JPH04343192A - Character segmenting method of character recognizing device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character segmentation method for a character recognition device, and more particularly to a character segmentation method for segmenting characters using a circumscribed rectangle connected to black pixels.

0002

2. Description of the Related Art When cutting out characters, it is necessary to consider contact between adjacent characters. To cut out this touching character, after extracting the circumscribed rectangles of connected black pixels and integrating those that are close to each other, if the circumscribed rectangle has a width exceeding 1.5 times the preset character width, it will be removed from its center position. JP-A-61-117670 describes a method of separating at a position where the projection in the direction perpendicular to the row has a minimum value within a certain range.

0003

[Problem to be Solved by the Invention] The method of cutting out characters by using a circumscribed rectangle connected with black pixels assumes that each character is always surrounded by white pixels.
If the characters are in contact, the two characters are extracted as one circumscribed rectangle. In actual printed manuscripts and copied manuscripts, many such touching characters appear. Furthermore, when the skew is large, the number of touching characters increases.

Therefore, in the conventional method described above, a circumscribed rectangle with a width exceeding 1.5 times the preset character width is determined to be a touching character, and the touching characters are separated. However, with this method, it is often not possible to achieve sufficient extraction accuracy for general Japanese documents.

[0005] In other words, Japanese documents contain a mixture of kanji, hiragana, and katakana, which have large differences in size and shape, and also have variable character sizes and irregular pitches, making it difficult to identify contact characters. This is because fixing the character width, which is the standard for this, would not be possible.

SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to provide a character cutting method that can reliably cut out touching characters even when the character size or character pitch changes. Another object of the present invention is to provide a method for cutting out characters, including touching characters, with excellent cutting accuracy for multi-sized Japanese documents with irregular pitches. Another object of the present invention is to provide a character extraction method that is resistant to document skew.

[0007]

[Means for Solving the Problem] The invention of claim 1 traces the connection of black pixels on a document image, extracts its circumscribed rectangle, cuts out lines by integrating the extracted circumscribed rectangles, and circumscribes each line. In the character cutting method that cuts out characters using rectangles, the standard character width is calculated for each line based on the size information of the circumscribing rectangle, and by comparing this standard character width with the width of each circumscribing rectangle, it is possible to cut out characters into contact characters. A method characterized by detecting a corresponding circumscribed rectangle, obtaining a projection of the detected circumscribed rectangle in a direction perpendicular to the row, and dividing the circumscribed rectangle at positions where the projection value is equal to or less than a predetermined value. It is.

The invention of claim 2 traces the connection of black pixels of a document image, extracts its circumscribed rectangle, cuts out lines by integrating the extracted circumscribed rectangles, and extracts a single circumscribed rectangle and a plurality of circumscribed rectangles for each line. In a character extraction method that generates character pattern candidates consisting of a combination of rectangles, performs character recognition, and determines whether each character pattern candidate should be finally extracted as a character pattern by evaluating the recognition results, for each line, A standard character width is calculated based on the size information of the circumscribed rectangle, and a circumscribed rectangle corresponding to the touching character is detected by comparing this standard character width with the width of each circumscribed rectangle.
For the detected circumscribed rectangle, the projection in the direction perpendicular to the line is obtained, the circumscribed rectangle is divided at positions where the projection value is less than or equal to a predetermined value, and the circumscribed rectangle after this processing is used as a character pattern candidate. It is characterized by being used to generate.

The invention of claim 3 is characterized in that in the character cutting method of the invention of claim 1 or 2, a standard character width is calculated based on the maximum height of a circumscribed rectangle in each line. be.

0010

[Operation] According to the invention of claim 1, the standard character width used as a criterion for determining the circumscribed rectangle of touching characters can be adapted to the character size of the document, so that even multi-sized Japanese documents can be used. Touch characters can be cut out reliably. In addition, since the standard character width is calculated for each line, it is possible to handle documents in which the font size varies from line to line.

