JPH11134439A - Method for recognizing word - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a word recognizing method with a reduced recognizing time. SOLUTION: A word pattern is divided into three areas of an upper pattern, middle pattern, and lower pattern by two horizontal division lines. Then, all character upper patterns are detected by classifying them into plural kinds based on the shape of the horizontal peripheral distribution and local vertical peripheral distribution of the upper pattern, and a word upper part pattern code in which codes indicating the number and kind of detection are arranged in the order of the positions of detection is obtained. Moreover, all character lower patterns are detected by classifying them into plural kinds based on the shape, and a word lower pattern code in which codes indicating the number and kind of the detection of the lower pattern are arranged in the order of the positions of detection is obtained. Then, more than one word candidates are selected by retrieving a word recognition dictionary according to the upper and lower pattern codes prepared from the inputted word pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition method for reading a character pattern of a printed European language.

[0002]

2. Description of the Related Art A conventional character recognition method of this type is disclosed in Japanese Patent Laid-Open Publication No. Sho 57-23185, "Character Recognition Method". FIG. 7 shows a configuration of a character recognition apparatus using the method. FIG. A character pattern for one character is input from an input terminal 1 to a character pattern memory 2 as a binary image. The character frame dividing unit 3 detects a circumscribed frame of the input character pattern, and obtains a dividing point for dividing the internal region into an M × N matrix.

The horizontal sub-pattern extracting section 4 extracts horizontal stroke components from an input character pattern to create a horizontal sub-pattern. FIG. 8A shows a horizontal sub-pattern (H
SP). Similarly, vertical, diagonally left, and diagonally right sub-pattern extraction units 5, 6, and 7 extract stroke components in each direction from the input character pattern, and generate sub-patterns as shown in FIGS. 8B, 8C, and 8D. create.

The feature matrix extraction unit 8 divides these four sub-patterns into a matrix using the division points obtained by the character frame division unit 3 and quantifies the stroke length in each area. (A) horizontal matrix H as shown,
(B) vertical matrix V, (c) left oblique matrix L,
(D) Obtain a right oblique matrix R.

A matrix of 4 × (M × N) elements obtained by merging these elements is called a feature matrix because it represents a feature that can identify what character (category) it is. This feature matrix expresses the characteristics peculiar to the category such as the direction, position, and length of the stroke constituting the character pattern, but the character size, line width, variation in the character shape, etc. do not affect the value of the feature matrix. Has been converted to

The character collation / identification unit 9 creates a feature matrix obtained from an input character pattern and a standard character pattern for each category (character) of the entire character set to be read in the same procedure as described above. A comparison is made with the feature matrix stored in the matrix dictionary 10 to select a category having the most similar matrix.

In order to do so, the distance between feature matrices is evaluated by the sum of squares (or the square root) of the difference between corresponding elements, and one or several categories whose distance is closest to 0 are selected as candidate characters. , And outputs the character code from the output terminal 11.

[0008]

However, in the above-described character recognition method, the input character pattern is input in units of one character, and a character matrix is obtained and collated with all the character patterns stored in the dictionary. In the case of recognizing one European word, it is necessary to repeat this character recognition process by the number of characters constituting the word, so that much time has been spent for word recognition.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a recognition method capable of shortening a word recognition time.

[0010]

In order to achieve the above object, according to the present invention, a word pattern having European characters on a printed body as a component is divided into two patterns by an upper pattern, a middle pattern, and a lower pattern. The upper pattern is divided into three regions, and from the horizontal peripheral distribution and the local vertical peripheral distribution of the upper pattern, the character upper patterns of the European characters are classified into a plurality of types according to their shapes and all are detected, and the number of detected upper patterns and the relevant type are detected. Is obtained by arranging the codes representing the characters in the order of the detection positions, and from the horizontal peripheral distribution and the local vertical peripheral distribution of the lower pattern, the character lower patterns of the European characters are classified into a plurality of types by shape. All words are detected, and a code representing the number of detections and the type of the lower pattern are arranged in the order of the detection positions to obtain a word lower pattern code, and the word upper pattern code and Using a word lower pattern code, a word recognition dictionary in which words to be recognized are classified is created in advance, and the upper and lower pattern codes of the word are obtained from the input word pattern. It is characterized in that one or more word candidates are selected by indexing a dictionary.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION

[Explanation of Embodiment] [Explanation of Configuration] The word recognition method of the present invention is not a conventional character-based recognition but a method of directly recognizing a word using a feature included in a pattern of the entire word. is there. FIG. 1 is an explanatory diagram of a word recognition method according to an embodiment of the present invention, showing an encoding method of an upper pattern of a word "information" and related data.

In this embodiment, as shown in FIG.
The rectangular frame surrounding the word pattern is divided by two horizontal lines into three rectangular areas of an upper (head) pattern, a middle (main body) pattern, and a lower (bottom) pattern. The width of each rectangular area is W in common, and the height is H1,
Divided into H2 and H3. Of these, the middle pattern is filled with the pattern of its constituent characters, but in this example, only the four characters "i", "f", "t", and "i" contribute to the upper pattern in this example. It becomes a sparse pattern. Also, since no character having a lower pattern is included, the entire lower pattern of an English word is empty.

In this embodiment, a clear and orderly printed European document is targeted, and it is assumed that each line is cut out from the document image and each word is cut out from the line to recognize the word. Therefore, at the word recognition stage, it is assumed that the font size is known, the word pattern has no inclination in the horizontal and vertical directions, and the word pattern can be correctly divided into three parts by the dividing lines.

In order to determine the height of the dividing line, for example, the vertical marginal distribution of an image including one line or one word
In other words, taking a distribution indicating how many black pixels are on each horizontal line, since the word has most of the black pixels in its middle pattern, this is close to a rectangular distribution, so Two upper and lower dividing lines are obtained from both edges of the rectangular distribution.

FIG. 2A shows nine characters having an upper pattern among the 26 lowercase letters of the alphabet, which are classified into four types according to the shape of the upper pattern, and numerical values from 0 to 3 are assigned as codes indicating the types. It is. That is, the characters “i” and “j” are 0, and the characters “b”, “d”, “h”, and “k”
An upper pattern code of 1 is assigned to “l”, 2 to character “t”, and 3 to character “f”. In FIG. 2B, five characters having a lower pattern are classified into three types according to the shape of the lower pattern, and 0, 1, and 2 are assigned as character lower pattern codes. That is, the characters "j" and "y" are 0,
Letters “p” and “q” are assigned a lower pattern code of 1 and letters “g” are assigned a lower pattern code of 2.

An upper pattern code and a lower pattern code of the English word are defined by a variable-length sequence of pattern codes of these characters appearing in the English word. Therefore, the English word "info
The upper pattern code of "rmation" is "i""f""t"
It is represented as 0320 using the character upper pattern code of “i”. This is prefixed with 4 for the number of character upper patterns detected,
Let 40320 be an English word upper pattern code. Also, in this example, the lower pattern code is represented by only the detected number of character lower patterns, 0.

In FIG. 1 (a), the horizontal positions of the four character upper patterns are represented by p1, p using the distance from the left end of the English word.
They are represented as 2, p3 and p4. These values may vary depending on the size of the character font and variations in the print interval, but the range of the approximate relative position within the word can be limited.

FIG. 1B shows the above-mentioned solid English word “info”.
The main features of the upper pattern of "rmation" are summarized in a table as data that can be used for English word recognition processing. Such data is created for all English words to be recognized, and an English word recognition dictionary is created using this data.

The English word recognition dictionary according to the present embodiment is indexed by upper and lower pattern codes of English words to obtain one.
More than one English word candidate is determined, and from the contents of the dictionary, the upper or lower pattern code of the English word candidate is identified based on the detection position, or some constituent characters are identified by pattern matching, and the The English word is determined, and the identification code or constituent character string of the English word is finally obtained.

