JPH03198180A - Post-processing method for character recognition - Google Patents

Abstract

PURPOSE:To improve the probability that a correct recognition result is obtained by putting the priority to a word candidate to a result of character recognition in accordance with its appearance frequency. CONSTITUTION:In a word dictionary (17), in addition to a notation of a word, information of an appearance frequency in a document of the word is also registered, and a post-processing part 16 retrieves its appearance frequency information, as well at every word candidate, and sorts the word candidate in order of a higher appearance frequency. That is, the priority of the word candidate is put in order of a higher appearance frequency. In such a way, even in the case the candidate order of a character candidate is not appropriate, the word candidate in which possibility of a correct answer is generally high is processed preferentially, therefore, the probability that a correct recognition result is obtained is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to character recognition devices such as kanji OCR, and particularly to a post-processing method for recognition results obtained by matching with a character dictionary.

[Conventional technology]

Because kanji and other characters have an extremely large number of similar characters, there is a limit to the accuracy rate of character-by-character recognition processing. Therefore, Kanji OCR
etc. have a post-processing section next to the character recognition section, and the character recognition results are often corrected by examining the validity of the word or sentence in this post-processing section.

In such post-processing, multiple word candidates are often obtained for the character recognition results, and in this case, it is necessary to prioritize the word candidates9. For each candidate word obtained by matching it with the word dictionary, weights are assigned according to the ranking, and the sum of the weights assigned to the candidate characters that make up the word is calculated, and the candidate words are assigned in descending order of the value. There is a method of prioritizing (Special Publication No. 1-19195).

[Problem to be solved by the invention]

However, the candidate ranking based on character recognition on a character by character basis does not necessarily correspond to the ranking of words, so words that appear extremely rarely are given priority, or even frequently occurring characters are given a low candidate ranking. However, there was a problem in that words containing this word were given a low ranking, and as a result, correct recognition results could not be obtained.

Therefore, an object of the present invention is to solve the above-mentioned problems in post-processing.

[Means to solve the problem]

The present invention registers not only the notation of words but also the frequency of appearance of each word in a document in a word dictionary, and when multiple word candidates are obtained by comparing the character recognition results with the word dictionary, It is characterized by prioritizing each word candidate according to appearance frequency information.

[For production]

Since word candidates are prioritized according to their frequency of appearance as words, even if the candidate ranking of character candidates is not appropriate, priority is given to word candidates that are generally more likely to be correct, increasing the probability of obtaining correct recognition results. improves.

〔Example〕

FIG. 1 is a schematic block diagram of Kanji OCR according to an embodiment of the present invention, and FIG. 2 is a schematic flowchart of processing.

In FIG. 1, image data of a document to be processed is input from an image input unit 10 and stored in an image memory 11 (
Processing step ■). This input image data is subjected to line cutting and character cutting by the line/character cutting unit 12, and the cut out character image data is stored in the character image memory 13.
(processing step ■), the character recognition unit 14 performs preprocessing such as size normalization and feature extraction, and stores the extracted feature vectors and the character dictionary memory 15 in the character dictionary memory 15. By matching with the obtained character dictionary, character candidates from the smallest distance to the highest number N are selected (processing step ■).

The processing of the post-processing unit 16 will be described using an example in which recognition and post-processing of a character string "...character..." included in a general document is performed.

For example, it is assumed that the recognition results shown in Table 1 are obtained by the processing of the character recognition unit 14 (character recognition on a character-by-character basis).

Table 1 (Character recognition results) The post-processing unit 16 selects word candidates that match the combination of character candidates by comparing the character recognition results with the word dictionary stored in the word dictionary memory 17 (processing step ■). Here, it is assumed that word candidates shown in Table 2 are obtained, for example.

Table 2 (Word Candidates/Frequency) However, when multiple word candidates are obtained, it is not possible to determine which of them is closest to the correct answer. Therefore, according to the present invention, in addition to the notation of the word in the word dictionary (17), information on the frequency of occurrence of the word in the document is also registered, and the post-processing unit 16 also records the frequency of occurrence information for each word candidate. Search and sort word candidates in descending order of frequency of appearance (processing step ■). That is, word candidates are prioritized in descending order of frequency of appearance. Assuming that the appearance frequencies are as shown in Table 2, the sorting results are as shown in Table 3.

Table 3 (word candidate ranking after sorting) Then, the post-processing unit 16 rewrites the character recognition result with the word candidate that has been given the highest priority (processing step ■) 1. In this example, the "r character" has been given the highest priority. Therefore, the first candidate for character number 1 is rewritten as "sentence", and the first candidate for character number 2 is rewritten as "character".

The recognition result corrected by such post-processing is stored in the result memory 19 and output by the result output unit 18 (processing step [phase]).

Note that if there is only one word candidate, that word candidate is automatically given top priority and correction is performed (processing steps ①, ①).

Note that the present invention can also be applied to post-processing of speech recognition.

[Effects of the Invention] As explained above, according to the present invention, by prioritizing word candidates for character recognition results according to their frequency of appearance, it is possible to avoid giving priority to words that rarely appear,
By prioritizing words that have the highest frequency of appearance and generally have a high probability of being correct, it is possible to increase the probability of obtaining correct recognition results.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram of Kanji OCR according to an embodiment of the present invention, and FIG. 2 is a schematic flowchart of processing. 10... Image input section, 12... Line/character cutting 13... Character image memo 15... Character dictionary memo 17... Word dictionary memo 19... Result memory. 11...Image memory, part, 14...Character recognition part. S, 16... Post-processing section, S, 18... Result output section,

Claims (1)

[Claims] (1) In a character recognition post-processing method in which word candidates are obtained by comparing the character recognition results for each character with a word dictionary in which the words are registered, and the character recognition results are corrected according to the word candidates, each word is added to the word dictionary. Information on the frequency of occurrence of words in documents is registered, and when multiple word candidates are obtained by comparing the character recognition results with the word dictionary, it is possible to prioritize each word candidate according to the frequency of occurrence information. Characteristic post-processing method for character recognition.