JPH09138834A - Character recognition device/method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly precisely execute a post-processing by generating again a pattern whose threshold of binarization is changed for a character which is not collated yet, executing individual character recognition and referring to the new candidate of a discriminated result obtained by character recognition to execute word collation. SOLUTION: A character pattern which is photo-electrically converted is inputted and it is stored in an image data storage means 301. A binarization/ normalization means 101 executes binarization/normalization by using the same threshold by the whole image for one character. A feature extraction means 102 extracts feature quantity from the input pattern and stores it in a feature data storage means 302. A character recognition means 103 discriminates the character by using the feature quantity. Then, a post-processing means 104 collates a recognized result with knowledge for the plural candidates obtained by the character recognition means 103. The pattern whose threshold of binarization is changed is generated again for the character which is not collated yet. Individual character recognition is executed and the word is collated by referring to the new candidate of the discriminated result obtained by character recognition.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device for automatically reading input characters.

[0002]

2. Description of the Related Art In the case of automatically recognizing a character with high accuracy, individual character recognition results are output as recognition result candidates.
By performing post-processing such as word matching, even if all characters are not completely readable, it may be possible to read in word units, and in the situation where knowledge such as name and address is available, positively improve reading accuracy. Is used for.

Two specific examples of address reading will be described below. First, in Japanese Unexamined Patent Application Publication No. 2-109187, "Post-character recognition post-processing method for solid address," the number of matches between the character recognition result and the address dictionary is calculated to match the individual character recognition result with the address dictionary, and the best match is obtained. A large number of address names are extracted from the address dictionary and are used as recognition results.

Next, in Japanese Patent Laid-Open No. 62-103785, "Character Reader", the candidates output by individual character recognition results are set as a collation range, and collation with an address dictionary is performed within the range, and collation is completed. The matching result is used as the recognition result.

[0005]

However, the above-mentioned 1
If you try to make a judgment based only on the number of coincidences as in the second example, for example, as shown in FIG. 6, “Higashimachi” is the correct answer and “East”
When it happened that individual characters could not be read and there were "Higashimachi" and "Honmachi" in the same area, the number of matches was 1 for both "Higashimachi" and "Honmachi" for the individual character recognition result "? Machi". It will be the same and the address cannot be specified. This is because the number of matches is don't care for the unreadable part of the individual character.

On the other hand, in the above-mentioned second method, the collation range is raised as a candidate for outputting the individual character recognition result, and if there is a candidate to be collated among them.
Compared to the first example, since the unreadable portion is searched from the lower candidates of the individual character recognition result, the collation accuracy can be expected to be improved. For example, if “Higashimachi” is the correct answer as in the previous example and “Higashi” does not appear in the first place of the individual character recognition result as shown in FIG. There is a high possibility that it is "Higashimachi", and the collation can be performed with higher accuracy than in the first example above. However, even in this case, if both “East” and “Book” exist in the lower candidates, the reliability of the rank of the lower candidates is generally unstable even if the upper candidate is “East”. If the "book" appears in the upper rank or if there is no correct character in the candidates, the correct collation result cannot be obtained as in the first case. This is particularly noticeable in the case of this example in which there are a plurality of candidate characters that can be embedded and they are similar to each other. Even when the characters are crushed, the feature extraction is not performed well, and it becomes unstable.

This is because, if a character pattern that cannot be read by the individual character recognition unit can be complemented as a result of collation, it is not confirmed by any means whether or not the character can be complemented.

An object of the present invention is to provide a character recognition device which solves the problems associated with the collating method as described above and can perform post-processing with higher accuracy.

[0009]

A character recognition method of the present invention is a character recognition method for binarizing an input character pattern and performing word matching, in which a character that has not been collated at the time of word matching is binarized. A feature is that word matching is performed by regenerating patterns with different thresholds, executing individual character recognition, and referring to new candidates of the discrimination result obtained there.

