JPH03225576A - Device for segmenting word - Google Patents

Abstract

PURPOSE:To highly accurately segment words from a character string by providing the word segmenting device with a preprocessing means, a pitch estimating means and a word segmenting means. CONSTITUTION:The preprocessing means 100 reads out recording information by an image reading part and stores black/white binary information in an image storage part 102. A character string extracting part 103 sends the character string pattern information S103 of one line from the storage part 102. The pitch estimating means 110 computes the positional information S111 of a black block along the character string by a block extracting part 111 and an estimated character pitch P which is the sum of the maximum width of the black block and the minimum width of a blank area adjacent to the black block is outputted from a pitch calculation part 112. In the segmenting means 120, a threshold calculating part 121 finds out a blank width deciding threshold TH based upon P, and when the blank width exceeds the threshold, a word segmenting signal S122 is generated from a word extracting part 122 based upon the block position information S111.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a word cutting device for cutting out words from image information such as a Roman document.

(Conventional technology) Conventionally, as a technology in this field, Japanese Patent Application Laid-Open No. 62-1
There was one described in Publication No. 33585.

Generally, in a device that processes document images, when dealing with European languages, it is often necessary to process not only characters but also words. For example, in an optical character reader (OCR), by performing word recognition processing after character recognition processing, even if the recognition result for each character is incorrect, the incorrect character can be corrected by word recognition processing. becomes possible. At this time, it is necessary to cut out character lines from the image information of the target document, cut out characters from each character line, and also cut out words.

As a conventional word extraction device of this type, there is one described in the above-mentioned literature. With this word extraction device,
Based on the peripheral distribution of the image information of the character string along the character string direction, a histogram (hist
o(lram; an occurrence frequency distribution diagram of a single random variable). Then, a threshold for determining the gap between words is determined along the maximum width corresponding to the peak of the histogram, and a break in the marginal distribution with a width greater than or equal to the threshold is determined to be a gap between words, and words are extracted. there were.

(Problems to be Solved by the Invention) However, the apparatus with the above configuration has the following problems.

(a) With conventional devices, when the number of words included in a character string is small, it is difficult to obtain a peak corresponding to an inter-word gap in the created histogram, making it difficult to determine a threshold for determining an inter-word gap. As a result, it is not possible to cut out the words correctly.

(b) In conventional devices, it is necessary to create a histogram in order to determine the threshold value for determining the inter-word gap, which has the problem of complicating the device configuration and slowing down the processing speed.

The present invention provides a word segmentation device that solves the problems of the prior art, such as the inability to accurately segment words and the fact that the device configuration becomes complicated and the processing speed decreases. be.

(Means for Solving the Problems) In order to solve the above problems, the present invention provides a word extraction device that performs a process of cutting out words of a document based on image information of a document such as a European language. a preprocessing means for extracting a column pattern; a pitch estimation means for detecting an estimated character pitch of the character string pattern; and a pitch estimation means for detecting a blank width in the character string pattern that is larger than a determination threshold determined according to the estimated character pitch. and cutting means for cutting out words according to the position of the blank width.

The pitch determining means is configured to detect an estimated character pitch based on, for example, a peripheral distribution of the character string pattern in a character string direction.

(Operation) According to the present invention, since the word extraction device is configured as described above, the preprocessing means extracts a character string pattern from image information of a document such as a Roman language, and provides it to the pitch estimating means. The pitch estimating means detects an estimated character pitch based on, for example, the peripheral distribution of the character string pattern in the character string direction, and supplies the estimated character pitch to the cutting means. The cutting means determines a judgment threshold based on the estimated character pitch, detects a blank width larger than the judgment threshold, and cuts out a word at the position of the blank width.

As a result, words can be extracted with high precision regardless of the number of words contained in a character string, and the creation of histograms as in the past can be avoided, simplifying the device configuration and increasing processing speed. I can figure it out. Therefore, the above problem can be solved.

(Embodiment) FIG. 1 is a block diagram of a word extraction device showing an embodiment of the present invention.

