JPH0652366A - Post-processing method for character recognition result - Google Patents

Abstract

PURPOSE:To retrieve a candidate character with high efficiency without consuming much time by setting dictionary side character classification and recognition side character classification as separate ones based on the asymmetric property of error trend. CONSTITUTION:The same classification name is attached on classification conformed as the error trend in the dictionary side character classification and the recognition side character classification. Difference between the dictionary side character classification with the same classification name and the recognition side character classification is generated by considering the asymmetric property of the error trend. For example, since such possibility that a character (k) is recognized erroneously as a character (t) in the classification name B, the character (t) is classified to the recognition side character classification with the same classification name as that of the dictionary side character classification including the character (k), however, since the character (t) is not recognized erroneously as the character (k), the character (t) is removed from the dictionary side character classification. Also. characters (a), (o), and (c) are included in the recognition side character classification, and the characters (a), (o) in the dictionary side character classification in the classification name A, while, the character (c) is included in the recognition side character classification, and the characters (c), (o) in the dictionary side character classification, which conforms to the asymmetric property.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a post-processing method for character recognition results, and more particularly to improving recognition performance by searching for words in consideration of error tendency of the recognition results.

[0002]

2. Description of the Related Art For example, as a machine translation system, there is a machine translation system in which a character recognition device is applied to an input means to simplify an input operation by a user. For example, in the character recognizing device applied in this way, of course character-by-character recognition,
Word-by-word recognition is also important. In the past, a method has been already proposed in which a correct recognition result can be obtained for each word even if there is an erroneous recognition result in recognition for each character.

For example, reference is made to “Numakura et al.,“ Information Retrieval System Retrievable with Incorrect Key ””, Journal of Information Processing Society of Japan, Vol. 30, No. 11, pp. 1468-1478, November 1989, for example. be able to.

Hereinafter, a method for searching for a word according to the method disclosed in the above-mentioned document for obtaining a more accurate recognition result for a recognition target word from the recognition result for each character constituting the recognition target word (after the character recognition result) Processing which constitutes a part of the processing method) will be described.

When applying this method, it is necessary to prepare in advance a character classification and a hierarchical word dictionary classified by the notation of each word. Here, the character classification is a classification of all characters into some classes based on the error tendency of the characters, and each class is given a class name. The word class name notation is a notation formed by arranging the class names to which the characters that make up the word belong.

When recognizing a word, first, a class name notation X is created by replacing the recognition result of each character forming the recognition target word with the class name of the character classification. Next, the class name notation X having the highest degree of coincidence with the class name notation X thus obtained is searched from the class name notation of the above-mentioned word dictionary. Finally, a word in the above-mentioned word dictionary having the class name notation Y is set as a search target, a word having the highest degree of coincidence with the above-described recognition result is searched, and the word is set as the search result.

[0007]

However, the above-mentioned conventional method exhibits a symmetrical error tendency that a certain character A may be erroneously recognized as the character B and vice versa. Works well in character recognition devices,
This is not always satisfactory for a character recognition device that exhibits an asymmetric error tendency in which a certain character A may be erroneously recognized as the character B, but the character B is not erroneously recognized as the character A.

The character recognition device obtains a recognition result based on the distance from the feature amount of the input pattern and the feature amount of the reference pattern. The pattern relating to the feature amount calculation of the reference pattern and the character of the recognition target document are obtained. When the pattern is different, an asymmetric error tendency is often exhibited.

The above-mentioned character B may be mistaken for the character C without being mistaken for the character A. Even if there is no error between the character A and the character C, the characters A, B and C are in the same class in the past. being classified. In this case, there are many words having the same class name notation in the word dictionary. As a result, the number of candidate words obtained by searching the word dictionary increases, and moreover, the candidate words include many unnecessary words that need not be actually considered. Therefore, the process of finding a candidate word and the process of determining an output word from the candidate word also take a lot of time, and even if the output word is determined, it is not a valid word in many cases, and the opportunity for the user to rework increases. The problem arises.

For example, if a recognition process is performed such that k may be erroneously recognized as t but t is not erroneously recognized as k, an error tendency of erroneously recognizing k as t is expressed. Must conventionally classify k and t into the same class. In this way, the word "t
When “ake” is erroneously recognized as “state”, “ta” is included in the candidate words retrieved from the reference word “state”.
ke ”is included. However, if such a character classification is created, even if the word “bake” is erroneously recognized as “kake”, t and k are classified into the same class, and therefore the reference word “kake” is searched. "Take" is included in the selected candidate words, and it is necessary to determine whether or not the candidate word "take" should be an output word.

The present invention has been made in consideration of the above points, and even when character recognition has an asymmetric error tendency, it is possible to efficiently search for candidate words in a short time, An object of the present invention is to provide a post-processing method for character recognition results that can improve both the processing time and accuracy of post-processing for recognition results.

