装置２３に保存する。また、平均ベクトル算出部１１２は、文字ごとに求めた平均ベクトルを主成分分析部１１３と、グループ化部１１４とに出力する。

The average vector calculation unit 112 adds vector elements of the same dimension of each feature vector as shown in Expression (2), and divides this by the learning sample number N to obtain an average vector.
For example, if the one-dimensional vector elements of the feature vector, ^x ₁ 1 shown in formula (1) ^above, _x 2 _^1, x 3 1, · · ^·, by adding the value of ^x _{N 1,} this The value divided by N becomes a one-dimensional vector element of the average vector shown in Equation (3).
The average vector calculation unit 112 calculates an average vector for each character, and stores the calculated average vector in the storage device 23 in association with the character code. The average vector calculation unit 112 outputs the average vector obtained for each character to the principal component analysis unit 113 and the grouping unit 114.

Next, the principal component analysis unit 113 will be described. The principal component analysis unit 113 acquires feature vectors of a plurality of learning samples generated by the feature vector generation unit 111 for each character. In addition, the principal component analysis unit 113 acquires the average vector obtained for each character by the average vector calculation unit 112.
The principal component analysis unit 113 uses the acquired information to generate eigenvalues and eigenvectors representing the pixel distribution of the pattern image data for each character. First, the principal component analysis unit 113 obtains a variance-covariance matrix from the average vector and the feature vector. The variance-covariance matrix Σ is shown in Equation (4).

The covariance value σ _ij of the variance covariance matrix Σ is calculated by the following equation (5).

主成分分析部１１３は、各学習サンプルの特徴ベクトルの要素ｉと平均ベクトルの要素ｉとの差と、特徴ベクトルの要素ｊと平均ベクトルの要素ｊとの差とを掛け合わせて平均を求めたものを、要素ｉと要素ｊの共分散値として算出する。
なお、ｉとｊは、１〜Ｄの任意の自然数である。

The principal component analysis unit 113 obtains the average by multiplying the difference between the element i of the feature vector and the element i of the average vector of each learning sample by the difference between the element j of the feature vector and the element j of the average vector. Is calculated as the covariance value of element i and element j.
Note that i and j are arbitrary natural numbers from 1 to D.

Next, the principal component analysis unit 113 performs eigenvalue calculation on the obtained variance-covariance matrix Σ, and obtains eigenvalues and eigenvectors from the variance-covariance matrix Σ.
The eigenvalues are D values of λi (i = 1, 2,..., D), and λ1>λ2>.
The eigenvectors are Φi = (Φ ⁱ ₁ , Φ ⁱ ₂ ,..., Φ ⁱ _D ) (i = 1, 2,..., D).
The eigenvectors are vectors of size 1 that are orthogonal to each other.
Further, the eigenvalue λi represents a standard deviation value of a plurality of learning samples when viewed in the direction indicated by the eigenvector Φi. Therefore, the entire distribution of a plurality of learning samples can be approximately expressed by a plurality of eigenvalues and eigenvectors.

It should be noted that eigenvalues and eigenvectors can usually be obtained from the D covariance matrix Σ of D rows and D columns, respectively. However, since eigenvalues having a small value generally do not significantly affect the distribution shape of the learning sample, the distribution of the learning sample can be sufficiently expressed by only the top M pieces having the largest eigenvalue.
Therefore,
The eigenvalue is λi (i = 1, 2,..., M)
The eigenvectors are Φ ⁱ = (Φ ⁱ ₁ , Φ ⁱ ₂ ,..., Φ ⁱ _D ) (i = 1, 2,..., M) (M <D). Note that M is an arbitrary natural number smaller than D.
The principal component analysis unit 113 obtains eigenvalues and eigenvectors for each character and stores the obtained eigenvalues and eigenvectors in the storage device 23 together with the character code. The principal component analysis unit 113 outputs the obtained eigenvalues and eigenvectors to the grouped character information registration unit 116 and the ungrouped character information registration unit 117 of the dictionary generation / registration unit 115.

