JPH0418749B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a region extraction method for extracting similar image regions from an image.

(Prior Art) As this type of region extraction method, the 12th Image Engineering Conference 7-2 "Recognition of Rectangular Fields and Handwritten Identification Numbers in Facsimile Documents" is known. In this method, the image is first divided into small areas, then binarized, the entire screen is combined into multiple large rectangular areas using a logic plate, and each large rectangular area is then divided into window areas. , the probability of occurrence of black pixels and the local bias within the window of black pixels are determined. Furthermore, the area projection of the data onto the coordinate axes (X, Y) is determined as the marginal distribution. Then, based on the probability of black pixel occurrence and the local bias within the window, we looked at the attributes of "photo", "diagram", and "text", examined the periodicity of the peripheral distribution, and determined that "text" and "diagram"
The area is extracted based on the judgment.

(Problems that the invention purports to solve) By the way, this method has poor compatibility with arbitrarily shaped regions because it is first grouped into large rectangular regions, and region extraction is difficult for non-periodic character units. The problem was that it was difficult. Therefore, a method that is adaptable to arbitrary shapes and arbitrary images has been desired.

The present invention was made in view of the above problems, and
The purpose is to provide a region extraction method that can extract image regions of arbitrary shapes.

(Means for Solving the Problems) The present invention, which solves the above problems, divides an image into a plurality of blocks and classifies each block into a predetermined pattern using binarized image data.
This method is characterized in that the connection between each block is checked using the pattern and similar image areas are extracted.

(Example) Hereinafter, the method of the present invention will be specifically explained using the drawings. FIG. 1 is a flowchart showing the steps in a specific example of the method of the present invention. In this example, the background level (concentration level) is first determined, the image is then blocked, and the threshold value determined based on the background level is The pixels in each block are binarized using the binarized image data, the texture of each block is analyzed, each block is classified into a predetermined pattern, the connection between each block is investigated using the pattern, and the Extracting image area. Note that the order of background level determination and image blocking may be reversed.

Next, the above procedure will be explained in detail step by step.

() Determination of background level In order to determine the background level in a short time, as shown in Figure 2, m x m (for example, 32
×32) Create a pixel window,
A density histogram as shown in FIG. 3 is created based on the image data in the window, and the background level is determined. FIG. 4 is a flowchart showing the procedure for determining the background level. First, a window is determined (position determined), a density histogram is created based on the window, and the average value M and standard deviation S are calculated. Next, the standard deviation S is set to a predetermined threshold T
If it is smaller, the background level B is selected as a value within the range of M-3S≦B≦M+3S. Usually, B=M+
3S. On the other hand, if the standard deviation S is equal to or greater than the threshold value T, the window position is moved in the horizontal or vertical direction in FIG. 2 to determine a new window, and the density histogram is created again. This is because when S≧T, the window does not exist at a position where the background can be captured well.
Note that the size of the threshold T may be any value as long as it is within a range that can capture the background, but it is preferably a size that allows the window to be determined a minimum number of times.

() Blocking an image Divide an image into multiple blocks. For example, when the image is N×M=2000×2500 pixels (10 dots/mm) as shown in FIG. 5, the size of each block is n×n=32×32. Of course, 16×16 or 64×
You can also choose 64 etc.

() Texture analysis of each block Next, the image data is binarized using the background level B as a threshold, and each block is classified into a predetermined pattern using the binarized image data obtained thereby. As the patterns prepared in advance, for example, 16 patterns (0 to G) as shown in FIG. 6 are used. This pattern changes as it changes from 0 to G, i.e. (0, 1) → (2~5) → (6,
The pattern has many black and white changes each time it changes from 7) to (8 to B) to (C to F) to G. To classify each block into the above patterns, for example, as shown in FIG. Pattern classification can be easily performed using a computer or the like by using a method of finding a corresponding pattern from two cumulative number distributions in the X and Y directions. In the example of FIG. 7, black pixels are concentrated at the bottom right of the block, so the cumulative number on the right side is large in the X direction, and the cumulative number on the bottom side is large in the Y direction. Since the same tendency is shown when projected in a two-dimensional direction, it is concluded that the block corresponds to pattern 8.

