JPH0337785B2 - - Google Patents

Description

[Detailed description of the invention]

(Technical Field) The present invention relates to a document image processing device that targets unknown input document images, including characters, figures,
This is a device that extracts simple structural information such as photos, titles, and text, facilitates interactive processing with humans, and also enables automatic extraction. (Background Art) A conventional image processing device is shown in FIG. 1-1 is an image input device that quantizes density values by 4 bits per pixel, and 1-2 is an image memory that stores an input image. The image stored in 1-2 is assumed to have (N x M) pixels as one element (in this case, N=M=12), and the average density value (1-3) within the element is assigned to each element, and only rough density information is obtained. A first abstract image consisting of is stored in the image memory 1 (1-4). Next, (1-4)
The difference in average density between elements is determined based on the images stored in (1-5). Here, the average concentration of the current element is D _ij , the average concentration values of the four neighbors are D _i,j-1 , D _i-1,j ,
When D _i,j+1 and D _i+1,j (see Figure 2), 1/2 |D _i,j −D _i-1,j |+1/2|D _i,j −D _{i ,j-1} ｜ to an appropriate value (1/4 of full scale, i.e. 4N ² )
Or normalize the value of max ^l,m=1,+1 (D _i,j −D _il,jn ) with an appropriate value (3/2N ² ) to establish the initial value of the main contour strength. It is expressed as P ⁽⁰⁾ _ij and stored in the image memory 2 (1-6) as a second abstract image.
However, when it takes a value of 1 or more, it is treated as 1. Based on the second abstract image, main contour elements are extracted using the relaxation method (1-7). From the values of each (i, j) element of the second abstract image, the initial image space is defined as P ⁽⁰⁾ = {P ⁽⁰⁾ _i,j | 1≦i≦n, 1≦m; i, j integer} Then, perform the iterative calculation using the following formula. That is, P ^(k+1) _ij =P _ij ^(k) q _ij ^(k) /P _i,j ^(k) q _ij ^(k) + (1−P _ij
^(k) (1−q _ij ^(k) )q ^(k) _ij =P ^(k) _ij −(1−Δ ^(k) _ij ) Δ ^(k) _ij =max{1/2|P ^(k) _{i -1,j} +P ^(k) _i+1,j −P ^(k) _i,
j-1 −P ^(k) _i,j+1 ｜, 1/2｜P ^(k) _i-1,j-1 ＋P ^(k) _i+1,j+1
−P ^(k) _i-1,j+1 −P ^(k) _i+1,j-1 |}. Here, the physical quantity Δ ^(k) _ij is related to the connectivity of each element of the second image, and the value P ^(k) _ij of each image element is its own contour strength P ^(k) _ij and connectivity Δ ^(k) It will be updated by _ij . This repeated operation is terminated when the number of elements with values greater than or equal to 0.9 or less than or equal to 0.1 reaches 0.95 or more of the total number of elements. Next, it is binarized and stored again in the image memory 2 (1-6). Second
The image is an image composed of main outline elements. Next, area segmentation is performed based on the second image (1
-8). The main contour elements are spatially integrated by 8 connections,
Divide by circumscribed rectangle. The representative points are stored in the image table (1-9). Each small area is labeled (1-10) with reference to the image table (1-9). Labeling is done as follows. Two physical quantities are calculated based on the first image and second image corresponding to each small area. That is, (1) find the difference in average density on both sides of the main contour element as a boundary, and calculate the average value of change values within the region (ΔD); and (2) the ratio of the number of main contour elements occupying the region (D _N ).
It is. ΔD vs. D When ΔD≧0.7 on _{the N} plane, “photo”;
“Graph” when ΔD<0.7 and D _N <0.4, and ΔD
<0.7 and D _N >0.4, each label of "character" is attached, and as a result, the image table shown in Figure 3 (1
-9) is obtained. Therefore, the conventional technology has the disadvantage that character strings and organized descriptions of each character string are not made from the area determined to be "character". (Problems to be solved by the invention) The present invention eliminates these drawbacks and makes it possible to perform search and text extraction by extracting character strings and organizing character strings. The present invention
Perform orthogonal transformation on the abstracted first image "character" area, calculate the rough pitch, prepare a corresponding mask, extract the character string by logical operation, and calculate the size of the extracted character string. An image processing apparatus is characterized in that a title and a main text are extracted based on the size, position, and number of lines with similar pitches. (Structure and operation of the invention) FIG. 4 shows an embodiment of the present invention.
-11 is 1-1 of the conventional image processing device (Fig. 1)
~1-11 is completely equivalent. Step 4-12 refers to the image table and performs orthogonal transformation on the region of the "character" label in the image memory (4-2) to determine the spatial frequency. Here, a discrete Fourier transform is employed as the orthogonal transform, and the two-dimensional orthogonal transform is approximated as follows. p _ij to image memory (4-
2), and the spatial frequency spectrum in the x and y directions is

