JP2001338264A - Device and method for preparing character recognition pattern dictionary, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognition pattern dictionary preparing device of which the accuracy of recognition will not be lowered, even when a feature vector to be used for selecting a feature amount is not suitable for character recognition. SOLUTION: This device is provided with an input part 110 for inputting a character image, feature amount extracting part 120 for extracting the feature amount from the character image, which is inputted by the input part 110, by a discriminating analysis method or main component analysis method, feature amount selection feature vector preparing part 130, for preparing a feature amount selection feature vector for reducing the number of dimensions of the feature amount extracted by the feature amount extracting part 120 through adjustment using the appearance frequency information of a character a feature amount selecting part 140 for reducing the number of dimensions of the feature amount extracted by the feature amount extracting part 120, using the feature amount selection feature vector prepared by the feature amount selection feature vector preparing part 130, and dictionary preparing part 150 for preparing a dictionary from the feature amount selected by the feature amount selecting part 140.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition pattern dictionary creation device for creating a pattern recognition dictionary used for character recognition / optical character recognition device / image processing, a method thereof, and a recording medium.

[0002]

2. Description of the Related Art Generally, in order to improve the accuracy of character recognition,
It is necessary to extract feature quantities with high separability. For such a purpose, there is a method of increasing the number of dimensions of the feature amount by subdividing the area division at the time of feature extraction. However, when the number of dimensions is increased, there is a problem that a calculation amount and a storage capacity required for identification increase. In order to solve this problem, it is necessary to reduce the number of dimensions of the feature amount without lowering the recognition rate, and this can be achieved while leaving effective features by a statistical method such as discriminant analysis.

[0003]

However, if a feature vector used for selecting a feature amount is not appropriate for character recognition, recognition accuracy may be reduced. An object of the present invention is to provide a character recognition pattern dictionary creation device, a character recognition pattern dictionary creation method, and a recording medium for creating a dictionary for pattern recognition that does not reduce the recognition accuracy.

[0004]

According to a first aspect of the present invention, there is provided a character recognition pattern dictionary creating apparatus, comprising: an input unit for inputting a character image; and a discriminant analysis method or principal component analysis based on the character image input by the input unit. A feature value extraction unit for extracting feature values by a method, and a feature value selection feature vector for reducing the number of dimensions of the feature value extracted by the feature value extraction unit by adjusting using character appearance frequency information. A feature quantity selection feature vector creation unit to be created, and a feature quantity that reduces the number of dimensions of the feature quantity extracted by the feature quantity extraction unit using the feature quantity selection feature vector created by the feature quantity selection feature vector creation unit. A feature is provided that includes a selection unit and a dictionary creation unit that creates a dictionary from the feature amounts selected by the feature amount selection unit. According to a second aspect of the present invention, there is provided the character recognition pattern dictionary creating apparatus according to the first aspect, wherein the feature amount selection feature vector creating unit assigns a character having a high appearance frequency to be recognized. A feature amount selection feature vector is created in which a high prior probability is adjusted for a character having a low appearance frequency to be recognized using a low prior probability to reduce the number of dimensions of the feature amount extracted by the feature amount extraction unit. It is characterized by the following. Further, the character recognition pattern dictionary creating apparatus according to claim 3 of the present invention is characterized by claim 1 or claim 2.
In the character recognition pattern dictionary creation device described above, the feature amount selection feature vector creation unit creates a feature amount selection feature vector after performing clustering of learning data for a character to be recognized. According to a fourth aspect of the present invention, there is provided the character recognition pattern dictionary creating apparatus according to the third aspect, wherein the learning data is clustered by a distance of a cluster member farthest from the center of the generated cluster. Is not exceeded by a combination clustering method using the clustering termination condition as a feature value selection feature vector. According to a fifth aspect of the present invention, there is provided a character recognition pattern dictionary creating method according to the first aspect, wherein a character image is input, a feature amount is extracted from the input character image, and the feature amount is adjusted using appearance frequency information to be recognized. A feature quantity selection feature vector for reducing the number of dimensions of the extracted feature quantity, and reducing the dimension number of the feature quantity extracted using the created feature quantity selection feature vector to select The feature is that a dictionary is created from the set feature amounts.

