JPH04163681A - Information processor and character recognizing device - Google Patents

Abstract

PURPOSE:To output each character block always in correct connection order by deciding connection relation between the character blocks, and outputting each character block from a document picture constituted of plural character blocks in this decided order. CONSTITUTION:An information processor 200 is provided with a character block extracting part 210, a character segmenting part 220, a character recognizing part 230, a word segmenting part 240, a feature word selecting part 250, a feature word storage part 260, a feature word collating part 270, and a character block selecting part 280. Then, the feature word of every character block extracted from the document picture is selected, and further, the feature word of each character block is compared and collated with each other, and the connection relation between the respective character blocks is decided on the basis of this result, and the character block is outputted in the connection order based on this. Namely, each character block is outputted in the order of the connection relation added with the contents of a sentence. Thus, each block is outputted always in the correct connection order.

Description

[Detailed Description of the Invention] (Field of Industrial Application) This invention is a method of outputting each character block from a document image composed of a plurality of character blocks in this order by determining the connections between them. The present invention relates to an information processing device that can be used as an information processing device, and a character recognition device that includes this information processing bag M.

(Conventional technique m) If it is possible to accurately obtain the information contained in the original document from the document image on a machine or medium, an information processing device with various functions,
For example, it becomes possible to construct a character recognition device, a device that outputs recognized characters as voice, a device that translates recognized characters, etc.

However, as shown in FIG. 2, for example, a document image contains ruled lines or blank spaces (the illustrated example is blank), and character strings or multiple character blocks arranged in grids and separated from the surroundings by the above-mentioned blank spaces.
There is a document image 11 composed of letters 1a to 11d, and in recognition processing of such document image 11, unlike processing of character images of known formats, character blocks are extracted and connections between each character block are determined. Special measures are required to determine the relationship.

Therefore, as a conventional device that can perform this, for example, the literature
(The Institute of Electronics, Information and Communication Engineers Technical Research Report PRU86-33) disclosed an automatic reading device for unknown format documents.

The device disclosed in this document uses marginal distribution features, line density features, and circumscribed rectangle features to extract document constituent elements such as heading character blocks and body character blocks from a document with an unknown format. Each character is extracted from each of these character blocks, and feature extraction is performed on each character to recognize each character.

In addition, at this time, the connection order of each character block can be determined, for example, in the literature (IEICE Transactions, (1986, 3),
As disclosed in J69-D, No. 3゜p, 404), the order evaluation value of each character block in the document is calculated according to the following formula (1), and the h is ranked in descending order. It was common to

h=x, +ε・Yc = (1) Here, XC is the x coordinate of the upper left corner point of each character block, Y
c represents the '1 mark of the upper left corner point of each character block, and ε represents a constant of approximately 0.1.

(Problem to be Solved by the Invention) However, in the automatic reading device disclosed in the above-mentioned document, each block in the document image shown in FIG.
Is the connection relationship between a and lid normal? 11a → 11C → 11b
→11d, it is determined that the connection is 11a → 11b → 11c → 11d because it follows the position evaluation value above, and it is not possible to accurately determine the order of each character block. There was a problem that it wasn't there.

Therefore, when this automatic reading device M is applied to a character recognition device, accurate recognition results cannot be obtained because character recognition cannot be performed in the correct order. Also,
In order to prevent this, it is necessary for the operator to perform correction work, but this work is not desirable because it is frequent and leads to a significant decrease in the processing performance of the character recognition device and worsening of errors. do not have.

The present invention has been made in view of the above points, and therefore, an object of the present invention is to provide an information processing device that can accurately determine the connection relationship between character blocks, and a character recognition device equipped with the same. There is a particular thing.

(Means for Solving the Problem) In order to achieve this object, according to the first invention of this application, in an information processing apparatus comprising a character block extracting section for extracting character blocks from a document image, A character extraction unit that extracts a character pattern from the character block that has been identified, a character recognition unit that recognizes the aforementioned character pattern, and a word extraction unit that extracts a word from the aforementioned recognized character pattern string, for each character block. , a feature word selection section that selects a feature word from the aforementioned cut out words, a feature word storage section that stores the aforementioned selected feature words for each character block, and features for each block. A feature word matching unit that compares and matches words between each character block, and determines the connection relationship between each character block based on the comparison and matching results of the feature words, and outputs a character block based on the determination result. The present invention is characterized in that it includes a character block selection section.

In carrying out the first invention, it is preferable that the feature word selection section described above selects feature words based on the number of characters that make up the word.

Alternatively, it is preferable to select a word among words in a character block that does not match a general word registered in advance as a characteristic word.