Next, processing of the grouping unit 114 will be described.
The grouping unit 114 acquires an average vector for each character from the average vector calculation unit 112. The grouping unit 114 performs hierarchical clustering based on the acquired average vector of each character, and divides the character into a character having a character with a similar average vector and a character having no similar character with an average vector. . The grouping unit 114 distributes characters having similar characters in the average vector to the same group.

Hierarchical clustering will be specifically described with reference to FIGS.
In FIG. 3, “large”, “dog”, “thick”, “tree”, “bird”, “eagle”, “soil”, “shi”, “±”, “sub”, “arm”, “ 13 characters of “right” and “day” and vector elements of the average vector are shown. In the example shown in FIG. 3, the dimension of the average vector is 20 dimensions (that is, 20 vector elements), but the dimension of the average vector is not limited to 20 dimensions. Further, the 13 characters shown in FIG. 3 have three similar character groups “Large, Dog, Thick, Tree”, “Bird, Spear”, and “Sat, Shi, ±”, and no similar characters. , And the other four characters.

The grouping unit 114 performs hierarchical clustering on these 13 characters. FIG. 4A shows a state in which 13 characters are arranged in a two-dimensional vector space. The average vector is actually a 20-dimensional vector, but is simplified and displayed in two dimensions.
The axis 1 that defines the vector space is an axis in which only the component of the first vector element has a value among the 20 vector elements of the average vector. An axis 2 is an axis in which only the second vector element has a value among the 20 vector elements.
The grouping unit 114 obtains the intervector distances of the 13 average vectors. The grouping unit 114 integrates, as a group, two vectors having the closest distances among the obtained distances between vectors. FIG. 4B shows the result of the first hierarchical clustering. In the example shown in FIG. 4B, the grouping unit 114 determines that the distance between the vectors of “Sat” and “±” is the shortest, and classifies these characters into the same group. The grouping unit 114 obtains an average of average vectors of characters classified into the same group as an average vector, and sets the obtained average vector as a representative vector of the group.

Next, the grouping unit 114 obtains an inter-vector distance for the average vector of 11 characters and the representative vector of the set group. The grouping unit 114 integrates, as a group, two vectors having the closest distances among the obtained distances between vectors. FIG. 5A shows the result of the second hierarchical clustering.
In the example illustrated in FIG. 5A, the grouping unit 114 determines that the inter-vector distance between the representative vector indicating the group of “Sat” and “±” and the average vector of “Shi” is the closest, “Shi” is integrated into the “Sat” and “±” groups.
The grouping unit 114 obtains an average vector that is an average of the average vectors of the characters “Sat”, “±”, and “Shi” classified into the same group, and uses the obtained average vector as a representative vector of the group. Here, the average of the previously obtained “Soil” and “±” representative vectors and the “shi” average vector may be obtained as an average vector, and the obtained average vector may be used as the group representative vector.

  Thereafter, the grouping unit 114 performs hierarchical clustering in the same procedure. FIG. 5B shows the result of the third hierarchical clustering, and FIG. 6A shows the result of the fourth hierarchical clustering. FIG. 6B shows the result of the fifth hierarchical clustering, and FIG. 7 shows the result of the sixth hierarchical clustering.

  As described above, the grouping unit 114 obtains the distance between vectors of the average vector and the average vector, or the representative vector and the average vector, and classifies the characters having the closest vector distance into the same group. Further, a threshold value is set in advance for the inter-vector distance, and the grouping unit 114 ends the classification into the same group when the calculated inter-vector distance becomes larger than the threshold value. In the example shown in FIGS. 4 to 7, classification is made into three groups and four characters having no similar characters by six times of hierarchical clustering.

