JP4328511B2 - Pattern recognition apparatus, pattern recognition method, program, and storage medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a pattern recognition apparatus, a pattern recognition method, a program, and a storage medium for character recognition and the like.
[0002]
[Prior art]
Character recognition is common as a practical example of pattern recognition. The character recognition device extracts features from the input characters, compares them with standard patterns of all target characters prepared in advance, and outputs the character having the highest similarity as a recognition result. In the identification operation for comparing the input character and the standard pattern, the amount of calculation increases as the character type increases, and the recognition accuracy decreases when the amount of calculation is reduced. Japanese is a language that has many character types along with Chinese, and has been devised to speed up processing while maintaining recognition accuracy.
[0003]
For example, a recognition process has been proposed in which a large classification process for narrowing the number of characters is performed before the identification calculation. In this case, with respect to the large classification, the original identification calculation may be referred to as detailed identification.
[0004]
In the large classification, an identification calculation with a smaller calculation amount than that of the detailed identification is performed, and a specific number of characters is selected from the top of the similarity, or a character whose similarity exceeds the threshold is selected. In the detailed identification, only the characters narrowed down by the large classification are discriminated and the character with the highest similarity is output as the recognition result.
[0005]
This can speed up the processing while maintaining the recognition accuracy.
[0006]
On the other hand, a character recognition process using similar character information has been proposed.
[0007]
In the character recognition processing described in Japanese Patent Laid-Open No. 5-20503 (referred to as Conventional Example 1), a candidate character dictionary in which characters having similar shapes are collected is provided, and characters similar to the recognition result are extracted from the candidate character dictionary and recognized. A similar character can be selected along with the result.
[0008]
In the character recognition processing described in Japanese Patent Application Laid-Open No. 8-147406 (conventional example 2), similar candidate characters that may become recognition results in the process of recognition processing are extracted, and characters as recognition results cannot be determined. In addition, a similar candidate character can be searched.
[0009]
[Problems to be solved by the invention]
Conventional examples 1 and 2 make it easy for the user to select a similar character when erroneously recognizing, and do not increase the recognition accuracy.
[0010]
The present invention was devised to solve such conventional problems, and improves the recognition accuracy while maintaining the processing speed.
[0011]
[Means for Solving the Problems]
A pattern recognition apparatus according to the present invention includes a large classification dictionary including a large classification category for performing large classification processing on a pattern to be recognized, and a large classification category for executing a large classification process that is more detailed than the large classification process. Including a large classification dictionary, a similar category table showing similar categories similar to each large classification category included in the large classification dictionary, a detailed identification dictionary for performing detailed identification processing more detailed than the large classification processing, With reference to the large classification dictionary, a large classification means for extracting a large classification category corresponding to the recognition target pattern, and each of a plurality of large classification categories classified into the large classification category extracted by the large classification means, A first major similarity calculating unit that calculates a similarity between the recognition target pattern and the first classification dictionary by referring to the major classification dictionary; Depending on the maximum similarity among the similarity degrees calculated by the class similarity calculation unit, selects the large classification category from among the plurality of large classification category the classified into large large classification category extracted in large rough classification means , Similar category extraction means for extracting similar categories similar to the selected large category category with reference to the similar category table, and similarities extracted by the similar category extraction means with reference to the large classification dictionary A second major classification similarity calculating unit that calculates a similarity between a category and the recognition target pattern; a similarity calculated by the first major classification similarity calculating unit; and the second major classification similarity Based on the similarity calculated by the calculation means, a plurality of large classification categories classified into the large classification categories extracted by the large classification means and the classes extracted by the similar category extraction means Major classification result extraction means for extracting a category of a major classification result from the categories, and detailed identification for recognition result candidates classified into the category of the major classification result extracted by the major classification result extraction means Detailed identification means for identifying a recognition result of the recognition target pattern from the recognition result candidates by performing detailed identification processing with reference to a dictionary.
[0012]
DETAILED DESCRIPTION OF THE INVENTION
Next, a first embodiment of a pattern recognition apparatus according to the present invention will be described with reference to the drawings.
[0013]
[Pattern recognition device]
FIG. 1 is a block diagram showing a first embodiment of a pattern recognition apparatus according to the present invention.
