JP2002183668A - Word-recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a word-recognizing device sharply reducing capacity of a character feature dictionary, and capable of speeding up recognizing processing. SOLUTION: The character feature dictionary is reduced in capacity, and is stored in a feature dictionary 3. The capacity is reduced by imparting an identification number of 1 to m by representing a column or row feature by m pieces of column vectors, by executing clustering processing of a feature vector in respective column or row units on a character feature. After executing the clustering processing, the column or row feature is expressed by the addition sum of the other column or row feature and a difference feature, and the character feature is dimensionally compressed, so that the dictionary can be reduced further in capacity. A word is recognized, by checking a word feature with a feature extracted from an input word by synthesizing the checking up word feature on the basis of a word list being a recognizing object. The word feature, different in the dimensional number, may be checked with an input word feature by using nonlinear extensible-contractible matching.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION In recent years, the demand for character recognition devices OCR and software OCR as document input devices has been increasing. The present invention relates to word recognition in the character recognition device. The word recognition is a method of recognizing a word such as "Tokyo" by recognizing the word itself instead of performing individual character recognition after separating the characters into individual characters. According to this method, it is possible to realize highly accurate recognition even when characters are in contact with each other, and this is one of the effective methods for recognizing a handwritten character string in a free pitch area. INDUSTRIAL APPLICABILITY The word recognition device of the present invention can be applied not only to a character recognition device for handwriting, but also to a character recognition device in a broad sense such as a print character recognition device and a character recognition device in a portable information terminal.

[0002]

2. Description of the Related Art When recognizing a handwritten word, a method of generating a word feature dictionary for matching by combining the features of each character constituting the word and matching the feature with the features of an input word is known as follows. By the time, for example, Japanese Patent Application No. 11-1137
No. 33 and Japanese Patent Application No. 11-330288 have been proposed. Japanese Patent Application No. 11-113733 discloses that a word feature dictionary is generated based on the characteristics of individual characters without individually recognizing the input word images, and the input word images are combined. Word recognition can be performed with high accuracy using a small-capacity individual character image dictionary. Also, the one disclosed in Japanese Patent Application No. 11-330288 makes it possible to cope with a change in the shape of an input word image by synthesizing a plurality of word features for one word to generate a word dictionary. Things.

[0003]

When a character feature dictionary for synthesizing a word feature is provided, the conventional method disclosed in Japanese Patent Application No. 11-330288 discloses a character image in which the position and width of each character are changed. We extracted features and kept them all. For example, as shown in FIG.
6, 2/6,..., 6/6 were extracted (hereinafter referred to as n / m features), and all of them were retained. In this case, the number of features per character is 21.
The characteristics of individual characters include, for example, the weighted direction exponential histogram feature (Tsuruoka et al., “Handwritten Kanji / Hiragana Recognition by Weighted Direction Exponential Histogram Method,” IEICE Transactions D Vol.J70-D No.7, pp.1390- 1397 July 1987)
Used. The weighted direction index histogram feature divides a character image into small regions and uses the direction index histogram of each small region as a feature vector. For example, as shown in FIG. Feature values are extracted for eight directions obtained by equally dividing the angle into eight. Each mesh has an eight-dimensional feature amount. For example, 3/7 of the character “East”
The features are as shown in FIG. Then, when synthesizing word features, the word features are synthesized so that the sum of the n / m fractions of the individual character features becomes one. For example, in the case of a two-letter word, 3/7 features + 4/7 features,
Word features are synthesized by synthesizing 7 features + 5/7 characteristics and the like.
For example, when synthesizing the word feature “Tokyo”, as shown in FIG.
Processing such as combining the seven features and combining "Tokyo" is performed. However, since it is necessary to have character characteristics in which the positions and widths of all the approximately 4000 character categories are changed, several hundred Mbytes are required in terms of capacity, and this has been a major problem in practical use. The present invention
In view of the above circumstances, an object of the present invention is to significantly reduce the capacity of the character feature dictionary,
Another object of the present invention is to provide a word recognition device capable of speeding up a recognition process.