[0012]

In order to solve such a problem, in the present invention, a character classification in which a class name is assigned to each classification in which a character is classified based on an error tendency at the time of recognition and a word are formed. A process for creating a reference word that is referred to when searching the word dictionary from the character recognition result, and a class name description of the reference word In the post-processing method of character recognition results including the process of obtaining, and a process of searching a word dictionary using the class name notation of the reference word as a key, a dictionary side character classification for creating the class name notation in the word dictionary, It is characterized in that the recognition-side character classification for obtaining the class name notation from the reference word is made different based on the asymmetry of the error tendency.

[0013]

According to the present invention, if a word dictionary can be searched based on an asymmetric error tendency, candidate words can be searched efficiently without spending time, and the processing time and accuracy of post-processing for recognition results can be improved. Was made based on. Then, in order to be able to search the word dictionary based on the asymmetrical error tendency, the dictionary side character classification for creating the class name notation in the word dictionary and the recognition side character for obtaining the class name notation from the reference word The classification and the two were made different based on the asymmetry of the error tendency.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A post-processing method for character recognition results according to the present invention will be described in detail below with reference to the drawings regarding an embodiment in which it is applied to the recognition of English words.

Although not shown in the drawings, this embodiment is practically realized in hardware, for example, by a workstation equipped with an optical character reader (OCR). The functional block is as shown in FIG.

In FIG. 2, the document reading means 11 reads a document on a sheet as binary data. The character recognition means 12 cuts out each character area, each word area, and the like based on the binary data, obtains the characteristic amount of the binary pattern of each character area, and creates the reference feature for each character in advance. The recognition result of each character is obtained according to the distance from the amount, and is stored in the recognition result storage means 13 for each word (region). The reference word creating means 14 creates a reference word to be referred to at the time of search based on the recognition result for each character. The character classification storage unit 15 stores a character classification in which all characters are classified by allowing duplication based on the error tendency of each character, together with a class name (classification name). The word dictionary 16 stores at least English words that appear to appear in general sentences, and is also hierarchized by class name notation. Candidate word search means 17
Is a word dictionary that obtains a class name notation that consists of the class names of the characters that make up the reference word and that has a class name notation that matches this class name notation and that has the largest number of matching characters with the reference word. 16 is searched, and the searched candidate word is temporarily stored in the recognition result storage means 13. The output word determining means 18 determines an output word from the candidate words using the initial recognition result, and causes the result output means 21 to print or display.

FIG. 3 shows an example of the structure of a word dictionary. In the word dictionary, words having the same word length are hardened, and further, the same kind name is hardened for each word. In addition, a hash key is provided for each class name notation in order to speed up the search. As the hash key, for example, the remainder obtained by dividing the sum of the ASCII codes of the class names (characters) in the class name notation by 32 is applied.

In the case of this embodiment, character classification (hereinafter referred to as dictionary side character classification) for creating class name notation in such a word dictionary and class name notation of reference word obtained from the recognition result are described. One of the characteristics is that the character classification for obtaining (hereinafter referred to as the recognition-side character classification) is different.

FIG. 1 shows the dictionary side character classification and the recognition side character classification. Regarding the classification corresponding to the error tendency in the dictionary side character classification and the recognition side character classification,
The same class name is attached. The difference between the dictionary side character classification having the same class name and the recognition side character classification is that the asymmetry of the error tendency is taken into consideration. The recognition-side character classification focuses on classifying characters having the same error tendency at the time of recognition. On the other hand, the dictionary-side character classification classifies characters that have the same error tendency at the time of recognition and exclude those that should not be regarded as belonging to the same classification when searching for candidate characters.

For example, in the class name B, since the character k may be erroneously recognized as the character t, the character t is classified into the recognition side character classification having the same class name as the dictionary side character classification including the character k. , The character t is not erroneously recognized as the character k, so the character t is excluded from the dictionary side character classification. Further, although the characters a and o may be erroneously recognized as the character c, the character c may be erroneously recognized with respect to the character o, but the character a is not erroneously recognized. In the class name A, the recognition-side character classification includes the characters a, o, and c, and in the dictionary-side character classification, the characters a and o are included. On the other hand, in the class-name C, the recognition-side character classification includes the character c, and the dictionary-side character. The letters c and o are included in the classification so as to comply with the asymmetry described above.

In order to retrieve and obtain candidate words,
It is sufficient that at least the class name and the recognition-side character classification are stored. The dictionary-side character classification is sufficient if it can form a class name notation in the word dictionary, and is not directly used when searching for candidate words in this embodiment.

The process flow of the embodiment method which is realized by the functional blocks as described above and uses the word dictionary and the character classification having the above-described configuration is as shown in FIGS. 4, 5 and 6.

First, a large processing flow of the embodiment method will be described with reference to FIG.