In a character recognition pattern dictionary creating method according to a sixth aspect of the present invention, a character image is input, and a characteristic amount is extracted from the input character image by a discriminant analysis method or a principal component analysis method and recognized. After clustering the learning data for the power character, it was extracted by adjusting using a high prior probability for the character with high appearance frequency that is the target of character recognition, and using a low prior probability for the character with low appearance frequency that is the target of character recognition. A feature amount selection feature vector for reducing the dimension number of the feature amount is created, the dimension number of the feature amount extracted using the created feature amount selection feature vector is reduced, and the feature amount is selected from the selected feature amount. It is characterized in that a dictionary is created. A recording medium according to claim 7 of the present invention is a computer-readable recording medium for causing a computer to function as a character recognition pattern dictionary creating device that creates a pattern dictionary for character recognition. An input unit for inputting an image, a feature amount extracting unit for extracting a feature amount from a character image input by the input unit by a discriminant analysis method or a principal component analysis method, and adjusting using character appearance frequency information, A feature amount selection feature vector creation unit that creates a feature amount selection feature vector for reducing the number of dimensions of the feature amount extracted by the feature amount extraction unit; and a feature amount created by the feature amount selection feature vector creation unit. A feature amount selecting unit for reducing the number of dimensions of the feature amount extracted by the feature amount extracting unit using the selected feature vector, and a feature amount selected by the feature amount selecting unit. The program for functioning as a character recognition pattern dictionary creating apparatus and a dictionary creation unit performing dictionary creation was recorded.

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The configuration and operation principle of the present invention will be described below with reference to the drawings. FIG. 1 is a functional block diagram of a character recognition pattern dictionary creation device showing an example of an embodiment of the present invention. The input unit 110, the feature amount extraction unit 12
0, a feature amount selection feature vector creation unit 130, a feature amount selection unit 140, a dictionary creation unit 150, and a recognition dictionary 200. The input unit 110 reads an image on a document sheet or the like with a scanner. This image reading may be performed not only by a scanner but also by reading an already read image from a file or the like. The read character image is sent to the feature amount extraction unit 120 character by character. The feature amount extraction unit 120 extracts a feature vector (n-order vector), which is a feature amount used for recognition, from a character image of one character sent from the input unit 110. This feature vector is created using a directional code histogram or the like in Document 1 (Japanese Patent Laid-Open No. 1-250184). The feature amount selection feature vector creation unit 130 creates a feature amount selection feature vector for selecting a feature amount by using an appropriate feature vector creation method with reference to the feature vector of the feature amount. The feature amount selection unit 140 reduces the feature amount extracted by the feature amount extraction unit 120 using the feature selection feature vector created by the feature amount selection feature vector creation unit 130. The dictionary creation unit 150 creates the recognition dictionary 200 from the feature amount whose dimensions have been reduced. The recognition dictionary 200 stores a character code for each registered character.
Information such as a standard pattern feature vector (n-dimensional) and a feature amount selection feature vector (n × k matrix, k <n) is held. For k, a numerical value smaller than n is selected by experiment in order to reduce the dictionary size.

The flow of processing in the embodiment of the present invention will be described with reference to the flowchart of FIG. The image of the learning pattern drawn on the manuscript paper or the like is read by a scanner. Alternatively, the already read image is read from a file or the like (step S100). The read image is displayed on a display, and a character portion is selected from the read image from a pointing device such as a mouse. The character portion may be determined using an area identification technique. Normalization of the size and noise removal are performed on the image of the selected character portion. A direction code histogram is extracted as a feature amount of the input character image (see Reference 1) (step S110). Here, the extracted feature amount is represented by an n-dimensional feature amount X. The characteristics of the extracted features are determined, and a feature selection feature vector is created by applying an appropriate feature selection method (step S120) (details will be described later). Here, the created feature amount selection feature vector is assumed to be a matrix A of n × k (k <n). The dimension of the extracted feature X is reduced using the created feature selection feature vector A (step S1).
30). This can be obtained by Y = A × X, where Y (k dimension) is a feature amount whose dimension is reduced by the feature amount selection. The feature amount Y in which the dimension of the feature amount is reduced is registered as the feature amount of the corresponding character in the recognition dictionary 200 (step S140).

