JPH02224085A - Character recognizing device - Google Patents

Abstract

PURPOSE:To correct erroneous reading and to secure high accuracy for reading by recognizing a character with the use of a character pattern itself in a slip. CONSTITUTION:Concerning the input pattern which is detected with a character detection part 2 at first and continuously processed with a first character recognition part 3 and a post-processing part 4, the final result of character recognition is obtained by executing again processing to include one or both of character detection and second character recognition by using a second dictionary storage part 12 for character recognition. Namely, after the execution of picture input, character detection, first character recognition and post-processing, concerning a part for which the reliability of the result is judged to be low, the detection and segmentation of the character and character recognition are executed again by using the character pattern on the input slip. Thus, even when a context or a word, which is not included in knowledge concerning context or monotone, is inputted or when the character of low quality is inputted, the character is not erroneously read but a high recognition rate can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a character recognition device that corrects misread portions of character recognition results and obtains correct recognition results.

(Prior art) A character recognition device has an image input section, a character detection section, two character recognition sections,
It consists of a post-processing unit that improves the final recognition rate by correcting ambiguity in character recognition results based on context and word knowledge.

However, when a context or word that is not known in the context or word is input, or when low-quality characters are input, no means has been developed so far to relieve the eventual misreading of characters.

(Problem to be solved by the invention) As described above, conventional character recognition devices have the disadvantage that there is no remedy for unknown words or low-quality characters. The object of the present invention is to provide a character recognition device capable of obtaining a high recognition rate as a whole without performing wA reading even when a word or a low-quality character is input.

[Structure of the invention]

(Means for Solving the Problem) The present invention includes an image input section, a single character detection section, a first character recognition section, a post-processing section, and a re-recognition section, and the re-recognition section is a first character recognition section. A part that estimates the correct character code of the input character, that is, a correct character, from the output result of the post-processing unit, a part that creates a dictionary for second character recognition based on this estimation result, and a part that executes the second character recognition. It consists of parts.

First, characters are detected, and then all or part of the input cover pattern that has been subjected to the first character recognition and post-processing is subjected to at least character detection and second character recognition using a second character recognition dictionary. The final character recognition result is obtained by performing the process including one or both of the above character recognitions again. .

(Function) In the present invention, after performing one image input, character detection, first character recognition, and post-processing, the second character pattern of the input form is used for parts where the reliability of the results is judged to be low. It becomes possible to detect and cut out characters and perform character recognition again.

(Embodiment) FIG. 1 is a conceptual block diagram for carrying out an embodiment of the present invention. Image data of a form on which characters are printed or written is converted into an electrical signal and inputted from the image input section 1. Next, character lines are extracted from this image data in the character detection section 2, and then the Characters are extracted from inside.

The extracted character patterns are each recognized by the first character recognition section 3, and as a result of the recognition, for example, a plurality of candidate characters are determined. A plurality of candidate characters obtained for each character are sent to the post-processing section 4. In step 4, a character string that is estimated to be correct based on knowledge regarding this context, such as a knowledge dictionary consisting of 11-word data, is estimated from the input character recognition result and output. This result is further sent to the re-recognition unit 5, where the processing described below is executed. The final result is output from 5 and becomes the final character recognition result.

Next, the operation in step 5 will be explained. FIG. 2 is a detailed block diagram of the second part.

FIG. 3 shows the flow of data in this embodiment. 20 is an input image, and the result of character recognition is 21. In this example, the top three candidates of the character recognition results are illustrated. The O mark in the figure is a correct answer candidate. By using knowledge of the context as post-processing, this result is modified as in 22. In this 4-, the word ``jin-sama'' is recognized as a human right, but because ``jin-sama'' is not registered in the @language dictionary, knowledge about the context cannot be used.
Not fixed. The data of 21 and 22 are input to the correct character estimation section in FIG.
Among the candidate characters in the lowest position, the characters that have been corrected as a result of post-processing are misread characters, and the characters resulting from post-processing are determined to be the correct characters. , and sends this to the second recognition dictionary creation section 11. The second dictionary creation section creates a new character recognition dictionary for the character, based on the input character pattern of the character that has been uploaded.

This dictionary is shown in FIG. 73, 23, and is stored in the second recognition dictionary storage section 12.