() Extraction of image region The connection between each block may be examined in units of blocks, or each block may be divided into two vertically and horizontally (four divisions in total) and examined.
Figure 8 is a simplified explanatory diagram to show the latter method. Figure 8A is a diagram in which the image is made into a block and a binarized image is superimposed on it (the shaded area is the black pixel area). B is a diagram in which the pattern numbers of FIG _. 6 are attached to _each block, and _{FIG .} FIG. 6 is a diagram in which the pattern numbers in FIG. 6 are assigned to each obtained small block. When each pattern in FIG. 6 is divided into four, it is possible to know in advance which patterns each pattern will be decomposed into, so it is extremely easy to obtain FIG. 8(b) to FIG. 8(c). Note that when each pattern in Figure 6 is divided into four, each pattern in Figure 6 may be selected so that it is accurately decomposed into the patterns in Figure 6, but it is possible to select each pattern in Figure 6 so that it is decomposed into the patterns in Figure 6 approximately. It is sufficient even if the Next, the connection between the small blocks is examined, and in order to determine that there is a connection relationship, for example, there is a criterion for determining whether there is a connection relationship between the patterns of adjacent small blocks, such as if there is at least a black part on both sides of the joint. Alternatively, the presence or absence of a connection relationship may be determined by taking into consideration the overall shape of the pattern. In any case, since the combinations of pattern connections are limited, it is possible to determine whether or not small blocks are connected by determining in what positional relationship each pattern should be adjacent to each other to determine if small blocks are connected. The decision is quite simple. Figure 8 D is a diagram showing an image area obtained by assuming that there is no connection relationship between small blocks when there is only a white part on the joint side of the pattern, and areas with the same number in the figure are image areas of the same type. be. Of course, the connections between blocks may be checked using the above method without performing the four-division process. However, in this case, the boundaries of the regions become rough. In order to solve this problem of roughness, it is conceivable to reduce the size of the block from the beginning, but doing so would require a considerable amount of time for the texture analysis.

As a result of the above, the image area has been extracted. Incidentally, the degree of appearance (frequency) of each pattern in gradation drawings, character drawings, and line drawings can be shown in FIG. Therefore, by knowing the pattern distribution within each image area, it is also possible to determine what kind of image each image area is.

Figures 10 to 12 are more specific illustrations of region extraction, with Figure 10 showing the original image, Figure 11 showing the pattern classification image, and Figure 1
FIG. 2 is a diagram showing a region extracted image. Furthermore, the 10th
The areas indicated by 1 and 2 in the figure are filled with characters, and the areas indicated by 3 and 4 are continuous tone images. As shown in FIG. 12, the image extraction is successful, but in FIG. 11, isolated areas (blocks) occur due to dirt etc. in the original image, and these create unique image areas in FIG. 12. There is.
In order to prevent this noise, an isolated block may be treated as being in the same image area as surrounding blocks.

(Effects of the Invention) As explained above, according to the method of the present invention, image regions of any shape can be extracted.
Therefore, great effects can be obtained when the method of the present invention is used when extracting and recording only necessary parts, or when filing or the like.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing the procedure in a specific example of the method of the present invention, FIG. 2 is an explanatory diagram of the window, FIG. 3 is an explanatory diagram of the density histogram, and FIG. 4 is a flowchart showing the procedure for determining the background level. ,
Fig. 5 is an explanatory diagram of image blocking, Fig. 6 is an explanatory diagram of block patterns, Fig. 7 is an explanatory diagram of pattern classification, Fig. 8 is an explanatory diagram of image area extraction, and Fig. 9 is an explanatory diagram of image area extraction.
FIG. 10 is an explanatory diagram of pattern frequency, FIG. 10 is a diagram showing an original image, FIG. 11 is a diagram showing a pattern classification image, and FIG. 12 is a diagram showing a region extraction image. α...Block, _α1 , _α2 , _α3 , _α4 ...Small block, 1, 2...Character image area, 3, 4...Continuous tone image area.

Claims (1)

[Claims] 1. An image is divided into a plurality of blocks, each block is classified into a predetermined pattern using binarized image data, and the connection between each block is checked using the pattern to identify similar image areas. A region extraction method characterized by extraction. 2. The area extraction method according to claim 1, wherein the binarization threshold value when obtaining the binarized image data is determined based on background information of the image. 3 When classifying the blocks into patterns,
3. The area extraction method according to claim 1, wherein data obtained by two-dimensionally projecting the binarized image data of each block is used. 4. When checking the connections between the blocks,
Claim 1, characterized in that each of the blocks is divided into two in the vertical and horizontal directions, and connections between the divided small blocks and the surrounding small blocks are examined.
3. The area extraction method according to item 2, item 3, or item 3. 5. The area extraction method according to any one of claims 1 to 4, characterized in that an isolated block is regarded as the same image area as surrounding blocks.