<Embodiment of Feature Amount Selection Feature Vector Creation Unit 130> FIGS. 3 and 4 are diagrams for explaining an example of an embodiment of the feature amount selection feature vector creation unit 130. FIG. Referring to FIG. 3, two categories “wo” and “ramie” having the same number of learning patterns in a two-dimensional feature space.
Slightly intersect in the feature space. Here, O1 and O2 are the average feature amounts of the category “を” and the category “ramie”. Now, let the total number of patterns in the learning pattern be N,
Assuming that the number of learning patterns of a category is Nl, the prior probability is estimated from P (wl) = Nl / N. From this equation, if the number of learning patterns is the same, the prior probabilities are considered to be the same. Reference 2 (Wakabayashi, Tsuruoka, Kimura, Miyake: "Feature Selection in Handwritten Digit Recognition", IEICE Transactions D-II Vol.J78-D-II No.11 pp.1627-1638,
When the feature quantity selection feature vector is calculated using the canonical discriminant analysis method (November 1995), Z1 in FIG. 3 is obtained. These two categories "" and the center O1 and O2 of the category "ram" are projected onto Z1 to obtain an intermediate straight line G1 of these two straight lines. The point on the O1 side of the straight line G1 is determined to belong to the category "", and the point on the O2 side of the straight line G1 is determined to belong to the category "ramie".

As shown in FIG. 3, a straight line G1 passes through the middle of the two categories, and the number of misidentified characters in each of the two categories is the same. Therefore, the appearance frequency of the learning pattern of the category “を” is increased, and the appearance frequency of the learning pattern of the category “ramie” is realized by adding a weight to the appearance frequency of the character so as to keep the conventional frequency. That is, the category ""
The prior probability of “ramie” is kept large. After such setting, the canonical discriminant analysis method (Reference 2)
) To calculate the feature selection feature vector,
The feature selection feature vector Z2 of FIG. 4 is obtained. Due to the nature of the canonical discriminant analysis method, the feature amount selection feature vector is inclined toward the convergence direction of a category having a large prior probability. Category ""
And the center O1 and the center O2 of the category “ramie” are projected onto the feature amount selection feature vector Z2, and an intermediate straight line G2 between these two straight lines is obtained. The point on the O1 side of the straight line G2 is the category ""
And the point on the O2 side of the straight line G2 is the category "
It is determined that the character belongs to “ramie”. By changing the feature amount selection feature vector without performing work such as adding learning data as described above, the number of misidentified characters in the category “
The feature vector can be changed in a direction where recognition of the category "" is advantageous. In general manuscripts, the category "ramie"
The frequency of occurrence of the category “を” is much higher, and calculating a feature selection feature vector for these two characters with the same prior probability is disadvantageous for improving recognition accuracy. It is also extremely difficult to collect learning data according to the appearance frequency of actual manuscripts. Therefore, by changing the appearance frequency of characters belonging to a certain category as in the present invention, it is possible to reduce the total number of misidentified characters.

<Another embodiment of feature amount selection feature vector creation unit 130> FIGS. 5 and 6 are diagrams for explaining another example of the feature amount selection feature vector creation unit 130. . Referring to FIG. 5, the two categories “pi” and “ze” in the two-dimensional feature amount space have similar portions in the character shapes of the two categories, and therefore intersect in the feature amount space. Further, since each category also includes different learning data such as fonts, the feature space has a long and elongated shape. For such features,
When the feature amount selection feature vector is calculated using the canonical discriminant analysis method (see Document 2), Z1 in FIG. 5 is obtained. The upper left of the straight line G1 in FIG. 5 (this straight line can be obtained by linear discriminant analysis or the like) is an area of the category "pi", and the straight line G1 in FIG.
Is in the area of the category "Z", but since these two categories "P" and "Z" intersect considerably, each of them has a large number of erroneous characters. Therefore, before performing feature amount feature selection by applying the canonical discriminant analysis method, clustering is performed. FIG. 6 shows the clustering result. Here, the dotted line indicates each category before clustering, and the solid circle indicates the clustering result. Referring to this, category "pi" and category "ze" are each divided into two clusters,
Each cluster does not intersect. After performing this clustering, a feature amount selection feature vector is created using the canonical discriminant analysis method (see Document 2) to obtain Z2. Straight line G2,
Each cluster is completely separated by G3 and G4, and there is no erroneous character.