[0014]
The pattern recognition apparatus includes an external storage device 104 such as a magnetic disk and a pointing device 107, and inputs image data from the external storage device 104 or the pointing device 107. The pattern recognition apparatus includes a CPU 101 for overall control, and a control program for the CPU 101 is stored in the ROM 102. Image data input from the external storage device 104 or the like is held in the RAM 103, and processing for pattern recognition is executed by a control program in the ROM. A large classification dictionary and a detailed identification dictionary for pattern recognition are stored in the ROM 102 together with the control program, or are read from the external storage device 104 into the RAM 103 for each necessary character type.
[0015]
In FIG. 1, 105 is a display and 106 is a keyboard. Reference numeral 109 denotes a network interface, which can communicate with a device located at a remote place (not shown) to transmit / receive a program, a recognition dictionary, data, and the like.
[0016]
Although not shown, the image scanner may be configured to include an image scanner for reading an image when applied to character recognition, and a microphone or a playback device when applied to voice recognition.
[0017]
Next, a pattern recognition method executed in the pattern recognition apparatus will be described.
[0018]
[First Embodiment of Pattern Recognition Method]
A first embodiment of a pattern recognition method will be described with reference to the drawings. In the first embodiment, large classification is performed in order to narrow down the categories to be processed for large classification. That is, a large classification process by the second recognition unit is executed on the large classification result by the first pattern recognition unit, and a third recognition is performed on the large classification result by the second pattern recognition unit. The detailed classification process by means is executed.
[0019]
2 is a flowchart showing a first embodiment of a pattern recognition method executed by the pattern recognition apparatus of FIG. 1, FIG. 3 is a flowchart showing an example of a similar category table creation method in the pattern recognition method of FIG. 2 is a diagram for explaining an example of a similar category table in the pattern recognition method of FIG. 2, FIG. 7 is a conceptual diagram showing a classification method for large classification, and FIG. 8 is a contour direction for another classification method of large classification FIG. 9 is a conceptual diagram showing the direction index used in FIG. 8.
[0020]
In FIG. 2, the flow of processing of the first embodiment will be described.
[0021]
Step S501: Large classification is executed using a large classification dictionary obtained by classifying the categories of the large classification dictionary into larger categories. The large classification uses a processing method with a smaller calculation amount than the large classification. For example, as one method used for large classification, the optical character recognition apparatus can perform classification based on the presence or absence of black pixels in a specific area. In (10-1) of FIG. 7, the feature region is a 3 × 3 region (a, b, c from the upper left to the right, d, e, f from the left of the second step surface to the right, and from the left of the third step surface. An example of division is shown on the right with g, h, and i. (10-2) in FIG. 7 shows the relationship between the pixel of the character “A” and these divided regions, and (10-3) shows the relationship between the pixel of the character “L” and the divided regions. The character “A” has black pixels in the regions b, d, e, f, g, and h, and the character “L” has black pixels in the regions a, b, e, g, h, and i. In this way, the category of characters can be classified according to the area where black pixels exist.
[0022]
Alternatively, the characters can be classified into large categories by extracting the contour of the image in the feature region and giving the direction index of FIG. 9 to the boundary pixels forming the contour. The boundary pixels in the feature region are evaluated in the order in which the left-to-right scanning is sequentially performed from the top to the bottom, and the directions of the pixels following the target boundary pixel P0 are the four non-opposing directions among the eight pixels around the target pixel. Representing any of them (for example, lower, lower right, right, upper right) is sufficient to represent the feature data. These directions are expressed by direction indices of C1 (lower), C2 (lower right), C3 (right), and C4 (upper right), for example. For example, in FIG. 8, when the width of the feature region 701 (indicating an image of one character of hiragana “i”) is w and the height is h, the feature region 701 is divided into a predetermined number of widths and heights. A plurality of small regions are formed. The number of divisions is, for example, the width direction division number n = 4 and the height direction division number m = 4.
[0023]
In FIG. 8, one small region 702 is representatively shown, and among the pixels in the small region, the boundary pixels are hatched, and the symbols P1, P2, P3, P4, P5, P6, P7, and P8 are denoted. Giving. These boundary pixels have a pixel P2 following the pixel P1 and a pixel P3 following the pixel P2, when viewed in the scanning order from left to right and from top to bottom. Further, the pixel P5 exists after the pixel P4, the pixel P6 exists after the pixel P5, the pixel P7 exists after the pixel P6, and the pixel P8 exists after the pixel P7. This continuous state is obtained by giving the direction index C2 to the pixel P1, the direction index C2 to the pixel P2, the direction index C2 to the pixel P4, the direction index C3 to the pixel P5, the direction index C2 to the pixel P6, and the direction index C2 to the pixel P7. Expressed.