The grouping unit 114 assigns identification information for identifying the group to the classified group, and stores it in the storage device 23 together with the character code classified into the group. Further, the grouping unit 114 adds the identification information indicating that the character is a character code that has not been grouped to the character code that has not been grouped, and stores it in the storage device 23.
The grouping unit 114 also outputs the character codes classified into groups and identification information indicating the groups to the grouped character information registration unit 116 and the non-grouped character information registration unit 117. The grouping unit 114 also displays a character code that has not been classified into a group and identification information indicating that the character code has not been classified into a group, and a grouped character information registration unit 116 and a non-grouped character information registration unit 117. And output.

Next, the dictionary creation / registration unit 115 will be described. The dictionary creation / registration unit 115 includes a grouped character information registration unit 116, an ungrouped character information registration unit 117, and a table storage unit 118.
The grouping character information registration unit 116 determines that there are similar characters in the grouping unit 114, and registers the eigenvalues and eigenvectors of the characters classified into groups in the index table. An example of the index table is shown in FIG. As shown in FIG. 8, the index table includes a character code for identifying a character, an index flag, an eigenvalue of the character indicated by the character code, and an eigenvector. Note that the index table and the code book table correspond to the aforementioned pattern recognition dictionary.
Since the characters classified into groups have similar characters and require a lot of information for character recognition, all of the eigenvalues and eigenvectors generated by the principal component analysis unit 113 are registered in the index table. . As described above, the eigenvector generated by the principal component analysis unit 113 is a D-dimensional vector having a plurality of vector elements. The grouped character information registration unit 116 registers all the vector elements of the D-dimensional eigenvector in the corresponding character column of the index table as eigenvectors for character recognition. Note that “0” is recorded as an index flag for characters in which all vector elements are registered in the index table.

The non-grouped character information registering unit 117 determines that there is no similar character in the grouping unit 114, and registers the eigenvalues and eigenvectors of characters not classified into groups in the index table and codebook table.
Characters that are not classified into groups do not have similar characters and do not require much information for character recognition. Therefore, all vector elements of eigenvectors generated by the principal component analysis unit 113 are accurately indexed. There is no need to register in the table. Therefore, the ungrouped character information registration unit 117 extracts an element sequence that partially matches the vector elements of the eigenvectors of other characters from the vector elements of the eigenvectors of the characters that are not grouped. The ungrouped character information registration unit 117 assigns an index number for identifying the element string to the extracted element string. Then, the ungrouped character information registration unit 117 records the assigned index number in the eigenvector recording column of the corresponding character in the index table. In addition, the ungrouped character information registration unit 117 records the index number and the corresponding element string in the codebook table (sub dictionary) in order to manage the association between the index number and the extracted element string. An example of the code book table is shown in FIG.

The processing of the ungrouped character information registration unit 117 will be specifically described.
The eigenvector generated by the principal component analysis unit 113 is a D-dimensional vector as described above, and is composed of D vector elements.
Φ ⁱ = (Φ ⁱ ₁ , Φ ⁱ ₂ ,..., Φ ⁱ _D )
Among these D elements, an element string that partially matches an element of an eigenvector of a character code that has not been grouped (hereinafter, this element string is referred to as a subvector) is extracted. For example, in the example shown in FIG. 10A, the eigenvector with the character code “0x8148” and the eigenvalue “12.5” and the eigenvector with the character code “0x7652” and the eigenvalue “25.3” are part of the subvector. There is a match.
The ungrouped character information registration unit 117 assigns an index number to the sub-vectors (21, 11, 45, 62) that are partially matched. The ungrouped character information registration unit 117 records the assigned index number in the corresponding column of the index table when the eigenvector having the character code “0x8148” and the eigenvalue “25.3” is recorded in the index table. Record. Similarly, the ungrouped character information registration unit 117 records the assigned index number in the corresponding column of the index table when the eigenvector having the character code “0x7652” and the eigenvalue “12.5” is recorded in the index table. Record. The ungrouped character information registration unit 117 registers the partially matched subvector and an index number for identifying the subvector in the codebook table (see FIG. 10B).