[0004]

In order to solve the above-mentioned problems, in the present invention, the capacity of a feature dictionary for synthesizing word features is reduced, and the dictionary capacity is set to a practical level.
In addition, the collation method of the synthesized word feature and the input word feature was improved, and the variation of the character shape was absorbed in the collation portion, thereby further reducing the dictionary capacity. FIG. 1 shows an outline of the present invention. In FIG. 1, reference numeral 1 denotes a character feature dictionary, which stores a feature vector extracted from an input character image. 2 is a character feature dictionary minimizing means, 3 is a feature dictionary, and the feature dictionary 3 has column features (in the case of horizontal writing) or row features reduced by the capacity minimizing means 2 during learning. (In the case of vertical writing) is stored. The capacity reduction of the character feature by the capacity reduction means 2 is performed as follows. (1) With respect to the character features stored in the character feature dictionary 1, clustering of feature vectors is performed on a column basis or on a row basis, and similar features are collectively represented by m columns or row vectors. 1 to each of
An identification number of m is given (hereinafter, giving an identification number up to m is referred to as coding). In addition, coding may be performed by clustering not only in one column or one row but also in a plurality of columns. Furthermore, if the coding of the column feature or the row feature is performed not in the unit of the column feature or the row feature but in the unit of the mesh, the coding based on more accurate feature approximation can be performed. (2) For m clustered feature vectors, a certain column feature can be represented by the sum of other column features, or a row feature can be represented by the sum of other row features. It checks whether there is a combination, and if there is a possible combination, stores the identification number and the combination coefficient of the column feature or the row feature in the dictionary. Also, a combination that allows a certain column feature to be represented by the sum of another column feature and a difference feature,
Or, check if there is a combination that can represent a certain row feature as the sum of another row feature and the difference feature, and if there is a possible combination, store the identification number of column feature or row feature and the combination coefficient in the dictionary I do. (3) Perform dimensional compression by performing feature conversion on character features in advance,
The feature-converted features are coded by performing a clustering process. In addition, by having index information for the feature vector in the feature dictionary 3, high-speed dictionary access is possible. In addition, by arranging column features or row features in descending order of frequency of use, access to index information can be performed at high speed. The word recognition is performed as follows using the feature dictionary 3 reduced in size as described above. The input word is normalized by the normalizing means 4, and the feature is extracted by the feature extracting means 5. On the other hand, a word feature for comparison is synthesized by the word feature synthesizing means 6 from the column or row features stored in the feature dictionary 3 based on a predetermined word list to be recognized. Next, the matching unit 7 checks the feature extracted from the input word and the synthesized word feature to perform word recognition. At the time of matching by the matching means 7, the word features having different dimensions and the input word features are matched by using non-linear expansion / contraction matching, so that variation in the character shape can be absorbed not by the feature vector in the dictionary but by the matching portion. Thus, it is possible to further reduce the dictionary capacity. That is, in the conventional word recognition, since the matching method itself is vulnerable to the change in the character shape of the input character, a plurality of features (e.g.
/ 6 features, 4/6 features, etc.). On the other hand, the above-mentioned nonlinear expansion / contraction matching has an effect of absorbing the variation of the character shape in the method itself, so that the number of features in the feature dictionary can be expected to be reduced, and the dictionary capacity can be reduced.

[0005]

Embodiments of the present invention will be described below. The present invention is realized by a normal computer system including a processing device, a main storage device, an external storage device, a keyboard, an input device such as a scanner for reading an image, a display device, an output device such as a printer, a communication interface, and the like. A program, data, and the like for performing the processing of the present invention are stored in an external storage device or the like, and at the time of execution, the program or the like is read into the main storage device, and the processing of the present invention is executed.

Although the present invention is effective irrespective of vertical writing / horizontal writing, the following description is directed to horizontal writing. When applied to vertical writing, "column vector" described below
May be referred to as a “row vector”. Although character features are classified into several types, the present invention targets features that can be divided into columns. Further, in the case of a feature that can be divided into mesh units, features in the mesh are arranged vertically and treated as column features. More specifically, the feature of each column includes an nth-order peripheral feature, a projection feature, and the like, and the mesh type feature includes the above-described weighted direction index histogram feature, a direction line element feature, a mesh feature, and the like. Hereinafter, the weighted direction index histogram feature will be described as an example. Note that the number of mesh divisions, the number of directions, and the like used in the following description are not necessarily limited to these numerical values.