The last pattern collation involves recognition in character units by the conventional method, and is used when there are few upper / lower patterns and many English word candidates or when there is no corresponding word.

[Explanation of Operation] FIG. 3 is a flowchart of English word recognition using only the upper pattern of English words.
The processing of the lower pattern must also be performed at the same time, but since the processing method is the same as that of the upper pattern, its flowchart is omitted for simplicity. It is assumed that an English word identification dictionary has already been created, and here, a recognition process in the case where a character pattern of an unknown English word is input will be considered.

In step S101, an upper pattern of an English word is cut out. In process S102, the horizontal peripheral distribution of the entire upper pattern of the English word is obtained. FIG. 4A shows the horizontal peripheral distribution of the upper pattern (the left part of the dotted line) of “information”. This indicates how many black pixels are in the vertical direction at each horizontal position of the English word upper pattern. In this example, since there are black pixels at the horizontal positions of the characters "i", "f", "t", and "i" having the upper pattern, the horizontal peripheral distribution of the character upper pattern appears at four locations.

In step S103, the character upper part pattern is detected as a non-zero section in which the horizontal peripheral distribution is continuous from left to right. For each character upper pattern, a local vertical marginal distribution is obtained in step S104. This is obtained by counting the number of black pixels in the horizontal direction at each vertical position of one character upper pattern. FIG. 4A shows the four local vertical marginal distributions thus obtained.

In step S105, the type (sign) of the character upper pattern is determined from the shapes of the horizontal peripheral distribution and the local vertical peripheral distribution, and the detection position of the character upper pattern (horizontal coordinates of the representative point) is determined. Since there are only four types of character upper patterns from FIG. 2, these can be easily discriminated. FIG.
The both peripheral distributions of the upper pattern of the letter “b” are shown on the right side of the dotted line in FIG.

A typical method of discriminating the upper pattern is based on whether the horizontal marginal distribution reaches the vicinity of the maximum value H1. If not, the pattern code is "0" or "2".
Or, it is “3”, and the pattern code is set to “0” if the vertical marginal distribution has a zero section (a section without a black point) near the lower part. When the pattern codes are “1” and “3”, if the horizontal position at which the horizontal marginal distribution takes the maximum value is near the left end of the non-zero section, it is set to “3”.

The lowercase letter "f" may be printed so as to overlap with the character on the right side. In such a case, the method of discrimination is as follows. Then, both peripheral distributions of the remaining character upper pattern may be obtained again and discriminated.

In step S106, the English word recognition dictionary is indexed using the upper pattern code of the English word obtained by combining the above character upper pattern codes together with the number thereof, and one or more candidate English words are selected. Actually, lower pattern codes are obtained in the same procedure as S101 to S105, and in S106, English words are narrowed down using both of these pattern codes. However, as described above, the description of the lower patterns is omitted. FIG. 4B shows the English word “jumpin”.
"g" shows three types of lower patterns and their peripheral distributions. The method of determining the lower pattern uses, for example, a horizontal distribution pattern width and increase / decrease information.

FIG. 5 shows an indexing method of the English word recognition dictionary. The set of all English words is one that has no upper character pattern in total, one that has only one character has four classifications, one that has two characters has sixteen classifications, three has 64 classifications, four has 256 classifications,
Five are classified into 1024 classifications. English words with 6 or more letter upper patterns are 40 words using only the first 6 words.
Classify into 96.

English words without a character upper pattern are "a" and "ac".
It consists of a large number of English words starting with "e". The English word `` i
"nformation" is a class with the upper pattern code of "40320", and a total of three English word candidates "infraction" and "infectio"
n ”and“ information ”.