Further, the ungrouped character information registration unit 117 detects a sub-vector that partially matches with an error.
For example, in the example shown in FIG. 10C, the character code “0x5241”, the subvector of “83, 55, 36, 22” of the eigenvalue “10.1”, the character code “0x3698”, the eigenvalue “9.6”. The sub-vectors “83, 55, 32, and 22” in FIG.
In the ungrouped character information registration unit 117, a threshold for the distance between vectors is set. The ungrouped character information registration unit 117 assigns subvectors whose subvectors are not equal to each other but whose vector distance is equal to or smaller than a threshold to the same index number.
At this time, either “83, 55, 36, 22” or “83, 55, 32, 22” must be registered in the codebook table. The ungrouped character information registration unit 117 Select and register the vector with the highest number of appearances. For example, if the number of appearances of “83, 55, 36, 22” is 5 and the number of appearances of “83, 55, 32, 22” is 3, the codebook table includes “83, 55, 36, 22 "are registered. An example of registration in the code book table is shown in FIG. In this way, the ungrouped character information registration unit 117 replaces all vector elements of eigenvectors of characters not classified into groups with index numbers. Then, the ungrouped character information registration unit 117 records only the assigned index number in the eigenvector recording column of the corresponding character in the index table.

  In this way, the amount of information registered in the index table can be reduced for characters that are not classified into groups, that is, for characters that do not have similar characters. In the code book table shown in FIG. 9, four-dimensional (four element strings) subvectors are registered in the code book table as a unit. However, the number of sub-vector elements registered in the code book table is not limited to four. For example, sub-vectors having a large number of elements such as five or six are registered in the code book table as a unit. Also good. Also, subvectors having different numbers of vector elements can be registered in the codebook table.

  The table storage unit 118 stores the index table created by the grouped character information registration unit 116, the ungrouped character information registration unit 117, and the code book table in the storage device 23.

Next, the processing flow of the dictionary creation unit 10 will be described with reference to the flowchart shown in FIG.
First, the dictionary creation unit 10 acquires pattern image data of characters for creating a pattern recognition dictionary from the RAM 13 (step S1). A plurality of pattern image data are prepared as learning samples for each character.

  Next, the dictionary creation unit 10 creates feature vectors of pattern image data prepared as a plurality of learning samples using the feature vector generation unit 111 (step S2), and obtains an average vector thereof (step S3). The average vector is calculated for each character.

  Next, the dictionary creation unit 10 performs principal component analysis in the principal component analysis unit 113 using the calculated average vector to obtain eigenvalues and eigenvectors (step S4). The eigenvector is obtained for each eigenvalue and has a plurality of vector elements.

  Next, the dictionary creation unit 10 divides the characters for which the grouping unit 114 creates the pattern recognition dictionary into characters that have similar characters and characters that have no similar characters (step S5). The grouping unit 114 classifies similar characters into the same group, and stores the information for identifying the group and the character codes of the characters belonging to the group in the storage device 23. Further, the grouping unit 114 stores, in the storage device 23, information indicating that the character does not belong to the group together with the character code of the character when the character does not have a similar character.

Next, the dictionary creation unit 10 creates a pattern recognition dictionary by registering the eigenvalues and eigenvectors of the characters in the index table and codebook table in the dictionary generation / registration unit 115 (step S6).
At this time, the grouped character information registration unit 116 registers all the eigenvalues and eigenvectors of the characters classified into the group in the index table. Further, the ungrouped character information registration unit 117 detects a subvector that partially matches a vector element of an eigenvector of another character among vector elements of eigenvectors of characters that are not classified into groups. At this time, the ungrouped character information registration unit 117 also extracts subvectors whose vector distance is within the error range as partially matching subvectors. The ungrouped character information registration unit 117 assigns an index number to the extracted subvector. Then, the ungrouped character information registration unit 117 registers the index number in the eigenvector registration field of the corresponding character in the index table, and registers the extracted subvector and index number in the codebook table to manage the correspondence. To do. In this way, the ungrouped character information registration unit 117 replaces all vector elements of eigenvectors of characters not classified into groups with index numbers. Then, the ungrouped character information registration unit 117 records only the assigned index number in the eigenvector recording column of the corresponding character in the index table.