The above-described clustering is performed in the following procedure. (Procedure 1) Let N be the total number of learning data within a category.
In the initial state, it is assumed that each of the N pieces of learning data forms one cluster. Therefore, the distance from the center of the cluster to the farthest learning data is D
IST = 0, and the number M of clusters is M = N. (Procedure 2) A pair having the highest similarity (smallest distance) among the M clusters is obtained, and the pairs are fused as one cluster. As a result, the number M of clusters becomes M−1. The distance from the center of the merged cluster to the farthest learning data is defined as DIST. Thereafter, the number of clusters M is 1 or more, and the distance DIST is equal to a predetermined threshold CO.
If it is NST, proceed to the next step 3, otherwise proceed to step 4. Here, the threshold value CONST is empirically determined by the method of calculating the properties of the learning data and the feature amount (for example, considering the variance within a category, the variance between categories, the variance ratio, and the like). (Step 3) The similarity (or distance) between the newly created cluster and another cluster is calculated, and the procedure returns to step 2. (Procedure 4) The number of features (data) in the cluster and the features in the cluster are output to a file or the like as clustering information, and the clustering ends.

<Embodiments as Program of the Present Invention> Further, the present invention is not limited to only the above embodiments. For example, the present invention can be realized by a computer having the hardware configuration shown in FIG.
That is, the input device 1 includes a keyboard, a mouse, a touch panel, and the like, and is used for inputting information. Further, a scanner may be connected to input a character image. The display device 2 displays and outputs various output information, information input from the input device 1, and the like. CPU (Ce
An ntral Processing Unit (Central Processing Unit) 3 operates various programs. The memory 4 holds the program itself, and also holds information temporarily created when the program is executed by the CPU 3.
The storage device 5 stores the recognition dictionary 200 and the program, temporary information at the time of program execution, and the like handled by the present device.
The medium drive device 6 is used to mount a recording medium storing programs, data, and the like, read them, and store them in the memory 4 or the storage device 5. Further, it may be used for directly inputting / outputting data (for example, character learning data) or executing a program. In such a computer device, each function constituting the character recognition pattern dictionary creating device shown in FIG.
M, etc., and the CD-ROM is mounted on a computer equipped with a medium drive device 6 such as a CD-ROM drive at each site, and the program is stored in the memory 4 or the storage device 5, By executing this, it is possible to realize the same function as that of the embodiment of the above-described embodiment. As a recording medium, a semiconductor medium (for example, ROM, IC memory card, etc.), an optical medium (for example, DVD-ROM, MO, MD, CD-R)
Etc.) and magnetic media (for example, magnetic tape, flexible disk, etc.). Further, the program for realizing the functions of the present invention can be distributed in the form of a medium. It is also possible to store a program for realizing the functions of the present invention in a storage device such as a magnetic disk, and distribute the program in a form such as download by a wired or wireless communication network. Further, a program for realizing the functions of the present invention may be provided by distributing it by broadcast waves. In addition, the functions of the present invention are implemented by a client /
You may make it apply to a server system. That is, an image is fetched from the scanner of the computer on the client side, and the fetched image is input to the input unit 1 on the server side.
10 to generate the recognition dictionary 200. Alternatively, the image may be sent to the input unit 110 on the server side after the client performs the initial processing of the captured image.

[0013]

As described above, according to the present invention,
Since the appearance frequency of the character to be recognized is introduced, and the clustering is further performed, the feature amount selection feature vector is appropriately set, so that when performing feature amount selection, recognition can be performed without lowering the recognition accuracy. I was able to create a recognition dictionary that can do it.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a functional configuration of the present invention.

FIG. 2 is a flowchart showing a functional process of the present invention.

FIG. 3 is a diagram illustrating a feature amount selection feature vector before introducing a character appearance frequency.

