JPH0340186A - Character recognizer - Google Patents

Abstract

PURPOSE:To perform character recognition with high likelihood by performing comparative collation with a second dictionary consisting of a reference pattern for a character with low frequency in use sequentially when the collation result of a character pattern is less than a prescribed level. CONSTITUTION:The character is classified at every frequency in use at a dictionary part 4, and is stored as individual dictionaries (1)-(3), respectively. Also, a character pattern collation part 3 is connected to a reject level judging part 5 which employs the collation result with a dictionary with high frequency in use when it exceeds the prescribed level, and performs collative comparison with the dictionary with low frequency in use when it is less than the prescribed level. Therefore, it is possible to prevent the character seldom used outputted accidentally as the recognition result, and to improve the accuracy of the recognition result, and furthermore, since collation is separated corresponding to the frequency in use, the collation of an ordinary character with the character with low frequency in use can be omitted. In such a way, the character recognition with high likelihood can be performed.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character recognition device that recognizes characters drawn on a recording medium such as paper.

[Background Art] In recent years, with the spread of word processors and the like, it has become common practice to encode and store characters.

If it is encoded and stored, the storage capacity can be reduced, and if a common code (for example, JIS code) is used, it can be easily recognized by different types of machines. Furthermore, if display is performed according to coded character data, desired display can be performed according to various demands.

In this way, various advantages can be obtained by encoding and memorizing characters, but this requires a means to specify codes such as a keyboard as a means of stringing characters. It is not possible to extract the code from the characters,
There was a problem in that the characters had to be written manually by operating a keyboard or the like while looking at the characters drawn on paper or the like.

Therefore, it has been proposed to automatically read the characters already drawn on a recording medium such as paper, recognize the characters, and encode them. If such reading and encoding were possible, the above-mentioned problems would be solved and it would be very convenient.

That is, conventional character recognition devices optically detect characters drawn on a recording medium such as paper, binarize the characters into a predetermined number of dots, and read the characters as a two-dimensional dot pattern consisting of digital data. This reading result is then compared with pre-stored character patterns, and the closest character pattern is detected as the corresponding character.

With such conventional devices, if the characters drawn on paper etc. are printed characters, the reading results for each character are quite accurate, and if the font etc. of the printed characters can be recognized, the reading results can be quite accurate. Verification can be performed.

Normally, in the above-mentioned character comparison, printed or handwritten characters are read as a two-dimensional dot pattern, the characteristics of each character are extracted, and the characteristics of each character are compared with the characteristics of all characters stored in a dictionary.

Therefore, the inter-pattern distances as a result of the comparison may actually be approximately equal for a plurality of characters.

In particular, for complex characters, there are cases where multiple characters with similar characteristics are approximately equidistant, and for a single input pattern, multiple candidates with approximately equal interpattern distances are often selected.

[Problem to be Solved by the Invention] Therefore, in the past, since the character with the closest distance between patterns is selected, the character with the closest distance selected from all dictionaries is output, and the characters that are close but completely different are output. This caused the characters to be selected, reducing the possibility that the real characters would be selected.

That is, in the conventional method, since the frequency of character use by the user is not taken into consideration, characters that are not normally used are unexpectedly selected as candidates, and even if the distance between the patterns is the minimum, Since these characters have a low probability of being actually used, they are often incorrect candidates, which has been a source of dissatisfaction for users.

SUMMARY OF THE INVENTION An object of the present invention is to provide a character recognition device that can perform highly accurate character recognition while taking into account differences in the frequency of use of characters.

[Means for Solving the Problems] In order to achieve the above object, the device of the present invention divides a dictionary into a plurality of parts according to the frequency of use of characters, and usually performs pattern comparison only in the frequently used dictionaries. Contains a matching section.

In the device of the present invention, the character pattern matching unit adopts the frequently used dictionary if the result of matching with the frequently used dictionary is equal to or higher than a predetermined level; It is connected to a rejection level determination unit that performs comparison with a dictionary.

[Operation] The character recognition device according to the present invention has the above-described configuration, and reads characters drawn on a recording medium such as paper as a two-dimensional dot pattern. The read results are then compared with standard pattern data for frequently used characters. In this comparison, if a character cannot be recognized, it is compared with a standard pattern for characters that are used less frequently.

Therefore, it is possible to prevent characters that are used very rarely from being erroneously output as recognition results, and it is possible to improve the accuracy of recognition results. Furthermore, since the matching is divided according to the frequency of use, there is an advantage that matching of ordinary characters with characters that are used less frequently can be omitted.

[Example 7] Hereinafter, an example of a character recognition device according to the present invention will be described based on the drawings.

In FIG. 1, an optical detection unit 1 irradiates light onto a character C drawn on a recording medium P such as paper, and detects the shading on the recording medium P based on the intensity of the reflected light. and convert it into a two-dimensional dot pattern. Next, this two-dimensional dot pattern is supplied to the read data processing section 2.