  The dictionary creation unit 10 causes the table storage unit 118 to store the grouped character information registration unit 116, the index table created by the ungrouped character information registration unit 117, and the code book table in the storage device 23 (step S1). S7).

As described above, in the present embodiment, in the case of a character having a similar character, a lot of information is required for character recognition. Therefore, all the eigenvalues and eigenvectors of this character are registered in the index table.
In addition, in the case of a character having no similar character, a large amount of character recognition information is not required as compared with a character having a similar character. For this reason, out of vector elements of eigenvectors, subvectors that partially match a vector element sequence of other eigenvectors are extracted while allowing an error in the distance between vectors. Then, an index number for identifying the subvector is registered in the index table instead of the extracted subvector. Thereby, the amount of data registered in the index table can be reduced. For similar characters, the eigenvalues and eigenvectors of the secondary identification method are directly registered in the pattern recognition dictionary, so that the character recognition accuracy can be improved.

For example, assume that 4,000 characters and each character has six eigenvectors (one eigenvector is 200 bytes).
If eigenvalues and eigenvectors of all characters are registered in the index table without grouping, the data amount of eigenvectors is 4,000 × 6 × 200, which is 4,800,000 bytes.
The eigenvalue is 192,000 bytes of 4,000 × 6 × 8 if one eigenvalue is a floating-point value of 8 bytes. Therefore, the total data amount of the eigenvalues and eigenvectors is 4,992,000 bytes.

  On the other hand, in this embodiment, as for eigenvectors, it is assumed that 1,000 groups are created by hierarchically clustering 4,000 × 6 vectors. Then, since the eigenvector of one character can be expressed by 6 subvectors, it becomes 6 × 2 bytes per character (one subvector is expressed by 2 bytes). Therefore, the total is 4,000 × 6 × 2 = 48,000 bytes. The data amount of the index table is 1,000 × 8 = 8,000 bytes. Furthermore, the sum of the eigenvalues and eigenvectors is 304,000 bytes, which may be about 6% of the original size.

Next, a case where pattern image recognition is performed using the created index table and codebook table will be described.
For example, assume that an image of a character pattern is input and character recognition of the input character pattern is performed. First, a feature vector of an input character pattern is generated and principal component analysis is performed to generate eigenvalues and eigenvectors of the input character pattern.
Next, character recognition is performed on the character pattern input by reading the eigenvalue and eigenvector of the character registered as a sample from the index table and codebook table. At this time, when reproducing the eigenvector of the character having “1” recorded in the index flag of the index table, the index number recorded in the corresponding character column of the index table is first extracted. Then, the subvector of the corresponding index number is extracted with reference to the code book table. The extracted subvectors are arranged in the order of the index numbers recorded in the index table to reproduce a D-dimensional eigenvector. Then, based on the reproduced eigenvector and the eigenvalue registered in the index table, it is determined whether or not the input character pattern matches the character registered as the sample.

[Modification]
Another method of hierarchical clustering will be described.
The grouping unit 114 first arranges the acquired average vector of characters in the vector space. The vector space is, for example, a space represented by 20 axes if the average vector is a 20-dimensional vector having 20 vector elements. FIG. 12 shows a two-dimensional vector space of axis 1 and axis 2 for simplification. The axis 1 is an axis in which only the component of the first element among the 20 elements of the average vector has a value. The axis 2 is an axis in which only the second element of the 20 elements has a value.