According to the invention of claim 2, since the character recognition result is fed back to the character extraction, the accuracy of character extraction for Japanese documents of essentially multi-size and irregular pitch is excellent, and the characters in each line are By detecting the circumscribed rectangle of the touching character using a standard character width adapted to the size and separating it, the accuracy of cutting out the touching character is further improved.

According to the third aspect of the present invention, the maximum height of the circumscribing rectangle, which is less affected by the skew of the document, unlike the width of the circumscribing rectangle, is used to calculate the standard character width. It is possible to cut out touching characters with high precision without any trouble.

[0013]

[Embodiment] An embodiment of the present invention will be explained along with the apparatus configuration shown in FIG. 1 and the processing flow shown in FIG. 2, respectively.

An image of a document (or form) is read by a binary image input device 11 such as a scanner, and the binary image data is stored in the image memory 12 (step 51). For each entire or partial area of this image data, a circumscribed rectangle extraction unit 13 traces connections of black pixels, extracts a circumscribed rectangle that includes the connected components, and stores data such as its coordinates in a circumscribed rectangle memory 14. Referring to this data, the line cutting unit 15 extracts the range of lines by integrating circumscribed rectangles with small distances in the line direction and in the direction perpendicular to the lines, cuts out the image line by line from the image memory 12, and extracts the line from the image memory 12. Store in image memory 16 (step 52
).

[0015] In rectangle extraction by tracing black pixel connections, components of black pixel connections are extracted as circumscribed rectangles, so extremely small items are included, and character pattern candidates are generated using the extracted circumscribed rectangles as they are. This results in poor processing efficiency. Therefore, as a pre-processing for character pattern candidate generation, the character chunk cutting unit 17 refers to the coordinate values and spacing of the circumscribed rectangles, and extracts close circumscribed rectangles that can be considered to belong to the same character or character part. A process of integrating the rectangles is performed, the rectangle after the integration process is cut out as a character block, and data such as its coordinates are stored in the character block memory 18 (step 53). A character block may be a character pattern itself or a part of a character pattern.

This integration process is performed, for example, as follows. A circumscribed rectangle with a width greater than two-thirds of the height of a line is defined as a full-width rectangle, and a value half the minimum value of the distance between adjacent full-width rectangles is determined as the minimum character spacing. Also, the maximum value of the width of the circumscribed rectangle is determined as the maximum character width. Then, among the extracted circumscribed rectangles, those whose spacing is less than or equal to the minimum character spacing are integrated. However, make sure that the width after merging does not exceed the maximum character width.

The standard character width calculation unit 19 refers to the data in the character block memory 18, detects the maximum height of character blocks in one line, and multiplies this value or this value by a correction coefficient. The corrected value is stored in the standard character width memory 20 as the standard character width for the line (step 54). Note that width means the size in the row direction, and height means the size in the direction perpendicular to the rows.

In the touching character discrimination section 21, the character block memory 18 is referred to, and the width of the character block in one line is compared with the double value of the standard character width of the line obtained in the standard character width memory 20. Then, a character block having a width exceeding twice the standard character width is detected as a character block of contact characters, and its image is cut out from line image memory 16 and stored in contact character image memory 22 (step 55). And for this image,
The touching character separation unit 23 takes a projection in the direction perpendicular to the line, divides the character block of the touching character at a position where the projection value (histogram of black pixels) is below a predetermined threshold value, and stores the character block in the character block memory 18.
Corresponding data is corrected (steps 56, 57). Figure 3
shows the separation of ``day'' and ``book'' that came into contact.

The character pattern generation unit 17 refers to the character block memory 18 and generates character pattern candidates by combining character blocks from the beginning of the line to the end of the line (from left to right), and each generated character pattern candidate The combination information of the character blocks constituting the character block and its image are stored in the character pattern memory 25 (step 58). However, make sure that the pattern width after combination does not exceed the maximum character width.

