JPH09251517A - Method and device for recognizing character - Google Patents

Abstract

PROBLEM TO BE SOLVED: To preferably use a learning dictionary while suppressing an influence of erroneous recognition to a basic dictionary. SOLUTION: A teacher code (S1) is given, a character to be learned is recognition-processed only by the basic dictionary (S2) and a category different from the character to be learned is scanned (S3). Then, establishment distance (S4) is obtained by distance from a nearest and different-kind category and the learning dictionary is generated (S5). Establishment distance corresponding to the learning dictionary is stored (S6) and the basic dictionary for recognizing a character is preferably used based on establishment distance.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a recognition method and a recognition apparatus for preferentially using a recognition learning dictionary created by a user for recognition, and more particularly to a character recognition method and a character recognition apparatus.

[0002]

2. Description of the Related Art In conventional character recognition, research and development has been conducted to pursue a high recognition rate.
Devices from single fonts to multi-fonts have been realized, and future development of omni fonts is expected. In addition, as for the online character recognition device, a character level device has been continuously developed from regular writing, and it is expected that the characters will be broken into characters in the future.

However, even though the recognition performance has been improved, it is difficult to accurately recognize the user's unique pattern (handwritten characters or abbreviations) with only the basic dictionary, and this problem is solved. For this reason, conventionally, character recognition using a learning dictionary unique to a user by learning a unique pattern (learning pattern) has been used.

However, since the act of learning is performed on the character recognition device, the user expects that it will be correctly recognized (correctly, a pattern close to it) after that. However, it is hard to say that it is a practical method if it can be easily recognized. That is, in the case of a conversational recognition system, the user expects a 100% correct answer for the learning pattern. Therefore, it is necessary to give some priority to the learning dictionary. For example, Japanese Patent Laid-Open No. 145789/1989 discloses a technique of preferentially outputting a unique pattern of a user registered in the learning dictionary. (Prior art 1).

However, when learning patterns are additionally registered in the learning dictionary, learning dictionary patterns similar to one another are created one after another, which rather lowers the character recognition rate. To solve this problem, for example, JP-A-1-
Japanese Patent No. 44586 discloses a technique of calculating and displaying a stability index of how stable a character pattern to be registered can be recognized compared with a registered pattern, and displaying and outputting the stability index. (Prior art 2).

[0006]

However, if the priority of the learning dictionary is simply raised as in the case of the conventional technique 1,
Misrecognition occurs.

Hereinafter, the adverse effects of misrecognition will be specifically described by taking the characters "king" and "ball" as an example. As a premise, it is assumed that the standard pattern for "king" and the standard pattern for "balls" in the basic dictionary are similar patterns (characteristic points are close in distance) in character-based categories. First, the user inputs a handwritten character "O", but it is difficult to recognize in the basic dictionary, so it is assumed that the user registered it as a learning pattern in the learning dictionary. In other words, the handwritten character "king" is a pattern that is far from the standard pattern of "king" in the basic dictionary. In addition, the handwritten character "O" is far from the standard pattern of "tama" in the basic dictionary.

[0008] Therefore, since the handwritten character "king" is learned, it is easy to recognize it. However, if a "ball" similar to the handwritten character "king" is input by handwriting, "ball" in the basic dictionary is input.
It is distant from, and because there is a similar "king" in the learning dictionary, the misrecognition harmful effect that the handwritten character "ball" is mistakenly recognized as "king" occurs. In other words, if the learned character is far from the basic dictionary and a similar character is input near the learned character, misrecognition will occur and it will become more prominent if the learning dictionary has some priority. .

On the other hand, the misrecognition adverse effect can be solved to some extent because the user determines whether or not to learn in the prior art 2, but the recognition rate cannot be uniformly improved and the operation becomes complicated. There is a problem.

An object of the present invention is to provide a character recognizing method and a character recognizing device capable of recognizing by preferentially using a learning dictionary while minimizing adverse effects of erroneous recognition at the time of recognition, without requiring the user to judge the presence or absence of learning. To provide.

[0011]

A character recognition method according to claim 1 is a character recognition method for recognizing an input character in units of categories using a basic dictionary and a learning dictionary. A distance shorter than the distance to the closest heterogeneous category is set in advance as the fixed distance of the learning dictionary of the character to be learned, and at the time of character recognition, the basic character dictionary and the learning dictionary are referred to and the input character is used as a candidate for recognition by matching processing. The distance is calculated, and when the distance is smaller than the fixed distance, the distance of the recognition candidate is reset so as to be shorter, and the order of the recognition candidates is changed based on the reset distance to perform character recognition. Characterize.