Calculate by [Formula] (Figure 5). Here, L _x and L _y are the spatial sizes of the regions in each x and y direction. P _k , P _l
Find the peak of the spectrum. The peak of the spectrum is the spectrum component excluding the DC component, and if it is above a certain threshold (10% of the magnitude of the DC component) or if there are multiple peaks, the peak of the spectrum on the lowest frequency side is adopted, and the spatial frequency (k _x ,
Find k _y ). If there is no equivalent, set k _x ork _y to 1. Next, character string extraction (4-13) is performed with reference to (k _x , k _y ) (L _x , L _y ). First, it is determined whether the writing is vertical or horizontal. When L _y /k _y <L _x /k _x , it is written vertically, and when L _y /k _y >L _x /k _x , it is written horizontally. Next, prepare a mask for vertical writing (L _y /k _y pixels x L _y /8k _y pixels) and for horizontal writing (L _x /8k _x pixels x L _x /k _x pixels), and use the logical sum (threshold value). is half of the full scale) calculation is performed in the image memory (4
-2) and extract the character string. The area of each character string is described by its circumscribed rectangle, and two points (FIG. 6) are stored in the image table (4-9). Each character string is numbered from the largest j _nax when written vertically, and from the smallest i _nio when written horizontally, and organized by line number. However, addresses of any two character strings (i _1nio ,
i _1nio ) (i _1nax , i _1nax ) and (i _2nio , j _2nio ) (i _2nax ,
j _2nax ) is given, in horizontal writing i _1nio <
(i _2nio ori _2nax ) <i _1nax , and in vertical writing j _1nio <
(j _2nio orj _2nax ) < j _1nax , give the same line number, and those with the same line number will be rewritten according to the new format.
Ranking is performed in the iorj direction. In this way, an image table as shown in FIG. 6 is obtained. Next, the character size (L _c ) and line pitch (L _p ) of each character string are estimated as average values from each character area. That is, It is. Therefore, i _onax , i _onio , j _onax , and j _onio are the representative addresses of each character string, and N is the number of character strings extracted in that character area. Also, (I _1nio , J _1nio ),
(I _1nax , J _1nax ) is the representative address of the character area, and N _L is the number of lines included in the character area. Based on L _c , L _p , N _L and coordinate values, the main text, headings, and flow of the main text are identified. First, L _c ,
Classified by L _p . Combine L _c and L _p within ±10%. After classification, the text with the largest number of lines is used as the main text, those with the same format are grouped together, and the others are placed in another group. Rank those with the same format according to the format. In the case of horizontal writing, compare two character areas (I _1nio , J _1nio ) (I _1nax , J _1nax ) and (I _2nio , J _2nio ) (I _2nax , J _2nax ), and if J _1nax
<J _2nio or I _1nax <I _2nio , the former is ranked first. If not, the order will be reversed. In the case of vertical writing, similarly, when J _1nio > J _2nax or I _1nax < I _2nio , the former is ranked first,
If not, the order will be reversed. As a result, items with the same format as those labeled “body” are arranged in order. Find a string labeled “body” that has a value greater than or equal to L _c . If so, check its ranking. That character area is the “body”
If the title is located between or below the character area labeled as "subtitle", it is labeled as "subtitle". If there is a character area higher than the character area labeled “body”, the highest
If the value of L _c is “body” or “subtitle”
When the value is greater than 1.1 times the value of L _c , it is considered a “title”. Furthermore, a search is made for a value that is equal to the L _c value within ±10% from the top, and is labeled as a "title." If you find something that is not, use what follows as the "subtitle." Regardless of the format, if the value of L _c is 1.5 times or more than the main text, a "title" label is given. (Effects of the Invention) The present invention extracts character strings, ranks each character area, and labels the text title and subtitle based on the nature of the character strings, making it possible to automatically file and edit free-format documents. Become. Furthermore, a flexible document image editing device can be realized by means of human interaction.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a conventional image processing device, FIGS. 2 and 3 are explanatory diagrams of the operation of the device in FIG.
This figure is a block diagram of an image processing apparatus according to the present invention, and FIGS. 5 and 6 are explanatory diagrams of the operation of the apparatus shown in FIG. 4. 4-12; Orthogonal transformation circuit; 4-13; Character string extraction circuit.

Claims (1)

[Scope of Claims] 1. means for converting the density of each pixel of the original image into an electrical signal and storing it; means for storing the average density of an element composed of a plurality of pixels as a first image;
The image is divided into regions, and each divided region is divided into two parameters: the proportion of the main contour elements (D _N ) and the average density change value on both sides of the main contour elements (ΔD). Therefore, the photographic field,
In an image processing apparatus having means for classifying into graphic areas and character areas, a first image processing apparatus configured with an average density for each character area
means for obtaining a spatial frequency (k) of the image by performing orthogonal transformation on the image; means for providing a mask of a size corresponding to the spatial frequency; and means for extracting a character string by a logical operation on the mask and the image; An image processing apparatus comprising means for determining vertical writing or horizontal writing based on spatial frequency and providing a label to the line number of each character string.