Roughly speaking, in implementing the first invention, is the feature word matching unit configured to count the number of words that appear in common between the character block of interest and other character blocks in the comparison and matching of the feature words? suitable.

Further, in carrying out the present invention, the character block selection section described above selects a character block having the maximum number of common words or more than a predetermined value, and selects a character block that is the same as the number of common words or other characters connected to the character block of interest. It is preferable to adopt a configuration in which it is determined that the block is a block.

Further, the character recognition device of the second invention of this application is characterized by comprising the information processing device of the first invention and a character recognition stage that recognizes the characters of the character block output from the information processing device. shall be.

(Operation) According to the information processing device of the first invention of this application, feature words are selected for each character block extracted from a document image,
The characteristic words of each character block are roughly compared and collated, and based on this result, the connection relationship of each character block is determined, and the character blocks are output in the order of connection based on this.

In other words, each character block is output in a connection relationship that takes into account the content of the text. Therefore, even when extracting and outputting character blocks from a document with no regularity in the layout of each character block, conventional methods that output each character block simply based on the coordinates of each character block in the document image Compared to , each block is output in the correct order.

Further, according to the character recognition device of the second invention, character blocks that are output in the correct order as described above are recognized. Therefore, character recognition of a document with irregular layout can be performed based on the document.

(Example) Examples of the information processing device of the first invention and the character recognition device of the second invention of this application will be described below using an example of a character recognition device to which the information processing bag M% of the first invention is applied.

1+@J7: No. 8th FIG. 1 is a block diagram schematically showing the configuration of a character recognition device 100 of an embodiment incorporating an information processing device 200 of the embodiment. In order to facilitate the explanation of this character recognition unit M10o, the following explanation will be given using an example of a document image to be recognized. Here, a document image 11 as shown in FIG. 2, which is composed of blank spaces and character blocks lla to 11d, is taken as an example of a document image to be recognized.

This character recognition device 100 includes a photoelectric conversion section 110, an information processing device 200 of the embodiment, a character recognition stage 120, and an output terminal 130. In addition, this information processing device 11!200
are a character block extraction unit 210 and a character extraction unit 220.
, character recognition unit 230, word extraction unit 24o, feature word selection unit 250, feature word storage unit 260, feature word matching unit 270, and character block selection unit 280! The character recognition stage 120 also includes a character segmentation section 122 and a character recognition section 124.

Next, details of each component will be explained.

The photoelectric conversion unit 110 can be configured with, for example, a conventionally known image sensor, etc., and can be configured with characters, figures, etc. (hereinafter referred to as characters).
A record! For example, a document image 11 as shown in FIG.
For example, the optical signal S% from 210.

The character block extraction unit 2 ] 0 extracts data corresponding to character blocks of a document image from the input image data input from the photoelectric conversion unit 710 (this is referred to as character block extraction).

) is extracted and outputted to the character block selection section 280 and character cutting section 220, respectively. For example, the second
To explain using the example of the document image 11 shown in the figure, an area 1 in which character strings are regularly and completely arranged and separated from the surroundings by spaces etc.
1a to ]1d (each of these areas becomes a character block) is extracted and output to the character block selection unit 280 and character cutting unit 220, respectively.

Here, character block data can be extracted by various conventionally known methods, but in this embodiment, it is performed by the method described below.

First, input image data is scanned in the tm direction (corresponding to the direction indicated by y in FIG. 2) to create a peripheral distribution. next,
In this peripheral distribution, an area defined by the number of black hits from the scanning line where the number changes from O to the value or more to the scanning line immediately before the number changes from the value or more to O is detected as one block in the column direction. Next, the input image data portion corresponding to one block in the column direction is scanned in the horizontal direction (corresponding to the direction indicated by the x in FIG. 2) to create a peripheral distribution in the horizontal direction. In this horizontal peripheral distribution, the width of the area defined by the number of black bits or the scanning line that changes from O to the value or more and the scanning line immediately before changing from the value or more to O or the threshold T
If it is larger than H, the data of the overlapping portion between this area and the aforementioned block in the column direction is extracted as character block data. Note that the threshold value TH is a value to prevent erroneous determination of a character block due to, for example, dirt in the document, and is set to 140 in this embodiment.

The character cutting unit 220 cuts out a character line from the character block data input from the character block extracting unit 210, further cuts out character patterns of each character from this character line, and then sequentially passes these character patterns to the character recognition unit 230.
This is what is output to. Here, cutting out a character line and cutting out a character pattern from this character line can be performed by, for example, extracting character block data from input image data based on the peripheral distribution and performing the same procedure as described above.