[0024]
In each small area, the direction index is statistically processed. For example, the number of each direction index is used as characteristic data of the small area. In the small area 702, C2 occurs 5 times and C3 occurs once. Therefore, if the feature data is {C1, C2, C3, C4}, {0, 5, 1, 0} is obtained. Alternatively, if the most frequent direction index is the feature data of the small area, the global direction of the image is expressed.
[0025]
In such feature data, for example, a category (referred to as a large classification category) is extracted based on the number of any of the four elements. As a result, the categories to be classified are reduced, and the classification process can be made more efficient.
[0026]
Step 502: Only for the large classification category selected in the large classification in step S501, the similarity between the pattern feature and the standard pattern is calculated with reference to the large classification dictionary.
[0027]
Step S503: From the large classification categories, a large classification category having the maximum similarity is extracted. The maximum similarity S is compared with predetermined threshold values T1, T2, T3, and T4 (T1>T2>T3> T4), and the next process is selected according to the value of S.
[0028]
When S> T1, it is determined that it is not necessary to add more categories, and the process proceeds to step S508. When T1 ≧ S> T2, the process proceeds to step S504. When T2 ≧ S> T3, the process proceeds to step S505. When T3 ≧ S> T4, the process proceeds to step S506. In these steps, as the degree of similarity is higher, a smaller number of large classification categories are extracted, and the processing efficiency can be improved while maintaining recognition accuracy.
[0029]
Step S504: It is determined that there is a possibility that a correct character is included in the similar category of the category with the first category of major classification, and the similar category of the category with the first category of major classification similarity is extracted from the similar category table. And add it to the result of large classification.
[0030]
Step S505: Extract the similar categories of the categories having the first and second categories of major classification from the similar category table and add them to the major classification result.
[0031]
Step S506: The similar category of the category having the first to the eighth rank in the large classification is extracted from the similar category table and added to the large classification result.
[0032]
Step S507: The similarity between the added category and the category of the major classification dictionary is calculated.
[0033]
Step S508: Category selection is performed using the large classification calculation results calculated in Step S502 and Step S507. The category selected in this step is the result of the major classification.
[0034]
Step S509: Detailed identification is executed, and the recognition result is determined.
[0035]
In FIG. 3, in the first embodiment, a similar category table is created by the following processing.
[0036]
Step S601: The first category of the standard pattern stored in the large classification dictionary is set as the attention category.
[0037]
Step S602: The processes in and after step S603 are performed until all categories are completed.
[0038]
Step S603: The similarity between the standard pattern of the target category and the standard patterns of all other categories is calculated.
[0039]
Step S604: The top K categories having high similarity are selected and stored.
[0040]
Step S605: The category next to the attention category is set as the attention category, and the process returns to step S602.
[0041]
Step S606: When all categories have been processed, for each of the selected top M categories, the similar category of that category is set as the target category.
[0042]
As shown in FIG. 4A, in the process of step S604, when K = 2, the similar categories for the attention category “A” are “B” and “C”, and the similar category for the attention category “B” is “A”. ”,“ D ”and the similar category for the attention category“ C ”are“ A ”and“ B ”, and the similar category for the attention category“ D ”is“ B ”and“ A ”.
[0043]
As shown in FIG. 4B, the similar category table has a similar category number “3” for the category “A”, a similar category number “3” for the similar categories “B”, “C”, “D”, and the category “B”. ”, Similar categories“ A ”,“ C ”,“ D ”, category“ C ”, similar category number“ 1 ”, similar categories“ A ”, category“ D ”, similar category number“ 1 ”, similar category“ B ” Is displayed.
[0044]
Note that the similar category table may be created using a detailed identification dictionary or data including all other standard patterns without using the large classification dictionary.
[0045]
In the first embodiment, when the category selected by the large classification includes a category with high reliability, the amount of calculation can be limited by reducing the added category, and the accuracy of the large classification is reduced. While supplementing, there is an effect that can be recognized at high speed.