A character recognition device according to a second aspect is characterized by using the character recognition method according to the first aspect.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows a handwritten character recognition device for performing character recognition on an input character basis in a matching process between a pattern stored in a basic dictionary or a learning dictionary and an input pattern of a handwritten input character. Here, 1 is an existing recognition system, 2 is an input device such as a tablet, 3 is a basic dictionary in which standard patterns are stored in character units, and 4 is a learning pattern registered by the user in character units. Learning dictionary, 5 is a fixed distance buffer that stores a fixed distance that is a feature of the present invention, 6 is a recognition candidate buffer that stores recognition candidates, 7 is a feature extraction of an input character, character recognition, and fixed distance calculation, A control unit composed of a CPU or the like for performing the priority processing of the learning dictionary in the processing described later, and 8 is a general-purpose register used in the processing described later to prioritize the learning dictionary.

The existing recognition system 1 is a recognition processing device or processing program for performing character matching processing such as pattern matching between a dictionary pattern and an input pattern to perform character recognition in input character units.

When handwriting is input to the tablet of the input device 2 with a pen, the coordinates of the tip of the pen are sampled at fixed time intervals, an input pattern is generated from the coordinate values, and the controller 7 operates to Characteristic data is extracted character by character. Here, the input device 2 may be a display-integrated tablet.

The basic dictionary 3 stores the characteristic data corresponding to the standard pattern, and the learning dictionary 4 stores the characteristic data corresponding to the learning pattern. Although it depends on the recognition method, for example, a segment of fixed length is divided into stroke units. Direction and length, coordinates and directions of variable points divided by variable length in radical units, or in translation OCR, the circumscribing frame of the character is set to 8
It has the black pixel density value in each mesh which is divided into 64 as characteristic data. It should be noted that the basic dictionary 3 and the learning dictionary 4 each store a pair of characteristic data corresponding to the pattern and its character code.

Characteristic data of the input pattern of the handwritten input character from the input device 2 is character-recognized by pattern matching with the characteristic data stored in the basic dictionary 3 or the learning dictionary 4, and the character corresponding to the correct code is recognized. Is output.

Further, since the character recognition device recognizes characters in character units, the category is a character. However, since categories depend on the recognition unit of the recognition method, other categories may be used. In the online character recognition, characters are stroked or copied. It may be divided into necks for recognition, in which case the category is stroke or radical. In addition, in European language recognition, words may be recognized word by word, in which case the category will be a word. In any case, the category is the dictionary and the unit for calculating the distance.

Next, a method for calculating the defined distance of the learning dictionary 4 will be described with reference to FIGS. 2 to 4. FIG. 2 is a flowchart of the fixed distance calculation, and FIG. 3 is a schematic diagram showing how to determine the fixed distance. The fixed distance of the learning dictionary 4 is obtained when the user receives the learning data from the input device 2, and the learning data to be stored in the learning dictionary 4 includes the characteristic data corresponding to the learning pattern to be learned by the character recognition device and its characteristic data. It is the correct code (character code that you want to recognize correctly).

First, when learning data is given, a teacher code is given (step S1). If the user interactively inputs the correct code, enter the teacher code.
This becomes a teacher code, and in the case of an automatic learning system, the system automatically generates a teacher code.

Next, for the characteristic data of the learning data,
The existing recognition system 1 performs matching processing using only the basic dictionary 4 for recognition and obtains recognition candidates (step S2).

Next, the recognition candidates obtained in step S2 are scanned based on the distance obtained by the matching process,
A category different from the teacher code and having the smallest distance (nearest neighbor heterogeneous category) is searched (step S3), and 1 / n (n> 1) of the distance is set as the definite distance (step S).
4). Here, n is a value that can be set arbitrarily, and n = 2 in the present embodiment. This is schematically shown in FIG.

Next, in order to create a learning dictionary (step S5), the characteristic data of the learning pattern and its correct answer code are stored in the buffer of the learning dictionary 4, and the fixed distance for this learning pattern is stored in the fixed distance buffer 5. Yes (step S6). The fixed-distance buffer 5 is a distance value buffer for the number of learning dictionaries, and the number and capacity thereof depend on the character recognition device. In the present embodiment, the number of learning dictionaries is 128, and since one learning dictionary is 1 byte, it is a buffer of 128 bytes. This is the end of the calculation of the definite distance of the learning dictionary created when the learning data is given.