Further, the character recognition device of the present invention stores the image data stored in the image data storage means and the image data storage means for storing the signal obtained by the photoelectric conversion means for converting the input character pattern into an electric signal. Binarization / normalization means for performing binarization, absorbing variations within the same character type, and emphasizing differences between different character types, and input processed by the binarization / normalization means A feature extraction unit that extracts a feature amount effective for discriminating characters from a character pattern, a feature data storage unit that stores the feature amount effective for discriminating characters extracted by the feature extraction unit, and the feature data storage unit. A character recognition unit that determines which character type the input character pattern corresponds to on the basis of the stored characteristic amount, and outputs a plurality of candidates in descending order of possibility as a determination result; The graphic information extraction means for extracting the graphic information of the input character pattern stored in the data storage means, and the input to the plurality of candidates obtained as the determination result for the input character pattern obtained by the character recognition means Post-processing means for performing collation based on judgment results of patterns before and after the input character pattern using knowledge of the character pattern, and for determining one character as a recognition result from the plurality of candidates of the input character pattern. When the recognition result cannot be confirmed by the means, the graphic information of the input character pattern stored in the image data storage unit is extracted by the graphic information extracting means, and the binary information is extracted until it is different from the current seed. For the pattern obtained by changing the binarization threshold value in the normalization / normalization means, the feature extraction procedure is performed until the recognition result is finalized in the post-processing means. Feature extraction in the determination by the character recognition means, and repeating the correction by the post-processing means.

[0011]

The operation of the present invention will be described with reference to FIG.

The graph of FIG. 4 is a density histogram of the character pattern "East" which has not been matched in the example quoted in FIG. 6 or 7 at the time of word matching. The horizontal axis represents the density level and the vertical axis represents the frequency. .

The threshold for binarization in the first individual character recognition is set to s1, and the binary pattern obtained at that time is shown in FIG.
Shown in Since the crushing has occurred, "East" or "Book" cannot be accurately determined by the individual character recognition means at this point. Therefore, the threshold for binarization is set to the one in which the line width becomes thin in FIG.
Change to 2. The binary pattern obtained at that time is shown in FIG. In this pattern, the characters are not crushed and the strokes are clear, so stable features are extracted, and as a result, the individual character recognition means can correctly determine "East".

As described above, according to the present invention, for a character which has not been collated at the time of word collation, a pattern in which the threshold for binarization is changed is generated again, individual character recognition is executed, and the discrimination obtained there is carried out. Since the collation is performed by referring to the new candidate of the result, the conventional technique enables the collation of the correct character even when the collation is incorrect.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION Next, a first embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a character recognition apparatus showing an embodiment of the present invention, and FIG. 2 is a flow chart showing its operation.

The character recognition apparatus of the present invention stores image data storage means 301 for storing a signal obtained by a photoelectric conversion means for converting an input character pattern into an electric signal, and image data stored in the image data storage means 301. Binarization / normalization means 101 that performs binarization, absorbs fluctuations within the same character type, and emphasizes differences between different character types, and the binarization / normalization means 101 performs the processing. Feature extraction means 102 for extracting a feature amount effective for character discrimination from the input pattern, and the feature extraction means 102.
Characteristic data storage means 302 for storing the characteristic amount effective for discrimination of the characters extracted by the above-mentioned characteristic data storage means 3
A character recognition unit 103 that determines which character type the input pattern corresponds to based on the feature amount stored in 02, and outputs a plurality of candidates in order of high possibility as a determination result;
Graphic information extraction means 10 for extracting graphic information of the input pattern stored in the image data storage means 301.
5 and the plurality of candidates obtained as the determination result for the input pattern obtained by the character recognition means 103 are collated from the determination results for the patterns before and after the input pattern using the knowledge of the input pattern, Post-processing means 1 for determining one character as a recognition result from the plurality of candidates of the input pattern
04, the binarization / normalization means 101, the image data storage means 301, the feature extraction means 102, the feature data storage means 302, the character recognition means 103, and the graphic information extraction means 105, and a data bus 401 for exchanging data.
And the binarization / normalization means 101, the image data storage means 301, the characteristic means means 102, the characteristic data storage means 302, the character recognition means 103, and the graphic information extraction means 10.
It is composed of a control means 201 for controlling the number 5.

First, the photoelectrically converted character pattern is input and stored in the image data storage means 301 (step 21).