(1) Embodiment 1 As described above, in the final stage of the feature extraction process, the weighted direction exponent histogram features are divided into 8 meshes, for example, in a 7 × 7 mesh segmented from the character-normalized image. It has the characteristic of the direction component. That is, it is a 7 × 7 × 8-dimensional feature. Here, the eight directions refer to directions obtained by equally dividing 360 ° into 45 ° units as shown in FIGS. In this embodiment, in order to reduce the size of the character feature dictionary, clustering of feature vectors is performed in column units. FIG. 2 shows a functional configuration of the present embodiment. In FIG. 2, reference numeral 11 denotes a character feature dictionary, which stores a feature vector extracted from an input character image during learning. Numeral 12 denotes a clustering means according to the present embodiment. In order to reduce the size of the character feature dictionary 11, the character feature stored in the character feature dictionary 11 is weighted in the weighting direction histogram as shown in FIG. The feature vectors are clustered for each feature column. That is, a feature vector (7 × 8 = 56 dimensions) in a 7 × 1 mesh is used as one unit, and similar features are collectively represented by m column vectors. Then, an identification number from 1 to m is assigned to each of the representative vectors. FIG. 3B shows a column vector number notation example of the weighted direction histogram feature. As shown in the figure, m identification numbers coded by performing clustering for each column are assigned to each column of the character feature vector. In this example, the identification numbers (32230, 13118,.
451).

In this way, according to the conventional method, in the conventional method, ((the number of character categories) × (the number of features per character) × (the number of horizontal column features)) minutes of column features (for example, the number of character categories = 400)
0, the number of features per character = 21, and the number of horizontal row features = 7, 4000 × 21 × 7 column features) were required. Can be completed. As a clustering method, a general clustering method such as hierarchical clustering, k-means, or LVQ can be used.
The m column vectors clustered as described above and assigned identification numbers are stored in the feature dictionary 13.

FIG. 4 shows an example of the configuration of the feature dictionary 13.
As shown in the figure, the reduced feature dictionary 13 has each identification number as index information and position information in the dictionary, and the position information is a storage position of m feature vectors corresponding to each identification number. Is shown. When arranging the identification numbers, a category (ex. Address, name) to be recognized in advance is used.
By examining the appearance frequency of the character features required for generating the word features for, and arranging them in descending order of frequency, access to the index information can be speeded up. Word recognition is performed as follows using the feature dictionary 13 reduced in size as described above.
First, the input word is normalized by the normalizing unit 14, and the feature is extracted by the feature extracting unit 15. On the other hand, based on a predetermined word list to be recognized (for example, a list of prefecture names when performing word recognition for prefecture names),
A word feature for matching is synthesized by the word feature synthesizing means 16 from the column features stored in the feature dictionary 13. Then
A feature extracted from the input word by the matching unit 17;
The synthesized word features are collated to perform word recognition. The collation between the input word feature and the synthesized word feature is performed using the Euclidean distance or the like. As described above, in the present embodiment, the feature vectors of the character feature dictionary 11 are clustered in units of columns, and similar features are collectively represented by m column vectors and coded. Can be greatly reduced, and the dictionary capacity can be reduced to a practical level. Also, since index information is provided for the feature vector in the feature dictionary, dictionary access can be performed at high speed.

In the above embodiment, as shown in FIG. 5A, when clustering column features, clustering of feature vectors is performed in units of columns having a width of 1. However, not only column features having a width of 1 but also widths of 2 are clustered. It is also possible to collectively cluster column features such as width and width 3. That is, FIG.
As shown in (b), when the width is n, the height is 7 × n × 8.
Clustering is performed using the dimensional features as one unit. Since the unit of the column feature is larger than that in the case of one horizontal row, word synthesis can be speeded up.