In the read data processing section 2, the detection result (two-dimensional dot pattern) obtained by the optical detection section 1 is extracted and normalized for each character. That is, in a case where a normal handwritten document is written on the recording medium P, each character is separated and extracted using the blank part of the character amount, and for each character (human handwritten character) that is extracted, For example, it is converted into a 60x60 two-dimensional dot pattern. In other words, depending on the size of the input character, the read data processing unit 2 performs enlargement or reduction processing (normalization).
It converts human data into a two-dimensional dot pattern of a desired size.

The two-dimensional dot pattern thus obtained is then supplied to the character pattern matching section 3. The character pattern matching section 3 performs matching processing by comparing standard pattern data for each character stored in advance in the dictionary section 4 with the two-dimensional dot pattern supplied from the reading pig processing section 2. Here, this matching process is not simply a comparison of whether the shading of each dot matches, but also extracts features of the data supplied from the read data processing section 2.
This is done by comparing the extracted feature data with the features of the standard character pattern data. To extract this feature data, for example, the two-dimensional pattern data of 60 x 60 dots is divided into 3 x 3 blocks, and each dot in the block consisting of 3 x 3 dots is extracted in the right direction as seen from that dot. Or 0"" depending on whether there is a line downwards or not.
This is done by making a set of digital data such as ``1''.Also, these feature data are compared by calculating the total sum of squares of the difference between the input character for each dot and the standard pattern, and then comparing the detected characters. This is done by calculating the inter-pattern distance between the standard pattern stored in the dictionary and the standard pattern stored in the dictionary.

Further, details of character recognition by extracting feature data as described above are described in "Introduction to Character Recognition" (edited by Shinichi Hashimoto, published by the Electrical Communication Association, March 20, 1982).

The collation data for each character in the character pattern collation unit 3 thus obtained is compared with a predetermined level in the jet level determination unit 5. That is, if the distance between patterns is smaller than a predetermined rejection level, the characters are taken up as candidate characters for characters recognized as having a high degree of similarity, and if the distance is larger than this reject level, they are considered to be different characters and are not adopted. When a plurality of candidate characters having a reject level or higher are picked up in this way, the closest one among them is supplied to the display section 6 as a recognized character. The display unit 6 displays the recognized characters on a display D or the like. In addition, a plurality of candidate characters may be displayed if necessary.

Here, what is characteristic of this invention is that the dictionary section 4 classifies characters according to their frequency of use and stores them as separate dictionaries (1) to (3). In the matching process, the character pattern matching unit 3 compares each dictionary classified by frequency of use, without comparing it with the character patterns stored in the entire dictionary unit 4.

Next, the character recognition procedure of this embodiment will be explained based on FIG. 2.

First, when processing starts, a dictionary (1) that stores standard character patterns for frequently used characters is used.
is adopted as the matching target (matching target is dictionary (1))
. Next, the feature data extracted from the two-dimensional dot pattern of the human-powered characters supplied from the read data processing unit 2 is compared with the feature data about the standard character pattern from the dictionary (1), and the distance between the patterns is determined (in the dictionary). ). The inter-pattern distance obtained in this way is then compared with a pre-stored rejection level to determine whether the input character feature data is sufficiently similar to the standard character pattern to be compared (reject level judgement).

Here, the comparison with this reject level is dictionary (1)
This is performed for all standard character patterns in the file, and a determination is made as to whether or not all of them have been rejected (rejected?). If all of them are rejected, it means that there are no candidate characters in dictionary (1), so next it is determined whether or not all dictionaries have been checked (whether all dictionaries have been checked) . In this case, since only the dictionary (1) is checked, the result is No, and a switch is made to the next dictionary (switch to the next dictionary).

Then, switch the dictionary to the next dictionary, that is, dictionary (2
) The characters in parentheses etc. are compared in the same way as the previous time. and,
If all the character patterns in the dictionary (2) are compared and there is no character pattern that is higher than the reject level, the next dictionary, that is, the dictionary (3) is switched to and the character patterns, etc. in this dictionary are compared. If there is no standard pattern of reject level or higher in this dictionary (3), all dictionaries are checked, but there are no candidate characters, so it is determined that there are no candidate characters, and this is displayed ( (no candidate characters).

On the other hand, if there is a standard pattern of reject level or higher in any of the dictionaries (1)', (2), and (3), this is recognized as a candidate character and stored (candidate character output). Then, the matching process for this input character ends.

As described above, in the present invention, if there is a similar standard pattern of a reject level or higher in the comparison with the dictionary (1), it is output as a candidate character and the matching process is ended.

That is, according to the apparatus of this embodiment, the dictionary (1), dictionary (2), and dictionary (3) are checked in order of frequency of use. Then, if a candidate character is listed in dictionary (1),
The following sequential pattern dictionary matching is not performed. Therefore, frequently used characters are preferentially selected as candidates, and characters that are rarely used are prevented from being selected as candidates, thereby improving the overall recognition accuracy.

Next, the overall processing for such character recognition will be explained based on FIG. 3.