This word cutting device uses character string pattern data 8103
, a pitch estimating means 110 that detects an estimated character pitch p based on the character string pattern data 5103, and a determination threshold T according to the estimated character pitch p.
H is determined, a blank width larger than the determination threshold TH is detected in the character string pattern data 5103, and a cutout means 120 is configured to cut out a word based on the blank position.

The preprocessing unit 100 includes an image reading unit 101 that reads black and white binary image information 5101 from a recording medium such as a European printed document.
have. This image reading unit 101 is composed of an image scanner or the like, and has image information 5101 on its output side.
An image storage unit 102 consisting of an image buffer or the like that stores , and a character string extraction unit 103 are connected. The character string extraction unit 103 extracts character string pattern data 5103 from the image information 5101 stored in the image storage unit 102,
It has a function of providing it to the pitch estimating means 110.

The pitch estimation means 110 uses character string pattern data 5103
Position information 5 of partial figures (this is called a black block) that constitutes a character string based on the peripheral distribution along the character string direction.
111, and a pitch calculation section 107 that detects the estimated character pitch p based on the position information 5111 of the black block, and a cutout means 120 is connected to the output side of the pitch calculation section 107.

The extraction means 120 includes a threshold calculation unit 121 that calculates a determination threshold TH based on the estimated character pitch p, and a word extraction unit 1.
It is composed of 22. The word extraction unit 122 detects that the blank width between adjacent black blocks calculated based on the black block position information 5111 is larger than the determination threshold TH, and extracts a word indicating that a word is to be extracted based on the blank position. It has a function of outputting a cutout signal 5122.

FIG. 2 is a diagram for explaining an example of the word extraction operation shown in FIG. 1, and the operation of the apparatus shown in FIG. 1 will be explained with reference to this diagram.

Note that 111a in FIG. 2 is character string pattern data 510.
Marginal distribution along the character string direction in 3, 81-B22
is a black block, W1 to W4 are blank areas, and H is the combined width of black block B8 and blank area W1.

First, recorded information such as a European printed document is stored in the image reading unit 101.
, and stored in the image storage unit 102 as black and white binary image information 5101.

Next, the character string extraction unit 103 extracts character string pattern data 5103 (for example, wordsegmen) corresponding to one line from the binary image information stored in the image storage unit 102.
tation method) and provides it to the block extraction unit 111.

The block extraction unit 111 extracts character string pattern data 5103
A peripheral distribution 111a along the character string direction is created by calculation, and each continuous area in which the peripheral distribution 111a has a positive value is extracted as black blocks B1 to B22.

Each black block B1 to B2 consists of a starting end position and a terminal end position along the character string direction of each black block B1 to B22.
The position information 5111 of No. 2 is calculated and the calculation result is saved.

Next, the pitch calculation unit 112 calculates the position information 5i of the black block.
Based on ll (starting end position and ending end position), calculate the sum of the maximum width of the black block and the minimum width of the blank area adjacent to the black block having the maximum width, and send this to the threshold calculation unit 121 as the estimated character pitch p. Output. That is, the starting and ending positions along the character string direction for the first black block are respectively S, , E, (where, Si<Ei<Si+1
, i=1.2,3. . . , N, N is the number of black blocks), the estimated character pitch p is expressed by the following equation.

However, Z represents i when the value (Ei S ) takes the maximum value, and m1n (A, B) represents the minimum value of A and B. The estimated character pitch value obtained in this manner has a value approximately equal to the character pitch within each word.

In other words, in the pitch calculation section 112, as shown in FIG.
Black block B having the largest width than black blocks 81 to B22
8 is detected, and the blank area W1 having the smallest width among the blank areas W1 and W2 adjacent to the black block B8 is detected, and the width H which is the combined width of the black block B8 and the blank area W1 is detected.
is calculated and used as the estimated character pitch p in the threshold calculation unit 1.
Output to 20.

The threshold calculation unit 121 calculates a determination threshold T regarding the blank width for cutting out words based on the estimated character pitch F) (=H).
H is calculated according to the following formula and output to the word extraction section 122.

TH=aX15 However, a is a constant, and in this example, a=0.2
5.

