JPH04120679A - Alphabetical character recognizing device - Google Patents

Abstract

PURPOSE:To reduce the processing time of the title device by finding the certainty of a word by adding a correcting value to the word on the basis of the product of the recognizing accuracy of each character and statistically obtained neighboring frequency of each character and giving priority to spell checking candidate character strings. CONSTITUTION:A picture inputting section 1 inputs a document to be recognized as a document picture. A sentence segmenting section 2 finds the histogram of black picture elements of the document picture in the vertical and horizontal directions and discovers a sentence area. A word area segmenting section 3 segments characters between spaces having widths larger than a certain value as a word area. A character area segmenting section 4 divides a train of characters in the word area into individual characters by regarding the part where the histogram of black picture elements in the word area is lower than a certain value as a punctuation of characters and segments the divided characters as character areas. A character recognizing section 5 performs character recognition on each character. The recognizing accuracy of a word is found by adding correcting values to each other on the basis of the product of the recognizing accuracy of each character and neighboring frequency 8 of each character. Then spell checking 7 is performed on the words in the order of the recognizing accuracy.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to recognition of English characters.

(Prior art) In recent years, there has been an increasing demand for character recognition devices to be used as input devices for computers, etc., and character recognition devices that can efficiently obtain stable recognition results are essential for improving the performance of systems such as computers. It becomes.

In conventional recognition devices, when multiple candidate characters are obtained as a result of recognition of one input, in order to determine the correct character from among the candidate characters, several recognition devices are used based on the recognition results of words consisting of the characters before and after the candidate character. A character string was generated, and a method such as a spell check was used to determine the correct spelling of the character string, which was then used as the recognition result.

(Problem to be Solved by the Invention) However, in the conventional character recognition device described above, when generating character strings as word candidates, candidate character strings are ranked only based on the accuracy of recognition results, so When multiple characters within a word have many candidate characters, the ranking of candidate character strings is not necessarily appropriate, which increases the number of times spell checking and other processes are performed, leading to an increase in processing time.

(Means for Solving the Problems) In order to solve the above problems, the present invention calculates, for each combination of word candidate character strings, the product of the recognition accuracy of each character and the statistically obtained adjacent frequency of the characters. By adding the correction value calculated based on the original word and finding its accuracy as a word, and effectively ranking candidate character strings when spell checking, the number of spell checking processes etc. can be reduced and the processing time can be reduced. This is intended to shorten the time.

(Function) Therefore, according to the present invention, by ranking candidate character strings when performing a spell check based on the recognition accuracy of each character and the statistically obtained adjacent frequency of characters, an appropriate ranking can be achieved. By doing so, it is possible to reduce the number of times of processing such as spell checking and shorten the processing time.

(Embodiment) FIG. 1 shows a block configuration of an alphabet g H device in an embodiment of the present invention. In Figure 1, 1
is an image input unit that inputs a document to be recognized as a document image;
2 finds a collection of character strings from the input document image,
3 is a word area extraction section that finds a word-based break from the text area and outputs a range of one word as a word area; 4 is a word-area cut-out section that finds character-based breaks from the word area; 5 is a character recognition unit that outputs a plurality of candidate characters based on the graphical features extracted from the character area; 6 is a candidate output from the character recognition unit; The characters are stored for one word, and a correction value calculated based on the product of the recognition accuracy of each character and the statistically determined adjacency frequency of the characters is added to the character string obtained by combining them to form a word. The word candidate generation unit generates word candidates ranked as candidates by determining the accuracy of the word candidates. 7 checks the spelling of the word candidates with the highest ranking, and outputs the correct character string as a recognition result. 8 is a character adjacency frequency list that stores the adjacency frequencies of the aforementioned characters; 9 is an internal bus that connects each part from 1 to 7; lO
is an internal bus that connects the word candidate generator and the character adjacent frequency list.

Regarding the English character recognition device of this embodiment configured as described above, Fig. 2 shows a flowchart of the overall processing, and Fig. 3 shows a frequency table of two adjacent characters, which statistically calculates the adjacency frequency of characters. ing. Next, the operation of the embodiment will be explained.

