JP2813601B2 - Tabular document recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention cuts out ruled lines and line-to-line rules in a table when image data of a tabular document is input, and outputs characters in a frame surrounded by these ruled lines. The present invention relates to a tabular document recognition device for accurately recognizing a document.

[Conventional technology]

In conventional table format document recognition, characters are recognized by cutting out only actual ruled lines (hereinafter, also referred to as real ruled lines) such as solid lines and dotted lines. As one of the proposals, the applicant has proposed the following (Japanese Patent Application No. 1-214934;
Hereinafter, it is referred to as a proposed method).

In this method, for a tabular document as shown in FIG. 10, the processing area 11 specified first is divided into strips with a constant band width 12, and each band is projected in the vertical (y) direction. Next, from the projection data, a part considered as a part of the ruled line (ruled line candidate) 13
Is extracted. After this is extracted for each band, the candidate having the largest overlap is searched for from a neighboring band with respect to a certain candidate of interest, and is connected as the same ruled line candidate. And
For the connected candidates, the operation of searching for the candidate having the largest overlap from the adjacent band is repeated, thereby creating a connected candidate group 14 considered to be the same ruled line. Here, the horizontal projection of these connection candidate groups is defined as a partial projection 15, and the coordinates 16 at both ends are determined by taking a projection in a direction perpendicular to the band projection to obtain a ruled line. A final ruled line group is obtained by performing processing for combining or integrating the ruled line groups thus obtained into one. Although FIG. 10 shows an example of extracting horizontal ruled lines, it goes without saying that vertical ruled lines are also extracted in the same manner as described above.

On the other hand, as for the line rule drawn between the character strings, the processing region is divided into bands for the image data of the tabular document, and the vertical (y) direction A blank portion which is considered as a part of the ruled line between the lines is projected and extracted as a candidate. Then, for each candidate group, an operation of finding the most overlapped candidate from the adjacent band is repeated to create a group of line ruled line connection candidates, and a line projection rule is also determined by taking a partial projection. At this time, unlike the actual ruled line, the line ruled line is different in size because it picks up all blank spaces that are originally empty as candidates, so that the line spacing candidate is divided as necessary, thereby consolidating the lines. Is easy to perform. This is shown in FIG.

First, calculate the slope from the already determined actual ruled line,
The ratio 20 of the band width to the slope is set as the candidate shift width. If there is no actual ruled line, the deviation value of the candidate is obtained while actually dividing the line, and the average value is used. Next, the number of overlapping candidates is counted from the band adjacent to the left and right with respect to a certain line spacing candidate 18 of interest, and as a result, the number of overlapping candidates is adjusted to the direction in which there are many overlapping candidates. The image is divided as shown in FIG. The candidates obtained in this manner are used as the division candidates 19 in the candidate connection process.

As described above, the actual ruled line or the line-to-line ruled line is obtained, and the coordinates of the frame surrounded by these ruled lines are sent to the character recognition unit, whereby the characters in the frame are recognized and the structure of the table can be grasped.

[Problems to be solved by the invention]

However, in the above-described method, if the input image is inclined, there is a problem that the accuracy of cutting out the ruled line between lines is reduced. In other words, in a method in which the image is divided into bands and the projection is performed in the orthogonal direction for each band to extract the ruled line between lines, a deviation occurs in the end point coordinates (y coordinate) of each connection candidate.
This is because the candidate deviation width used when dividing the candidate is the value of the band width with respect to the gradient, but the gradient is a real value, whereas the candidate deviation width must be expressed by an integer value. As a result, the value of the slope is not accurately reflected. That is, there is a difference between the inclination of the line ruled line 23 obtained from the candidate group 21 obtained by connecting the division candidates as shown in FIG. 12 and the inclination of the actual ruled line 22 obtained before the extraction of the line ruled line. ,
It becomes necessary to correct the inclination of the ruled line.

From the above, it can be understood that the accuracy of the ruled line where the ruled line is drawn in the blank portion where there is no original becomes a problem especially when the image is inclined.

Accordingly, it is an object of the present invention to accurately extract a ruled line between lines even when an image is inclined.

[Means for solving the problem]

An image input unit for inputting image data, a real ruled line extracting unit for cutting out solid ruled lines from the image data, a tilt detecting unit for detecting a tilt of the image from the extracted ruled lines, and an image having ruled lines deleted from the image. A projecting unit that obtains an image in the inclination direction obtained by using the image, a line rule extracting unit that extracts a line rule from the projection data, and a character that recognizes a character in a frame surrounded by actual rule lines and a group of line rules between lines And a recognition unit.

[Action]

The actual ruled lines are extracted, their inclinations are calculated, and projection is performed to cut out the line ruled lines in the direction of the inclination, thereby realizing the cutout of the line ruled lines independent of the image inclination, and further improving the extraction accuracy of the line ruled lines. This improves the recognition accuracy of the table structure.

〔Example〕

 FIG. 1 is a block diagram showing an embodiment of the present invention.

The tabular document 1 including the actual ruled lines and the line ruled lines is transferred to the host CPU 2
Is converted into binary image data by the image input unit 3 controlled by the control unit 2 and stored in the image memory 4. The ruled line extracting unit 5 includes a solid line ruled line extracting unit 51, an inclination detecting unit 52, a projecting unit 53, and a line ruled line extracting unit 54, and determines a processing area on the image memory 4 to extract a real ruled line and a line ruled line. The coordinates of the frame surrounded by the obtained ruled line group are sent to the character recognition unit 6, and the characters in the frame are recognized.

Hereinafter, a specific description will be given with reference to FIGS. 2 to 9. It should be noted that the actual ruled lines are extracted in the same manner as in the prior art, and therefore the details are omitted.

