JPS62224871A - Document picture processing system - Google Patents

Abstract

PURPOSE:To automatically extract the identification information of document from an input document by eliminating noises from an input picture, executing a specific preprocessing, thereafter extracting an area. CONSTITUTION:The picture of the document 3 is converted by a photoelectric converter 4, and stored in a memory 51. The said picture in the memory 51 is normalized, and stored in a memory 52. The picture in the memory 52 is eliminated from its noises, and stored in a memory 53. The picture in the memory 53 is subjected to a continuous filtering process, and stored in a memory 54. The picture in the memory 54 is compressed, and stored in a memory 55. The contour extraction processing is applied to the picture in the memory 55, and the result is stored in a memory 57. Since the input picture is simplified by the above-said continuous filtration, the processing can be speeded up. In such a way, the identification information of document can automatically extracted from the input document.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a document image processing method, and particularly to a document image processing method suitable for automatically extracting index information of the image from a document image to be stored in an image file. Regarding image processing methods.

[Conventional technology]

In a conventional document image file device, when storing a document image, it is necessary to input identification information of the inputted document image from a keyboard or the like. For example, in the case of a patent publication, the identification information of a document image includes the patent publication number and the non-pull (page number) in a combined publication.The task of inputting identification information from a keyboard is not only complicated. However, there was a problem in that the processing speed was reduced when continuously inputting a large number of documents using an automatic paper feeding mechanism. It is conceivable to enter identification information only for the first page, and then automatically increment the counter internally to generate identification information.

There were problems such as misalignment occurring when a certain page in a document was dropped.

Therefore, a specific area containing document identification information (hereinafter referred to as a heading area) can be extracted from a document image and the image of this part can be registered separately as a heading image, or characters in the heading image can be recognized and registered as code information. It is desired to do so.

In the prior art, published patent publication No. 17566/1983 (
2914 (January 1985), a method for obtaining a search keyword by recognizing characters in a specified area of a document is proposed. Also, published patent publication Showa fio-
In No. 17571 (published on January 29, 1985),
A method has been proposed that extracts a specific closed curve from an input image and recognizes characters inside the closed curve. Furthermore, Showa 6
Papers included in the 2008 IEICE General Conference National Conference Proceedings,
810-2 "Document image understanding using r format definition language" (
(Published on March 5, 1985) proposes a method of extracting character components from a document image by contour extraction and comparing them with a predetermined document format definition to identify each partial region of the document. Also, IBM Journal of Research and Development, No. 26
In the paper published in Volume 6, Pages 647 to 656, "Document Analysis System J (published in November 1982), a group (series) of continuous white pixels within horizontal and vertical scanning lines from an input image is used.
, and compare the length of the run with a predetermined threshold,
When the length of the run is smaller than the threshold, the process of replacing the run of white pixels with a run of black pixels is performed in the horizontal and vertical directions, respectively, and which logical product is made with the pixels of the image obtained.
A method has been proposed in which features are extracted from the resulting image and the original image is segmented into regions.

[Problems to be solved by the present invention]

The first conventional method is not suitable for automation because the area on the document is specified manually. The second method is limited to cases where the identification information of one document exists inside a closed curve, and is not common.The third method is powerful, but requires processing time and processor internal processing because it extracts the outline of a single document image. There is a problem that the amount of memory increases. The fourth method is effective for dividing documents containing a mixture of text and photos, but because it processes one document image horizontally and vertically, it increases the amount of internal memory of the processor and the processing time. It is easy to malfunction due to the small noise above, and it is difficult to describe the image characteristics of the heading area, so it cannot be immediately applied to automatic extraction of heading areas for filing. There's a problem.

SUMMARY OF THE INVENTION An object of the present invention is to realize a method of identifying a partial area containing identification information of a document using a simple configuration.

[Means to solve the problem]

The third method described above is considered to be basically suitable for the purpose described in the above [Prior Art], that is, automatic extraction of a heading area. The problem with the third method above is that
This problem can be solved by applying relatively simple preprocessing to the input image to simplify the input image and then extracting the region. Preprocessing suitable for this purpose involves removing noise from the input image, extracting a run of white pixels, and replacing the run of white pixels with a run of black pixels when the length of that run is smaller than a threshold. can be used.

[Effect]

It has been experimentally confirmed that by applying the above preprocessing, the number of contours can be reduced to approximately 1/20 for concave document images, and processing can be significantly simplified and speeded up. Furthermore, since the image becomes simpler after preprocessing is applied, data reduction through image thinning processing, which could not be applied to the original image, can be applied, making it possible to achieve even higher speeds. For example, 1 each in the vertical and horizontal directions.
If the pixels are thinned out to /4, the total number of pixels will be reduced to 1/16.