Based on the position information 5111 of each black block stored in the block extraction unit 111, the word extraction unit 122 determines whether the width of the blank area between adjacent black blocks (for example, W3 and W4) exceeds the determination threshold TH. Depending on the blank position,
A word cutting signal 5122 indicating that a word is to be cut out is output. That is, the value (Si, 1-E, -1> is the value T
If it is larger than H, the first black block and the (i+
1) The space between the block and the black block is determined to be a word break and a word cutout signal 111 is output (i=1°2, .
..., N-1). Therefore, in this embodiment, each time a blank width exceeding 0.25 times the estimated character pitch p is detected,
The cutting position (starting end position and ending position) of one word is determined.

For example, in FIG. 2, black blocks 81 to B4 ("wor
d"), black block 85 to B16 ("segm
), and black blocks B17 to B
22 ("method") constitutes a word. When such a word extraction signal 5122 is output from the word extraction unit 122, the word extraction process for the character string pattern data 5103 is completed.

As described above, this embodiment has the following advantages.

(a) Detect the estimated character pitch p of the character string pattern data 5103, detect a blank width larger than the determination threshold TH determined according to the estimated character pitch β in the character string pattern data 5103, and determine the word pattern based on the blank position. Since the structure is configured to extract words, words can be extracted with high precision regardless of the number of words included in a character string.

(b) The pitch estimating means 110 calculates the position information 5l11 of the black block based on the peripheral distribution 111a in the character string direction of the character string pattern data 5103, and calculates the maximum width of the black block and the adjacent black block having the maximum width. The estimated character pitch p is determined by the sum (H) of the minimum width of the blank area
Since the word segmentation device is configured to detect , the device configuration is simpler and the processing speed can be further increased compared to a conventional word segmentation device that requires creating a histogram of the width of the break in the peripheral distribution 111a.

Note that the present invention is not limited to the illustrated embodiment, and various modifications are possible. Examples of such modifications include the following.

(i> In the above embodiment, the pitch estimating means 110
is the position information 5 of the black block based on the peripheral distribution 111a in the character string direction of the character string pattern data 5103.
111, and the sum (H) of the maximum width of the black block and the minimum width of the blank area adjacent to the black block having the maximum width is used to detect the estimated character Pitchiro. You can. For example, string pattern data 5
There are various methods, such as obtaining the positional information of the black block based on the peripheral distribution 111a in the character string direction of 103, and detecting the estimation and p by the sum of the maximum width of the black block and the minimum width of the blank area. It is possible to detect the estimated character pitch p.

(ii) Each block in FIG. 1 may be constructed by individual circuits or by software processing using a microprocessor or the like.

(Effects of the Invention) As described above in detail, according to the present invention, the pitch estimation means detects the estimated character pitch based on the character string pattern extracted by the preprocessing means, and then the cutting means detects the character string pattern. The system is configured to detect a blank width larger than a judgment threshold determined according to the estimated character pitch and cut out word patterns based on the blank position, so it is possible to extract words with high accuracy regardless of the number of words included in the character string. It can be cut out.

Furthermore, since the pitch estimating means is configured by detecting the estimated character pitch based on the peripheral distribution in the character string direction of the character string pattern, for example, it is not necessary to create a histogram as in the past, which simplifies the device configuration. It can also be expected to have the effect of increasing speed and processing speed.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the configuration of a word extraction device showing an embodiment of the present invention, and FIG. 2 is a diagram for explaining an example of the word extraction operation in FIG. 1. 100... Preprocessing means, 101... Image reading unit, 102... Image storage unit, 103...
...Character string extraction unit, 110...Pitch estimation means, 111...Block extraction unit, 112...
... Pitch calculation section, 120 ... Cutting means, 121 ... Threshold calculation section, 122 ...
・Word extraction part.

Claims (1)

[Scope of Claims] 1. In a word extraction device that performs processing to extract words from a document based on image information of the document, a preprocessing means for extracting a character string pattern from the image information of the document; pitch estimating means for detecting an estimated character pitch; and cutting means for detecting a blank width larger than a determination threshold determined according to the estimated character pitch in the character string pattern and cutting out a word according to the position of the blank width, A word cutting device characterized by: 2. The word cutting device according to claim 1, wherein the pitch estimating means is configured to detect an estimated character pitch based on a peripheral distribution of the character string pattern in the character string direction.