JPH0337781A - Post-processing method for character recognition results - Google Patents

Abstract

PURPOSE:To curtail the capacity of a storage memory by setting a unit of the processing to one line, and also, shifting a delimiter part of the last word and sentence clause, etc., of a line being an object of the processing to the head of the next line. CONSTITUTION:When the character recognition of a one-line portion is ended, a recognition result reading-in part 21 reads in recognition of a one-line portion from a recognition result memory 8, and stores it in a line recognition result memory 22. Subsequently, a temporary sentence clause segmenting part 23 delimits a recognition result character-string of one line in the line recognition result memory 22 into plural parts being analytic units by a position in which the kind of a character is varied, etc., and an analytic range determining part 24 considers a position of a temporary sentence clause, and sets a head character position of the line recognition result memory 22 to a character position of this side of the temporary sentence clause to an analytic object range. Next, a sentence analyzing/correcting part 25 rewrites correctly a result of recognition in the line recognition result memory 22, and a final sentence clause writing part 26 executes an operation for shifting a result of recognition (the part excluded from an object to be analyzed and corrected) of the temporary sentence clause of the line end in the line recognition result memory 22, to the head. In such a way, the storage capacity can be curtailed.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to single character recognition, and particularly to post-processing for modifying character recognition results.

[Conventional technology]

It is extremely difficult to completely eliminate misrecognition by character recognition in units of royal characters, so in kanji OCR,
Error correction (post-processing) of character recognition results is often performed by word matching, morphological analysis, etc.

Conventionally, the post-processing method for general documents is to post-process all character strings in the area after the recognition of the area to be recognized (for example, a paragraph) is completed, assuming that there are no line breaks in that area. It is supposed to be done. Alternatively, there is also a method in which a sentence is the unit of post-processing, in which the character string from the beginning of the sentence to a period or comma is read and post-processed, ignoring line breaks.

Publicly known materials regarding post-processing of character recognition results of Japanese documents using word matching and morphological analysis include, for example, "Nishino et al.: Realization of post-processing for Japanese document readers, Natural Language Processing 64-6 (1987, November 2013). 20), PP, 45-
There is 52J.

[Problem to be solved by the invention]

However, with post-processing methods that treat paragraphs or sentences as processing units, if the paragraphs or sentences are long, the post-processing has to wait until the recognition results are finished, resulting in poor processing efficiency. The problem is that it requires a large amount of memory.

Therefore, an object of the present invention is to provide a post-processing method for character recognition results that can solve such problems.

[Means to solve the problem]

In the post-processing of character recognition results to correct errors by word matching, morphological analysis, etc., the unit of processing is one line, and the delimiter of words, phrases, etc. at the end of the line to be processed is It is characterized by moving the line to the next line, or moving the part at the end of the line that cannot be parsed to the beginning of the next line.

[For production]

The character string in the upper line is divided into words, phrases, etc. and analyzed, but one word may be split into the end of a line and the beginning of the next line. Words split by line breaks like this cannot be analyzed correctly if they are handled without taking this into consideration, and if there is a misrecognition in that part, correction will fail.

However, according to the present invention, post-processing is performed on a line-by-line basis, but the word or clause separation part at the end of a line, or the unanalyzable part at the end of a line, is moved to the beginning of the next line, and the next line is post-processed. , it is possible to correctly correct misrecognitions of words that are split due to line breaks.

Further, in general, in a character recognition device, processing proceeds in the following order: ``area extraction or area specification → ``line extraction'' → ``character extraction'' → ``character recognition'' → ``post-processing.'' Therefore, if the unit of post-processing is one line, the post-processing of one line and the character recognition process of the next line can be executed in parallel, reducing the waiting time for post-processing and speeding up the entire process. can do. Furthermore, since the number of characters printed or written in the upper line is approximately fixed, the amount of memory required to store recognition results can be reduced.

〔Example〕

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows a block diagram of Kanji OCR according to the present invention.
The overall structure and processing flow of this Kanji OCR is the same as the conventional one. That is, one image input section reads a target image from a scanner (or other input device 2) and stores it in the image memory 3.

The range specifying unit 4 automatically specifies the range to be recognized from the image input to the image memory 3, or specifies it by a specifying means. A cutting unit 5 cuts out a line image from the specified range, and cuts out individual character images from this line image.

The recognition unit 6 uses the dictionary stored in the character dictionary memory 7, performs character recognition on each cut-out character image, and stores the recognition results in the recognition result memory 8.

The post-processing unit 9 reads the recognition result character string from the recognition result memory 7, analyzes it with reference to the word dictionary and grammar dictionary stored in the word dictionary/grammar dictionary memory 10, and corrects portions that are inappropriate as sentences. do. This post-processing section 9 is a section directly related to the present invention, and its details will be described later for each embodiment. The output unit 11 outputs the post-processing results to a CRT or an output file 12.

4L The block diagram of the post-processing unit 9 in this embodiment is shown in the second @. When the character recognition for the upper line is completed, the recognition result reading unit 21 reads the recognition result for one line from the recognition result memory 8. One line recognition result memory 22 is stored. Note that if the end analysis part of the end of the previous line has been written to the beginning of the line recognition result memory 22, it is written from the next character position.

Next, the provisional phrase cutting unit 23 selects the t in the line recognition result memory 22.
The line recognition result character string is divided into a plurality of parts as units of analysis, depending on the position where the character type changes (for example, switching from hiragana to non-hiragana). The parts separated in this way will be called "kari bunsetsu".

The analysis range determining unit 24 takes into account the position of the pseudo clause and sets the range from the first character position in the line recognition result memory 22 to the character position before the last pseudo clause as the analysis target range.