When recognizing characters recorded on a recording medium, the recording medium is first set in a predetermined position 2 and optically read. Then, each character is extracted from the data read in this way, and this is compared for each character with standard character patterns in the dictionary to perform a matching process. Then, the verification process for the entire input document is performed as described above (the verification process is performed for the input document). Then, based on such recognition results, the first candidate (the character with the closest pattern distance) is displayed for each character, and the recognition results are displayed on a display or the like (recognition result display). Then, the user looks at this display and determines whether there are any characters that have been erroneously recognized (are there any erroneously recognized characters?). If there are no misrecognized characters, all recognition results are correct, and there is no need to perform correction processing, so the processing ends (end).

On the other hand, if there is a misrecognized character, the misrecognized character must be corrected. Therefore, the erroneously recognized character is specified by moving the cursor to the erroneously recognized character, and candidate characters are then displayed (candidate character display).

In other words, in the above-mentioned character-by-character recognition process, each character that has similarity equal to or higher than the reject level is stored as a candidate character, so if an incorrectly recognized character is specified, a candidate character for this character can be output. I can do it. Then, the user determines whether or not the correct characters are included in these candidate characters (are the corresponding characters included?). If the corresponding character is included, that character is selected from the candidate characters (candidate character selection). When this selection is made, the selected character is replaced with the first candidate character, which is the misrecognized character, and correction is performed using this character (correction using the selected character).

Then, the process moves on to the correction operation for the next character.

Furthermore, if the corresponding character is not included in the candidate character display, this means that the corresponding character was not output in the previous matching process, and therefore the dictionary is switched and matching is performed (switching dictionary and matching).

This is because in the apparatus of this embodiment, the dictionaries are classified into multiple categories and provided separately, so the matching process is performed for each dictionary. If a candidate character is found in a frequently used dictionary, it is not compared with the next dictionary. Therefore, if the corresponding character is not among the candidate characters, it is assumed that the next dictionary contains the corresponding character,
The dictionary with the next rank after the dictionary in which the candidate character was selected last time is selected and matched with the individually stored character patterns.

Then, a comparison process is performed with all the character patterns in the next dictionary, and if there are some candidate characters, the display returns to the candidate character display and it is determined whether or not the corresponding character is included. On the other hand, if there is no candidate character in the results of matching with all character patterns even though the dictionary has been switched and matching is performed, it is determined whether there is a dictionary to match next (all dictionaries are checked). Has the verification been completed?)

Then, if there is a remaining dictionary, the system switches to it and performs the matching process again, and if no candidate character is output even in this process, the character is input from the keyboard (the character is entered manually from the keyboard).

Then, correct the characters manually from this keyboard 5 Manually edit the characters and move on to correcting the next character.

In the manner described above, it is possible to correct an error in the initial character recognition.

In addition, in the above embodiment, dictionary (1) to dictionary (3)
Although three dictionaries were adopted, any number of dictionaries greater than or equal to two can be used.

Furthermore, the classification of the dictionary 4 according to usage frequency may be determined in advance, but may be changed at any time using a learning function or the like.

Furthermore, in the embodiment, classified dictionaries (1) to dictionaries (3) are physically separate, but there is only one dictionary 4, and dictionaries (1) to dictionaries (1) to dictionaries (3) are provided for each stored standard character pattern. A code may be provided to identify the classification of the dictionary (3). Furthermore, in character recognition, a function may be added to perform matching processing as an idiom by comparing not only the matching result for each character but also the preceding and succeeding recognized characters. Furthermore, the rejection level determined by the feature comparison after the feature data extraction described above may be set to a very small distance or a large distance depending on the character, so it is recommended to change the rejection level by 6 for each character.

[Effects of the Invention] As explained above, according to this invention, frequently used characters can be compared preferentially, which reduces the possibility that rarely used characters will be selected as candidates. , it is possible to improve the overall recognition system. Furthermore, in the case of a dictionary of characters that are used infrequently, the number of times of matching is reduced, so that the overall recognition speed can be improved.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of an embodiment of the present invention, Fig. 2 is a flowchart showing the flow of character recognition processing in the embodiment, and Fig. 3 shows the overall processing flow of character recognition in the embodiment. It is a flow chart for explanation. 1... Optical detection section 2... Read data processing section 3... Character pattern matching section 4... Dictionary section Rejection level judgment section Display section

Claims (1)

[Claims] (1) An optical detection unit that optically detects characters drawn on a recording medium such as paper and reads them as a two-dimensional dot pattern, and normalizes the two-dimensional dot pattern read by this optical detection unit. a read data processing unit that performs pre-processing such as the above, a dictionary unit that classifies and stores standard pattern data for each character into a plurality of dictionaries according to the frequency of use, and the optical detection unit and the read data processing unit. a character pattern matching section that compares and matches the read results obtained with a first dictionary consisting of standard patterns for frequently used characters; a rejection level determination unit that determines the character as a recognition candidate character, and the character pattern matching unit sequentially rejects characters with low usage frequency when the result in the character pattern matching unit is below a predetermined level. A character recognition device characterized by performing comparison verification with a second dictionary consisting of standard patterns.