Further, in the above embodiment, when a word feature is synthesized from a character feature, the number of dimensions of the synthesized word feature and
The dimensions of the input word features were the same. That is, in both cases, the collation is performed using the Euclidean distance, etc., based on the dimension of the dimension of 7 × 7 × 8 dimension, but as shown in FIG. You can also That is, by performing the matching using the non-linear expansion / contraction matching, the matching can be performed even when the two have different dimensions. As an example of nonlinear expansion / contraction matching, DP matching can be used (DP
For the matching, see, for example, “Pattern Recognition” by Noboru Funakubo, published by Kyoritsu Shuppansha, pp. 62-67. As a result, as shown in FIG.
× horizontal 8 × 8 direction dimension, the characteristics of the input word is vertical 7 × horizontal 7 × 8
Even if it is the direction dimension or the like, both can be collated. By performing the above-described collation, the fluctuation of the character shape can be absorbed not by the feature vector in the dictionary but by the collation part, so that the dictionary capacity can be further reduced.

Further, in the above description, clustering in units of column vectors has been described. However, clustering processing can be performed in units of meshes by looking at column vectors in more detail in units of meshes. That is, as shown in FIG. 7, clustering processing is performed in units of 1 × 1 in-mesh features (eight dimensions), and coding of features in one mesh is performed. Then, since the column vector is represented by a 7 × 1 mesh, it is represented by seven identification numbers. In the example of FIG. 7, an identification number (432, 123,..., 351) is assigned to each mesh (t).
Represents transposition). As described above, by performing clustering and coding for each mesh, it becomes possible to assign identification numbers based on more accurate feature approximation. Further, as described above, it is also possible to perform clustering on the in-mesh features to code each mesh, and perform a clustering process for each coded mesh on a column basis. That is, it is also possible to perform a clustering process on the in-mesh features and code it as shown in FIG. 7, perform the clustering process on each coded column as described above, and assign an identification number to each column.

(2) Second Embodiment Next, a description will be given of a second embodiment of the present invention in which the capacity of a feature dictionary is reduced by using a combination coefficient after clustering column features. Assuming that the number of coded column vectors is m, the p-th column vector is f _p , and the synthesis coefficient is k _i , whether there is a combination of the synthesis coefficient k and the column vector represented by the following equation (1): Are checked, and if there is a possible combination, the identification number of the column vector and the combination coefficient are recorded.

[0014]

(Equation 1)

[0015] With this, it is only necessary to have the synthesis coefficient k in the feature dictionary instead of the column vector, so that the dictionary capacity can be reduced. FIG. 8 shows a functional configuration of the present embodiment.
In FIG. 8, reference numeral 11 denotes a character feature dictionary, which stores a feature vector extracted from an input character image during learning as described above. Reference numeral 21 denotes a capacity reducing unit according to the present embodiment, which comprises the clustering unit 12 described above and a combining coefficient calculating unit 22. As described above, the clustering unit 12 performs clustering of the feature vectors for each column of the weighted direction histogram features, and assigns an identification number from 1 to m to each of the representative vectors. The synthesis coefficient calculation unit 22 calculates the synthesis coefficient k expressed by the above equation (1).
It is checked whether there is a combination of column vector and column vector. If there is a possible combination, the identification number of the column vector and the combination coefficient are recorded. The composite coefficient or column vector obtained by the composite coefficient calculation unit 22 as described above is stored in the feature dictionary 13.

The word recognition processing in this embodiment can be performed in the same manner as in the first embodiment. That is, the input word is normalized by the normalizing means 14, and the feature extracting means 1
5 to perform feature extraction. On the other hand, based on a predetermined list of words to be recognized, the word feature synthesizing means 16 synthesizes word features for collation from the column features stored in the feature dictionary 13. Next, the matching unit 17 compares the feature extracted from the input word with the combined word feature to perform word recognition. The above-mentioned matching is performed not only by matching the number of dimensions of the synthesized word feature and the number of dimensions of the input word feature, but also by matching column features having different numbers of dimensions as described in the first embodiment. Is also good. In addition, the clustering process is performed, as described in the first embodiment, for the width 1
Column features such as width 2 and width 3 may be collectively clustered, or clustering may be performed in mesh units.

In the above description, the case where a certain column vector is represented by the sum of other column vectors has been described. However, not only is a certain column vector represented by the sum of other column vectors, but also the difference between the other column vector and the difference vector. It may be represented by a sum. That is, assuming that the difference vector is g _j and the coefficient is l _j , it is checked whether there is a combination of the combination coefficient k, l, the column vector and the difference vector that can be expressed by the following equation (2). Record them, if any. The difference vector g _j is a difference between any of the feature vectors.