The character recognition unit 230 extracts the features of the character pattern input from the character cutting unit 220, roughly creates a feature matrix, and further calculates the similarity between this feature matrix and a dictionary Mad1 Nox (not shown) prepared in advance. Calculate,
The character name of the dictionary Mad 1 Nox with the highest degree of similarity is output to the word extraction unit 240 as the recognition result of the character pattern. Here, the above-mentioned feature extraction is performed using the conventionally known +
! In the case of this example (
This is done in the manner described below.

First, for a character pattern, a rectangular frame tS, for example, circumscribing the character line portion is drawn, and this is used as a character frame. Furthermore, the line width W of the character pattern is calculated using, for example, a well-known approximation formula as shown in equation (2) below.

W=1/ (1-(Q/A))...(2) However, (2
) In the formula, Q is the number of windows in which all the points within this window are black bits when each point constituting the character pattern is scanned with a window that can be seen in (2 x 2) nodes at a time. and is the total number of black bits in the character frame.

Furthermore, this character pattern is scanned in a plurality of directions to detect the number of consecutive black bits in each scanning line, and based on this number of consecutive black bits and the line width W described above, the character pattern is scanned in a plurality of directions. The corresponding sub-patterns are extracted for each. Then, the above-mentioned character frame of this character pattern is divided into (NXM) regions (M and N are constants) for each sub-pattern, and the feature amount representing the character line length in each divided region is calculated for each divided region. This feature quantity is normalized by the size of the character frame to obtain a feature matrix. In this embodiment, the feature amount is normalized by dividing it by a value of (dX+dY)/2. Here, d× is the horizontal length of the character frame, dY
is the vertical length of the character frame.

In addition, the above-mentioned similarity is shown below in this example (
3) Calculate based on the formula.

However, in equation (3), day is the similarity, f is the element value of the feature matrix of the character pattern of the character to be recognized, 9I is the element value of the dictionary matrix, and NXM is the dimension of the feature matrix of the character to be recognized and the dictionary matrix. Indicate the number of each.

The word cutting section 240 cuts out a word (sometimes referred to as a word character string) from the character pattern string inputted from the character cutting section and outputs it to the characteristic word selection section. Here, words can be cut out using various conventionally known methods, and in the case of this embodiment, the following method is used.

First, a word area is extracted from the start point coordinates and end point coordinates of each character pattern in the row direction (direction corresponding to the direction indicated by X in FIG. 2). Specifically, if the character pattern corresponding to the first character of the line satisfies the following formula (4), that character is determined to be the start character of a word, and the next word starts from this character. The word area extends up to the character immediately before the start character.

X −i X −+ −+ ＞■ ...(4)
X! In X sJ: Character pattern starting point coordinates X@IT
X@J Character pattern end point coordinates 1, j: Character number in the line n: Number of characters in the line Next, the word cutting unit 240 receives input from the character recognition unit 230 of each character pattern in this word area. A string made up of the recognized character names is output to the feature word selection unit 250 as a word character string (sometimes referred to as a word).

The word selection section 250 selects a word string that is characteristic of each character block from the word string inputted from the word cutting section 240, and the selected word string is referred to as a characteristic word. ), which is output to the feature word storage section 260.

Various methods can be considered for selecting characteristic words. In this example, as explained below, characteristic words are selected by focusing on the string length of the word string to be selected. shall be.

Specifically, if the character string length of a word character string is β, then a word character string with a character string length that satisfies the condition of equation (5) below, that is, a word character string whose character string length is equal to or larger than a predetermined threshold value. Select as the characteristic word of the character block. However, k is set to 6 in this embodiment.

β≧k (5) The feature word storage unit 260 stores the feature word input from the feature word selection unit 250 in association with the character block number to which it belongs. Feature word storage unit 2
60 can be constructed from a conventionally known memory. In this case, for convenience of explanation, the character block numbers are assumed to be 11a to 11d shown in FIG.

Each character block 17a~ of the document image 11 shown in FIG.
In this embodiment, from 11d@, words with character string length β or 6 or more are selected by the characteristic word selection unit 250 as characteristic words. Therefore, characteristic words are stored in the characteristic word storage section 260 for each character block as shown in Table 1. In Table 1, only some of the characteristic words are shown.The characteristic word matching unit 270 reads the characteristic words of each character block from the characteristic word storage unit and compares and collates the characteristic words of each block with each other. It is. In particular, the feature word matching unit 270 of this embodiment tests whether or not a common feature word exists in different character blocks, and also tests the number of common feature words between the character block of interest and each other character block. The structure is such that the numbers are counted for each character block in each location.