[0046]
[program]
The control program in the pattern recognition apparatus of FIG. 1 includes a program for executing the first embodiment of the pattern recognition method. The program for executing the first embodiment includes a routine for executing the processes of steps S501 to S509 and steps S601 to S606.
[Storage medium]
The pattern recognition method according to the present invention can be implemented by an apparatus other than the pattern recognition apparatus of FIG. 1, for example, a general-purpose computer. In this case, a control program provided by a storage medium or the like is set in an external storage device of the general-purpose computer. However, it may be executed by a CPU or the like according to an instruction from an operator or the like.
[0047]
FIG. 5 is a memory map showing a storage medium storing a program for executing the first embodiment of the pattern recognition method.
[0048]
In FIG. 5, directory information is stored at addresses A0 to A1, and a program for executing the flowcharts of FIGS. 2 and 3 in the first embodiment is stored in the area from the address next to A1 to address A2. The
[0049]
Although not particularly illustrated, data depending on various programs may be managed as directory information, and a program for installing various programs in a computer may be stored in the directory information.
[0050]
[Second Embodiment of Pattern Recognition Apparatus]
FIG. 6 is a block diagram showing a first embodiment of a pattern recognition apparatus that executes a pattern recognition method according to the present invention using a plurality of devices.
[0051]
In this embodiment, in the Internet service provider, a program including a recognition dictionary for pattern recognition is stored in the server SV1. The Internet service provider is connected to the Internet through communication means, and by accessing the server SV1 from the Internet, a plurality of users can use the pattern recognition program.
[0052]
The user has a scanner SC1, a personal computer PC1, and the like, and transmits an image of a document read by the scanner SC1 to the server SV1. In the server, pattern recognition processing is executed using an image of the document as an input image.
[0053]
Note that the large classification dictionary, the large classification dictionary, and the detailed identification dictionary do not necessarily exist on the same computer as the pattern recognition program, and may be referred to via a communication unit or the like. For example, it is possible to provide a pattern recognition program as a separate service by providing a server SV1 and a dictionary such as a detailed identification dictionary in a separate server SV2. In this case, the owner of the pattern recognition program and the owner of various dictionaries can own the intellectual assets such as programs and dictionaries, and illegal copying can be prevented.
[0054]
【The invention's effect】
According to the present invention, the recognition accuracy is improved without reducing the processing speed.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a first embodiment of a pattern recognition apparatus according to an embodiment of the present invention.
FIG. 2 is a flowchart showing a first embodiment of a pattern recognition method executed by the pattern recognition apparatus of FIG. 1;
FIG. 3 is a flowchart showing an example of a similar category table creation method in the pattern recognition method of FIG. 2;
4 is a diagram for explaining an example of a similar category table in the pattern recognition method of FIG. 2; FIG.
5 is a memory map of a storage medium for storing various data processing programs executed by the pattern recognition apparatus of FIG. 1. FIG.
FIG. 6 is a block diagram showing a pattern recognition system for executing a pattern recognition method according to the present invention by a plurality of devices.
FIG. 7 is a conceptual diagram illustrating a classification method for large and large classifications.
FIG. 8 is a conceptual diagram showing contour direction codes for another classification method of large classification.