[0018]

(Equation 2)

According to the above, the column feature can be represented by the sum of the other column features or the sum of the sum of the other column features and the difference feature. Can be expressed more frequently than in the case of expressing the sum of, and the capacity of the feature dictionary can be further reduced. FIG. 9 shows a configuration example of the feature dictionary 13 generated according to the present embodiment. As shown in the figure, the reduced feature dictionary 23 has each identification number as index information and position information in the dictionary, and the position information stores m feature vectors corresponding to each identification number. Indicates the position or the storage position of the combined coefficients k and l.
In addition, when the column feature is represented by the sum of the other column features as described above without using the difference feature, the synthesis coefficient 1 may be set to 0. To read a feature vector from the feature dictionary, a feature vector is obtained from position information corresponding to the identification number.
Alternatively, the storage position of the combined coefficient k, l is obtained, and when the feature vector is stored in the storage position, the feature vector is read as it is, and the combined coefficient k, l is stored in the storage position. In this case, a feature vector is calculated from the combined coefficients k and l according to the above-described equation (1) or (2). In the present embodiment, as in the first embodiment, when arranging the identification numbers, the appearance frequency of the character feature required for generating the word feature for the category to be recognized is checked in advance. By arranging them in descending order of frequency, access to index information can be speeded up.

(3) Embodiment 3 In the weighted direction index histogram feature, in order to eliminate the redundancy of the information included in the feature, the extracted 7 × 7 × 8-dimensional original feature is subjected to canonical discrimination analysis or the like. Dimensional compression is performed. As a result, the number of feature dimensions is reduced from, for example, 392 dimensions to about 100 dimensions. In this way, by having features that have been subjected to feature conversion such as principal component analysis and canonical discriminant analysis in advance, and by clustering and coding these features, the dictionary capacity can be reduced. FIG. 10 shows a functional configuration of the present embodiment. In FIG. 10, reference numeral 11 denotes a character feature dictionary, which stores a feature vector extracted from an input character image during learning as described above. Numeral 31 denotes a capacity reducing means according to the present embodiment, which comprises a dimensional compressing means 32 for performing the dimensional compression and the clustering means 12 described above. The dimensional compression unit 32 performs feature conversion such as canonical discrimination analysis as described above, and performs dimensional compression. When performing the feature conversion of the original feature, the feature conversion is performed not on the column vector but on the original feature itself. Here, assuming that the original feature is f, the feature-transformed feature is w, and the feature conversion matrix obtained by canonical discriminant analysis or the like is A, the feature conversion matrix A is obtained by the following equation (3). _{A * f i = w i ...} (3) clustering means 12 will feature transformation as above, w _i was dimensional compression (i = 0, ..., M , M: Hara number features) performs clustering processing on As described above, m (m ≦ M) feature vectors are represented, and an identification number from 1 to m is assigned to each of these representative vectors. Each coded column vector obtained as described above is stored in the feature dictionary 13.

The word recognition process in this embodiment is performed as follows. The input word is normalized by the normalizing unit 14, the feature is extracted by the feature extracting unit 15, and the extracted feature vector is subjected to feature conversion (dimensional compression) by the above (3). On the other hand, based on a predetermined list of words to be recognized, the word feature synthesizing means 16 synthesizes word features for collation from the column features stored in the feature dictionary 13. Next, the matching unit 17 compares the feature extracted from the input word and subjected to feature conversion (dimensional compression) with the synthesized word feature to perform word recognition. In the present embodiment, since the dimensionally compressed feature amount is subjected to the clustering process, the dictionary capacity can be further reduced. In the embodiment shown in FIG. 10, after performing the cluster rig processing, as described in the second embodiment, a combination coefficient may be obtained and the combination coefficient may be stored in the dictionary. As a result, it is possible to further reduce the capacity of the dictionary.