Next, the binarization / normalization means 101 performs binarization and further normalization (step 2).
2). For binarization, the same threshold may be used for the entire image for one character. As a threshold determination method at this time, for example,
"Otsu Nobuyuki: Mathematical Research on Feature Extraction in Pattern Recognition, IEICE Research Report No. 818 (1991.
7) ”, which can be realized by using the discriminant analysis method. For normalization, in order to facilitate identification, it is sufficient to suppress fluctuations within the same character type and clarify differences between different character types. For example, "J. Tsukumo,
et al. : Classification of
handprinted China Chara
cters Using Non-linear No
rmalization and Correlati
on Methods, Proc. of9th ICP
R (1990.7) ”.

Next, feature extraction is performed on the character pattern that has been binarized and normalized (step 23). It suffices that a feature amount effective for character discrimination can be extracted. For example, “Hamachu et al .: Consistency between nonlinear normalization and feature extraction in handwritten kanji recognition,” IEICE Technical Report, PRU 91-85 (199).
1.11) ”. The feature for the extracted character pattern is stored in the feature data storage unit 302.

Next, the character recognizing means 103 discriminates the character by using the characteristic amount stored in the characteristic data storing means 302 (step 24). From the extracted characteristic data, it is determined which character it is. If the feature quantity exemplified in step 23 is used, for example, “Atsushi Tsukumo: Improvement of direction pattern matching method and application to handwritten Chinese character recognition, IEICE Technical Report, PUR 90-20 (199).
0.7) ”is suitable. A plurality of recognition results, for example, 10 candidates are output in descending order of reliability.

Next, the post-processing means 104 collates the recognition result with the knowledge (step 25). Here 1
The recognition result for each word is verified using knowledge, and the individual recognition result for each character is corrected. At this time, a character that does not match the knowledge is detected, and finally the collation result is output (step 26).

An example of the processing procedure in step 25 will be described with reference to FIG. FIG. 3 is a flow chart showing an example of realizing an individual recognition result for each character and knowledge.

First, the candidate output from the individual recognition result is collated with the knowledge (step 31).

Next, it is determined whether or not there are all matching characters in the candidates output from the individual character recognition result (step 32).
That is, when matching is performed based on knowledge, it is checked whether all the characters are included in the candidates for the determination result of the individual character recognition. Here, if all the characters are in the candidates of the judgment result of the individual character recognition, the processing is ended. On the contrary, as a result of matching, knowledge and uncorrespondence,
That is, when the character to be corrected by knowledge does not exist in the recognition result candidates, graphic information is extracted from the character pattern of the corresponding portion (step 33). As this graphic information, for example, the line width of a character is extracted.

Next, the threshold for binarizing the character pattern of the corresponding portion is changed according to the character line width (step 3).
4). If the line width is thick, the line width is reduced, and conversely, if the line width is thin, the line width is changed to be thick to extract the line width (step 35), and the line width is detected again (step 36). This is repeated until the line width changes (loop 1). The step of changing the threshold may be arbitrary. When a change in line width is detected, the pattern is again subjected to individual character recognition (step 37), candidates for the determination result are output (step 38), word matching is performed again (step 31), and it is desired to match in the candidates. Repeat until a character appears (loop 2). When the character you want to match appears, the matching result is output. If the character you want to match does not appear in the candidates for the judgment result until the end even if you change the threshold value, that part is output as unrecognizable and is also output. (Step 39).

As the graphic information, the number of connecting line segments, the number of loops, or a combination thereof may be used regardless of the line width of the character.

Next, a second embodiment of the present invention will be described. In the first embodiment, if all the characters collated by knowledge in step 32 are present in the candidates for the determination result of the individual character recognition, the collation is ended and the result is output. On the other hand, in the second embodiment, the binarization and re-recognition of the above pattern are repeated until all the collated characters reach the first place in the judgment result of the individual character recognition. As the graphic information detected to change the binarization threshold value when repeating the individual character recognition, as in the first embodiment,
Any of the character line width, the number of connected components, the number of loops, and a combination thereof may be used.

Next, a third embodiment of the present invention will be described. In the present embodiment, in step 32, when a plurality of characters that can be collated by knowledge appear in a candidate, any one of the above characters is binarized until it reaches the first position in the judgment result of individual character recognition. It is to repeat recognition again. As in the first embodiment, the graphic information detected to change the binarization threshold value when repeating individual character recognition includes any of the line width of the character, the number of connected components, the number of loops, and a combination thereof. May be used.