Character regions, word regions, etc. are cut out from the document data converted into binary data, the characteristic amount of the binary pattern (character pattern) of each character region is obtained, and the reference feature created in advance for each character. The recognition result of each character is obtained according to the distance from the amount, and the recognition result of each character is arranged for each word area (step 100). FIG. 7 shows an example of the character recognition result, and is an example in which the input recognition target words (correctly, their patterns) are “take” and “bake”. Here, even if the character to be read this time and the character at the time of creating the reference feature amount are the same character, the feature amount differs due to the difference in the character pattern such as the font, so the distance is not always the minimum. Therefore, an incorrect character recognition result may be obtained. Here, all the characters whose distance is less than or equal to a certain threshold value are extracted as the recognition result.

When the recognition result is obtained for each character forming the word in this way, a reference word is formed by connecting the first candidate characters for each character (step 101).

FIG. 8 is a table showing information reference words, class name notations, and candidate words) obtained as the processing steps progress. As shown in FIG. 8, when the recognition result as shown in FIG. 7 is obtained, the reference words are “state” and “kak”.
e ”.

When such a reference word is obtained, a class name notation is obtained by using the stored recognition-side character classification, and a word dictionary is searched using this class name notation as a key to obtain candidate words ( Step 102). The processing of this step also has a feature of this embodiment, and the details will be described later.

When one or more candidate words are obtained in this way, the word to be output is determined using the recognition result obtained in step 100 (step 103). For this output word determination processing, for example, the method described in Japanese Patent Application No. 3-196509 and drawings can be used.
That is, the reference feature amount of each character of the candidate word is used to obtain the distance from each character of the recognition target word, and the sum of the distances obtained for each character is used as the evaluation value of this candidate word. Then, the candidate word having the smallest evaluation value is set as the word to be output.

Then, the determined output word is output by printing or displaying and a series of processing is completed (step 1).
04).

Next, the word dictionary search processing in step 102 will be described in detail with reference to FIGS. 5 and 6.

When the process of searching the word dictionary is started, first, all class name notation in which each character of the reference word is replaced with the class name of the recognition side character classification to which it belongs is created (step 200).
This creation process will be described later.

When one or more class name notations are obtained in this way, the following loop processing is repeated for each class name notation. First,
One class name notation is taken out (step 201), a hash value of the class name notation is calculated, and the hash value is used as a key to target the word group of the word dictionary related to the class name notation, in which the reference word All words having the maximum number of characters that match with are extracted (step 202). Then, the number of matching characters this time is compared with the number of matching characters of the words registered as candidate words up to now (step 203). If there is a match, the word extracted for this class name notation is added to the candidate words so far (step 204), and if there is a large number of matching characters this time, the previous word candidate is discarded and the word extracted this time. Is registered as a candidate word (step 205).
When the number of matching characters is small this time, or when the addition of the candidate word or the change of discarding the candidate word is completed, it is confirmed whether there is any class name not yet searched (step 20).
6) If it remains, the process returns to step 201 described above, and if the search process is completed for all the class name notations, the process proceeds to the above-described output word determination process (step 103).

In the case of this embodiment, since a character may belong to a plurality of character classes on the recognition side, the process of creating the class name notation in step 200 is not as simple as the conventional one, and is as shown in FIG. is there. Note that the class name notation creating process shown in FIG. 6 is a recursive process that calls the class name notation creating process shown in FIG.

First, it is judged whether or not the number of characters of the reference word is 0 (step 300), and if it is not 0, one character is deleted from the end of the reference word to make it the reference word x (step 302). Then, for the reference word x, the process proceeds to a self-processing routine for creating the class name notation x in FIG. 6 (step 302). On the other hand, if the number of characters in the reference word is 0,
An empty character string is set as the class name notation and the process ends (step 307).

When step 302 described above is completed, the following processing loop is repeated for each class of recognition-side character classification. One recognition-side character classification is extracted and it is determined whether or not the last character of the reference word belongs to that classification (steps 303 and 304). When it belongs, the class name is added to the end of the class name notation x (step 30).
5). When such additional processing is completed, or when the last character of the reference word does not belong to the classification,
It is confirmed whether or not the processes for all the recognition-side character classifications have been completed (step 306). If not completed, the process returns to step 303 described above, and if completed, a series of processes is completed.

FIG. 9 shows an example of changes in processing steps until the processing of FIG. 6 is completed, and shows a case where the initial reference word is “state”.

When the process shown in FIG. 6 is first executed, the number of characters of the reference word is 4, so that the reference word x is "t".
At "is created, and the processing shown in FIG. At this time, since the reference word “tat” has three characters, “ta” is created as the reference word x,
The process shown in FIG. 6 is reentered by step 302. At this time, the number of characters of the reference word "ta" is 2, so "t" is created as the reference word x, and the process shown in FIG. At this time, since the number of characters of the reference word “t” is 1, a blank character is created as the reference word x, and the process shown in FIG. At this time, since the number of characters of the reference word is 0, the class name is converted into an empty character string in step 307, and the number of recursion is one step before in step 3 in FIG.
Return to 03.