(Supplementary Note 1) A word recognition device for recognizing a word image, comprising a capacity reducing means for reducing the capacity of a character feature dictionary used for word feature synthesis, and a feature reduced in capacity by the capacity reducing means. Synthesizing means for synthesizing word features for matching based on a word list to be recognized from a column or row feature of a dictionary, feature extracting means for extracting features of an input word, and features extracted by the feature extracting means. A word recognition device comprising: a matching unit that matches a feature of an input word with the synthesized word feature. (Supplementary Note 2) The capacity reducing unit clusters column features or row features having similar features for each feature of a mesh or for each feature of a column or a row divided into meshes. Alternatively, the word recognition device according to appendix 1, further comprising: means for assigning an identification number for identification to the line feature; and means for holding the feature with the identification number in a dictionary. (Appendix 3) When performing clustering, one row or one
The word recognition device according to claim 2, wherein clustering is performed not only on the rows but also on a plurality of columns or a plurality of rows. (Supplementary Note 4) The word recognition device according to Supplementary Note 2, wherein a certain column feature or row feature is described by a coefficient sum of a plurality of other column features or row features among the clustered column features or row features. (Supplementary note 5) The word recognition device according to supplementary note 2, wherein a certain column feature or row feature is described as a sum of coefficients of another column feature or row feature and a difference feature among the clustered column features or row features. . (Supplementary note 6) The word recognition device according to Supplementary note 2, wherein the feature of each mesh in the column or the row is encoded before clustering the feature of the column or the row. (Supplementary Note 7) The word recognition device according to Supplementary Note 2, wherein the capacity reducing unit performs clustering using character features that have been subjected to feature conversion and dimension compression in advance when reducing the size of the character feature dictionary. (Supplementary Note 8) When constructing a feature dictionary with a reduced capacity, the identification number of each column feature or each row feature and the position in the dictionary are held as index information, and the partial features are arranged after the index information. 2. The word recognition device according to claim 1, wherein the word recognition device comprises a dictionary. (Supplementary Note 9) The word of Supplementary Note 8 in which, when arranging the column feature or the row feature, a column feature or a row feature that is frequently used is checked in advance, and the column feature or the row feature is arranged in descending order of the use frequency. Recognition device. (Supplementary Note 10) The word recognition device according to Supplementary Note 1, wherein the matching unit matches the feature of the input word and the synthesized word feature by non-linear expansion matching. (Supplementary Note 11) A recording medium recording a word recognition program for recognizing a word image, wherein the program includes:
The capacity of the character feature dictionary used for word feature synthesis is reduced, and a word feature for matching is synthesized from the column or row features of the reduced feature dictionary based on a word list to be recognized. A recording medium storing a word recognition program for extracting a feature and performing word recognition by comparing the feature of the extracted input word with the synthesized word feature. (Supplementary Note 12) A word recognition program for recognizing a word image, the program including a process of reducing the size of a character feature dictionary used for word feature synthesis, and a process of reducing the size of a column or a row feature of the reduced feature dictionary. Processing for synthesizing word features for matching based on a list of words to be recognized, extracting features of input words, and comparing the extracted features of the input words with the synthesized word features. A word recognition program characterized by causing a computer to execute word recognition processing.

[0023]

As described above, the following effects can be obtained in the present invention. (1) Since each character feature is clustered and coded on a column feature or row feature basis, the capacity of the character feature dictionary can be significantly reduced, and the dictionary capacity can be reduced to a practical level. If clustering and coding are performed not only for one column but also for a plurality of columns, it is possible to speed up the synthesis of word features. Furthermore, if coding of column features or row features is performed not in units of column features but in units of meshes, coding based on more accurate feature approximation becomes possible. (2) By comparing the word features having different dimensions with the input word features using nonlinear expansion / contraction matching, it is possible to absorb variation in the character shape not in the feature vector in the dictionary but in the matching portion. There is no need to register feature vectors to absorb fluctuations, and it is possible to further reduce dictionary capacity. (3) After clustering in units of column features or row features, a certain column feature can be represented by a sum of other column features, or a row feature can be represented by a sum of other row features. It is checked whether there is any combination, and if there is a possible combination, the identification number of the column feature or the row feature and the combination coefficient are stored in the dictionary, so that the dictionary capacity can be further reduced. It also checks whether there is a combination that can represent a certain column feature by the sum of another column feature and a difference feature, or a combination that can represent a certain row feature by the sum of another row feature and a difference feature. If there is a possible combination, storing the identification number of the column feature or the row feature and the combination coefficient in the dictionary increases the frequency of expression, as compared with the case of simply expressing the sum of the other column features. The capacity can be increased. (4) If a character feature is subjected to feature conversion in advance to perform dimensional compression, and the feature-converted feature is coded by performing clustering processing, it is not necessary to perform feature conversion after word feature synthesis. Processing can be speeded up. At the same time, the dictionary capacity can be reduced. (5) By having index information for a feature vector in a dictionary, high-speed dictionary access is possible.
Further, by arranging the column features or the row features in descending order of use frequency, it is possible to access the index information at high speed.