FIG. 4 shows a state in which the basic dictionary is a white circle and the learning dictionary is a black circle, and two-dimensionally plotted for each category. The defined distance of the present invention is small in an area where the basic dictionaries are dense and large in an area where the basic dictionaries are distant. Further, since the fixed distance of each learning dictionary is a value less than the distance between each of the nearest heterogeneous categories, there is no mixture of heterogeneous category dictionaries within this fixed distance, and there is no adverse effect on the basic dictionary. Is set for each learning dictionary.

Next, a method of prioritizing the use of the learning dictionary 4 in which the fixed distance is set in order to prevent the adverse effect of erroneous recognition will be described with reference to FIGS. 5 and 6. FIG. 5 shows a flowchart of a method of preferentially using the learning dictionary.

First, a recognition process is performed on an unknown input character using the basic dictionary 3 and the learning dictionary 4, and a recognition candidate is obtained in the recognition candidate buffer 6 (step S11). As shown in FIG. 6, for each recognition candidate, the character code of each candidate, the dictionary number, the unknown input character and the basic dictionary 3 or the learning dictionary 4
The recognition candidates are arranged in ascending order of distance, including the distance obtained as a result of the matching calculation with and the information of the discrimination flag of the dictionary of the basic dictionary 3 or the learning dictionary 4. Recognition candidate buffer 6
Usually has a configuration in which the basic dictionary 3 and the learning dictionary 4 are mixed.

Next, the obtained recognition candidates are scanned, and it is determined whether they are candidates that hit the learning dictionary by the discrimination flag (discrimination between basic dictionary and learning dictionary) (step S1).
2). In the case of the basic dictionary, the discrimination flag is used to skip to the next recognition candidate. In the case of a learning dictionary, it is determined whether the distance of the recognition candidate buffer 6 is less than the fixed distance of the corresponding learning dictionary (step S13).

Next, when the distance of the recognition candidate buffer 6 is equal to or larger than the fixed distance, the process skips to the next recognition candidate. If the distance in the recognition candidate buffer 6 is less than the fixed distance, the value obtained by multiplying the distance by the weight (w) (step S14) is stored in the recognition candidate buffer 6 as a new distance, and the process moves to the next recognition candidate. This weight (w) is a weight for giving priority to the learning dictionary, and is a value that can be set arbitrarily, and in this embodiment, w =
It is 0.5.

After performing the same processing on all recognition candidates, the recognition candidates including the basic dictionary and the learning dictionary are sorted again by the distance of the recognition candidate buffer 6 (step 15), and the distance is short. Can be rearranged in order.

By the above processing, the recognition candidates corresponding to the learning dictionary can be preferentially obtained.

Although the distance is used as the recognition scale in the above description, a recognition method using the similarity in addition to the distance may be used. Further, the present invention can be applied to any recognition device that can use OCR, online, voice, etc. as inputs and has a learning function for these inputs.

[0032]

As described above, according to the present invention,
The learning dictionary can be preferentially used for character recognition while automatically minimizing the adverse effect on the basic dictionary instead of making individual judgments by the user.

[Brief description of drawings]

FIG. 1 is a diagram showing a character recognition device according to the present invention.

FIG. 2 is a flowchart of a fixed distance calculation according to the present invention.

FIG. 3 is a diagram for explaining how to determine a fixed distance in a learning dictionary.

FIG. 4 is a diagram for explaining the concept of a fixed distance in a learning dictionary.

FIG. 5 is a flowchart for prioritizing a learning dictionary as a recognition candidate.

FIG. 6 is a diagram showing a structure of a recognition candidate buffer.

[Explanation of symbols]

 1 Existing recognition system 2 Input device 3 Basic dictionary 4 Learning dictionary 5 Definite distance buffer 6 Recognition candidate buffer 7 Control unit 8 General-purpose register

Claims (2)

[Claims] 1. A character recognition method for recognizing an input character in a predetermined category unit using a basic dictionary and a learning dictionary, wherein a distance shorter than a distance between a learning category of a character to be learned and a different category is It is set in advance as the fixed distance of the learning dictionary of the character to be learned, and at the time of character recognition, the distance of the recognition candidate is found by the matching process from the input characters by referring to the basic dictionary and the learning dictionary. When the distance is smaller than the fixed distance A character recognition method characterized in that the distance of the recognition candidates is reset so as to be short, and the order of the recognition candidates is changed based on the reset distance for character recognition. 2. A character recognition device using the character recognition method according to claim 1.