[0020] The recognition unit 22 extracts feature quantities for all character pattern candidates, performs matching with standard feature quantities in the recognition dictionary, obtains recognition result candidates, and selects the character code, distance, and total number of contours. Send to Department 27 (Step 5
9).

The selection unit 27 determines the character pattern to be cut out by evaluating the character recognition results of the character pattern candidates (step 60). That is, by referring to the character block combination information in the character pattern memory 25, an evaluation value is calculated by dividing the distance of the recognition result candidate by the total number of contours for a character pattern candidate whose character blocks overlap with other character pattern candidates. Calculate. This evaluation value has a high probability of being a small bullet character. Then, one character pattern candidate is determined as a character pattern from among a plurality of character pattern candidates with overlapping character blocks based on the respective evaluation values, and the recognition result data is output. for example,
Overlapping of character blocks occurs between a character partial pattern and its integrated pattern, so if the maximum evaluation value of each partial pattern is greater than or equal to the evaluation value of the integrated pattern, a character pattern candidate that is an integrated pattern is cut out as a character pattern, and in the opposite case, each character pattern candidate that is a partial pattern is cut out as a character pattern. Character pattern candidates with no overlapping character blocks do not require such evaluation, so they are determined as character patterns as they are, and the recognition result data is output.

[0022]

According to the invention as claimed in claim 1, it is possible to reliably cut out touching characters even in multi-sized Japanese documents, and it is also possible to cope with documents in which the character size varies from line to line.

According to the invention of claim 2, the multi-size
Characters, including touching characters, can be extracted with high precision from Japanese documents with irregular pitches.

According to the third aspect of the invention, it is possible to cut out touching characters with high precision without being affected by skew.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an apparatus according to an embodiment of the present invention.

FIG. 2 is a flow diagram showing the processing contents of an embodiment of the present invention.

FIG. 3 is an explanatory diagram of touching character separation.

[Explanation of symbols]

11 Binary image input device
12 Image memory 13 Circumscribed rectangle extraction section
14 Circumscribed rectangle memory 15 Line cutting section
16 line image memory 17 character block extraction section
18 Character block memory 19 Standard character width calculation unit
20 Standard character width memory 21 Touch character discrimination section
22 Contact character image memory 23 Exhortation character separation section
24 Character pattern generation unit 25 Character pattern memory
26 Recognition section 27 Selection section

Claims (3)

[Claims] Claim 1: A character extraction method that traces the connection of black pixels on a document image, extracts its circumscribed rectangle, extracts lines by integrating the extracted circumscribed rectangles, and extracts characters using the circumscribed rectangle for each line. , for each line, calculate the standard character width based on the size information of the circumscribed rectangle, detect the circumscribed rectangle corresponding to the touching character by comparing this standard character width with the width of each circumscribed rectangle, and calculate the detected circumscribed rectangle. For a rectangle, a character cutting method is characterized in that a projection in a direction perpendicular to a line is obtained and the circumscribed rectangle is divided at positions where the projection value is less than or equal to a predetermined value. 2. Track the connection of black pixels on a document image to extract its circumscribed rectangle, cut out lines by integrating the extracted circumscribed rectangles, and extract each line from a single circumscribed rectangle or a combination of multiple circumscribed rectangles. In a character extraction method that generates character pattern candidates, performs character recognition, and evaluates the recognition results to determine whether or not to finally extract each character pattern candidate as a character pattern, for each line,
Calculate the standard character width based on the size information of the circumscribed rectangle,
By comparing this standard character width with the width of each circumscribed rectangle, the circumscribed rectangle corresponding to the touching character is detected. For the detected circumscribed rectangle, the projection value is determined by calculating the projection in the direction perpendicular to the line. 1. A character cutting method characterized by performing a process of dividing the circumscribed rectangle at a position where the value is equal to or less than a predetermined value, and using the circumscribed rectangle after the process to generate a character pattern candidate. Claim 3: Claim 1, wherein the standard character width is calculated based on the maximum height of a circumscribed rectangle in each line.
Or the character cutting method described in 2.