FIG. 9 is a conceptual diagram showing the direction code direction used in FIG. 8;
[Explanation of symbols]
101 ... CPU
102 ... ROM
103 ... RAM
104 ... External storage device 105 ... Display 106 ... Keyboard 107 ... Pointing device 108 ... Network interface

Claims (6)

A large classification dictionary including large classification categories for performing large classification processing of patterns to be recognized;
A large classification dictionary including a large classification category for performing detailed large classification processing than the large classification processing;
A similar category table showing similar categories similar to each major classification category included in the major classification dictionary;
A detailed identification dictionary for executing detailed identification processing more detailed than the large classification processing;
A large and large classification means for extracting a large and large classification category corresponding to the recognition target pattern with reference to the large and large classification dictionary;
First major similarity that calculates a similarity between the recognition target pattern and each of a plurality of major classification categories that are classified into major classification categories extracted by the major classification means Degree calculation means;
According to the maximum similarity among the similarities calculated by the first large classification similarity calculating means, a large classification is selected from a plurality of large classification categories classified into the large classification categories extracted by the large classification means. A similar category extracting means for selecting a category and referring to the similar category table to extract a similar category similar to the selected major category;
A second large category similarity calculating unit that calculates a similarity between the similar category extracted by the similar category extracting unit and the recognition target pattern with reference to the large category dictionary;
Based on the similarity calculated by the first large classification similarity calculating means and the similarity calculated by the second large classification similarity calculating means, the large classification categories extracted by the large classification means are classified. A large classification result extracting unit for extracting a category of a large classification result from the plurality of large classification categories and the similar category extracted by the similar category extracting unit;
By performing detailed identification processing with reference to the detailed identification dictionary for recognition result candidates classified into the category of the major classification result extracted by the major classification result extraction means, the recognition result candidates are selected from the recognition result candidates. A pattern recognition apparatus comprising: detailed identification means for identifying a recognition result of a recognition target pattern.   2. The similar category table is created by calculating a similarity between categories stored in the large classification dictionary and setting categories having high similarity as similar categories. The pattern recognition apparatus of description. In the similar category extracting means, the number of major classification categories selected from a plurality of major classification categories classified into the major classification categories extracted by the major classification means is the first major classification similarity calculating means. The pattern recognition apparatus according to claim 1, wherein the pattern recognition apparatus changes according to the maximum similarity among the similarities calculated in step (1). A large classification step in which the large classification means refers to the large classification dictionary and extracts a large classification category corresponding to the recognition target pattern;
A first major classification similarity calculating unit for executing a major classification process more detailed than the major classification process for each of a plurality of major classification categories classified into the major classification category extracted in the major classification step; A first large classification similarity calculating step of calculating a similarity with the recognition target pattern with reference to a large classification dictionary;
The similar category extracting means has a plurality of large categories classified into the large classification categories extracted in the large classification step according to the maximum similarity among the similarities calculated in the first large classification similarity calculation step. A similar category similar to the selected major category is selected by selecting a major category from the category categories and referring to a similar category table showing similar categories similar to each major category included in the major dictionary A similar category extracting step for extracting
A second major classification similarity calculator calculates a similarity between the similar category extracted in the similar category extraction step and the recognition target pattern with reference to the major classification dictionary. Similarity calculation step;
The large classification result extraction means is extracted in the large classification step based on the similarity calculated in the first large classification similarity calculation step and the similarity calculated in the second large classification similarity calculation step. A large classification result extracting step for extracting a category of a large classification result from a plurality of large classification categories classified into the large classification category and the similar category extracted in the similar category extraction step;
A detailed identification dictionary for executing a detailed identification process more detailed than the large classification process for a recognition result candidate classified into a category of the large classification result extracted by the large classification result extraction step. And a detailed identification step of identifying a recognition result of the recognition target pattern from the recognition result candidates by performing detailed identification processing with reference to FIG. Computer
A large classification means for extracting a large classification category corresponding to the recognition target pattern with reference to the large classification dictionary,
For each of a plurality of large classification categories classified into the large classification categories extracted by the large classification means, refer to a large classification dictionary for executing a large classification process more detailed than the large classification process, and the recognition target First major similarity calculating means for calculating the similarity between the patterns;
According to the maximum similarity among the similarities calculated by the first large classification similarity calculating means, a large classification is selected from a plurality of large classification categories classified into the large classification categories extracted by the large classification means. Similar category extraction for selecting a category and extracting a similar category similar to the selected major category by referring to a similar category table showing similar categories similar to each major category included in the major category dictionary means,
Second major classification similarity calculating means for calculating a similarity between the similar category extracted by the similar category extracting means and the recognition target pattern with reference to the large classification dictionary;
Based on the similarity calculated by the first large classification similarity calculating means and the similarity calculated by the second large classification similarity calculating means, the large classification categories extracted by the large classification means are classified. A large classification result extracting unit for extracting a category of a large classification result from a plurality of large classification categories and the similar category extracted by the similar category extracting unit;
Refer to a detailed identification dictionary for performing a detailed identification process more detailed than the large classification process for recognition result candidates classified into the category of the large classification result extracted by the large classification result extraction unit. Detailed identification means for identifying the recognition result of the recognition target pattern from the recognition result candidates by performing identification processing;
A computer-executable program that functions as a computer. A computer-readable storage medium in which the program according to claim 5 is stored.

Images