On the other hand, the recognition result 21 and the post-processing result 22 are stored in the reading result storage section 13. After or during the recognition of the form, the recognition results stored in 13 are sent to the second recognition execution determination section 14, where it is determined whether or not second recognition is necessary for each character. In the example shown in Figure 3, the “people” part for which post-processing was unsuccessful is the second
Selected as the execution part of recognition. Therefore, the image pattern of the "person" portion is extracted again from the second character detection section and sent to the ztirR section. The second edict recognition section uses the 23 already registered dictionaries to correctly recognize the part of "person" and sends it to the output section. The output unit outputs the reading results of the portion not sent to the second recognition unit and the second recognition unit.
The recognition results of the recognition parts are integrated and the final output result is output.

Here, if the characters input to the second recognition unit are unreadable, or if it is determined that the second reading result is less reliable than the first reading result, the first character recognition unit Output the reading result as the final result.

As the second recognition method, recognition is performed using the following simple method.

First, a pattern 32 is created by thickening the pattern 31 from which a dictionary is created, and this pattern 32 is used as a dictionary.

For recognition, a pattern 34 is created by thickening the input cover pattern 33, and the exclusive OR of the stiffness and the dictionary pattern 32 is calculated for each pixel to create a pattern 38.
The area of the unmatched part of 4, that is, the part of 35 is calculated, and the ratio of this to the area of the matched part, that is, 37 is calculated, and the character corresponding to the dictionary with the smaller unmatched part is taken as the recognition result.

In this method, the fattening operation and the exclusive OR for each pixel can be directly performed even on a pattern of run-1 non-guess expression, so the expression of a pattern or dictionary can also be expressed as a run-length expression.

FIG. 5 is a diagram for explaining another embodiment.

Here, 42, 43, and 44 are the - parts of the nine patterns stored in the second recognition dictionary storage section, respectively. 41, the three characters are in contact [7], so individual characters cannot be cut out,
This is a pattern where the data is incorrectly read due to bending or forced reading. The re-recognition unit inputs this pattern, superimposes each dictionary pattern on the left side, examines the degree of matching, and identifies the overlapping portion 45 with the dictionary pattern 42 with a high degree of matching as a character in the pattern. 4], cut out pattern 4 in the same way sequentially.
Each character is cut out from the pattern 41 by cutting out 4G that matches 3 and cutting out 47 that matches the pattern 44.

Using such a method, it is possible to reliably extract even complicated touching characters. ,': For confirmation, the first character recognition method may be applied to the cut out pattern to check the recognition results in 1.8.

In the first example, the correct character was estimated based on the first character recognition result and post-processing, but in the second example, the correct character was estimated based on the first character recognition result and the post-processing, but in the second example, the recognition result that is 1 inch smaller than the character recognition result of The estimation may be performed using an evaluation value regarding reliability, for example, similarity.

Note that the present invention is not limited to the above embodiments.

The recognition method of the first recognition section and the recognition method of the second recognition section may be of any type, and are not limited to the above embodiments.

The first i,W recognition method and the second recognition method may be the same, and any post-processing method may be used. Re-recognition can be performed at any time.3 Also, when determining which part to re-recognize, we do not use only the post-processing results, but also use the reliability of the first character recognition result as part of the evaluation value. can do.

Alternatively, all characters may be subject to re-recognition. The dictionary used for recognition may be the original pattern itself without making it thicker. The second recognition dictionary may be stored over multiple forms.

〔Effect of the invention〕

According to the present invention, the character pattern itself in the form is used to prevent misreading that occurs when the quality of input characters is low or when a word that does not exist in the word dictionary used in post-processing is input. Since the recognition is performed by
As a result, high reading accuracy can be ensured, making it suitable for practical use.
-, it can have a tremendous effect.

[Brief explanation of the drawing]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention, FIG. 2 is a configuration diagram of the re-recognition section of the present invention, FIG. 3 is a diagram showing the data flow of the present invention, and FIG. 4 is the original FIG. 5 shows how patterns and dictionary patterns are matched. FIG. 5 shows how a contact character pattern is cut out. 1... Image input section 2... Character detection section 3.
...1st character recognition section 4...Post-processing section 5...
Re-recognition Department Agent Patent Attorney Nori Chika Ken Yudo Mitsuyuki Matsuyama Tadashi

Claims (1)

[Claims] an image input section for inputting an image including a character pattern; a character detection section for detecting and cutting out characters from the input image; a first character recognition section for recognizing the cut out characters; a post-processing unit that corrects misread parts using knowledge of the above; an estimation unit that estimates misread characters by comparing the post-processing results with the recognition results; and an estimation unit that estimates misread characters by comparing the post-processing results with the recognition results; A character that is characterized by comprising: a dictionary creation unit that creates a second character recognition dictionary; and means that uses the second character recognition dictionary to perform at least one of character detection and second character recognition. recognition device.