FIG. 4 is a diagram illustrating a feature amount selection feature vector after introducing a character appearance frequency.

FIG. 5 is a diagram illustrating a feature amount selection feature vector before clustering.

FIG. 6 is a diagram illustrating a feature amount selection feature vector after clustering.

FIG. 7 is a diagram illustrating a hardware configuration of a computer on which the present invention operates.

[Explanation of symbols]

 1: Input device 2: Display device 3: CPU 4: Memory 5: Storage device 6: Medium drive device 100: Character recognition pattern dictionary creation device 110: Input unit 120: Feature extraction unit 130: Feature selection feature vector creation unit 140: feature quantity selection unit 150: dictionary creation unit 200: recognition dictionary

Claims (7)

[Claims] An input unit for inputting a character image, a characteristic amount extracting unit for extracting a characteristic amount from the character image input by the input unit by a discriminant analysis method or a principal component analysis method, and character appearance frequency information A feature amount selection feature vector creation unit that creates a feature amount selection feature vector for reducing the number of dimensions of the feature amount extracted by the feature amount extraction unit,
A feature amount selection unit that reduces the number of dimensions of the feature amount extracted by the feature amount extraction unit using the feature amount selection feature vector created by the feature amount selection feature vector creation unit; And a dictionary creating unit for creating a dictionary based on the obtained feature amount. 2. The character recognition pattern dictionary creation device according to claim 1, wherein the feature amount selection feature vector creation unit has a high prior probability for a character having a high appearance frequency to be recognized and a low appearance frequency for a character to be recognized. Character recognition pattern dictionary creation characterized by creating a feature selection feature vector that reduces the dimensionality of the feature extracted by the feature extraction unit by adjusting using a low prior probability for the character. apparatus. 3. The character-recognition pattern dictionary creating apparatus according to claim 1, wherein the feature-amount selecting feature vector creating unit performs clustering of learning data for a character to be recognized, and then performs a feature-amount selecting feature vector. A character recognition pattern dictionary creating apparatus characterized by creating a character recognition pattern dictionary. 4. The character recognition pattern dictionary creating apparatus according to claim 3, wherein the learning data is clustered on condition that a distance of a cluster member farthest from a center of a generated cluster does not exceed a predetermined numerical value. A character recognition pattern dictionary creation apparatus characterized in that a feature amount selection feature vector is calculated by performing the combination clustering method described above. 5. A feature for inputting a character image, extracting a feature amount from the input character image, adjusting the feature amount using character appearance frequency information, and reducing the number of dimensions of the extracted feature amount. A feature is that a quantity selection feature vector is created, a dimension is reduced from the feature quantity extracted using the created feature quantity selection feature vector, and a dictionary is created from the selected feature quantity. How to create a character recognition pattern dictionary. 6. A character image is input, a characteristic amount is extracted from the input character image by a discriminant analysis method or a principal component analysis method, clustering of learning data for a character to be recognized is performed, and then a character recognition target is extracted. A feature amount selection feature vector for reducing the number of dimensions of the extracted feature amount by adjusting using a high prior probability for characters with a high appearance frequency and a low prior probability for characters with a low appearance frequency to be recognized A character recognition pattern characterized in that the number of dimensions of the extracted feature amount is reduced using the created feature amount selection feature vector, and a dictionary is created from the selected feature amount. Dictionary creation method. 7. A computer-readable recording medium in which a program for causing a computer to function as a character recognition pattern dictionary creating device for creating a character recognition pattern dictionary is stored. An input unit, a feature amount extracting unit that extracts a feature amount from a character image input by the input unit by a discriminant analysis method or a principal component analysis method, and adjusting the feature amount by using character appearance frequency information; A feature amount selection feature vector creation unit that creates a feature amount selection feature vector for reducing the number of dimensions of the feature amount extracted by the unit; and a feature amount selection feature vector created by the feature amount selection feature vector creation unit. A feature amount selecting unit that reduces the number of dimensions of the feature amount extracted by the feature amount extracting unit using the feature amount extracting unit; A computer-readable recording medium having recorded thereon a program for functioning as a character recognition pattern dictionary creation device having a dictionary creation unit for creating a letter.