A document to be recognized is input as a document image using the image input unit 1 (Sll). The input document image is sent to the text area clipping unit 2, which calculates a histogram of black pixels in the vertical and horizontal directions of the document image, finds a text area based on this, and extracts the text area. The location information of is stored as internal data (S12). The position information of the text area is sent to the word area extraction unit 3, and word area extraction processing is performed for the inside of the text area. The word area cutting unit 3 pays attention to the fact that the space before and after a word is larger than the space between characters within the word, and cuts out characters sandwiched between spaces north of a certain width as a word area. The positional information of all the word areas within the character area found by the character area cutting unit 4 is obtained and stored as internal data (S13). The position information of the word area is sent to the character area cutting unit 4, and character area cutting processing for the word area is performed. The character area extraction unit 4 pays attention to the changes in the histogram of black pixels in the word area, uses the portion of the histogram below a certain value as a character delimiter, and separates the series of characters in the word area into character units to create a character area. break the ice. The positional information of all character areas within the word area found by the word area extraction unit is obtained and stored as internal data (S14).

Based on the position data of the text area, word area, and character area obtained in steps 312 to 314 above, the character recognition unit 5
Recognition processing is performed on a character-by-character basis, and the recognition candidate characters and recognition accuracy, which are the results of the recognition processing, are stored as internal data for one word (815-517).

The following explanation will be given assuming that the recognition candidate characters and recognition accuracy for one word are determined as shown in the table below.

Word candidates are obtained by combining the obtained recognition candidate characters (Sl, 8).

In this example, the first, second, and third characters have only one candidate, and the fourth and fifth characters have two candidates, so the following four combinations are possible as word candidates.

Word candidate 1: appic Word candidate 2: appie Word candidate 3: appQc Word candidate 4: appQe Next, the accuracy as a word, that is, the word accuracy, is determined for all the determined word candidates (818-521). The word accuracy is obtained by summing the product of the recognition accuracy of each character that makes up the word and the correction value calculated based on the adjacent frequency of the characters, and the recognition accuracy is the one obtained in the character recognition process described above. , and the adjacency frequency of letters is determined by the frequency table of adjoining two-letter pairs (CorneW, RlW, : A 5t
atical Method of Spelli
ng Correction, Inf, Control
l, Vol, ]2°No, 2. (quoted from pp. 79-93). Note that the character adjacent frequency data is stored in a character frequency list.

First, the product of the recognition accuracy of each word candidate is calculated (Sl, 9).

The closer the recognition accuracy is to 1, the higher the probability that the recognition result is correct, and the recognition accuracy of the entire word is the product of the accuracy of all the constituent characters. The results obtained are shown below.

Word candidate 1: 1. OXl, OXl, OXo, 7X0
．． 6=0.42 word candidate 2: 1. Ox1. Ox1.
oxo, 7xo, 4=0.28 word candidate 3: 1. O
XI, OXI, OXo, 3X0.6=0.18 Word candidate 4: 1. OXl, OXI, OXo, 3X0.4=0
．． 12 Next, find the correction value based on the adjacent frequency (S 2
0).

Now, for a character that has multiple candidates in a word candidate (in this example, the fourth or fifth character), calculate the product of the adjacency frequency with the previous character and the adjacency frequency with the next character. If the character adjacency value is the word adjacency value, and the word adjacency value is the sum of the adjacency values of all characters that have multiple candidates in the word candidate, then the correction value is the word adjacency value of each word candidate. It is divided by the sum of adjacent values.

The results of determining the word adjacency values for each word candidate are shown below.