The function of the feature word collation unit 270 will be explained below using operations for feature words for each character block shown in Table 1.

When character block lla@character block of interest is selected, the feature word matching unit 270 selects rlasers4, rweapons J between character blocks 11a and 11c.
and r 1ntelliqence" are found as common characteristic words, and the number of common characteristic words is 3.
Count. Similarly, r president J is found as a common feature word between character blocks lla and lld, and the number of common feature words is counted. Furthermore, since no common feature words can be found between character blocks 11a and 11b, the number of common feature words is set to 0. The feature word matching unit 270 compares character blocks llb, l
For lc, the number of common feature words is counted in the same manner as for the character block 11a. However, in the case of the document image example in FIG. 2, character block 11b and character block 1
1c and 11d, and there is no common feature word between character block 11c and character block lld. Therefore, the common feature words of each character block in the document image shown in FIG. The counting results are shown in Table 2 below. This counting result is output to character block selection section 280.

The second table character block selection unit 280 determines the connection relationship between each character block based on the comparison/matching results input from the feature word matching unit 270, and outputs a character block based on the determination result. be. In particular, the character block selection unit 280 of this embodiment uses the number of common characteristic words as shown in Table 2 inputted from the characteristic word matching unit 270 and each character block data inputted from the character block extraction unit 210. Based on this, the connection order of each character block is determined, and each character block data is output to the character cutting section 122 of the character recognition stage 120 in this order. The connection order determination procedure here is as follows.

First, a temporary connection order of each character block is determined based on the position coordinates of each character block. In this embodiment, this provisional connection order is determined by the position evaluation value E of each character block in the document image according to equation (6) below (where 1 is the character block number, and in this example, Inside lla~
11d. ), and perform the calculation in descending order of El.

E+ =y□+F−x, ・・・(6) However, (
6) In the formula, X and l are the X coordinates of the upper left corner of the rectangle circumscribing the character block, Yet is the y coordinate of the upper left corner, and f is a predetermined constant. In the example case, it is 10. Note that the x and y coordinates are based on the x and y coordinate system shown in Figure 2 of the stem.

Therefore, the tentative connection order of each character block in the document image shown in FIG. 2 is lla→11b→11C→11d.

Next, the character block selection unit 280 selects the feature word matching unit 2.
Based on the number of common feature words outputted from 70, the tentative connection order is modified as described below.

First, it is tested whether the number of common feature words of the character block of interest with respect to other character blocks is greater than a predetermined specific @P, and other character blocks that have a value greater than the value are added to the character block of interest. However, it is determined that there is a connection relationship.

However, if there are multiple other character blocks with the number of common feature words equal to or greater than the value, the character block with the largest number of common feature words among these is determined to have the most connected relationship with the character block of interest. shall be.

Next, the provisional connection order is corrected by setting the other character blocks that have been determined to have a connection relation to the character block of interest as the character blocks positioned next to the character block of interest.

Note that in this embodiment, P=2. Further, the upper character block in the tentative connection order is sequentially determined as the character block of interest, and the above determination and correction are performed in order.

In the case of the document image example shown in FIG. 2, as shown in Table 2, the number of common characteristic words between the character block 11a and the character block 11C is 3, which is greater than the value.
The character block selection unit 280 determines that there is a connection relationship between the character block lla and the character block 11C, and roughly changes the tentative connection order of 11a → 11b → 11c → 11d to 11a → 11C → 11b → Correct the connection order to 11d. Then, this character block selection section 28
0 sequentially outputs each character block data to the character cutting section 122 of the character recognition stage 120 according to this modified connection order.

The character cutting section 122 of the character recognition stage 120 cuts out a character line from the character block data inputted from the character block selection section 280, cuts out a character pattern of each character from this character line, and sends this data to the character recognition section 124.
The data is output sequentially. Note that this character cutting section 122 may have a configuration similar to that of the character cutting section 220 in the information processing device 1200, for example.

The character recognition unit 124 of the character recognition stage 120 recognizes the character pattern input from the character cutting unit 122, and outputs the recognition result character name from the output terminal 130 to the device I (for example, a display, computer, etc.) at the edge stage. This is what is output to. Note that this character recognition section 124 may have a configuration similar to that of the literary magazine I1 section 230 in the information processing device 200, for example.

According to the character recognition device 100 of this embodiment, when the information processing device 200 cuts out each character block from a document image and outputs them to the character recognition stage 1201, it is determined that a word with good luck is present in each character block. Each block is output in order so that character blocks containing related words are connected to each other. For this reason, the character recognition stage 120 can perform recognition in accordance with the flow of the document, so that, for example, when using recognition results for speech synthesis or translation, meanings will not suddenly be understood between character blocks. This enables accurate speech synthesis and translation.

