JPS60142791A - Dictionary forming method for character recognition - Google Patents

Abstract

PURPOSE:To reduce the man-hour by forming a dictionary being a base in advance and collecting additional learning data while selecting them to decrease the required time for forming the dictionary. CONSTITUTION:The feature code relating to the learning data being a base is edited to generate a dictionary, the additional learning data is collected by utilizing a character reader, the data is re-edited and the dictionary 8 is formed. When the additional learning data is collected, a character of the additional modification mode written on a form is applied sequentially to a pattern reader 3, an A/D converter 4, a pre-processing circuit 5 and a feature extraction circuit 6 to extract a feature code of the character. The feature code is collated with the feature code in the dictionary 8 by a collation circuit 7 and only the feature code dissident with collation is selected and collected in an RAM9 as the additional learning data.

Description

[Detailed Description of the Invention] <Technical Field of the Invention> The present invention provides a character recognition device that specifies an unknown character by storing the characteristic code of an unknown character pattern in a dictionary in advance and comparing it with the characteristic code of a standard pattern. In particular, the present invention relates to a method for creating a dictionary for character recognition, in which a dictionary is created by editing the feature codes of the standard pattern.

<Background of the Invention> Conventionally, this type of dictionary was created by having an unspecified number of people write down each character to be recognized, and then sequentially reading these written characters using a character reading device to identify the characteristics of each of the seven characters. After coding and collecting the feature codes in memory as learning data, this learning data is edited to create a dictionary.

However, in the case of such a method, in order to improve dictionary accuracy, it is necessary to collect as many deformed written characters as possible at once. However, since the collected written characters include characters with common feature codes, it is necessary to select the learning data when organizing the dictionary. As a result, a great deal of effort and time is expended in data collection and selection, which significantly reduces the efficiency of dictionary creation work. Furthermore, since learning data obtained from written characters tends to be biased toward specific feature codes depending on the number of scribes, content, etc., it is necessary to prepare other types of written characters and collect additional learning data. In this case, additional training data is collected in memory in the same way, so
There have been many problems, such as duplication with learning data already stored in the dictionary, and it takes more effort and time to select additional learning data.

<Objective of the Invention> The present invention shortens the time required to create a dictionary and reduces labor by creating a basic dictionary in advance and collecting data while selecting additional learning data using this dictionary. The purpose of this paper is to provide a new method for creating a dictionary for character recognition that achieves the following.

<Structure and Effects of the Invention> In order to achieve the above object, the present invention first creates a dictionary by editing each feature code of basic learning data, and then extracts character feature codes to be used as additional data. hand,
The basic learning data of the storage material is compared with the code in the dictionary, and only the feature codes that do not match are collected and used as additional learning data of the dictionary.

The present invention does not collect a lot of learning data for dictionary creation at once, but instead creates a basic dictionary and then adds learning data one by one. Since only necessary additional learning data that is not duplicated is collected, a series of processes for creating or reorganizing a dictionary can be carried out efficiently, and the purpose of the invention can be achieved, such as shortening time and reducing labor.

<Description of Embodiments> FIG. 1 shows a schematic configuration of a character recognition device. In the illustrated example, unknown character 2 written on form 1 is a CCD (Char
The optical pattern is optically read by a pattern reading device 3 such as a GE-Coupled Device, and the optical pattern is converted into a time-series digital signal by an A/D converter 4. The preprocessing circuit 5 performs noise removal and binarization processing on the digital signal, and stores the character pattern in an image memory (not shown). The feature extraction circuit 6 encodes and extracts character features required for character recognition, such as the number of intersections, the number of end points, the number of branch points, and the number of loops, from the character pattern stored in the image memory. The dictionary matching circuit 7 stores the feature code generated by the feature extraction circuit 6 in a ROM (Rea).
d 0nly Memory ) is compared with the feature code of the standard pattern stored in the dictionary 8, and when the two match, the ID code of the standard pattern with that feature code is output, and when they do not match, reject processing is performed. be done.

The dictionary 8 stores a plurality of feature codes for each character, which encode the features of each deformation of the character. Figure 2 shows a 5-bit configuration of the characteristics of each deformed form of the katakana character "A" (in the illustrated example, the presence or absence of a connection, the presence or absence of a loop, the presence or absence of a dent on the left side, the presence or absence of 3 end points, and the presence or absence of a branch point). This is shown in the code below. For example, the modified form "A" shown in FIG. 2(2) has only the feature of "a dent on the left side" and is given the feature code of roolooJ.

Dictionary 8 above is created by first collecting basic learning data by a specific person, using a character reading device for this learning data to collect additional learning data, and then reorganizing the data. be completed. To collect this additional training data,
Characters of the transformation required for addition are written on the form, these characters are optically read by the pattern reading device 3, and after predetermined processing is performed by the A/D converter 4 and preprocessing circuit 5, features are extracted. A circuit 6 extracts character feature codes. Then, as shown in FIG. 3, this feature code is compared with the feature code stored in the dictionary 8 in a dictionary matching circuit 7, and only feature codes that do not match are selected and stored in RAM (Ra).
ndom Access Memory ) 9 is collected as additional learning data.

FIG. 4 shows a flowchart of such additional learning data collection processing. First, in step 11, the character of the transformation required for addition is read, followed by step 12.
In step 13, preprocessing and feature extraction processing are executed. Next, in step 14, the feature code of that character and the dictionary 8
Each feature code related to the basic learning data stored in
ES” and the process proceeds to step 16. If there is no code match, step 14 becomes “No” and the process proceeds to step 16.
In the next step 15, the feature code related to the read character is stored in the RAM 9 as additional learning data. If there is a next character to be read, the determination at step 16 becomes -YES'' and the process returns to step 11, where the same process is executed. In this way, only the feature codes that are not stored in the dictionary 8 are selected and collected one after another in the RAM 9, and the additional learning data after collection is re-edited together with the previous learning data and stored in the dictionary 8.

[Brief explanation of the drawing]

Figure 1 is a block diagram showing the schematic configuration of a character recognition device, Figure 2 is an explanatory diagram showing the characteristics of each deformation of the character "a" converted into data, and Figure 3 is a circuit used to collect additional learning data. FIG. 4 shows a flowchart showing the process of collecting additional learning data.

Claims (1)

[Claims] After creating a dictionary by editing each feature code of the basic learning data, extract the feature codes of characters that should be added data, match the codes with the basic learning data already stored in the dictionary, and A method for creating a dictionary for character recognition, characterized in that only unmatched feature codes are collected and used as additional learning data for the dictionary.