After the actual ruled line is extracted, its inclination is calculated and the result is stored in a predetermined memory. Here, the line rule is extracted using an image in which vertical and horizontal actual rule lines as shown in FIG. 2 are deleted from the image. (The dotted line in FIG. 2 indicates that the portion where the actual ruled line originally existed was deleted). This is to make it easy to extract a candidate for a line spacing rule (a portion considered as a part of the line spacing rule) by clearly distinguishing a blank portion between lines and a character portion. Next, the inside of the frame 11 surrounded by the outermost frame ruled line is set as the processing region of the line spacing rule (in the case of a tabular document having no outermost frame ruled line, the processing region specified first is used as it is) and saved first. The projection method is determined based on the value of the tilt, and the processing area 11 is divided with a constant bandwidth as shown in FIG. 3, and the projection in the tilt direction is performed. Then, in order to extract a line rule candidate from each band, the projection data is binarized again at a certain binarization level 24 as shown in FIG. 4, and "0" is constant from the binarization data. A portion continuous over the width is extracted as a line spacing candidate 17, and a portion where “1” is continuous over a certain width as a character candidate 25.

In the line spacing candidates 17 thus obtained, all blank portions are detected, and the sizes of the candidates are various as shown in FIG.
Therefore, at the time of connection, an operation of dividing candidates is performed as necessary. That is, when a plurality of line spacing candidates are connected when looking at a band adjacent to the left and right of a certain candidate of interest, the line spacing candidates are divided. For example, as shown in FIG. 6, for the line spacing candidate 18 of interest, the number of candidates overlapping with the candidate is counted from the candidate group of the band adjacent thereto.
Then, in accordance with the direction in which the count number was large, the sixth
In the case of the figure, the count number in the forward direction (right side of the line spacing candidate of interest) = 2, while the count number in the backward direction (left side of the line spacing candidate of interest) = 3. Divided into At this time, since the division is performed based on the inclination of the image and the projection in the direction perpendicular to the image, positional displacement does not occur as compared with the case of dividing by the ruled line displacement width as in the proposed method.

Next, the resulting candidate is referred to as a division candidate 19, and the line spacing candidates and the character candidates are connected together including this division candidate. First, for the line-of-interest candidate 18, a candidate that is the most overlapped in the forward direction is searched for from adjacent band candidates, and the candidates are connected as the same ruled line candidate. Repeat the operation to find the most overlapping candidate from the list. The same process is performed in the opposite direction (backward direction) to extract a group of connection candidates 14 having the same ruled line as shown in FIG.

Next, it is determined whether or not it can be extracted as a ruled line from the connection candidate group 14 by examining the validity of each candidate in the connection candidate group. For example, the number of line spacing candidates 17, the number of division candidates 19, and the number of character candidates 25 are counted for the connection candidate group 14 as shown in FIG. 8, and the number of line spacing candidates + division candidate> character candidate × P (P is an arbitrary level) When satisfies, it is determined that it is suitable as a line spacing candidate.
That is, the fact that the ratio of including character candidates is high means that the line spacing candidates in the concatenation candidate group are often character strings and margins between the character strings, and drawing a ruled line there causes inconvenience.
Therefore, as shown in FIG. 9, for example, as shown in FIG. 9, a partial projection in a direction parallel to the inclination is performed on the connection candidate group that satisfies the above condition. By determining the coordinates of both ends, the ruled line 23 can be obtained.

As described above, the table formed by the actual ruled lines and the line spacing ruled lines is recognized.

〔The invention's effect〕

According to the present invention, since the actual ruled line is extracted, its inclination is obtained, and the projection for cutting out the line ruled line is taken in the direction of the inclination, the ruled line can be drawn at an accurate position. Further, since the accuracy of extracting ruled lines is improved, characters surrounded by the ruled lines can be correctly recognized, and an advantage that a table format document can be accurately grasped can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention. FIG. 2 is an explanatory diagram for explaining an example of an image in which solid ruled lines are deleted.
FIG. 3 is an explanatory diagram for explaining band projection, and FIG.
FIG. 5 is an explanatory diagram for explaining a candidate extracting method, FIG. 6 is an explanatory diagram for explaining a candidate dividing method, and FIG. 7 is a candidate connecting method. , FIG. 8 is an explanatory diagram for explaining a connected state of candidates, FIG. 9 is an explanatory diagram for explaining a partial projection, and FIG. 10 is a conventional method for extracting a solid ruled line. FIG. 11 is an explanatory diagram for explaining a conventional candidate dividing method, and FIG. 12 is an explanatory diagram for explaining an error due to image inclination. Description of symbols 1... Table format document 2... Host CPU 3... Image input unit 4... Image memory 5... Ruled line extraction unit 6... Character recognition unit 11. ... Band width, 13 ... Rule line candidate, 14 ... Connection candidate group, 15 ... Partial projection, 16 ... Rule line end point, 17 ... Line spacing candidate, 18 ... Target line spacing candidate, 19 ... Division candidate, 20 ... … Candidate shift width, 21 …… Line spacing ruled line connection candidate group,
22 ... actual ruled line, 23 ... line spacing ruled line, 24 ... re-binarization level, 25 ... character candidate, 51 ... solid line ruled line extraction unit, 52 ... inclination detection unit, 53 ... projection unit, 54 ... ... Line spacing rule extraction unit.

Claims (1)

(57) [Claims] An image input unit for inputting image data; an actual ruled line extracting unit for cutting out a solid ruled line from the image data; an inclination detecting unit for detecting an image inclination from the extracted ruled line; A projection unit for projecting the image in the inclination direction obtained by using the image deleted from the image, a line rule extraction unit for extracting the line rule from the projection data, and a frame surrounded by the actual rule and the rule group between the lines. A tabular document recognition device, comprising: a character recognition unit for recognizing characters.