For word candidates - Adjacent frequency of the fourth character's character adjacent values p and 1 - 8 Adjacent frequency of 1 and C = 55 In other words, character adjacent value = 440 - Adjacent frequency of the fifth character's character adjacent values 1 and C = 55 Frequency of adjacency between C and space = 7 In other words, character adjacency value = 385 - Word adjacency value (character adjacency value for the fourth character) x (character adjacency value for the fifth character) = J6940Q Word candidate 2: - No. Adjacent frequency of the fourth character's character adjacent values p and i = 8 Adjacent frequency of 1 and e = 37 In other words, character adjacent value = 296 - Adjacent frequency of the fifth character's character adjacent values 1 and e = 37 e and space Adjacent frequency = 446 In other words, character adjacency value = 16502 - Word adjacency value (4th character adjacency value) x (5th character adjacency value) = 4884592 Word candidate 3: - 4th character Adjacent frequency between adjacent values p and Q = 29 Adjacent frequency between 2 and C = 8 In other words, character adjacent value = 232 ・Adjacent frequency between character adjacent values Q and C for the fifth character = 8 Adjacent frequency between C and space = 7 That is, character adjacency value = 56 - Word adjacency value (character adjacency value of the fourth character) x (character adjacency value of the fifth character) = 12992 Word candidate 4: - Character adjacency value p and Ω of the fourth character Adjacent frequency of = 29 Adjacent frequency of Q and e = 72 In other words, character adjacency value = 2088 ・Adjacent frequency of 5th character character adjacency value 2 and e = 72 Adjacent frequency of e and space = 446 In other words, character adjacency value = 32112 - Word adjacency value (character adjacency value of the fourth character) x (character adjacency value of the fifth character) = 67049856 Next, the results of determining the correction value based on the word adjacency values are shown.

The correction value is calculated as (adjacent value of each word candidate) / (sum of adjacency values of each word candidate), and the sum of adjacency values of each word candidate = 72116840, so the correction value of each word candidate is word candidate 1 Correction value: 169400/7211684
0=0.0023 Correction value of word candidate 2=4884592/7211684
0=0.0677 Correction value of word candidate 3= 12992772116840
= 0.0002 Correction value of word candidate 4 = 67049856/721168
40=0.9297.

Find the word probability of each word candidate.

Since the word accuracy is the recognition accuracy plus the correction value, the word accuracy of word candidate 1 word candidate 2 word candidate 3 word candidate 4 = 0.42 + 0.0023 = word accuracy of 0.4223 = 0.28 + 0.0677 = Word accuracy of 0.3477=0.18+O, word accuracy of 0O02=0.1802=0.12+0.9297=1.0497.

Next, a spell check process is performed for each word candidate obtained through the above process in descending order of word accuracy.

Spell check processing determines whether the character string of a word candidate is correct as an English word. If the checked result is correct, that word candidate is output as a recognition result, and if it is incorrect, the next word candidate is checked. do. If all the word candidates are incorrect, the word candidate with the highest recognition accuracy is output as the recognition result (822-8
26).

Therefore, the order of spell checking is as follows: ■ Word candidate 4 (a p p Q e) ■ Word candidate 1 (appic) ■ Word candidate 2 (appie) ■ Word candidate 3 (a p p Q c) Therefore, word candidate 4 (appQe) with rank ■ is spelled correctly, so “a”
l) pQe” is output as the recognition result.

Character recognition processing for a given document image is performed by repeating steps S15 to S26 for all word regions as explained above. (Effects of the Invention) As explained above, this invention allows This has the effect of reducing the number of times of processing and reducing processing time.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of an English character recognition device according to an embodiment of the present invention, Fig. 2 is a flowchart of the entire character recognition process, and Fig. 3 is a character adjacency frequency table showing statistical adjacency frequencies of characters. . l...Image input unit, 2...Character area cutting unit, 3...Word area cutting unit, 4...Character area cutting and comparing unit, 5...Character recognition unit, 6...Word Candidate generation section, 7... Spell check section, 8.
...Character adjacent frequency list, 9.10...Internal bus. Patent applicant Matsushita Electric Industrial Co., Ltd.

Claims (1)

[Claims] an image input unit that inputs a recognition target document; a text area cutting unit that outputs a text area from the input document image;
A word area extraction unit that outputs a word area from a text area, a character area extraction unit that outputs a character area from the word area, and a character recognition unit that outputs a plurality of corresponding candidate characters based on the graphical characteristics of the character area. , the candidate characters output from the character recognition unit are stored for one word, and for the character string obtained by combining them, the recognition accuracy of each character is integrated and the adjacency frequency of the characters determined statistically is calculated. A word candidate generator generates word candidates that are ranked as candidates by adding correction values calculated based on the original word to find the accuracy of the word, and a spelling check is performed starting from the word candidates with the highest ranking. This is an English character recognition device that includes a spell check unit that performs a spell check and outputs the correct character string as a recognition result.