In the above, embodiments of the information processing device and character recognition device of this application have been described, but the present invention is not limited to the above-mentioned embodiments and can be modified in various ways as described below. .

For example, in the embodiment described above, the feature word selection unit 250
Was there a configuration in which a word that satisfies the conditional expression β≧ by focusing on the character string length β of the word string extracted from the word extraction unit 240 is selected as a characteristic word?For example, a general word to be exported into a general document It is also possible to prepare in advance a general word dictionary in which are registered, and select words that are not registered in the general word dictionary from among the words in the target character block and use them as characteristic words of the target character block.

In addition, in the above embodiment, the information processing bag! 200 is provided with a character extraction section 220 and a character recognition section 230, and a character recognition stage 120 is provided with a character extraction section 122 and a character recognition section 124.
The character recognition unit 23 may be configured so that the character extraction units 220 and 124 are shared by one character extraction unit.
Of course, a configuration in which 0 and 124 are shared by one character recognition unit is also possible.

Furthermore, extraction of character blocks, extraction of character lines, extraction of character patterns, creation of feature matrices for character recognition, similarity calculation, word extraction, etc. are not limited to the methods explained in the examples. Obviously, other suitable methods may be used.

Furthermore, in the above embodiment, the information processing bag M of the first invention was applied to a character recognition device, but this information processing device determines the connection order of character blocks and sequentially outputs the character blocks according to this order. The present invention can be applied to image processing devices, and also to devices that extract the layout structure of a document by extracting connection relationships between character regions.

(Effects of the Invention) As is clear from the above description, the information processing device of the first invention of this application selects characteristic words for each character profile extracted from a document image, and The feature words of each character block are then compared and collated, and based on this result, the connection relationship of each character block is determined, and the character blocks can be output in the connection order based on this. Therefore, even when extracting and outputting character blocks from a document with no regularity in the layout of each character block, conventional devices that output each character block simply based on the position M coordinate of each character block in the document image can be used. By comparing, each block can be output in the correct connection order.

Further, according to the character recognition device of the second aspect of the invention, character recognition can be performed on character blocks output in the correct order as described above. Therefore, for example, character recognition of a document with irregular layout can be performed in accordance with the flow of the document, so that a high-performance character recognition section M can be realized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram for explaining the embodiment, and is a block diagram schematically showing the configuration of the character recognition device of the embodiment incorporating the information processing bag M of the embodiment. FIG. 2 is a diagram showing an example of a document image. S... Optical signal from document image 11... Document images 11a to lld... Character block 100... Character recognition device, 110... Photoelectric conversion unit 1
20...Character recognition stage, 122...Character cutting unit 124...Character recognition unit, 130...Output terminal 200...Information processing device 210...Character block extraction unit 220...Character cutting out Units 230...Character recognition unit, 240...Word extraction unit 250...Feature word selection unit 260...Feature word storage unit 270...Feature word collation unit 280...Character block selection unit. Patent applicant: Oki Electric Industry Co., Ltd. 11 Document images 11a-lid Figure 2 shows an example of a document image subject to character block processing

Claims (6)

[Claims] (1) An information processing device comprising a character block extraction section for extracting character blocks from a document image, comprising: a character extraction section for cutting out a character pattern from the extracted character block; and a character for recognizing the cut out character pattern. a recognition unit; a word extraction unit that extracts a word from the recognized character pattern string; a feature word selection unit that selects a characteristic word from among the extracted words for each character block; a feature word storage section for storing feature words for each character block; a feature word matching section for comparing and matching the feature words for each block between each character block; 1. An information processing device comprising: a character block selection unit that determines the connection relationship between each character block by using the above method, and outputs a character block based on the determination result. (2) The information processing apparatus according to claim 1, wherein the characteristic word selection section selects the characteristic word based on the number of characters composing the word. (3) The information processing device according to claim 1, wherein the characteristic word matching unit counts the number of words that appear in common between the character block of interest and other character blocks in comparing and collating the characteristic words. Information processing equipment. (4) In the information processing device according to claim 3, the character block selection unit connects the character block whose number of common words is greater than or equal to a predetermined value and exhibits the largest value to the character block of interest. An information processing device characterized in that the information processing device determines that the character block is another character block. (5) In the information processing device according to claim 1, the feature word selection unit selects, as a feature word, a word that does not match a pre-registered general word among the words in the character block. An information processing device characterized by: (6) Character recognition comprising the information processing device according to any one of claims 1 to 5 and a character recognition stage that recognizes characters in a character block output from the information processing device. Device.