Next, the grouping unit 114 pays attention to one average vector and obtains coordinate values on each axis. Further, the grouping unit 114 obtains a range that is enlarged by a certain value on the left and right on the axis around the obtained coordinate value. Then, the characters within this range are obtained, and the obtained character group is set as a character set on this axis.
In the example shown in FIG. 12, for example, when focusing on the character “large”, on the axis 1, four characters “thick, dog, large, tree” are within a range of size 7 centered on “large”. It can be seen that exists. Similarly, on the axis 2, it can be seen that there are six characters “dog, sub, large, wood, thick, arm” within a range of size 7 centered on “large”. The grouping unit 114 calculates a logical product of the character sets obtained on each axis, and detects a character positioned in the vicinity of “large” on all axes. In the example illustrated in FIG. 12, “thick, dog, large, tree” is extracted by the grouping unit 114 by obtaining a logical product of the character set obtained from the axis 1 and the character set obtained from the axis 2. The grouping unit 114 extracts the extracted character set as a similar character set. The grouping unit 114 performs the above processing for all characters, and obtains a character set that is similar for all characters. Then, the grouping unit 114 sets a similar character set that is completely matched among the obtained similar character sets to the group. The grouping unit 114 outputs the set group information to the storage device 23, the grouped character information registration unit 116, and the non-grouped character information registration unit 117.

  The embodiment described above is a preferred embodiment of the present invention. However, the present invention is not limited to this, and various modifications can be made without departing from the scope of the present invention.

(Appendix)
(Appendix 1)
A feature vector generating means for obtaining a feature amount based on an image pattern of an input image and generating a feature vector having the feature amount as an element;
Grouping means for comparing feature vectors of each image to extract images having high similarity, and grouping the images having high similarity as a group;
Feature data generating means for generating a plurality of feature data including information on pixel distribution of an image based on the feature vector generated by the feature vector generating means;
In the case where the identification information for identifying each image and the feature data of each image are stored in the storage means to create a pattern recognition dictionary, the group is more than the data amount of the feature data of the image not included in the group. Recognizing dictionary creating means for creating the pattern recognizing dictionary so that the amount of image feature data included in the image data is increased, and storing the created pattern recognizing dictionary in the storage means;
A pattern recognition dictionary creation device having
(Appendix 2)
The recognition dictionary creating means extracts feature data common to a plurality of feature data of other images when creating the pattern recognition dictionary of images not included in the group, and the plurality of extracted features Registering identification information for identifying data in the pattern recognition dictionary instead of the plurality of extracted feature data, and extracting the plurality of extracted feature data and identification information for identifying the plurality of extracted feature data The pattern recognition dictionary creation device according to appendix 1, which creates a sub-dictionary to be associated.
(Appendix 3)
The pattern recognition dictionary creation according to appendix 1, wherein the feature data generation means is a means for performing principal component analysis based on the feature vector and generating an eigenvalue indicating a pixel distribution of the image and an eigenvector having a plurality of vector elements. apparatus.
(Appendix 4)
When the recognition dictionary creating means creates the pattern recognition dictionary for images not included in the group, a vector that partially matches a vector element of an eigenvector of another image among vector elements of the eigenvector of the image At least one of an element sequence and a vector element sequence in which the difference between vector elements is a predetermined value or less is extracted, and identification information for identifying the extracted vector element sequence is stored in the storage unit, and the extracted vector The pattern recognition dictionary creation device according to appendix 3, which creates a sub-dictionary that associates a column with the identification and identification.
(Appendix 5)
Computer
Means for obtaining a feature quantity based on an image pattern of an input image and generating a feature vector having the feature quantity as an element;
Means for comparing feature vectors of each image to extract images having high similarity, and grouping the images having high similarity as a group;
Means for generating a plurality of feature data including pixel distribution information of the image based on the feature vector generated by the feature vector generation means;
In the case where the identification information for identifying each image and the feature data of each image are stored in the storage means to create a pattern recognition dictionary, the group is more than the data amount of the feature data of the image not included in the group. The pattern recognition dictionary is created so that the amount of image feature data included in the image data is increased, and the program is made to function as means for storing the created pattern recognition dictionary in the storage means.
(Appendix 6)
The means for storing the pattern recognition dictionary in the storage means extracts feature data common to a plurality of feature data of other images when creating the pattern recognition dictionary for images not included in the group. The identification information for identifying the extracted feature data is registered in the pattern recognition dictionary instead of the extracted feature data, and the extracted feature data and the extracted feature data are registered. The program according to appendix 5, which creates a sub-dictionary that associates identification information for identifying
(Appendix 7)
The program according to claim 5, wherein the means for generating the feature data is a means for performing principal component analysis based on the feature vector and generating an eigenvalue indicating a pixel distribution of the image and an eigenvector having a plurality of vector elements.
(Appendix 8)
The means for storing the pattern recognition dictionary in the storage means, when creating the pattern recognition dictionary for images not included in the group, includes eigenvectors of other images among vector elements of eigenvectors of the image. At least one of a vector element sequence partially matching the vector element and a vector element sequence in which the difference between the vector elements is equal to or less than a predetermined value is extracted, and identification information for identifying the extracted vector element sequence is stored in the storage unit And creating a sub-dictionary that associates the extracted vector string with the identification and identification.