[Brief description of the drawings]

FIG. 1 is a diagram showing an outline of the present invention.

FIG. 2 is a diagram illustrating a functional configuration of a first embodiment of the present invention.

FIG. 3 is a diagram illustrating clustering of feature vectors in units of columns.

FIG. 4 is a diagram illustrating a configuration example of a feature dictionary according to the first embodiment;

FIG. 5 is a diagram illustrating clustering of a plurality of column features.

FIG. 6 is a diagram illustrating matching when the number of dimensions of a synthesized word feature is different from the number of dimensions of an input word feature.

FIG. 7 is a diagram illustrating a case in which clustering is performed for each feature in a mesh.

FIG. 8 is a diagram showing a functional configuration of a second embodiment of the present invention.

FIG. 9 is a diagram illustrating a configuration example of a feature dictionary generated according to the second embodiment;

FIG. 10 is a diagram showing a third embodiment of the present invention.

FIG. 11 is a diagram illustrating an example of a reduced character image (large characters) for character characteristics.

FIG. 12 is a diagram illustrating an example of a weighted direction index histogram feature.

FIG. 13 is a diagram illustrating an example of combining word features.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 character feature dictionary 2 miniaturization means 3 feature dictionary 4 normalization means 5 feature extraction means 6 word feature synthesis means 7 collation means 11 character feature dictionary 12 clustering means 13 feature dictionary 14 normalization means 15 feature extraction means 16 word feature synthesis Means 17 Matching means 22 Synthesis coefficient calculating means 32 dimensional compression means

Claims (6)

[Claims] 1. A word recognition apparatus for recognizing a word image, comprising: a capacity reducing means for reducing a capacity of a character feature dictionary used for word feature synthesis; and a feature dictionary of a feature dictionary reduced by the capacity reducing means. Synthesizing means for synthesizing word features for matching based on a word list to be recognized from a column or row feature, feature extracting means for extracting features of an input word, and input words extracted by the feature extracting means The features of
A word recognizing device comprising: a matching unit that matches the synthesized word feature. 2. The capacity reducing means clusters column features or row features having similar features for each feature of a mesh or for each feature of a column or a row divided into meshes. 2. The word recognition device according to claim 1, further comprising: means for assigning an identification number for identification to a feature or a line feature; and means for holding the feature with the identification number in a dictionary. 3. The word recognition according to claim 2, wherein the capacity reducing means performs clustering using character features which have been subjected to feature conversion and dimension compression in advance when reducing the capacity of the character feature dictionary. apparatus. 4. The word recognition apparatus according to claim 1, wherein the matching means matches the feature of the input word with the synthesized word feature by nonlinear expansion / contraction matching. 5. A recording medium on which a word recognition program for recognizing a word image is recorded, wherein the program reduces a size of a character feature dictionary used for word feature synthesis, and a sequence of the reduced feature dictionary or From the line features,
A word feature for matching is synthesized based on the word list to be recognized, a feature of the input word is extracted, and the feature of the extracted input word is compared with the synthesized word feature to obtain a word. A recording medium storing a word recognition program characterized by performing recognition. 6. A word recognition program for recognizing a word image, the program comprising: a process for reducing the size of a character feature dictionary used for word feature synthesis; and a column or row feature of the reduced feature dictionary. From
A process of synthesizing a word feature for matching based on a word list to be recognized; extracting a feature of an input word; and matching the extracted feature of the input word with the synthesized word feature. A word recognition program for causing a computer to execute a process of performing word recognition on a computer.