〔Example〕

Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 shows the flow of processing in an embodiment of the document image processing method of the present invention.
ys, tsDiagram). Second
The figure is a block diagram showing the configuration of an apparatus that implements the process of FIG. 1. Each part of the device is connected to a bus 1, and the overall operation is controlled by a controller 2.

101 in FIG. 1 is initialization of various variables used during processing.

This is the input of the document image 102 in FIG. 1, where the information (document image) on the document 3 shown in FIG.
Corresponds to storing in .

The memory 51 forms a part of the memory 5 together with 52 to 59, which will be described later.In the following explanation, it is assumed that one image is binarized into one bit, and the value "0" represents a white pixel, and II II II represents a black pixel. However, other expressions may be used.

103 in FIG. 1 performs normalization such as position correction processing and tilt correction processing on the document image.
An image obtained by normalizing the image of No. 1 is stored in the memory 52. If the positioning accuracy of the input image is sufficiently high, step 103 may be omitted.

At 104 in FIG. 1, a noise removal process is performed. This noise removal process removes small clusters of isolated black pixels in a white background, using a known method 1, for example, by scanning an image with a 3×3 mask. , a method such as replacing the 111" pixel surrounded by 10" with "0" may be used. 2nd I
In step iii1, the image from which noise has been removed from the image in the memory 52 is stored in the memory 53.

105 in FIG. 1 is a run length filtering process, and in FIG. 2, an image that has been subjected to a run length filtering process from an image in the memory 54 is stored in the memory 54. The detailed contents of the run length filtering process will be explained later with reference to the PAD diagram in FIG.

The first yA 106 is a part that performs a thinning process for one image, and in FIG. 55. The compression process in step 106 can be omitted if processing speed is not much of an issue.

107 in FIG. 1 is a part that performs contour extraction processing and extracts a sequence of coordinate points on the contour. In FIG. 2, the number and type of contour extracted from the compressed image in the memory 55 and the coordinates of each contour are shown. A sequence of points is obtained in memory 56.

This coordinate point sequence is obtained as many times as there are contours. Here, it is assumed that the contour is distinguished into an outer contour and an inner contour. The inner contour (outer contour) is the boundary of the black background surrounding the white background (the white background surrounding the black background), and when you go around the outline with the white background on the right and the black background on the left, the boundary changes clockwise (counterclockwise). Distinguishing between an outer contour and an inner contour, which are defined as contours, can be performed simultaneously with contour tracking using a known method.

FIG. 1 108 shows loop control in which the processes 109 to 113 are repeated for contours, where N is the number of contours and j is the contour number. In FIG. 2, processing is performed on the contour data in memory 56, and the processing results are stored in memory 57. 1
09 is a determination unit that determines whether the i-th contour is an outer contour or an inner contour, and when it is an outer contour, processes 1.10 to 113 are performed.

]10 is the process of calculating the width W(i) and height H(i) of the first contour. Specifically,
Maximum and minimum values of coordinates and Y coordinates X wax.

Find X max , Y max , Y min and calculate W (i) = Xmax - XmaxH(i)
=Ymax-Ymin. 111 is a determination unit that determines whether the i-th contour satisfies the conditions as a character line. For example, Wl<W (i) <W2H1<H(i
) <H2 Kl<H(i)/W (i) <K2,
Here, Wl, W2. Hl, H2゜Kl, and 2 are parameters for extracting a long and narrow region, and the above equation is the first
It specifies that for an outline to be a character line, its size and aspect ratio must fall within a certain range. At step 112, the number K of extracted character lines is increased by 1. At step 113, the representative position coordinates (for example, X win, etc.) of the i-th axis are registered in the character line coordinate registration area. Initialization process 10
1 is set to 0.

The processing shown in FIG. 1 114 to 125 is a portion in which the extracted character line area is compared with format data regarding the target character. The format data consists of M statements.
Each sentence includes M1 sentences regarding the position and size of the heading area to be extracted, and M2 sentences regarding the mutual relative positions (M.

Although M1 and M2 change depending on the format data number j, the subscript j is omitted below for simplicity.)

114 in FIG. 1 shows 115 to 1 for five format data.
This is a loop control that repeats 20 processes, and j represents the number of format data.