DESCRIPTION OF SYMBOLS 1 Pattern recognition dictionary creation apparatus 10 Dictionary creation part 21 Scanner apparatus 22 Operation part 23 Storage device 24 Display apparatus 30 External storage device 111 Feature vector generation part 112 Average vector calculation part 113 Principal component analysis part 114 Grouping part 115 Dictionary generation / registration Part 116 Grouped character information registration part 117 Ungrouped character information registration part 118 Table storage part

Claims (5)

A feature vector generating means for obtaining a feature amount based on an image pattern of an input image and generating a feature vector having the feature amount as an element;
Grouping means for comparing feature vectors of each image to extract images having high similarity, and grouping the images having high similarity as a group;
Feature data generating means for generating a plurality of feature data including information on pixel distribution of an image based on the feature vector generated by the feature vector generating means;
When creating a pattern recognition dictionary by storing the identification information for identifying each image and the feature data of each image in the storage means, without increasing the amount of feature data of the images included in the group By reducing the amount of feature data of images not included in the group, the amount of feature data of images included in the group becomes larger than the amount of feature data of images not included in the group. Creating the pattern recognition dictionary as described above, and storing the created pattern recognition dictionary in the storage means;
A pattern recognition dictionary creation device having   The recognition dictionary creating means extracts feature data common to a plurality of feature data of other images when creating the pattern recognition dictionary of images not included in the group, and the plurality of extracted features Registering identification information for identifying data in the pattern recognition dictionary instead of the plurality of extracted feature data, and extracting the plurality of extracted feature data and identification information for identifying the plurality of extracted feature data The pattern recognition dictionary creation device according to claim 1, which creates a sub-dictionary to be associated.   The pattern recognition dictionary according to claim 1, wherein the feature data generation unit is a unit that performs principal component analysis based on the feature vector and generates an eigenvalue indicating a pixel distribution of an image and an eigenvector having a plurality of vector elements. Creation device.   When the recognition dictionary creating means creates the pattern recognition dictionary for images not included in the group, a vector that partially matches a vector element of an eigenvector of another image among vector elements of the eigenvector of the image At least one of an element sequence and a vector element sequence in which the difference between vector elements is a predetermined value or less is extracted, and identification information for identifying the extracted vector element sequence is stored in the storage unit, and the extracted vector 4. The pattern recognition dictionary creation device according to claim 3, wherein a sub-dictionary that associates a column with the identification and identification is created. Computer
Means for obtaining a feature quantity based on an image pattern of an input image and generating a feature vector having the feature quantity as an element;
Means for comparing feature vectors of each image to extract images having high similarity, and grouping the images having high similarity as a group;
Means for generating a plurality of feature data including pixel distribution information of the image based on the feature vector generated by the feature vector generation means;
When creating a pattern recognition dictionary by storing the identification information for identifying each image and the feature data of each image in the storage means, without increasing the amount of feature data of the images included in the group By reducing the amount of feature data of images not included in the group, the amount of feature data of images included in the group becomes larger than the amount of feature data of images not included in the group. The pattern recognition dictionary is created as described above, and the program is made to function as means for storing the created pattern recognition dictionary in the storage means.

Images