Next, FIG. 6 shows a method of identifying between the three candidate English words. FIG. 7A shows a method of identifying a specific upper pattern based on the detected position, and FIG. 7B shows a method of recognizing only one constituent character and determining the result based on the result.

In the identification method shown in FIG. 2A, the following identification method is specified in the dictionary. The ratio between the detection position p2 of the second character upper pattern and the width w is calculated, and if 3/16 <p2 / W <5/16, “infarction” 2/8 <p2 / W <3/8, “infection” If 1/8 <p2 / W <2/8, "information"

This is to determine from the relative position of the letter "f" in the English word. However, the three tolerances overlap and cannot be completely identified in this example.

In the identification method shown in FIG. 3B, the dictionary specifies the following identification method. Recognizes the next character in the upper pattern of the second character. If that character is "o", it is "information". "E" is "infection". "A" is "infarction"

Since this cannot be distinguished at the position of "f", the character next to "f" is recognized by the conventional method and is to be determined by that character. In this example, the character can be completely identified.

Returning to the flowchart of FIG. 3, processing S107 is performed.
In step S108, for all the English word candidates in one classification, the detection position of the upper pattern of the input English word and the character recognition result are obtained in accordance with the instruction of the identification method specified there, and the dictionary is obtained. Check whether the conditions described in the above are met. In processing S109, a target English word is determined from the result.

[Explanation of Usage Form] Since a recognition method for each character can be arbitrarily selected, a usage form in which the method of the present invention is integrated into an existing character recognition system is considered. Since the present invention speeds up the English sentence recognition, it can be used for creating a book database or connected to a text-to-speech conversion system or a communication system.

[0037]

As described above, according to the English word recognition method of the present invention, a word pattern having European characters on a printed body as a component is divided into two parts by an upper pattern, a middle pattern, and a lower pattern. The upper pattern is divided into three regions, and from the horizontal peripheral distribution and the local vertical peripheral distribution of the upper pattern, the character upper patterns of the European characters are classified into a plurality of types according to their shapes and all are detected, and the number of detected upper patterns and the relevant type are detected. Is obtained by arranging the codes representing the characters in the order of the detection positions, and from the horizontal peripheral distribution and the local vertical peripheral distribution of the lower pattern, the character lower patterns of the European characters are classified into a plurality of types by shape. All the words are detected, and a code indicating the number of detections and the type of the lower pattern are arranged in the order of the detection positions to obtain a word lower pattern code, and the word upper pattern code and the word lower pattern are obtained. Using a tan code, a word recognition dictionary in which the words to be recognized are classified is created in advance, the upper and lower pattern codes of the word are obtained from the input word pattern, and the word recognition dictionary is obtained by the two pattern codes. A conventional method is used in which one or more word candidates are selected by indexing, and upper and lower patterns, that is, portions of the English word that protrude from the body of the English word are detected and the entire English word is recognized. The procedure of cutting out and recognizing characters individually as in the above can be omitted.

The upper and lower patterns can be easily and uniquely obtained by analyzing the horizontal peripheral distribution and the local vertical peripheral distribution, and the entire English word recognition dictionary is classified without duplication by these pattern codes. English word candidates can be narrowed down by indexing instead of pattern matching.

Further, when there are a plurality of English word candidates, the determination is made based on the data obtained by the feature extraction such as the detection positions of the upper pattern and the lower pattern, so that the minimum processing is required. Even when English characters need to be recognized, only a few specific characters need to be recognized instead of all characters. When there are many English word candidates, more detailed classification can be performed using the recognition result of a specific character or other information.

As described above, according to the present invention, the recognition processing can be made much more efficient and faster than the conventional method.

It should be noted that the method of the present invention can be applied to most typefaces by slightly changing the method of determining the character upper / lower pattern.
For example, in an italic (tilted) body, it can be dealt with by correcting the tilt or ignoring the protrusion below the "f". However, in this case, it is necessary to cut out the upper pattern of a typeface with a small protrusion to the upper part of “t”.