At this time, since the reference word is “t” and the class name notation x is an empty character string, a class name notation in which the class names B and D of the classification having the character t as an element are added to the class name notation x is created. Then, step 30 in the process of FIG.
Return to 3.

At this time, since the reference word is "ta" and the class name notation x is B or D, the class name A having the last character a of the reference word as an element is changed to the class name notation x B or D. The added class name notations BA and DA are created, and the process returns to step 303 in the process of FIG.

At this time, since the reference word is “tat” and the class name notation x is BA or DA, the class names B and D of the classification having the last character t of the reference word as an element are the class name notation x.
Class name notation added to A and DA respectively BAB, DAB,
BAD and DAD are created, and the process returns to step 303 in the process of FIG.

At this time, since the reference word is "state" and the class name notation x is BAB, DAB, BAD, DAD, the class name C of the classification having the last character e of the reference word as an element is the class name notation x. Name designations BABC, DABC, BADC, D added to BAB, DAB, BAD, DAD, respectively
The ADC is created, and the creation process is completed for a series of class name notations.

Based on the class name notation thus created, the word dictionary is searched as described above to obtain candidate words.

In FIG. 8 described above, the reference word is "state".
Also, the obtained class name notation and the candidate word are also shown depending on whether the reference word is “ake” or.

According to the above embodiment, in consideration of the asymmetry of the error tendency, the dictionary side character classification relating to the creation of the class name notation of the word dictionary and the recognition side character classification creating the class name notation from the reference word are performed. Since they are separate, the word dictionary can be searched based on the asymmetric error tendency.

In the case of the reference word "ate" described above, the word "take" in which the second t is replaced with k is also included in the candidate words. On the other hand, in the case of the reference word “kake”, “take” obtained by replacing the first k with t is not included in the candidate words. That is, a search is performed based on an error tendency that k can be erroneously recognized as t, but t is not erroneously recognized as k.

By executing the search based on such an asymmetric error tendency, it is possible to prevent unnecessary words from becoming candidate words, improve the accuracy of the determined output word, and increase the accuracy of the output word. The correction operation can be reduced, and the overall recognition processing time can be shortened on average.

In the above embodiment, the case where an English word is a recognition target has been described, but the present invention is also applicable when a word in another language is a recognition target. For example, in the case of Japanese, the extraction of words and the recognition of words are performed in parallel, but the present invention can be applied.

In the above embodiment, one reference word is set to obtain the candidate word, but two or more reference words are set from the character recognition result to obtain the candidate word. good.

[0049]

As described above, the dictionary side character classification for creating the class name notation in the word dictionary and the recognition side character classification for obtaining the class name notation from the reference word are asymmetric with respect to the error tendency. Since they are separate based on each other, even if there is an asymmetric error tendency in character recognition, it is possible to efficiently search candidate words without spending time, and the processing time and accuracy of post-processing for the recognition result can be improved. It is possible to realize a post-processing method of a character recognition result that can improve both of the above.

[Brief description of drawings]

FIG. 1 is an explanatory diagram showing dictionary-side character classification and recognition-side character classification according to an embodiment.

FIG. 2 is a functional block diagram showing a configuration to which an embodiment is applied.

FIG. 3 is an explanatory diagram showing a word dictionary configuration of the embodiment.

FIG. 4 is a flowchart showing a schematic process of an embodiment.

FIG. 5 is a flowchart showing dictionary search processing according to the embodiment.

FIG. 6 is a flowchart showing a process of creating a class name notation according to the embodiment.

FIG. 7 is an explanatory diagram illustrating an example of a recognition result of the example.

FIG. 8 is an explanatory diagram showing information obtained at each processing stage of the embodiment.

FIG. 9 is an explanatory diagram of a class name notation creating process according to the embodiment.

[Explanation of symbols]

 15 ... Character classification storage means, 16 ... Word dictionary.

Claims (1)

[Claims] 1. Hierarchization by a class name notation in which a class name is assigned to each class in which characters are classified based on an error tendency at the time of recognition, and a class name of a class to which a character forming a word belongs is arranged. And a process of creating a reference word to be referred to when searching the word dictionary from the character recognition result, and
In a post-processing method of character recognition results including a process of obtaining a class name notation of a reference word and a process of searching a word dictionary using the class name notation of a reference word as a key, a class name notation in the word dictionary is created. A post-processing method for character recognition results, characterized in that the character classification on the dictionary side and the character classification on the recognition side for obtaining the class name notation from the reference word are separate based on the asymmetry of the error tendency.