115 in FIG. 1 is a part for extracting the jIF-th format data, and specifically, this is done by writing format data from the control section 2 into the memory 58 in FIG.

FIG. 1 116 shows a loop control in which the processes 117 to 119 are repeated for Ml sentences, where m represents the number of the heading area. 117 for row areas 118-119
118 is a determination unit that determines whether the row area satisfies the conditions for the m-th heading area, and k represents the line area number. The condition that the representative position coordinates are included in the m-th sentence,
For example, the determination is made by comparing L ] < Xmjn < L 2 D 1 < Ymin < D 2 . If the determination result is that the matching is successful, k is registered as a candidate C(m) for the m-th heading area in step 119.

When the loop from 116 to 119 is completed, it is determined in 120 whether candidates C(m) have been found for all the heading areas, and if all have been found, the processes from 121 onwards are performed.

121 is a loop control that repeats the process of 122 for M2 sentences, and n represents the sentence number. 122, the relative positional relationship between the two heading areas I (a) and I (b), for example, R1<Xmjn (b) -Xma
x (a) is described, where a and b are specific area numbers. Therefore, the extracted candidate regions C(a), C
It is sufficient to check whether the representative coordinates for (b) satisfy the above conditions.

When the loop from 121 to 122 is completed, it is checked in 123 whether the conditions regarding all relative relationships are satisfied, and if the conditions are satisfied, j is identified as the format number of the input document in 124, and j and Candidate area C(m) (m=1.ML)
and their representative coordinates are registered in the memory 57, and the loop 114 is exited at step 125. In this embodiment, when the determination at step 123 is "no", the process proceeds to the next step through the loop at step 109 without doing anything, but it is also possible to try again.

That is, the loop 116 and subsequent steps are executed multiple times, and from the second time onwards, the previously determined candidate area (k=C(m)) is excluded when determining the heading area (m) in 118. In order to make this process possible, the correspondence between m and k is stored in the form of a matrix or a list, and then used during the first trial.

All you have to do is clear this correspondence.

If the format number to be collated cannot be found even after processing all formats j in step 114, an error is displayed in step 126 and the processing for this input document is stopped.

FIG. 1 127 shows 128 to 1 for Ml heading areas.
128 is a loop control that repeats the process of 29, where m represents the number of the heading area, 128 extracts the representative coordinates of this area from the registered candidate area C(m) number in the memory 59, and 129 represents this area. A partial area specified by the representative coordinates is extracted from the original image stored in the memory 51 in FIG.
This is the part that recognizes the characters in it, and single character recognition is executed by the character recognition unit 6 in FIG. 2, 0 character l! The ITa results are stored in memory 59 in FIG. Since the normalization process 103 and the thinning process 106 are performed, the changes caused by these processes are corrected when extracting the partial area specified by the representative coordinates.

As with normal character recognition, it goes without saying that characters that fail to be recognized can be corrected using the console of the control unit 2, etc.

130 is a process for outputting the input document image and identification information (ie, the character string resulting from character recognition) to the file 7 in FIG. 2.

Next, details of the run length filtering process will be explained.

FIG. 3 is a PAD diagram showing the flow of run length filtering processing, and FIGS. 4 and 5 are diagrams explaining the principle thereof. Note that the size of the input image is assumed to be xxY pixels.

301 in FIG. 3 is an initialization process such as clearing the work area 6 302 in FIG.
310 is a loop control for repeating processing, and y represents a scanning line number. 303 is initialization for the y-th scanning line, in which the consecutive length of white pixels is set to zero. 304 is a loop control that repeats the processes 305 to 310 from left to right for pixels in the y-th scanning line, or represents a pixel number.

In 305, it is determined whether the pixel P(x, y) is white or black. P
If (xt y) is a white pixel, the white pixel is increased by 1 in 306. In the case of black pixels, set the white continuous length to 307.
Compare with. When L=O, the previous pixel is also a black pixel, and nothing is done in this case. When L≠0, in 308, the continuous length of white is compared with a threshold value LO. L<L
When O.

When the white image elements are continuous for a long enough time, nothing is done in this case either. When L<LO, white's length is short and 3
By the processing of 09 to 310, the own speed of and is changed to black. 3
09 is a loop control that repeats the process of 310 for L white pixels whose X coordinates are from wl to w2 (that is, preceding the current pixel), and W represents a pixel number. 3
10 indicates that the pixel P (W?y) is changed from white to black.

The run length filtering process AD diagram of FIG. 3 explained above,
FIGS. 4 and 5 are diagrams explaining the principle. Note that the size of the input image is assumed to be X X Y 1iIiii elements.