[0030]

As described above, according to the present invention, when the character to be collated does not appear in the candidate of the judgment result of the individual character recognition, the graphic information of the pattern of the corresponding portion is extracted, and the pattern information is changed until it changes. By changing the threshold of binarization and repeating the recognition until the judgment result candidate is reached, it is possible to completely correspond to the knowledge, and there is an effect that the recognition accuracy of the character is significantly improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a character recognition device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing an example of the operation of the character recognition device shown in FIG.

FIG. 3 is a flowchart showing an example of an operation of the character recognition device shown in FIG. 1 at the time of knowledge matching.

FIG. 4 is an explanatory diagram for explaining the principle of the present invention.

FIG. 5 is an explanatory diagram for explaining the principle of the present invention.

FIG. 6 is a diagram showing a determination method of a conventional character recognition device.

FIG. 7 is a diagram showing another determination method of the conventional character recognition device.

[Explanation of symbols]

 101 Binarization / Normalization Means 102 Feature Extraction Means 103 Character Recognition Means 104 Post-Processing Means 105 Graphic Information Extraction Means 201 Control Means 301 Image Data Storage Means 302 Feature Data Storage Means 401 Data Bus

Claims (6)

[Claims] 1. A character recognition method in which an input character pattern is binarized and word matching is performed, a pattern in which a binarization threshold value is changed is regenerated for a character that has not been matched at the time of word matching, and an individual character is generated. A character recognition method characterized in that character recognition is performed, and word matching is performed by referring to a new candidate of a discrimination result obtained there. 2. The character recognition method according to claim 1, wherein the binarization threshold value is changed according to the character line width, the number of connecting line segments, or the number of loops. 3. An image data storage means for storing a signal obtained by a photoelectric conversion means for converting an input character pattern into an electric signal, and the image data stored in the image data storage portion are binarized to have the same character type. To distinguish the characters from the binarization / normalization means that absorbs the fluctuations in the above and emphasizes the difference between different character types, and the input character pattern processed by the binarization / normalization means. Feature extraction means for extracting an effective feature amount, feature data storage means for storing the feature amount effective for discrimination of the character extracted by the feature extraction means, and based on the feature amount stored in the feature data storage means A character recognition unit that determines which character type the input character pattern corresponds to and outputs a plurality of candidates in descending order of possibility as a determination result; The graphic information extraction means for extracting the graphic information of the input character pattern that is present, and the knowledge of the input character pattern for the plurality of candidates obtained as the determination result for the input character pattern obtained by the character recognition means. The post-processing means for performing collation based on the determination results of the patterns before and after the input character pattern and for determining one character as the recognition result from the plurality of candidates of the input character pattern, the post-processing means can determine the recognition result. If not, the graphic information of the input character pattern stored in the image data storage unit is extracted by the graphic information extracting means, and the binarizing / normalizing means outputs 2 until the graphic information is different from the current seed. For the pattern obtained by changing the threshold for binarization, the feature extraction by the feature extraction means, the sentence until the recognition result is confirmed by the post-processing means. Determination by the recognition unit, a character recognition apparatus characterized by repeating the correction by the post-processing means. 4. The input character pattern stored in the image data storage unit when the post-processing unit adopts a judgment result other than the first candidate of the judgment result output by the character recognition unit. The graphic information is extracted by the graphic information extraction means, the binarization threshold value is changed in the binarization / normalization means until it is different from the current value, and the pattern obtained in the feature extraction means is used for the obtained pattern. 4. The feature extraction, the determination by the character recognition means, and the correction by the post-processing means are repeated until the first candidate of the determination result is determined as the recognition result by the post-processing means. Character recognizer. 5. When the candidate of the judgment result output by the character recognition means is adopted by the post-processing means, if there are a plurality of collable characters among the candidates, the candidate is stored in the image data storage section. The graphic information of the input character pattern is extracted by the graphic information extraction means, the binarization threshold value is changed in the binarization / normalization means until it is different from the current value, and the obtained pattern is The feature extraction in the feature extraction unit, the determination by the character recognition unit, and the correction by the post-processing unit are repeated until the first candidate of the determination result is confirmed as the recognition result by the post-processing unit. The character recognition device according to claim 3. 6. The character recognition device according to claim 3, 4 or 5, wherein said graphic information extracting means extracts the line width of the image data, the number of connected components or the number of loops as the graphic information.