The sentence analysis/correction unit 25 uses the word dictionary and grammar dictionary in the word dictionary/grammar dictionary memory 10 for each temporary clause in the set analysis target range.

Word matching and morphological analysis are performed to correct erroneous recognition, and the recognition results in the one-line recognition result memory 22 are rewritten correctly.

The thus corrected results are output by the output unit 1.

When the analysis and correction are completed up to the last pseudo clause in the analysis target range, the final clause writing unit 26 writes the recognition result of the pseudo clause at the end of the line in the line recognition result memory 22 (the part excluded from the analysis and correction target) to Perform the operation to move to the top. In other words, the pseudo clause at the end of the line may be separated by a line break, but
This will be treated as a provisional phrase at the beginning of the next line, and will be subject to post-processing on the next line.

Note that while the post-processing for the upper line is being executed, character recognition processing for the primary line and subsequent lines is executed in parallel.

Let's take the following target sentence as an example, and assume that the following recognition result is obtained for the I-th line.

“Character recognition technology is modern information processing.” This is divided into temporary clauses as follows by character type change (hiragana → non-hiragana).

``Character recognition technology is / modern / information processing'' And the scope of analysis is as follows 6 ``Character recognition technology is / modern / information processing'' Now, the word that was split by line breaks ``information processing (processing)''
is excluded from the analysis target in the Ith line.

When the analysis and correction of the analysis target range of the first line is finished and the results are output, the unprocessed pseudo clause part at the end of the line is transferred to the beginning of the line recognition result memory 22, and the subsequent character positions are is cleared.

The process moves to the second line, and the recognition result is stored in the line recognition memory 2.
It will be written following the part moved from the previous line of 2, so
The recognition result in the second line is as follows.

``This is the star technology of information processing.'' With this, the word ``information processing'', which was split across two lines, has been unified into one, making it possible to analyze and modify it.

Then, divide it into pseudo clauses in the same way, set the target range, analyze and modify. However, since the last character of this line is a period, it is determined that there is no word splitting at the end of the line. The last pseudo clause is treated in the same way even if the last character of the line to be analyzed is a comma.

The block diagram of the post-processing unit 9 in this embodiment is shown in FIG. 3. When character recognition for one line is completed, the recognition result reading unit 31 reads the recognition results for the upper line from the recognition result memory 8. , are stored in the line recognition result memory 32. Note that if the unanalyzable part at the end of the previous line has been written to the beginning of the line recognition result memory 32, the data is written from the next character position.

Next, the sentence analysis/correction unit 33 uses the word dictionary and grammar dictionary in the word dictionary/grammar dictionary memory 10 to perform correction processing on the contents of the line recognition result memory 32 through word matching and morphological analysis, and performs line recognition. The recognition results in the result memory 32 are rewritten correctly. The result corrected in this way is outputted by the output unit 11. However, if there is a part at the end of a line that cannot be parsed (a character string for which the corresponding word cannot be found), the end of the line is marked as ``continued to the next line.''

The next line mochikoshi part writing unit 34 writes the end of the line marked with the 7th line mochikoshi mark to the beginning of the line recognition result memory 32, and clears subsequent character positions. Now, the unanalyzable portion at the end of the current line is moved to the beginning of the next line, and is subjected to post-processing of the next line.

It should be noted that while the post-processing for one line is being executed, the character L& recognition process for the primary line and subsequent lines is executed in parallel.

Let's take the following target sentence as an example and assume that the following recognition result is obtained for the first line.

``Character recognition technology is the modern information department.'' If this character string is analyzed from the beginning by word matching or morphological analysis, it will look like this (
/ is a word break).

``Character/Recognition/Technology/is/Modern/'s/Information/Processing This last character'' process cannot be parsed because there is no corresponding word, so this ``processing is carried over to the next line, and the corrected result output is It will look like this:

``Character recognition technology is modern information'' Next, the ``processing is moved to the beginning of the next line.

The recognition result of the second line is ``It is a star processing technology.

The words that were split across two lines are now connected. Since the last character of this line is a period, there is no part that cannot be parsed until the end of the line, and the line is processed until the end.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to correct recognition results including words separated by line breaks,
It is possible to improve the efficiency of the entire process by performing post-processing and recognition processing in parallel row by row, and since the unit of post-processing is one row, the memory capacity for storing recognition results for post-processing is reduced. reduction is possible.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a Kanji OCR according to the present invention, FIG. 2 is a block diagram of a post-processing unit in one embodiment of the present invention, and FIG. 3 is a block diagram of a post-processing unit in another embodiment of the present invention. It is. 6... Recognition unit, 8... Recognition result memory, 9...
Post-processing unit, 10... word dictionary/grammar dictionary memory,
21... Recognition result reading unit, 22... Line recognition result memory, 23... Provisional phrase cutting unit, 24...
Analysis range determination unit, 25...Text analysis/correction unit, 2
6... Final phrase writing unit, 31... Recognition result reading unit, 32... Line recognition result memory, 33...
Sentence analysis/correction section, 34... Next line sticky part writing section. Bus Figure 2

Claims (2)

[Claims] (1) In post-processing to correct errors in character recognition results,
A post-processing method for character recognition results characterized in that the unit of processing is one line, and a delimiter such as the last word or clause of the line being processed is moved to the beginning of the next line. (2) In post-processing to correct errors in character recognition results,
A post-processing method for character recognition results characterized in that the unit of processing is one line, and the part at the end of a line that cannot be analyzed is moved to the beginning of the next line.