301 in FIG. 3 shows initialization processing such as clearing the work area.

302 in FIG. 3 shows the number of scanning lines 303 for the input image.
This is a loop control that repeats the processes from 310 to 310, and y represents the scanning line number. 303 is initialization for the y-th scanning line, where O is the continuous length of white pixels. 304 is a loop control that repeats the processes 305 to 310 from left to right for the small pixel of the y-th scanning line, or represents a pixel number.

In 305, it is determined whether the pixel P(xty) is white or black. P
If (xt y) is a white pixel, the white run length is increased by 1 in 306. In the case of black pixels, set the white continuous length to 307.
Compare with. When L=O, the previous pixel was also a black pixel, and nothing is done in this case. When L≠0, the white run length r of 308 is compared with the threshold LO. L
> When LO.

When white pixels are continuous for a long enough time, nothing is done at this time either. When L<LO, white's length is short and 3
This own speed is changed to black by the processing of 09 to 310. 3
09 is a loop control in which the process 310 is repeated for L white pixels whose X coordinates are from Wl to W2 (that is, preceding the current pixel), where W represents a pixel number. 31
0 indicates that the pixel P(wty) is changed from white to black.

By the run length filtering process in FIG. 3 explained above,
For example, in the image illustrated in FIG. 4(A), a group of white pixels 41
, 42, and 43 are converted into a black pixel group as shown in FIG. However, hatched pixels in FIG. 4 represent black, blank pixels represent white, and the parameter LO is set to 5. In addition, in Figure 4, it is assumed that there is a sufficiently long white area on both the left and right sides, and the white lines at both the left and right ends are not converted to black (although they are short in this example). may not be converted to black.

The results of applying the run length filtering process shown in FIG. 3 are schematically shown.1 For example, the image shown in FIG.
Adjacent characters are connected to integrate each character line into one area. In this process, the noise removal shown in FIG. 1 104 is time consuming, and if this is not done, two character lines will be connected by the noise, and character line extraction may fail.

Although one embodiment of the present invention has been described above, it goes without saying that various changes can be made without affecting the essence of the invention. For example, without recognizing the characters in the extracted heading area, It is also effective to store the image of the heading area in a file together with a pointer to the original image, and when displaying, display the image of the heading area first, and then display the original image connected to the heading area specified by the user. be.

Further, in the embodiment, a horizontally written document is assumed, and the scanning direction for finding runs is horizontal, but a vertically written document may be scanned vertically. Furthermore, for white characters, a series of black pixels may be used instead of a series of white pixels.

Further, in the embodiment, the entire document image is processed, but a range to be processed may be specified separately, and the index area may be extracted only within that range.

In addition, the extraction of the heading area is performed at the same time as the photoelectric conversion of the document, but first, only the photoelectric conversion of the document and storage in a file is performed, and then the document image is read out from the file and the heading area is extracted. You can do it like this.

Furthermore, when format identification or heading area extraction fails,
It is also possible to display intermediate results of processing on the console to request user assistance.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to automatically extract document identification information from an input document, and the conventional human keyboard work is not required or is significantly reduced.

[BRIEF DESCRIPTION OF THE DRAWINGS] FIG. 1 is a diagram showing the flow of processing in the document processing method of the present invention, FIG. 2 is a block diagram showing the configuration of a device that implements the processing contents of FIG. 1, and FIG. FIG. 1 is a diagram showing the flow of the run length filtering process in FIG. 1, and FIGS. 4 and 5 are diagrams explaining the processing contents in FIG. 3. 1...Bus, 2...Control unit, 3...Document, 4...
・Photoelectric conversion device, 5...Memory, 6...Character recognition unit, 7...Faiyu 4 diagram, 5 mouth

Claims (1)

[Claims] 1. Means for inputting an image converted into a digital format through photoelectric conversion, sampling, and quantization, and generating a second image from which minute noises in the input image are removed. Then, extract a continuous white or black pixel group within a horizontal or vertical scanning line from the second image, and compare the continuous length of the white or black pixel group with a predetermined threshold. , when the value is smaller than the threshold, a process of replacing the white or black pixel group with a pixel group with inverted black and white is performed on all or part of the input image to generate a third image, and from the third image, means for extracting a characteristic region of a document image, and the extraction of the characteristic region involves extracting connected regions from the image and determining the size, shape, and absolute position of the connected regions, and the relative sizes of the regions. A document image processing method comprising: detecting that a positional relationship is included in a predetermined range.