Since the upper and lower pattern codes are easy-to-understand expressions that match the human visual recognition, it is easy to analyze the cause of erroneous reading and improve the recognition rate.

Also, since the upper and lower pattern codes are included in about 20 bits even when combined with the identification codes of the candidate characters,
It is sufficiently practical to be used as a variable length or fixed length code for English words, and can be used for English word expression in English texts, grammar dictionaries, idiom dictionaries, and the like.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of an English word recognition method of the present invention.

FIG. 2 is an explanatory diagram of encoding of upper and lower patterns of a character.

FIG. 3 is an English word recognition flowchart using an upper pattern.

FIG. 4 is an explanatory diagram of a peripheral distribution of upper and lower patterns.

FIG. 5 is an explanatory diagram of a dictionary indexing method.

FIG. 6 is an explanatory diagram of a method of identifying between English word candidates.

FIG. 7 is a configuration diagram of a conventional character recognition device.

FIG. 8 is a diagram illustrating an example of a sub-pattern.

FIG. 9 is a diagram illustrating an example of a feature matrix.

[Explanation of symbols]

W 1 word rectangular area width P1, P2, P3, P4 Horizontal position of upper pattern H1 Upper pattern height H2 Middle pattern height H3 Lower pattern height

Claims (3)

[Claims] 1. A word pattern having European characters on a printed body as constituent elements is divided into three areas of an upper pattern, a middle pattern, and a lower pattern by two horizontal dividing lines, and a horizontal peripheral distribution of the upper pattern and a local pattern are locally divided. From the vertical peripheral distribution, the upper character patterns of the European characters are classified into a plurality of types according to their shapes, all are detected, and the number of detected upper patterns and codes representing the types are arranged in the order of the detected positions. From the horizontal peripheral distribution and the local vertical peripheral distribution of the lower pattern, character lower patterns of the European characters are classified into a plurality of types according to their shapes and all are detected, and the number of detected lower patterns and a code representing the type are detected. Are determined in the order of the detection positions to obtain a word lower pattern code, and the word to be recognized is simply classified using the word upper pattern code and the word lower pattern code. A word recognition dictionary is created in advance, upper and lower pattern codes of the word are obtained from the input word pattern, and the word recognition dictionary is indexed with the two pattern codes to select one or more word candidates. A word recognition method characterized by the following. 2. An English word pattern composed of 26 characters from lowercase letters "a" to "z" of a printed material is divided into three areas of an upper pattern, a middle pattern, and a lower pattern by two horizontal dividing lines. From the horizontal peripheral distribution and the local vertical peripheral distribution of the upper pattern, the upper part of “i” and “j”, “b”, “d”, “h”, and “k”
All upper character patterns classified into at least four types: upper part of "l", upper part of "t", and upper part of "f" are detected, and the number of detected upper patterns and the code indicating the type are determined in the order of the detected positions. From the horizontal peripheral distribution and the local vertical peripheral distribution of the lower pattern, the lower part of "j" and "y", the lower part of "p" and "q"
Detecting all lower character patterns classified into at least three types below "g" and obtaining an English word lower pattern code in which the number of detected lower patterns and codes representing the types are arranged in the order of the detection positions, Using an English word upper pattern code and an English word lower pattern code, an English word recognition dictionary in which English words to be recognized are classified is created in advance, and upper and lower pattern codes of the English word are input from the input English word pattern. A method for recognizing an English word, comprising: searching for the English word recognition dictionary using the two pattern codes to select one or more English word candidates. 3. An identification condition relating to the detection position of the upper character pattern and the lower character pattern or an identification condition relating to a constituent character of an English word is recorded in advance in the English word recognition dictionary as identification data of an English word candidate. Then, from among the selected English word candidates of the input English word pattern, an English word that meets the identification condition for the detection position or the identification condition for the constituent characters of the English word is selected. 3. The method for recognizing English words according to claim 2.